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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14
  15 #include <linux/cpu.h>
  16 #include <linux/percpu-defs.h>
  17 #include <linux/slab.h>
  18 #include <linux/tick.h>
  19 #include "cpufreq_governor.h"
  20
  21 /* On-demand governor macros */
  22 #define DEF_FREQUENCY_UP_THRESHOLD              (80)
  23 #define DEF_SAMPLING_DOWN_FACTOR                (1)
  24 #define MAX_SAMPLING_DOWN_FACTOR                (100000)
  25 #define MICRO_FREQUENCY_UP_THRESHOLD            (95)
  26 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE         (10000)
  27 #define MIN_FREQUENCY_UP_THRESHOLD              (11)
  28 #define MAX_FREQUENCY_UP_THRESHOLD              (100)
  29
  30 static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info);
  31
  32 static struct dbs_governor od_dbs_gov;
  33 static struct od_ops od_ops;
  34
  35 static unsigned int default_powersave_bias;
  36
  37 /*
  38  * Not all CPUs want IO time to be accounted as busy; this depends on how
  39  * efficient idling at a higher frequency/voltage is.
  40  * Pavel Machek says this is not so for various generations of AMD and old
  41  * Intel systems.
  42  * Mike Chan (android.com) claims this is also not true for ARM.
  43  * Because of this, whitelist specific known (series) of CPUs by default, and
  44  * leave all others up to the user.
  45  */
  46 static int should_io_be_busy(void)
  47 {
  48 #if defined(CONFIG_X86)
  49         /*
  50          * For Intel, Core 2 (model 15) and later have an efficient idle.
  51          */
  52         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
  53                         boot_cpu_data.x86 == 6 &&
  54                         boot_cpu_data.x86_model >= 15)
  55                 return 1;
  56 #endif
  57         return 0;
  58 }
  59
  60 /*
  61  * Find right freq to be set now with powersave_bias on.
  62  * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
  63  * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
  64  */
  65 static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
  66                 unsigned int freq_next, unsigned int relation)
  67 {
  68         unsigned int freq_req, freq_reduc, freq_avg;
  69         unsigned int freq_hi, freq_lo;
  70         unsigned int index = 0;
  71         unsigned int delay_hi_us;
  72         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
  73                                                    policy->cpu);
  74         struct policy_dbs_info *policy_dbs = policy->governor_data;
  75         struct dbs_data *dbs_data = policy_dbs->dbs_data;
  76         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
  77
  78         if (!dbs_info->freq_table) {
  79                 dbs_info->freq_lo = 0;
  80                 dbs_info->freq_lo_delay_us = 0;
  81                 return freq_next;
  82         }
  83
  84         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
  85                         relation, &index);
  86         freq_req = dbs_info->freq_table[index].frequency;
  87         freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
  88         freq_avg = freq_req - freq_reduc;
  89
  90         /* Find freq bounds for freq_avg in freq_table */
  91         index = 0;
  92         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
  93                         CPUFREQ_RELATION_H, &index);
  94         freq_lo = dbs_info->freq_table[index].frequency;
  95         index = 0;
  96         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
  97                         CPUFREQ_RELATION_L, &index);
  98         freq_hi = dbs_info->freq_table[index].frequency;
  99
 100         /* Find out how long we have to be in hi and lo freqs */
 101         if (freq_hi == freq_lo) {
 102                 dbs_info->freq_lo = 0;
 103                 dbs_info->freq_lo_delay_us = 0;
 104                 return freq_lo;
 105         }
 106         delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
 107         delay_hi_us += (freq_hi - freq_lo) / 2;
 108         delay_hi_us /= freq_hi - freq_lo;
 109         dbs_info->freq_hi_delay_us = delay_hi_us;
 110         dbs_info->freq_lo = freq_lo;
 111         dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
 112         return freq_hi;
 113 }
 114
 115 static void ondemand_powersave_bias_init(struct cpufreq_policy *policy)
 116 {
 117         unsigned int cpu = policy->cpu;
 118         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 119
 120         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
 121         dbs_info->freq_lo = 0;
 122 }
 123
 124 static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
 125 {
 126         struct policy_dbs_info *policy_dbs = policy->governor_data;
 127         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 128         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 129
 130         if (od_tuners->powersave_bias)
 131                 freq = od_ops.powersave_bias_target(policy, freq,
 132                                 CPUFREQ_RELATION_H);
 133         else if (policy->cur == policy->max)
 134                 return;
 135
 136         __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
 137                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 138 }
 139
 140 /*
 141  * Every sampling_rate, we check, if current idle time is less than 20%
 142  * (default), then we try to increase frequency. Else, we adjust the frequency
 143  * proportional to load.
