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 od_ops od_ops;
  33
  34 static unsigned int default_powersave_bias;
  35
  36 static void ondemand_powersave_bias_init_cpu(int cpu)
  37 {
  38         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
  39
  40         dbs_info->freq_table = cpufreq_frequency_get_table(cpu);
  41         dbs_info->freq_lo = 0;
  42 }
  43
  44 /*
  45  * Not all CPUs want IO time to be accounted as busy; this depends on how
  46  * efficient idling at a higher frequency/voltage is.
  47  * Pavel Machek says this is not so for various generations of AMD and old
  48  * Intel systems.
  49  * Mike Chan (android.com) claims this is also not true for ARM.
  50  * Because of this, whitelist specific known (series) of CPUs by default, and
  51  * leave all others up to the user.
  52  */
  53 static int should_io_be_busy(void)
  54 {
  55 #if defined(CONFIG_X86)
  56         /*
  57          * For Intel, Core 2 (model 15) and later have an efficient idle.
  58          */
  59         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
  60                         boot_cpu_data.x86 == 6 &&
  61                         boot_cpu_data.x86_model >= 15)
  62                 return 1;
  63 #endif
  64         return 0;
  65 }
  66
  67 /*
  68  * Find right freq to be set now with powersave_bias on.
  69  * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
  70  * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
  71  */
  72 static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
  73                 unsigned int freq_next, unsigned int relation)
  74 {
  75         unsigned int freq_req, freq_reduc, freq_avg;
  76         unsigned int freq_hi, freq_lo;
  77         unsigned int index = 0;
  78         unsigned int delay_hi_us;
  79         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
  80                                                    policy->cpu);
  81         struct policy_dbs_info *policy_dbs = policy->governor_data;
  82         struct dbs_data *dbs_data = policy_dbs->dbs_data;
  83         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
  84
  85         if (!dbs_info->freq_table) {
  86                 dbs_info->freq_lo = 0;
  87                 dbs_info->freq_lo_delay_us = 0;
  88                 return freq_next;
  89         }
  90
  91         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
  92                         relation, &index);
  93         freq_req = dbs_info->freq_table[index].frequency;
  94         freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
  95         freq_avg = freq_req - freq_reduc;
  96
  97         /* Find freq bounds for freq_avg in freq_table */
  98         index = 0;
  99         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 100                         CPUFREQ_RELATION_H, &index);
 101         freq_lo = dbs_info->freq_table[index].frequency;
 102         index = 0;
 103         cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
 104                         CPUFREQ_RELATION_L, &index);
 105         freq_hi = dbs_info->freq_table[index].frequency;
 106
 107         /* Find out how long we have to be in hi and lo freqs */
 108         if (freq_hi == freq_lo) {
 109                 dbs_info->freq_lo = 0;
 110                 dbs_info->freq_lo_delay_us = 0;
 111                 return freq_lo;
 112         }
 113         delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
 114         delay_hi_us += (freq_hi - freq_lo) / 2;
 115         delay_hi_us /= freq_hi - freq_lo;
 116         dbs_info->freq_hi_delay_us = delay_hi_us;
 117         dbs_info->freq_lo = freq_lo;
 118         dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
 119         return freq_hi;
 120 }
 121
 122 static void ondemand_powersave_bias_init(void)
 123 {
 124         int i;
 125         for_each_online_cpu(i) {
 126                 ondemand_powersave_bias_init_cpu(i);
 127         }
 128 }
 129
 130 static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
 131 {
 132         struct policy_dbs_info *policy_dbs = policy->governor_data;
 133         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 134         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 135
 136         if (od_tuners->powersave_bias)
 137                 freq = od_ops.powersave_bias_target(policy, freq,
 138                                 CPUFREQ_RELATION_H);
 139         else if (policy->cur == policy->max)
 140                 return;
 141
 142         __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
 143                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 144 }
 145
 146 /*
 147  * Every sampling_rate, we check, if current idle time is less than 20%
 148  * (default), then we try to increase frequency. Else, we adjust the frequency
 149  * proportional to load.
