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
  31 /*
  32  * dbs is used in this file as a shortform for demandbased switching
  33  * It helps to keep variable names smaller, simpler
  34  */
  35
  36 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  37 #define MIN_FREQUENCY_UP_THRESHOLD              (0)
  38 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  39
  40 #define DEF_FREQUENCY_DOWN_THRESHOLD            (20)
  41 #define MIN_FREQUENCY_DOWN_THRESHOLD            (0)
  42 #define MAX_FREQUENCY_DOWN_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 static unsigned int                             def_sampling_rate;
  55 #define MIN_SAMPLING_RATE                       (def_sampling_rate / 2)
  56 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  57 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
  58 #define DEF_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 DECLARE_MUTEX    (dbs_sem);
  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 };
  83
  84 static struct dbs_tuners dbs_tuners_ins = {
  85         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
  86         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
  87         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
  88 };
  89
  90 /************************** sysfs interface ************************/
  91 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
  92 {
  93         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
  94 }
  95
  96 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
  97 {
  98         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
  99 }
 100
 101 #define define_one_ro(_name)                                    \
 102 static struct freq_attr _name =                                 \
 103 __ATTR(_name, 0444, show_##_name, NULL)
 104
 105 define_one_ro(sampling_rate_max);
 106 define_one_ro(sampling_rate_min);
 107
 108 /* cpufreq_ondemand Governor Tunables */
 109 #define show_one(file_name, object)                                     \
 110 static ssize_t show_##file_name                                         \
 111 (struct cpufreq_policy *unused, char *buf)                              \
 112 {                                                                       \
 113         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 114 }
 115 show_one(sampling_rate, sampling_rate);
 116 show_one(sampling_down_factor, sampling_down_factor);
 117 show_one(up_threshold, up_threshold);
 118 show_one(down_threshold, down_threshold);
 119 show_one(ignore_nice, ignore_nice);
 120
 121 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
 122                 const char *buf, size_t count)
 123 {
 124         unsigned int input;
 125         int ret;
 126         ret = sscanf (buf, "%u", &input);
 127         if (ret != 1 )
 128                 return -EINVAL;
 129
 130         down(&dbs_sem);
 131         dbs_tuners_ins.sampling_down_factor = input;
 132         up(&dbs_sem);
 133
 134         return count;
 135 }
 136
 137 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 138                 const char *buf, size_t count)
 139 {
 140         unsigned int input;
 141         int ret;
 142         ret = sscanf (buf, "%u", &input);
 143
 144         down(&dbs_sem);
 145         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
 146                 up(&dbs_sem);
 147                 return -EINVAL;
 148         }
 149
 150         dbs_tuners_ins.sampling_rate = input;
 151         up(&dbs_sem);
 152
 153         return count;
 154 }
 155
 156 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 157                 const char *buf, size_t count)
 158 {
 159         unsigned int input;
 160         int ret;
 161         ret = sscanf (buf, "%u", &input);
 162
 163         down(&dbs_sem);
 164         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 165                         input < MIN_FREQUENCY_UP_THRESHOLD ||
 166                         input <= dbs_tuners_ins.down_threshold) {
 167                 up(&dbs_sem);
 168                 return -EINVAL;
 169         }
 170
 171         dbs_tuners_ins.up_threshold = input;