 144  */
 145 static void od_update(struct cpufreq_policy *policy)
 146 {
 147         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 148         struct policy_dbs_info *policy_dbs = dbs_info->cdbs.policy_dbs;
 149         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 150         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 151         unsigned int load = dbs_update(policy);
 152
 153         dbs_info->freq_lo = 0;
 154
 155         /* Check for frequency increase */
 156         if (load > dbs_data->up_threshold) {
 157                 /* If switching to max speed, apply sampling_down_factor */
 158                 if (policy->cur < policy->max)
 159                         policy_dbs->rate_mult = dbs_data->sampling_down_factor;
 160                 dbs_freq_increase(policy, policy->max);
 161         } else {
 162                 /* Calculate the next frequency proportional to load */
 163                 unsigned int freq_next, min_f, max_f;
 164
 165                 min_f = policy->cpuinfo.min_freq;
 166                 max_f = policy->cpuinfo.max_freq;
 167                 freq_next = min_f + load * (max_f - min_f) / 100;
 168
 169                 /* No longer fully busy, reset rate_mult */
 170                 policy_dbs->rate_mult = 1;
 171
 172                 if (od_tuners->powersave_bias)
 173                         freq_next = od_ops.powersave_bias_target(policy,
 174                                                                  freq_next,
 175                                                                  CPUFREQ_RELATION_L);
 176
 177                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
 178         }
 179 }
 180
 181 static unsigned int od_dbs_timer(struct cpufreq_policy *policy)
 182 {
 183         struct policy_dbs_info *policy_dbs = policy->governor_data;
 184         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 185         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 186         int sample_type = dbs_info->sample_type;
 187
 188         /* Common NORMAL_SAMPLE setup */
 189         dbs_info->sample_type = OD_NORMAL_SAMPLE;
 190         /*
 191          * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
 192          * it then.
 193          */
 194         if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
 195                 __cpufreq_driver_target(policy, dbs_info->freq_lo,
 196                                         CPUFREQ_RELATION_H);
 197                 return dbs_info->freq_lo_delay_us;
 198         }
 199
 200         od_update(policy);
 201
 202         if (dbs_info->freq_lo) {
 203                 /* Setup timer for SUB_SAMPLE */
 204                 dbs_info->sample_type = OD_SUB_SAMPLE;
 205                 return dbs_info->freq_hi_delay_us;
 206         }
 207
 208         return dbs_data->sampling_rate * policy_dbs->rate_mult;
 209 }
 210
 211 /************************** sysfs interface ************************/
 212 static struct dbs_governor od_dbs_gov;
 213
 214 static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf,
 215                 size_t count)
 216 {
 217         unsigned int input;
 218         int ret;
 219
 220         ret = sscanf(buf, "%u", &input);
 221         if (ret != 1)
 222                 return -EINVAL;
 223         dbs_data->io_is_busy = !!input;
 224
 225         /* we need to re-evaluate prev_cpu_idle */
 226         gov_update_cpu_data(&od_dbs_gov, dbs_data);
 227
 228         return count;
 229 }
 230
 231 static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf,
 232                 size_t count)
 233 {
 234         unsigned int input;
 235         int ret;
 236         ret = sscanf(buf, "%u", &input);
 237
 238         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 239                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 240                 return -EINVAL;
 241         }
 242
 243         dbs_data->up_threshold = input;
 244         return count;
 245 }
 246
 247 static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data,
 248                 const char *buf, size_t count)
 249 {
 250         struct policy_dbs_info *policy_dbs;
 251         unsigned int input;
 252         int ret;
 253         ret = sscanf(buf, "%u", &input);
 254
 255         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 256                 return -EINVAL;
 257
 258         dbs_data->sampling_down_factor = input;
 259
 260         /* Reset down sampling multiplier in case it was active */
 261         list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
 262                 /*
 263                  * Doing this without locking might lead to using different
 264                  * rate_mult values in od_update() and od_dbs_timer().
