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 };
  82
  83 static struct dbs_tuners dbs_tuners_ins = {
  84         .up_threshold           = DEF_FREQUENCY_UP_THRESHOLD,
  85         .down_threshold         = DEF_FREQUENCY_DOWN_THRESHOLD,
  86         .sampling_down_factor   = DEF_SAMPLING_DOWN_FACTOR,
  87 };
  88
  89 /************************** sysfs interface ************************/
  90 static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
  91 {
  92         return sprintf (buf, "%u\n", MAX_SAMPLING_RATE);
  93 }
  94
  95 static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
  96 {
  97         return sprintf (buf, "%u\n", MIN_SAMPLING_RATE);
  98 }
  99
 100 #define define_one_ro(_name)                                    \
 101 static struct freq_attr _name =                                 \
 102 __ATTR(_name, 0444, show_##_name, NULL)
 103
 104 define_one_ro(sampling_rate_max);
 105 define_one_ro(sampling_rate_min);
 106
 107 /* cpufreq_ondemand Governor Tunables */
 108 #define show_one(file_name, object)                                     \
 109 static ssize_t show_##file_name                                         \
 110 (struct cpufreq_policy *unused, char *buf)                              \
 111 {                                                                       \
 112         return sprintf(buf, "%u\n", dbs_tuners_ins.object);             \
 113 }
 114 show_one(sampling_rate, sampling_rate);
 115 show_one(sampling_down_factor, sampling_down_factor);
 116 show_one(up_threshold, up_threshold);
 117 show_one(down_threshold, down_threshold);
 118
 119 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused,
 120                 const char *buf, size_t count)
 121 {
 122         unsigned int input;
 123         int ret;
 124         ret = sscanf (buf, "%u", &input);
 125         if (ret != 1 )
 126                 return -EINVAL;
 127
 128         down(&dbs_sem);
 129         dbs_tuners_ins.sampling_down_factor = input;
 130         up(&dbs_sem);
 131
 132         return count;
 133 }
 134
 135 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 136                 const char *buf, size_t count)
 137 {
 138         unsigned int input;
 139         int ret;
 140         ret = sscanf (buf, "%u", &input);
 141
 142         down(&dbs_sem);
 143         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
 144                 up(&dbs_sem);
 145                 return -EINVAL;
 146         }
 147
 148         dbs_tuners_ins.sampling_rate = input;
 149         up(&dbs_sem);
 150
 151         return count;
 152 }
 153
 154 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 155                 const char *buf, size_t count)
 156 {
 157         unsigned int input;
 158         int ret;
 159         ret = sscanf (buf, "%u", &input);
 160
 161         down(&dbs_sem);
 162         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 163                         input < MIN_FREQUENCY_UP_THRESHOLD ||
 164                         input <= dbs_tuners_ins.down_threshold) {
 165                 up(&dbs_sem);
 166                 return -EINVAL;
 167         }
 168
 169         dbs_tuners_ins.up_threshold = input;
 170         up(&dbs_sem);
 171
 172         return count;
 173 }
 174
 175 static ssize_t store_down_threshold(struct cpufreq_policy *unused,
 176                 const char *buf, size_t count)
 177 {
 178         unsigned int input;
 179         int ret;
 180         ret = sscanf (buf, "%u", &input);
 181
 182         down(&dbs_sem);
 183         if (ret != 1 || input > MAX_FREQUENCY_DOWN_THRESHOLD ||
 184                         input < MIN_FREQUENCY_DOWN_THRESHOLD ||
 185                         input >= dbs_tuners_ins.up_threshold) {
 186                 up(&dbs_sem);
 187                 return -EINVAL;
 188         }
 189
 190         dbs_tuners_ins.down_threshold = input;
 191         up(&dbs_sem);
 192
 193         return count;
 194 }
 195
 196 #define define_one_rw(_name) \
 197 static struct freq_attr _name = \
 198 __ATTR(_name, 0644, show_##_name, store_##_name)
 199
 200 define_one_rw(sampling_rate);
 201 define_one_rw(sampling_down_factor);
 202 define_one_rw(up_threshold);
 203 define_one_rw(down_threshold);
 204
 205 static struct attribute * dbs_attributes[] = {
 206         &sampling_rate_max.attr,
 207         &sampling_rate_min.attr,
 208         &sampling_rate.attr,
 209         &sampling_down_factor.attr,
 210         &up_threshold.attr,
 211         &down_threshold.attr,
 212         NULL
 213 };
 214
 215 static struct attribute_group dbs_attr_group = {
 216         .attrs = dbs_attributes,
 217         .name = "ondemand",
 218 };
 219
 220 /************************** sysfs end ************************/
 221
 222 static void dbs_check_cpu(int cpu)
 223 {
 224         unsigned int idle_ticks, up_idle_ticks, down_idle_ticks;
 225         unsigned int total_idle_ticks;
 226         unsigned int freq_down_step;
 227         unsigned int freq_down_sampling_rate;
 228         static int down_skip[NR_CPUS];
 229         struct cpu_dbs_info_s *this_dbs_info;
 230
 231         struct cpufreq_policy *policy;
 232         unsigned int j;
 233
 234         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 235         if (!this_dbs_info->enable)
 236                 return;
 237
 238         policy = this_dbs_info->cur_policy;
 239         /*
 240          * The default safe range is 20% to 80%
 241          * Every sampling_rate, we check
 242          *      - If current idle time is less than 20%, then we try to
 243          *        increase frequency
 244          * Every sampling_rate*sampling_down_factor, we check
 245          *      - If current idle time is more than 80%, then we try to
 246          *        decrease frequency
 247          *
 248          * Any frequency increase takes it to the maximum frequency.
