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_jiffies,
  70  * freq_lo, and freq_lo_jiffies 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 jiffies_total, jiffies_hi, jiffies_lo;
  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_jiffies = 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_jiffies = 0;
 111                 return freq_lo;
 112         }
 113         jiffies_total = usecs_to_jiffies(dbs_data->sampling_rate);
 114         jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
 115         jiffies_hi += ((freq_hi - freq_lo) / 2);
 116         jiffies_hi /= (freq_hi - freq_lo);
 117         jiffies_lo = jiffies_total - jiffies_hi;
 118         dbs_info->freq_lo = freq_lo;
 119         dbs_info->freq_lo_jiffies = jiffies_lo;
 120         dbs_info->freq_hi_jiffies = jiffies_hi;
 121         return freq_hi;
 122 }
 123
 124 static void ondemand_powersave_bias_init(void)
 125 {
 126         int i;
 127         for_each_online_cpu(i) {
 128                 ondemand_powersave_bias_init_cpu(i);
 129         }
 130 }
 131
 132 static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
 133 {
 134         struct policy_dbs_info *policy_dbs = policy->governor_data;
 135         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 136         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 137
 138         if (od_tuners->powersave_bias)
 139                 freq = od_ops.powersave_bias_target(policy, freq,
 140                                 CPUFREQ_RELATION_H);
 141         else if (policy->cur == policy->max)
 142                 return;
 143
 144         __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
 145                         CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
 146 }
 147
 148 /*
 149  * Every sampling_rate, we check, if current idle time is less than 20%
 150  * (default), then we try to increase frequency. Else, we adjust the frequency
 151  * proportional to load.
 152  */
 153 static void od_check_cpu(int cpu, unsigned int load)
 154 {
 155         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu);
 156         struct policy_dbs_info *policy_dbs = dbs_info->cdbs.policy_dbs;
 157         struct cpufreq_policy *policy = policy_dbs->policy;
 158         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 159         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 160
 161         dbs_info->freq_lo = 0;
 162
 163         /* Check for frequency increase */
 164         if (load > dbs_data->up_threshold) {
 165                 /* If switching to max speed, apply sampling_down_factor */
 166                 if (policy->cur < policy->max)
 167                         dbs_info->rate_mult = dbs_data->sampling_down_factor;
 168                 dbs_freq_increase(policy, policy->max);
 169         } else {
 170                 /* Calculate the next frequency proportional to load */
 171                 unsigned int freq_next, min_f, max_f;
 172
 173                 min_f = policy->cpuinfo.min_freq;
 174                 max_f = policy->cpuinfo.max_freq;
 175                 freq_next = min_f + load * (max_f - min_f) / 100;
 176
 177                 /* No longer fully busy, reset rate_mult */
 178                 dbs_info->rate_mult = 1;
 179
 180                 if (!od_tuners->powersave_bias) {
 181                         __cpufreq_driver_target(policy, freq_next,
 182                                         CPUFREQ_RELATION_C);
 183                         return;
 184                 }
 185
 186                 freq_next = od_ops.powersave_bias_target(policy, freq_next,
 187                                         CPUFREQ_RELATION_L);
 188                 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
 189         }
 190 }
 191
 192 static unsigned int od_dbs_timer(struct cpufreq_policy *policy)
 193 {
 194         struct policy_dbs_info *policy_dbs = policy->governor_data;
 195         struct dbs_data *dbs_data = policy_dbs->dbs_data;
 196         struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, policy->cpu);
 197         int delay, sample_type = dbs_info->sample_type;
 198
 199         /* Common NORMAL_SAMPLE setup */
 200         dbs_info->sample_type = OD_NORMAL_SAMPLE;
 201         if (sample_type == OD_SUB_SAMPLE) {
 202                 delay = dbs_info->freq_lo_jiffies;
 203                 __cpufreq_driver_target(policy, dbs_info->freq_lo,
 204                                         CPUFREQ_RELATION_H);
 205         } else {
 206                 dbs_check_cpu(policy);
 207                 if (dbs_info->freq_lo) {
 208                         /* Setup timer for SUB_SAMPLE */
 209                         dbs_info->sample_type = OD_SUB_SAMPLE;