 265                  */
 266                 mutex_lock(&policy_dbs->timer_mutex);
 267                 policy_dbs->rate_mult = 1;
 268                 mutex_unlock(&policy_dbs->timer_mutex);
 269         }
 270
 271         return count;
 272 }
 273
 274 static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data,
 275                 const char *buf, size_t count)
 276 {
 277         unsigned int input;
 278         int ret;
 279
 280         ret = sscanf(buf, "%u", &input);
 281         if (ret != 1)
 282                 return -EINVAL;
 283
 284         if (input > 1)
 285                 input = 1;
 286
 287         if (input == dbs_data->ignore_nice_load) { /* nothing to do */
 288                 return count;
 289         }
 290         dbs_data->ignore_nice_load = input;
 291
 292         /* we need to re-evaluate prev_cpu_idle */
 293         gov_update_cpu_data(&od_dbs_gov, dbs_data);
 294
 295         return count;
 296 }
 297
 298 static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
 299                 size_t count)
 300 {
 301         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 302         struct policy_dbs_info *policy_dbs;
 303         unsigned int input;
 304         int ret;
 305         ret = sscanf(buf, "%u", &input);
 306
 307         if (ret != 1)
 308                 return -EINVAL;
 309
 310         if (input > 1000)
 311                 input = 1000;
 312
 313         od_tuners->powersave_bias = input;
 314
 315         list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list)
 316                 ondemand_powersave_bias_init(policy_dbs->policy);
 317
 318         return count;
 319 }
 320
 321 gov_show_one_common(sampling_rate);
 322 gov_show_one_common(up_threshold);
 323 gov_show_one_common(sampling_down_factor);
 324 gov_show_one_common(ignore_nice_load);
 325 gov_show_one_common(min_sampling_rate);
 326 gov_show_one_common(io_is_busy);
 327 gov_show_one(od, powersave_bias);
 328
 329 gov_attr_rw(sampling_rate);
 330 gov_attr_rw(io_is_busy);
 331 gov_attr_rw(up_threshold);
 332 gov_attr_rw(sampling_down_factor);
 333 gov_attr_rw(ignore_nice_load);
 334 gov_attr_rw(powersave_bias);
 335 gov_attr_ro(min_sampling_rate);
 336
 337 static struct attribute *od_attributes[] = {
 338         &min_sampling_rate.attr,
 339         &sampling_rate.attr,
 340         &up_threshold.attr,
 341         &sampling_down_factor.attr,
 342         &ignore_nice_load.attr,
 343         &powersave_bias.attr,
 344         &io_is_busy.attr,
 345         NULL
 346 };
 347
 348 /************************** sysfs end ************************/
 349
 350 static int od_init(struct dbs_data *dbs_data, bool notify)
 351 {
 352         struct od_dbs_tuners *tuners;
 353         u64 idle_time;
 354         int cpu;
 355
 356         tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
 357         if (!tuners) {
 358                 pr_err("%s: kzalloc failed\n", __func__);
 359                 return -ENOMEM;
 360         }
 361
 362         cpu = get_cpu();
 363         idle_time = get_cpu_idle_time_us(cpu, NULL);
 364         put_cpu();
 365         if (idle_time != -1ULL) {
 366                 /* Idle micro accounting is supported. Use finer thresholds */
 367                 dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 368                 /*
 369                  * In nohz/micro accounting case we set the minimum frequency
 370                  * not depending on HZ, but fixed (very low). The deferred
 371                  * timer might skip some samples if idle/sleeping as needed.