 150  */
 151 static void od_update(struct cpufreq_policy *policy)
 152 {
 153         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 154         struct policy_dbs_info *policy_dbs = dbs_info->cdbs.policy_dbs;
 155         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 156         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 157         unsigned int load = dbs_update(policy);
 158
 159         dbs_info->freq_lo = 0;
 160
 161         /* Check for frequency increase */
 162         if (load > dbs_data->up_threshold) {
 163                 /* If switching to max speed, apply sampling_down_factor */
 164                 if (policy->cur < policy->max)
 165                         policy_dbs->rate_mult = dbs_data->sampling_down_factor;
 166                 dbs_freq_increase(policy, policy->max);
 167         } else {
 168                 /* Calculate the next frequency proportional to load */
 169                 unsigned int freq_next, min_f, max_f;
 170
 171                 min_f = policy->cpuinfo.min_freq;
 172                 max_f = policy->cpuinfo.max_freq;
 173                 freq_next = min_f + load * (max_f - min_f) / 100;
 174
 175                 /* No longer fully busy, reset rate_mult */
 176                 policy_dbs->rate_mult = 1;
 177
 178                 if (!od_tuners->powersave_bias) {
 179                         __cpufreq_driver_target(policy, freq_next,
 180                                         CPUFREQ_RELATION_C);
 181                         return;
 182                 }
 183
 184                 freq_next = od_ops.powersave_bias_target(policy, freq_next,
 185                                         CPUFREQ_RELATION_L);
 186                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
 187         }
 188 }
 189
 190 static unsigned int od_dbs_timer(struct cpufreq_policy *policy)
 191 {
 192         struct policy_dbs_info *policy_dbs = policy->governor_data;
 193         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 194         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 195         int sample_type = dbs_info->sample_type;
 196
 197         /* Common NORMAL_SAMPLE setup */
 198         dbs_info->sample_type = OD_NORMAL_SAMPLE;
 199         /*
 200          * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
 201          * it then.
 202          */
 203         if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
 204                 __cpufreq_driver_target(policy, dbs_info->freq_lo,
 205                                         CPUFREQ_RELATION_H);
 206                 return dbs_info->freq_lo_delay_us;
 207         }
 208
 209         od_update(policy);
 210
 211         if (dbs_info->freq_lo) {
 212                 /* Setup timer for SUB_SAMPLE */
 213                 dbs_info->sample_type = OD_SUB_SAMPLE;
 214                 return dbs_info->freq_hi_delay_us;
 215         }
 216
 217         return dbs_data->sampling_rate * policy_dbs->rate_mult;
 218 }
 219
 220 /************************** sysfs interface ************************/
 221 static struct dbs_governor od_dbs_gov;
 222
 223 static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf,
 224                 size_t count)
 225 {
 226         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 227         unsigned int input;
 228         int ret;
 229         unsigned int j;
 230
 231         ret = sscanf(buf, "%u", &input);
 232         if (ret != 1)
 233                 return -EINVAL;
 234         od_tuners->io_is_busy = !!input;
 235
 236         /* we need to re-evaluate prev_cpu_idle */
 237         for_each_online_cpu(j) {
 238                 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 239                                                                         j);
 240                 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 241                         &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
 242         }
 243         return count;
 244 }
 245
 246 static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf,
 247                 size_t count)
 248 {
 249         unsigned int input;
 250         int ret;
 251         ret = sscanf(buf, "%u", &input);
 252
 253         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 254                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 255                 return -EINVAL;
 256         }
 257
 258         dbs_data->up_threshold = input;
 259         return count;
 260 }
 261
 262 static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data,
 263                 const char *buf, size_t count)
 264 {
 265         struct policy_dbs_info *policy_dbs;
 266         unsigned int input;
 267         int ret;
 268         ret = sscanf(buf, "%u", &input);
 269
 270         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 271                 return -EINVAL;
 272
 273         dbs_data->sampling_down_factor = input;
 274
 275         /* Reset down sampling multiplier in case it was active */
 276         list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
 277                 /*
 278                  * Doing this without locking might lead to using different
 279                  * rate_mult values in od_update() and od_dbs_timer().