 249          * Frequency reduction happens at minimum steps of
 250          * 5% of max_frequency
 251          */
 252
 253         /* Check for frequency increase */
 254         total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
 255                 kstat_cpu(cpu).cpustat.iowait;
 256         idle_ticks = total_idle_ticks -
 257                 this_dbs_info->prev_cpu_idle_up;
 258         this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
 259
 260
 261         for_each_cpu_mask(j, policy->cpus) {
 262                 unsigned int tmp_idle_ticks;
 263                 struct cpu_dbs_info_s *j_dbs_info;
 264
 265                 if (j == cpu)
 266                         continue;
 267
 268                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 269                 /* Check for frequency increase */
 270                 total_idle_ticks = kstat_cpu(j).cpustat.idle +
 271                         kstat_cpu(j).cpustat.iowait;
 272                 tmp_idle_ticks = total_idle_ticks -
 273                         j_dbs_info->prev_cpu_idle_up;
 274                 j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
 275
 276                 if (tmp_idle_ticks < idle_ticks)
 277                         idle_ticks = tmp_idle_ticks;
 278         }
 279
 280         /* Scale idle ticks by 100 and compare with up and down ticks */
 281         idle_ticks *= 100;
 282         up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) *
 283                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
 284
 285         if (idle_ticks < up_idle_ticks) {
 286                 __cpufreq_driver_target(policy, policy->max,
 287                         CPUFREQ_RELATION_H);
 288                 down_skip[cpu] = 0;
 289                 this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 290                 return;
 291         }
 292
 293         /* Check for frequency decrease */
 294         down_skip[cpu]++;
 295         if (down_skip[cpu] < dbs_tuners_ins.sampling_down_factor)
 296                 return;
 297
 298         total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
 299                 kstat_cpu(cpu).cpustat.iowait;
 300         idle_ticks = total_idle_ticks -
 301                 this_dbs_info->prev_cpu_idle_down;
 302         this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 303
 304         for_each_cpu_mask(j, policy->cpus) {
 305                 unsigned int tmp_idle_ticks;
 306                 struct cpu_dbs_info_s *j_dbs_info;
 307
 308                 if (j == cpu)
 309                         continue;
 310
 311                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 312                 /* Check for frequency increase */
 313                 total_idle_ticks = kstat_cpu(j).cpustat.idle +
 314                         kstat_cpu(j).cpustat.iowait;
 315                 tmp_idle_ticks = total_idle_ticks -
 316                         j_dbs_info->prev_cpu_idle_down;
 317                 j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
 318
 319                 if (tmp_idle_ticks < idle_ticks)
 320                         idle_ticks = tmp_idle_ticks;
 321         }
 322
 323         /* Scale idle ticks by 100 and compare with up and down ticks */
 324         idle_ticks *= 100;
 325         down_skip[cpu] = 0;
 326
 327         freq_down_sampling_rate = dbs_tuners_ins.sampling_rate *
 328                 dbs_tuners_ins.sampling_down_factor;
 329         down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) *
 330                         usecs_to_jiffies(freq_down_sampling_rate);
 331
 332         if (idle_ticks > down_idle_ticks ) {
 333                 freq_down_step = (5 * policy->max) / 100;
 334
 335                 /* max freq cannot be less than 100. But who knows.... */
 336                 if (unlikely(freq_down_step == 0))
 337                         freq_down_step = 5;
 338
 339                 __cpufreq_driver_target(policy,
 340                         policy->cur - freq_down_step,
 341                         CPUFREQ_RELATION_H);
 342                 return;
 343         }
 344 }
 345
 346 static void do_dbs_timer(void *data)
 347 {
 348         int i;
 349         down(&dbs_sem);
 350         for_each_online_cpu(i)
 351                 dbs_check_cpu(i);
 352         schedule_delayed_work(&dbs_work,
 353                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 354         up(&dbs_sem);
 355 }
 356
 357 static inline void dbs_timer_init(void)
 358 {
 359         INIT_WORK(&dbs_work, do_dbs_timer, NULL);
 360         schedule_delayed_work(&dbs_work,
 361                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 362         return;
 363 }
 364
 365 static inline void dbs_timer_exit(void)
 366 {
 367         cancel_delayed_work(&dbs_work);