 372                 */
 373                 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 374         } else {
 375                 dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
 376
 377                 /* For correct statistics, we need 10 ticks for each measure */
 378                 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
 379                         jiffies_to_usecs(10);
 380         }
 381
 382         dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
 383         dbs_data->ignore_nice_load = 0;
 384         tuners->powersave_bias = default_powersave_bias;
 385         dbs_data->io_is_busy = should_io_be_busy();
 386
 387         dbs_data->tuners = tuners;
 388         return 0;
 389 }
 390
 391 static void od_exit(struct dbs_data *dbs_data, bool notify)
 392 {
 393         kfree(dbs_data->tuners);
 394 }
 395
 396 static void od_start(struct cpufreq_policy *policy)
 397 {
 398         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 399
 400         dbs_info->sample_type = OD_NORMAL_SAMPLE;
 401         ondemand_powersave_bias_init(policy);
 402 }
 403
 404 define_get_cpu_dbs_routines(od_cpu_dbs_info);
 405
 406 static struct od_ops od_ops = {
 407         .powersave_bias_target = generic_powersave_bias_target,
 408 };
 409
 410 static struct dbs_governor od_dbs_gov = {
 411         .gov = {
 412                 .name = "ondemand",
 413                 .governor = cpufreq_governor_dbs,
 414                 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 415                 .owner = THIS_MODULE,
 416         },
 417         .kobj_type = { .default_attrs = od_attributes },
 418         .get_cpu_cdbs = get_cpu_cdbs,
 419         .gov_dbs_timer = od_dbs_timer,
 420         .init = od_init,
 421         .exit = od_exit,
 422         .start = od_start,
 423 };
 424
 425 #define CPU_FREQ_GOV_ONDEMAND   (&od_dbs_gov.gov)
 426
 427 static void od_set_powersave_bias(unsigned int powersave_bias)
 428 {
 429         struct cpufreq_policy *policy;
 430         struct dbs_data *dbs_data;
 431         struct od_dbs_tuners *od_tuners;
 432         unsigned int cpu;
 433         cpumask_t done;
 434
 435         default_powersave_bias = powersave_bias;
 436         cpumask_clear(&done);
 437
 438         get_online_cpus();
 439         for_each_online_cpu(cpu) {
 440                 struct policy_dbs_info *policy_dbs;
 441
 442                 if (cpumask_test_cpu(cpu, &done))
 443                         continue;
 444
 445                 policy_dbs = per_cpu(od_cpu_dbs_info, cpu).cdbs.policy_dbs;
 446                 if (!policy_dbs)
 447                         continue;
 448
 449                 policy = policy_dbs->policy;
 450                 cpumask_or(&done, &done, policy->cpus);
 451
 452                 if (policy->governor != CPU_FREQ_GOV_ONDEMAND)
 453                         continue;
 454
 455                 dbs_data = policy_dbs->dbs_data;
 456                 od_tuners = dbs_data->tuners;
 457                 od_tuners->powersave_bias = default_powersave_bias;
 458         }
 459         put_online_cpus();
 460 }
 461
 462 void od_register_powersave_bias_handler(unsigned int (*f)
 463                 (struct cpufreq_policy *, unsigned int, unsigned int),
 464                 unsigned int powersave_bias)
 465 {
 466         od_ops.powersave_bias_target = f;
 467         od_set_powersave_bias(powersave_bias);
 468 }
 469 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
 470
 471 void od_unregister_powersave_bias_handler(void)
 472 {
 473         od_ops.powersave_bias_target = generic_powersave_bias_target;
 474         od_set_powersave_bias(0);
 475 }
 476 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
 477
 478 static int __init cpufreq_gov_dbs_init(void)
 479 {
 480         return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
 481 }
 482
 483 static void __exit cpufreq_gov_dbs_exit(void)
 484 {
 485         cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
 486 }
 487
 488 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 489 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 490 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 491         "Low Latency Frequency Transition capable processors");
 492 MODULE_LICENSE("GPL");
 493
 494 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 495 struct cpufreq_governor *cpufreq_default_governor(void)
 496 {
 497         return CPU_FREQ_GOV_ONDEMAND;
 498 }
 499
 500 fs_initcall(cpufreq_gov_dbs_init);
 501 #else
 502 module_init(cpufreq_gov_dbs_init);
 503 #endif
 504 module_exit(cpufreq_gov_dbs_exit);