 210                         delay = dbs_info->freq_hi_jiffies;
 211                 } else {
 212                         delay = delay_for_sampling_rate(dbs_data->sampling_rate
 213                                                         * dbs_info->rate_mult);
 214                 }
 215         }
 216
 217         return delay;
 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         unsigned int input, j;
 266         int ret;
 267         ret = sscanf(buf, "%u", &input);
 268
 269         if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
 270                 return -EINVAL;
 271         dbs_data->sampling_down_factor = input;
 272
 273         /* Reset down sampling multiplier in case it was active */
 274         for_each_online_cpu(j) {
 275                 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 276                                 j);
 277                 dbs_info->rate_mult = 1;
 278         }
 279         return count;
 280 }
 281
 282 static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data,
 283                 const char *buf, size_t count)
 284 {
 285         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 286         unsigned int input;
 287         int ret;
 288
 289         unsigned int j;
 290
 291         ret = sscanf(buf, "%u", &input);
 292         if (ret != 1)
 293                 return -EINVAL;
 294
 295         if (input > 1)
 296                 input = 1;
 297
 298         if (input == dbs_data->ignore_nice_load) { /* nothing to do */
 299                 return count;
 300         }
 301         dbs_data->ignore_nice_load = input;
 302
 303         /* we need to re-evaluate prev_cpu_idle */
 304         for_each_online_cpu(j) {
 305                 struct od_cpu_dbs_info_s *dbs_info;
 306                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 307                 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 308                         &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
 309                 if (dbs_data->ignore_nice_load)
 310                         dbs_info->cdbs.prev_cpu_nice =
 311                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 312
 313         }
 314         return count;
 315 }
 316
 317 static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
 318                 size_t count)
 319 {
 320         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 321         unsigned int input;
 322         int ret;
 323         ret = sscanf(buf, "%u", &input);
 324
 325         if (ret != 1)
 326                 return -EINVAL;
 327
 328         if (input > 1000)
 329                 input = 1000;
 330
 331         od_tuners->powersave_bias = input;
 332         ondemand_powersave_bias_init();
 333         return count;
 334 }
 335
 336 gov_show_one_common(sampling_rate);
 337 gov_show_one_common(up_threshold);
 338 gov_show_one_common(sampling_down_factor);
 339 gov_show_one_common(ignore_nice_load);
 340 gov_show_one_common(min_sampling_rate);
 341 gov_show_one(od, io_is_busy);
 342 gov_show_one(od, powersave_bias);
 343
 344 gov_attr_rw(sampling_rate);
 345 gov_attr_rw(io_is_busy);
 346 gov_attr_rw(up_threshold);
 347 gov_attr_rw(sampling_down_factor);
 348 gov_attr_rw(ignore_nice_load);
 349 gov_attr_rw(powersave_bias);
 350 gov_attr_ro(min_sampling_rate);
 351
 352 static struct attribute *od_attributes[] = {
 353         &min_sampling_rate.attr,
 354         &sampling_rate.attr,
 355         &up_threshold.attr,
 356         &sampling_down_factor.attr,
 357         &ignore_nice_load.attr,
 358         &powersave_bias.attr,
 359         &io_is_busy.attr,
 360         NULL
 361 };
 362
 363 /************************** sysfs end ************************/
 364
 365 static int od_init(struct dbs_data *dbs_data, bool notify)
 366 {
 367         struct od_dbs_tuners *tuners;
 368         u64 idle_time;
 369         int cpu;
 370
 371         tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
 372         if (!tuners) {
 373                 pr_err("%s: kzalloc failed\n", __func__);
 374                 return -ENOMEM;
 375         }
 376
 377         cpu = get_cpu();
 378         idle_time = get_cpu_idle_time_us(cpu, NULL);
 379         put_cpu();
 380         if (idle_time != -1ULL) {
 381                 /* Idle micro accounting is supported. Use finer thresholds */
 382                 dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 383                 /*
 384                  * In nohz/micro accounting case we set the minimum frequency
 385                  * not depending on HZ, but fixed (very low). The deferred
 386                  * timer might skip some samples if idle/sleeping as needed.