 368         return;
 369 }
 370
 371 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 372                                    unsigned int event)
 373 {
 374         unsigned int cpu = policy->cpu;
 375         struct cpu_dbs_info_s *this_dbs_info;
 376         unsigned int j;
 377
 378         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 379
 380         switch (event) {
 381         case CPUFREQ_GOV_START:
 382                 if ((!cpu_online(cpu)) ||
 383                     (!policy->cur))
 384                         return -EINVAL;
 385
 386                 if (policy->cpuinfo.transition_latency >
 387                                 (TRANSITION_LATENCY_LIMIT * 1000))
 388                         return -EINVAL;
 389                 if (this_dbs_info->enable) /* Already enabled */
 390                         break;
 391
 392                 down(&dbs_sem);
 393                 for_each_cpu_mask(j, policy->cpus) {
 394                         struct cpu_dbs_info_s *j_dbs_info;
 395                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 396                         j_dbs_info->cur_policy = policy;
 397
 398                         j_dbs_info->prev_cpu_idle_up =
 399                                 kstat_cpu(j).cpustat.idle +
 400                                 kstat_cpu(j).cpustat.iowait;
 401                         j_dbs_info->prev_cpu_idle_down =
 402                                 kstat_cpu(j).cpustat.idle +
 403                                 kstat_cpu(j).cpustat.iowait;
 404                 }
 405                 this_dbs_info->enable = 1;
 406                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
 407                 dbs_enable++;
 408                 /*
 409                  * Start the timerschedule work, when this governor
 410                  * is used for first time
 411                  */
 412                 if (dbs_enable == 1) {
 413                         unsigned int latency;
 414                         /* policy latency is in nS. Convert it to uS first */
 415
 416                         latency = policy->cpuinfo.transition_latency;
 417                         if (latency < 1000)
 418                                 latency = 1000;
 419
 420                         def_sampling_rate = (latency / 1000) *
 421                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 422                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 423
 424                         dbs_timer_init();
 425                 }
 426
 427                 up(&dbs_sem);
 428                 break;
 429
 430         case CPUFREQ_GOV_STOP:
 431                 down(&dbs_sem);
 432                 this_dbs_info->enable = 0;
 433                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 434                 dbs_enable--;
 435                 /*
 436                  * Stop the timerschedule work, when this governor
 437                  * is used for first time
 438                  */
 439                 if (dbs_enable == 0)
 440                         dbs_timer_exit();
 441
 442                 up(&dbs_sem);
 443
 444                 break;
 445
 446         case CPUFREQ_GOV_LIMITS:
 447                 down(&dbs_sem);
 448                 if (policy->max < this_dbs_info->cur_policy->cur)
 449                         __cpufreq_driver_target(
 450                                         this_dbs_info->cur_policy,
 451                                         policy->max, CPUFREQ_RELATION_H);
 452                 else if (policy->min > this_dbs_info->cur_policy->cur)
 453                         __cpufreq_driver_target(
 454                                         this_dbs_info->cur_policy,
 455                                         policy->min, CPUFREQ_RELATION_L);
 456                 up(&dbs_sem);
 457                 break;
 458         }
 459         return 0;
 460 }
 461
 462 static struct cpufreq_governor cpufreq_gov_dbs = {
 463         .name           = "ondemand",
 464         .governor       = cpufreq_governor_dbs,
 465         .owner          = THIS_MODULE,
 466 };
 467
 468 static int __init cpufreq_gov_dbs_init(void)
 469 {
 470         return cpufreq_register_governor(&cpufreq_gov_dbs);
 471 }
 472
 473 static void __exit cpufreq_gov_dbs_exit(void)
 474 {
 475         /* Make sure that the scheduled work is indeed not running */
 476         flush_scheduled_work();
 477
 478         cpufreq_unregister_governor(&cpufreq_gov_dbs);
 479 }
 480
 481
 482 MODULE_AUTHOR ("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 483 MODULE_DESCRIPTION ("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 484                 "Low Latency Frequency Transition capable processors");
 485 MODULE_LICENSE ("GPL");
 486
 487 module_init(cpufreq_gov_dbs_init);
 488 module_exit(cpufreq_gov_dbs_exit);