 172         up(&dbs_sem);
 173
 174         return count;
 175 }
 176
 177 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
 178                 const char *buf, size_t count)
 179 {
 180         unsigned int input;
 181         int ret;
 182         ret = sscanf (buf, "%u", &input);
 183
 184         down(&dbs_sem);
 185         if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD ||
 186                         input < MIN_FREQUENCY_DOWN_THRESHOLD ||
 187                         input >= dbs_tuners_ins.up_threshold) {
 188                 up(&dbs_sem);
 189                 return -EINVAL;
 190         }
 191
 192         dbs_tuners_ins.down_threshold = input;
 193         up(&dbs_sem);
 194
 195         return count;
 196 }
 197
 198 static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
 199                 const char *buf, size_t count)
 200 {
 201         unsigned int input;
 202         int ret;
 203
 204         unsigned int j;
 205
 206         ret = sscanf (buf, "%u", &input);
 207         if ( ret != 1 )
 208                 return -EINVAL;
 209
 210         if ( input > 1 )
 211                 input = 1;
 212
 213         down(&dbs_sem);
 214         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
 215                 up(&dbs_sem);
 216                 return count;
 217         }
 218         dbs_tuners_ins.ignore_nice = input;
 219
 220         /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
 221         for_each_cpu_mask(j, policy->cpus) {
 222                 struct cpu_dbs_info_s *j_dbs_info;
 223                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 224                 j_dbs_info->cur_policy = policy;
 225
 226                 j_dbs_info->prev_cpu_idle_up =
 227                         kstat_cpu(j).cpustat.idle +
 228                         kstat_cpu(j).cpustat.iowait +
 229                         ( !dbs_tuners_ins.ignore_nice
 230                           ? kstat_cpu(j).cpustat.nice : 0 );
 231                 j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
 232         }
 233         up(&dbs_sem);
 234
 235         return count;
 236 }
 237
 238 #define define_one_rw(_name) \
 239 static struct freq_attr _name = \
 240 __ATTR(_name, 0644, show_##_name, store_##_name)
 241
 242 define_one_rw(sampling_rate);
 243 define_one_rw(sampling_down_factor);
 244 define_one_rw(up_threshold);
 245 define_one_rw(down_threshold);
 246 define_one_rw(ignore_nice);
 247
 248 static struct attribute * dbs_attributes[] = {
 249         &sampling_rate_max.attr,
 250         &sampling_rate_min.attr,
 251         &sampling_rate.attr,
 252         &sampling_down_factor.attr,
 253         &up_threshold.attr,
 254         &down_threshold.attr,
 255         &ignore_nice.attr,
 256         NULL
 257 };
 258
 259 static struct attribute_group dbs_attr_group = {
 260         .attrs = dbs_attributes,
 261         .name = "ondemand",
 262 };
 263
 264 /************************** sysfs end ************************/
 265
 266 static void dbs_check_cpu(int cpu)
 267 {
 268         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
 269         unsigned int total_idle_ticks;
 270         unsigned int freq_down_step;
 271         unsigned int freq_down_sampling_rate;
 272         static int down_skip[NR_CPUS];
 273         struct cpu_dbs_info_s *this_dbs_info;
 274
 275         struct cpufreq_policy *policy;
 276         unsigned int j;
 277
 278         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 279         if (!this_dbs_info->enable)
 280                 return;
 281
 282         policy = this_dbs_info->cur_policy;
 283         /*
 284          * The default safe range is 20% to 80%
 285          * Every sampling_rate, we check
 286          *      - If current idle time is less than 20%, then we try to
 287          *        increase frequency
 288          * Every sampling_rate*sampling_down_factor, we check
 289          *      - If current idle time is more than 80%, then we try to
 290          *        decrease frequency
 291          *
 292          * Any frequency increase takes it to the maximum frequency.