 387                 */
 388                 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 389         } else {
 390                 dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
 391
 392                 /* For correct statistics, we need 10 ticks for each measure */
 393                 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
 394                         jiffies_to_usecs(10);
 395         }
 396
 397         dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
 398         dbs_data->ignore_nice_load = 0;
 399         tuners->powersave_bias = default_powersave_bias;
 400         tuners->io_is_busy = should_io_be_busy();
 401
 402         dbs_data->tuners = tuners;
 403         return 0;
 404 }
 405
 406 static void od_exit(struct dbs_data *dbs_data, bool notify)
 407 {
 408         kfree(dbs_data->tuners);
 409 }
 410
 411 define_get_cpu_dbs_routines(od_cpu_dbs_info);
 412
 413 static struct od_ops od_ops = {
 414         .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
 415         .powersave_bias_target = generic_powersave_bias_target,
 416         .freq_increase = dbs_freq_increase,
 417 };
 418
 419 static struct dbs_governor od_dbs_gov = {
 420         .gov = {
 421                 .name = "ondemand",
 422                 .governor = cpufreq_governor_dbs,
 423                 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 424                 .owner = THIS_MODULE,
 425         },
 426         .governor = GOV_ONDEMAND,
 427         .kobj_type = { .default_attrs = od_attributes },
 428         .get_cpu_cdbs = get_cpu_cdbs,
 429         .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
 430         .gov_dbs_timer = od_dbs_timer,
 431         .gov_check_cpu = od_check_cpu,
 432         .gov_ops = &od_ops,
 433         .init = od_init,
 434         .exit = od_exit,
 435 };
 436
 437 #define CPU_FREQ_GOV_ONDEMAND   (&od_dbs_gov.gov)
 438
 439 static void od_set_powersave_bias(unsigned int powersave_bias)
 440 {
 441         struct cpufreq_policy *policy;
 442         struct dbs_data *dbs_data;
 443         struct od_dbs_tuners *od_tuners;
 444         unsigned int cpu;
 445         cpumask_t done;
 446
 447         default_powersave_bias = powersave_bias;
 448         cpumask_clear(&done);
 449
 450         get_online_cpus();
 451         for_each_online_cpu(cpu) {
 452                 struct policy_dbs_info *policy_dbs;
 453
 454                 if (cpumask_test_cpu(cpu, &done))
 455                         continue;
 456
 457                 policy_dbs = per_cpu(od_cpu_dbs_info, cpu).cdbs.policy_dbs;
 458                 if (!policy_dbs)
 459                         continue;
 460
 461                 policy = policy_dbs->policy;
 462                 cpumask_or(&done, &done, policy->cpus);
 463
 464                 if (policy->governor != CPU_FREQ_GOV_ONDEMAND)
 465                         continue;
 466
 467                 dbs_data = policy_dbs->dbs_data;
 468                 od_tuners = dbs_data->tuners;
 469                 od_tuners->powersave_bias = default_powersave_bias;
 470         }
 471         put_online_cpus();
 472 }
 473
 474 void od_register_powersave_bias_handler(unsigned int (*f)
 475                 (struct cpufreq_policy *, unsigned int, unsigned int),
 476                 unsigned int powersave_bias)
 477 {
 478         od_ops.powersave_bias_target = f;
 479         od_set_powersave_bias(powersave_bias);
 480 }
 481 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
 482
 483 void od_unregister_powersave_bias_handler(void)
 484 {
 485         od_ops.powersave_bias_target = generic_powersave_bias_target;
 486         od_set_powersave_bias(0);
 487 }
 488 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
 489
 490 static int __init cpufreq_gov_dbs_init(void)
 491 {
 492         return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
 493 }
 494
 495 static void __exit cpufreq_gov_dbs_exit(void)
 496 {
 497         cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
 498 }
 499
 500 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 501 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 502 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 503         "Low Latency Frequency Transition capable processors");
 504 MODULE_LICENSE("GPL");
 505
 506 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 507 struct cpufreq_governor *cpufreq_default_governor(void)
 508 {
 509         return CPU_FREQ_GOV_ONDEMAND;
 510 }
 511
 512 fs_initcall(cpufreq_gov_dbs_init);
 513 #else
 514 module_init(cpufreq_gov_dbs_init);
 515 #endif
 516 module_exit(cpufreq_gov_dbs_exit);