 280                  */
 281                 mutex_lock(&policy_dbs->timer_mutex);
 282                 policy_dbs->rate_mult = 1;
 283                 mutex_unlock(&policy_dbs->timer_mutex);
 284         }
 285
 286         return count;
 287 }
 288
 289 static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data,
 290                 const char *buf, size_t count)
 291 {
 292         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 293         unsigned int input;
 294         int ret;
 295
 296         unsigned int j;
 297
 298         ret = sscanf(buf, "%u", &input);
 299         if (ret != 1)
 300                 return -EINVAL;
 301
 302         if (input > 1)
 303                 input = 1;
 304
 305         if (input == dbs_data->ignore_nice_load) { /* nothing to do */
 306                 return count;
 307         }
 308         dbs_data->ignore_nice_load = input;
 309
 310         /* we need to re-evaluate prev_cpu_idle */
 311         for_each_online_cpu(j) {
 312                 struct od_cpu_dbs_info_s *dbs_info;
 313                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 314                 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 315                         &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
 316                 if (dbs_data->ignore_nice_load)
 317                         dbs_info->cdbs.prev_cpu_nice =
 318                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 319
 320         }
 321         return count;
 322 }
 323
 324 static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
 325                 size_t count)
 326 {
 327         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 328         unsigned int input;
 329         int ret;
 330         ret = sscanf(buf, "%u", &input);
 331
 332         if (ret != 1)
 333                 return -EINVAL;
 334
 335         if (input > 1000)
 336                 input = 1000;
 337
 338         od_tuners->powersave_bias = input;
 339         ondemand_powersave_bias_init();
 340         return count;
 341 }
 342
 343 gov_show_one_common(sampling_rate);
 344 gov_show_one_common(up_threshold);
 345 gov_show_one_common(sampling_down_factor);
 346 gov_show_one_common(ignore_nice_load);
 347 gov_show_one_common(min_sampling_rate);
 348 gov_show_one(od, io_is_busy);
 349 gov_show_one(od, powersave_bias);
 350
 351 gov_attr_rw(sampling_rate);
 352 gov_attr_rw(io_is_busy);
 353 gov_attr_rw(up_threshold);
 354 gov_attr_rw(sampling_down_factor);
 355 gov_attr_rw(ignore_nice_load);
 356 gov_attr_rw(powersave_bias);
 357 gov_attr_ro(min_sampling_rate);
 358
 359 static struct attribute *od_attributes[] = {
 360         &min_sampling_rate.attr,
 361         &sampling_rate.attr,
 362         &up_threshold.attr,
 363         &sampling_down_factor.attr,
 364         &ignore_nice_load.attr,
 365         &powersave_bias.attr,
 366         &io_is_busy.attr,
 367         NULL
 368 };
 369
 370 /************************** sysfs end ************************/
 371
 372 static int od_init(struct dbs_data *dbs_data, bool notify)
 373 {
 374         struct od_dbs_tuners *tuners;
 375         u64 idle_time;
 376         int cpu;
 377
 378         tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
 379         if (!tuners) {
 380                 pr_err("%s: kzalloc failed\n", __func__);
 381                 return -ENOMEM;
 382         }
 383
 384         cpu = get_cpu();
 385         idle_time = get_cpu_idle_time_us(cpu, NULL);
 386         put_cpu();
 387         if (idle_time != -1ULL) {
 388                 /* Idle micro accounting is supported. Use finer thresholds */
 389                 dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 390                 /*
 391                  * In nohz/micro accounting case we set the minimum frequency
 392                  * not depending on HZ, but fixed (very low). The deferred
 393                  * timer might skip some samples if idle/sleeping as needed.
