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/init.h>
  16 #include <linux/cpufreq.h>
  17 #include <linux/cpu.h>
  18 #include <linux/jiffies.h>
  19 #include <linux/kernel_stat.h>
  20 #include <linux/mutex.h>
  21
  22 /*
  23  * dbs is used in this file as a shortform for demandbased switching
  24  * It helps to keep variable names smaller, simpler
  25  */
  26
  27 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  28 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  29 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  30
  31 /*
  32  * The polling frequency of this governor depends on the capability of
  33  * the processor. Default polling frequency is 1000 times the transition
  34  * latency of the processor. The governor will work on any processor with
  35  * transition latency <= 10mS, using appropriate sampling
  36  * rate.
  37  * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
  38  * this governor will not work.
  39  * All times here are in uS.
  40  */
  41 static unsigned int def_sampling_rate;
  42 #define MIN_SAMPLING_RATE_RATIO                 (2)
  43 /* for correct statistics, we need at least 10 ticks between each measure */
  44 #define MIN_STAT_SAMPLING_RATE                  (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10))
  45 #define MIN_SAMPLING_RATE                       (def_sampling_rate / MIN_SAMPLING_RATE_RATIO)
  46 #define MAX_SAMPLING_RATE                       (500 * def_sampling_rate)
  47 #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER    (1000)
  48 #define TRANSITION_LATENCY_LIMIT                (10 * 1000)
  49
  50 static void do_dbs_timer(void *data);
  51
  52 struct cpu_dbs_info_s {
  53         cputime64_t prev_cpu_idle;
  54         cputime64_t prev_cpu_wall;
  55         struct cpufreq_policy *cur_policy;
  56         struct work_struct work;
  57         unsigned int enable;
  58 };
  59 static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
  60
  61 static unsigned int dbs_enable; /* number of CPUs using this policy */
  62
  63 /*
  64  * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
  65  * lock and dbs_mutex. cpu_hotplug lock should always be held before
  66  * dbs_mutex. If any function that can potentially take cpu_hotplug lock
  67  * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
  68  * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
  69  * is recursive for the same process. -Venki
  70  */
  71 static DEFINE_MUTEX(dbs_mutex);
  72
  73 static struct workqueue_struct  *kondemand_wq;
  74
  75 struct dbs_tuners {
  76         unsigned int sampling_rate;
  77         unsigned int up_threshold;
  78         unsigned int ignore_nice;
  79 };
  80
  81 static struct dbs_tuners dbs_tuners_ins = {
  82         .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
  83         .ignore_nice = 0,
  84 };
  85
  86 static inline cputime64_t get_cpu_idle_time(unsigned int cpu)
  87 {
  88         cputime64_t retval;
  89
  90         retval = cputime64_add(kstat_cpu(cpu).cpustat.idle,
  91                         kstat_cpu(cpu).cpustat.iowait);
  92
  93         if (dbs_tuners_ins.ignore_nice)
  94                 retval = cputime64_add(retval, kstat_cpu(cpu).cpustat.nice);
  95
  96         return retval;
  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(up_threshold, up_threshold);
 126 show_one(ignore_nice_load, ignore_nice);
 127
 128 static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
 129                 const char *buf, size_t count)
 130 {
 131         unsigned int input;
 132         int ret;
 133         ret = sscanf(buf, "%u", &input);
 134
 135         mutex_lock(&dbs_mutex);
 136         if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) {
 137                 mutex_unlock(&dbs_mutex);
 138                 return -EINVAL;
 139         }
 140
 141         dbs_tuners_ins.sampling_rate = input;
 142         mutex_unlock(&dbs_mutex);
 143
 144         return count;
 145 }
 146
 147 static ssize_t store_up_threshold(struct cpufreq_policy *unused,
 148                 const char *buf, size_t count)
 149 {
 150         unsigned int input;
 151         int ret;
 152         ret = sscanf(buf, "%u", &input);
 153
 154         mutex_lock(&dbs_mutex);
 155         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 156                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 157                 mutex_unlock(&dbs_mutex);
 158                 return -EINVAL;
 159         }
 160
 161         dbs_tuners_ins.up_threshold = input;
 162         mutex_unlock(&dbs_mutex);
 163
 164         return count;
 165 }
 166
 167 static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
 168                 const char *buf, size_t count)