 293          * Frequency reduction happens at minimum steps of
 294          * 5% of max_frequency
 295          */
 296
 297         /* Check for frequency increase */
 298         total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
 299                 kstat_cpu(cpu).cpustat.iowait;
 300           /* consider 'nice' tasks as 'idle' time too if required */
 301           if (dbs_tuners_ins.ignore_nice == 0)
 302                 total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
 303         idle_ticks = total_idle_ticks -
 304                 this_dbs_info->prev_cpu_idle_up;
 305         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
 306
 307
 308         for_each_cpu_mask(j, policy->cpus) {
 309                 unsigned int tmp_idle_ticks;
 310                 struct cpu_dbs_info_s *j_dbs_info;
 311
 312                 if (j == cpu)
 313                         continue;
 314
 315                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 316                 /* Check for frequency increase */
 317                 total_idle_ticks = kstat_cpu(j).cpustat.idle +
 318                         kstat_cpu(j).cpustat.iowait;
 319                   /* consider 'nice' too? */
 320                   if (dbs_tuners_ins.ignore_nice == 0)
 321                            total_idle_ticks += kstat_cpu(j).cpustat.nice;
 322                 tmp_idle_ticks = total_idle_ticks -
 323                         j_dbs_info->prev_cpu_idle_up;
 324                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
 325
 326                 if (tmp_idle_ticks < idle_ticks)
 327                         idle_ticks = tmp_idle_ticks;
 328         }
 329
 330         /* Scale idle ticks by 100 and compare with up and down ticks */
 331         idle_ticks *= 100;
 332         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
 333                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 334
 335         if (idle_ticks < up_idle_ticks) {
 336                 __cpufreq_driver_target(policy, policy->max,
 337                         CPUFREQ_RELATION_H);
 338                 down_skip[cpu] = 0;
 339                 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 340                 return;
 341         }
 342
 343         /* Check for frequency decrease */
 344         down_skip[cpu]++;
 345         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
 346                 return;
 347
 348         total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
 349                 kstat_cpu(cpu).cpustat.iowait;
 350           /* consider 'nice' too? */
 351           if (dbs_tuners_ins.ignore_nice == 0)
 352                   total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
 353         idle_ticks = total_idle_ticks -
 354                 this_dbs_info->prev_cpu_idle_down;
 355         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 356
 357         for_each_cpu_mask(j, policy->cpus) {
 358                 unsigned int tmp_idle_ticks;
 359                 struct cpu_dbs_info_s *j_dbs_info;
 360
 361                 if (j == cpu)
 362                         continue;
 363
 364                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 365                 /* Check for frequency increase */
 366                 total_idle_ticks = kstat_cpu(j).cpustat.idle +
 367                         kstat_cpu(j).cpustat.iowait;
 368                   /* consider 'nice' too? */
 369                   if (dbs_tuners_ins.ignore_nice == 0)
 370                         total_idle_ticks += kstat_cpu(j).cpustat.nice;
 371                 tmp_idle_ticks = total_idle_ticks -
 372                         j_dbs_info->prev_cpu_idle_down;
 373                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 374
 375                 if (tmp_idle_ticks < idle_ticks)
 376                         idle_ticks = tmp_idle_ticks;
 377         }
 378
 379         /* Scale idle ticks by 100 and compare with up and down ticks */
 380         idle_ticks *= 100;
 381         down_skip[cpu] = 0;
 382
 383         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
 384                 dbs_tuners_ins.sampling_down_factor;
 385         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
 386                         usecs_to_jiffies(freq_down_sampling_rate);
 387
 388         if (idle_ticks > down_idle_ticks ) {
 389                 freq_down_step = (5 * policy->max) / 100;
 390
 391                 /* max freq cannot be less than 100. But who knows.... */
 392                 if (unlikely(freq_down_step == 0))
 393                         freq_down_step = 5;
 394
 395                 __cpufreq_driver_target(policy,
 396                         policy->cur - freq_down_step,
 397                         CPUFREQ_RELATION_H);
 398                 return;
 399         }
 400 }
 401
 402 static void do_dbs_timer(void *data)
 403 {
 404         int i;
 405         down(&dbs_sem);
 406         for_each_online_cpu(i)
 407                 dbs_check_cpu(i);
 408         schedule_delayed_work(&dbs_work,
 409                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 410         up(&dbs_sem);
 411 }
 412
 413 static inline void dbs_timer_init(void)
 414 {