 394                 */
 395                 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 396         } else {
 397                 dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
 398
 399                 /* For correct statistics, we need 10 ticks for each measure */
 400                 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
 401                         jiffies_to_usecs(10);
 402         }
 403
 404         dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
 405         dbs_data->ignore_nice_load = 0;
 406         tuners->powersave_bias = default_powersave_bias;
 407         tuners->io_is_busy = should_io_be_busy();
 408
 409         dbs_data->tuners = tuners;
 410         return 0;
 411 }
 412
 413 static void od_exit(struct dbs_data *dbs_data, bool notify)
 414 {
 415         kfree(dbs_data->tuners);
 416 }
 417
 418 define_get_cpu_dbs_routines(od_cpu_dbs_info);
 419
 420 static struct od_ops od_ops = {
 421         .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
 422         .powersave_bias_target = generic_powersave_bias_target,
 423         .freq_increase = dbs_freq_increase,
 424 };
 425
 426 static struct dbs_governor od_dbs_gov = {
 427         .gov = {
 428                 .name = "ondemand",
 429                 .governor = cpufreq_governor_dbs,
 430                 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 431                 .owner = THIS_MODULE,
 432         },
 433         .governor = GOV_ONDEMAND,
 434         .kobj_type = { .default_attrs = od_attributes },
 435         .get_cpu_cdbs = get_cpu_cdbs,
 436         .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
 437         .gov_dbs_timer = od_dbs_timer,
 438         .gov_ops = &od_ops,
 439         .init = od_init,
 440         .exit = od_exit,
 441 };
 442
 443 #define CPU_FREQ_GOV_ONDEMAND   (&od_dbs_gov.gov)
 444
 445 static void od_set_powersave_bias(unsigned int powersave_bias)
 446 {
 447         struct cpufreq_policy *policy;
 448         struct dbs_data *dbs_data;
 449         struct od_dbs_tuners *od_tuners;
 450         unsigned int cpu;
 451         cpumask_t done;
 452
 453         default_powersave_bias = powersave_bias;
 454         cpumask_clear(&done);
 455
 456         get_online_cpus();
 457         for_each_online_cpu(cpu) {
 458                 struct policy_dbs_info *policy_dbs;
 459
 460                 if (cpumask_test_cpu(cpu, &done))
 461                         continue;
 462
 463                 policy_dbs = per_cpu(od_cpu_dbs_info, cpu).cdbs.policy_dbs;
 464                 if (!policy_dbs)
 465                         continue;
 466
 467                 policy = policy_dbs->policy;
 468                 cpumask_or(&done, &done, policy->cpus);
 469
 470                 if (policy->governor != CPU_FREQ_GOV_ONDEMAND)
 471                         continue;
 472
 473                 dbs_data = policy_dbs->dbs_data;
 474                 od_tuners = dbs_data->tuners;
 475                 od_tuners->powersave_bias = default_powersave_bias;
 476         }
 477         put_online_cpus();
 478 }
 479
 480 void od_register_powersave_bias_handler(unsigned int (*f)
 481                 (struct cpufreq_policy *, unsigned int, unsigned int),
 482                 unsigned int powersave_bias)
 483 {
 484         od_ops.powersave_bias_target = f;
 485         od_set_powersave_bias(powersave_bias);
 486 }
 487 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
 488
 489 void od_unregister_powersave_bias_handler(void)
 490 {
 491         od_ops.powersave_bias_target = generic_powersave_bias_target;
 492         od_set_powersave_bias(0);
 493 }
 494 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
 495
 496 static int __init cpufreq_gov_dbs_init(void)
 497 {
 498         return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
 499 }
 500
 501 static void __exit cpufreq_gov_dbs_exit(void)
 502 {
 503         cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
 504 }
 505
 506 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 507 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 508 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 509         "Low Latency Frequency Transition capable processors");
 510 MODULE_LICENSE("GPL");
 511
 512 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 513 struct cpufreq_governor *cpufreq_default_governor(void)
 514 {
 515         return CPU_FREQ_GOV_ONDEMAND;
 516 }
 517
 518 fs_initcall(cpufreq_gov_dbs_init);
 519 #else
 520 module_init(cpufreq_gov_dbs_init);
 521 #endif
 522 module_exit(cpufreq_gov_dbs_exit);