 169 {
 170         unsigned int input;
 171         int ret;
 172
 173         unsigned int j;
 174
 175         ret = sscanf(buf, "%u", &input);
 176         if ( ret != 1 )
 177                 return -EINVAL;
 178
 179         if ( input > 1 )
 180                 input = 1;
 181
 182         mutex_lock(&dbs_mutex);
 183         if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
 184                 mutex_unlock(&dbs_mutex);
 185                 return count;
 186         }
 187         dbs_tuners_ins.ignore_nice = input;
 188
 189         /* we need to re-evaluate prev_cpu_idle */
 190         for_each_online_cpu(j) {
 191                 struct cpu_dbs_info_s *dbs_info;
 192                 dbs_info = &per_cpu(cpu_dbs_info, j);
 193                 dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
 194                 dbs_info->prev_cpu_wall = get_jiffies_64();
 195         }
 196         mutex_unlock(&dbs_mutex);
 197
 198         return count;
 199 }
 200
 201 #define define_one_rw(_name) \
 202 static struct freq_attr _name = \
 203 __ATTR(_name, 0644, show_##_name, store_##_name)
 204
 205 define_one_rw(sampling_rate);
 206 define_one_rw(up_threshold);
 207 define_one_rw(ignore_nice_load);
 208
 209 static struct attribute * dbs_attributes[] = {
 210         &sampling_rate_max.attr,
 211         &sampling_rate_min.attr,
 212         &sampling_rate.attr,
 213         &up_threshold.attr,
 214         &ignore_nice_load.attr,
 215         NULL
 216 };
 217
 218 static struct attribute_group dbs_attr_group = {
 219         .attrs = dbs_attributes,
 220         .name = "ondemand",
 221 };
 222
 223 /************************** sysfs end ************************/
 224
 225 static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
 226 {
 227         unsigned int idle_ticks, total_ticks;
 228         unsigned int load;
 229         cputime64_t cur_jiffies;
 230
 231         struct cpufreq_policy *policy;
 232         unsigned int j;
 233
 234         if (!this_dbs_info->enable)
 235                 return;
 236
 237         policy = this_dbs_info->cur_policy;
 238         cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
 239         total_ticks = (unsigned int) cputime64_sub(cur_jiffies,
 240                         this_dbs_info->prev_cpu_wall);
 241         this_dbs_info->prev_cpu_wall = cur_jiffies;
 242         if (!total_ticks)
 243                 return;
 244         /*
 245          * Every sampling_rate, we check, if current idle time is less
 246          * than 20% (default), then we try to increase frequency
 247          * Every sampling_rate, we look for a the lowest
 248          * frequency which can sustain the load while keeping idle time over
 249          * 30%. If such a frequency exist, we try to decrease to this frequency.
 250          *
 251          * Any frequency increase takes it to the maximum frequency.
 252          * Frequency reduction happens at minimum steps of
 253          * 5% (default) of current frequency
 254          */
 255
 256         /* Get Idle Time */
 257         idle_ticks = UINT_MAX;
 258         for_each_cpu_mask(j, policy->cpus) {
 259                 cputime64_t total_idle_ticks;
 260                 unsigned int tmp_idle_ticks;
 261                 struct cpu_dbs_info_s *j_dbs_info;
 262
 263                 j_dbs_info = &per_cpu(cpu_dbs_info, j);
 264                 total_idle_ticks = get_cpu_idle_time(j);
 265                 tmp_idle_ticks = (unsigned int) cputime64_sub(total_idle_ticks,
 266                                 j_dbs_info->prev_cpu_idle);
 267                 j_dbs_info->prev_cpu_idle = total_idle_ticks;
 268
 269                 if (tmp_idle_ticks < idle_ticks)
 270                         idle_ticks = tmp_idle_ticks;
 271         }
 272         load = (100 * (total_ticks - idle_ticks)) / total_ticks;
 273
 274         /* Check for frequency increase */
 275         if (load > dbs_tuners_ins.up_threshold) {
 276                 /* if we are already at full speed then break out early */
 277                 if (policy->cur == policy->max)
 278                         return;
 279
 280                 __cpufreq_driver_target(policy, policy->max,
 281                         CPUFREQ_RELATION_H);
 282                 return;
 283         }
 284
 285         /* Check for frequency decrease */
 286         /* if we cannot reduce the frequency anymore, break out early */
 287         if (policy->cur == policy->min)
 288                 return;
 289
 290         /*
 291          * The optimal frequency is the frequency that is the lowest that
 292          * can support the current CPU usage without triggering the up
 293          * policy. To be safe, we focus 10 points under the threshold.