 415         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
 416         schedule_delayed_work(&dbs_work,
 417                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 418         return;
 419 }
 420
 421 static inline void dbs_timer_exit(void)
 422 {
 423         cancel_delayed_work(&dbs_work);
 424         return;
 425 }
 426
 427 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 428                                    unsigned int event)
 429 {
 430         unsigned int cpu = policy->cpu;
 431         struct cpu_dbs_info_s *this_dbs_info;
 432         unsigned int j;
 433
 434         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 435
 436         switch (event) {
 437         case CPUFREQ_GOV_START:
 438                 if ((!cpu_online(cpu)) ||
 439                     (!policy->cur))
 440                         return -EINVAL;
 441
 442                 if (policy->cpuinfo.transition_latency >
 443                                 (TRANSITION_LATENCY_LIMIT * 1000))
 444                         return -EINVAL;
 445                 if (this_dbs_info->enable) /* Already enabled */
 446                         break;
 447
 448                 down(&dbs_sem);
 449                 for_each_cpu_mask(j, policy->cpus) {
 450                         struct cpu_dbs_info_s *j_dbs_info;
 451                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 452                         j_dbs_info->cur_policy = policy;
 453
 454                         j_dbs_info->prev_cpu_idle_up =
 455                                 kstat_cpu(j).cpustat.idle +
 456                                 kstat_cpu(j).cpustat.iowait +
 457                                 ( !dbs_tuners_ins.ignore_nice
 458                                   ? kstat_cpu(j).cpustat.nice : 0 );
 459                         j_dbs_info->prev_cpu_idle_down
 460                                 = j_dbs_info->prev_cpu_idle_up;
 461                 }
 462                 this_dbs_info->enable = 1;
 463                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
 464                 dbs_enable++;
 465                 /*
 466                  * Start the timerschedule work, when this governor
 467                  * is used for first time
 468                  */
 469                 if (dbs_enable == 1) {
 470                         unsigned int latency;
 471                         /* policy latency is in nS. Convert it to uS first */
 472
 473                         latency = policy->cpuinfo.transition_latency;
 474                         if (latency < 1000)
 475                                 latency = 1000;
 476
 477                         def_sampling_rate = (latency / 1000) *
 478                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 479                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 480                         dbs_tuners_ins.ignore_nice = 0;
 481
 482                         dbs_timer_init();
 483                 }
 484
 485                 up(&dbs_sem);
 486                 break;
 487
 488         case CPUFREQ_GOV_STOP:
 489                 down(&dbs_sem);
 490                 this_dbs_info->enable = 0;
 491                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 492                 dbs_enable--;
 493                 /*
 494                  * Stop the timerschedule work, when this governor
 495                  * is used for first time
 496                  */
 497                 if (dbs_enable == 0)
 498                         dbs_timer_exit();
 499
 500                 up(&dbs_sem);
 501
 502                 break;
 503
 504         case CPUFREQ_GOV_LIMITS:
 505                 down(&dbs_sem);
 506                 if (policy->max < this_dbs_info->cur_policy->cur)
 507                         __cpufreq_driver_target(
 508                                         this_dbs_info->cur_policy,
 509                                         policy->max, CPUFREQ_RELATION_H);
 510                 else if (policy->min > this_dbs_info->cur_policy->cur)
 511                         __cpufreq_driver_target(
 512                                         this_dbs_info->cur_policy,
 513                                         policy->min, CPUFREQ_RELATION_L);
 514                 up(&dbs_sem);
 515                 break;
 516         }
 517         return 0;
 518 }
 519
 520 static struct cpufreq_governor cpufreq_gov_dbs = {
 521         .name           = "ondemand",
 522         .governor       = cpufreq_governor_dbs,
 523         .owner          = THIS_MODULE,
 524 };
 525
 526 static int __init cpufreq_gov_dbs_init(void)
 527 {
 528         return cpufreq_register_governor(&cpufreq_gov_dbs);
 529 }
 530
 531 static void __exit cpufreq_gov_dbs_exit(void)
 532 {
 533         /* Make sure that the scheduled work is indeed not running */
 534         flush_scheduled_work();
 535
 536         cpufreq_unregister_governor(&cpufreq_gov_dbs);
 537 }
 538
 539
 540 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 541 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 542                 "Low Latency Frequency Transition capable processors");
 543 MODULE_LICENSE ("GPL");
 544
 545 module_init(cpufreq_gov_dbs_init);
 546 module_exit(cpufreq_gov_dbs_exit);