 294          */
 295         if (load < (dbs_tuners_ins.up_threshold - 10)) {
 296                 unsigned int freq_next;
 297                 freq_next = (policy->cur * load) /
 298                         (dbs_tuners_ins.up_threshold - 10);
 299
 300                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L);
 301         }
 302 }
 303
 304 static void do_dbs_timer(void *data)
 305 {
 306         unsigned int cpu = smp_processor_id();
 307         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
 308
 309         if (!dbs_info->enable)
 310                 return;
 311
 312         lock_cpu_hotplug();
 313         dbs_check_cpu(dbs_info);
 314         unlock_cpu_hotplug();
 315         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work,
 316                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 317 }
 318
 319 static inline void dbs_timer_init(unsigned int cpu)
 320 {
 321         struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu);
 322
 323         INIT_WORK(&dbs_info->work, do_dbs_timer, 0);
 324         queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work,
 325                         usecs_to_jiffies(dbs_tuners_ins.sampling_rate));
 326         return;
 327 }
 328
 329 static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
 330 {
 331         dbs_info->enable = 0;
 332         cancel_delayed_work(&dbs_info->work);
 333         flush_workqueue(kondemand_wq);
 334 }
 335
 336 static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
 337                                    unsigned int event)
 338 {
 339         unsigned int cpu = policy->cpu;
 340         struct cpu_dbs_info_s *this_dbs_info;
 341         unsigned int j;
 342
 343         this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
 344
 345         switch (event) {
 346         case CPUFREQ_GOV_START:
 347                 if ((!cpu_online(cpu)) || (!policy->cur))
 348                         return -EINVAL;
 349
 350                 if (policy->cpuinfo.transition_latency >
 351                                 (TRANSITION_LATENCY_LIMIT * 1000)) {
 352                         printk(KERN_WARNING "ondemand governor failed to load "
 353                                "due to too long transition latency\n");
 354                         return -EINVAL;
 355                 }
 356                 if (this_dbs_info->enable) /* Already enabled */
 357                         break;
 358
 359                 mutex_lock(&dbs_mutex);
 360                 dbs_enable++;
 361                 if (dbs_enable == 1) {
 362                         kondemand_wq = create_workqueue("kondemand");
 363                         if (!kondemand_wq) {
 364                                 printk(KERN_ERR "Creation of kondemand failed\n");
 365                                 dbs_enable--;
 366                                 mutex_unlock(&dbs_mutex);
 367                                 return -ENOSPC;
 368                         }
 369                 }
 370                 for_each_cpu_mask(j, policy->cpus) {
 371                         struct cpu_dbs_info_s *j_dbs_info;
 372                         j_dbs_info = &per_cpu(cpu_dbs_info, j);
 373                         j_dbs_info->cur_policy = policy;
 374
 375                         j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j);
 376                         j_dbs_info->prev_cpu_wall = get_jiffies_64();
 377                 }
 378                 this_dbs_info->enable = 1;
 379                 sysfs_create_group(&policy->kobj, &dbs_attr_group);
 380                 /*
 381                  * Start the timerschedule work, when this governor
 382                  * is used for first time
 383                  */
 384                 if (dbs_enable == 1) {
 385                         unsigned int latency;
 386                         /* policy latency is in nS. Convert it to uS first */
 387                         latency = policy->cpuinfo.transition_latency / 1000;
 388                         if (latency == 0)
 389                                 latency = 1;
 390
 391                         def_sampling_rate = latency *
 392                                         DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 393
 394                         if (def_sampling_rate < MIN_STAT_SAMPLING_RATE)
 395                                 def_sampling_rate = MIN_STAT_SAMPLING_RATE;
 396
 397                         dbs_tuners_ins.sampling_rate = def_sampling_rate;
 398                 }
 399                 dbs_timer_init(policy->cpu);
 400
 401                 mutex_unlock(&dbs_mutex);
 402                 break;
 403
 404         case CPUFREQ_GOV_STOP:
 405                 mutex_lock(&dbs_mutex);
 406                 dbs_timer_exit(this_dbs_info);
 407                 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
 408                 dbs_enable--;
 409                 if (dbs_enable == 0)
 410                         destroy_workqueue(kondemand_wq);
 411
 412                 mutex_unlock(&dbs_mutex);
 413
 414                 break;
 415
 416         case CPUFREQ_GOV_LIMITS:
 417                 mutex_lock(&dbs_mutex);
 418                 if (policy->max < this_dbs_info->cur_policy->cur)
 419                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 420                                                 policy->max,
 421                                                 CPUFREQ_RELATION_H);
 422                 else if (policy->min > this_dbs_info->cur_policy->cur)
 423                         __cpufreq_driver_target(this_dbs_info->cur_policy,
 424                                                 policy->min,
 425                                                 CPUFREQ_RELATION_L);
 426                 mutex_unlock(&dbs_mutex);
 427                 break;
 428         }
 429         return 0;
 430 }
 431
 432 static struct cpufreq_governor cpufreq_gov_dbs = {
 433         .name = "ondemand",
 434         .governor = cpufreq_governor_dbs,
 435         .owner = THIS_MODULE,
 436 };
 437
 438 static int __init cpufreq_gov_dbs_init(void)
 439 {
 440         return cpufreq_register_governor(&cpufreq_gov_dbs);
 441 }
 442
 443 static void __exit cpufreq_gov_dbs_exit(void)
 444 {
 445         cpufreq_unregister_governor(&cpufreq_gov_dbs);
 446 }
 447
 448
 449 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 450 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 451 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 452                    "Low Latency Frequency Transition capable processors");
 453 MODULE_LICENSE("GPL");
 454
 455 module_init(cpufreq_gov_dbs_init);
 456 module_exit(cpufreq_gov_dbs_exit);