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 = 0, 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                 }
 212         }
 213
 214         if (!delay)
 215                 delay = delay_for_sampling_rate(dbs_data->sampling_rate
 216                                 * dbs_info->rate_mult);
 217
 218         return delay;
 219 }
 220
 221 /************************** sysfs interface ************************/
 222 static struct dbs_governor od_dbs_gov;
 223
 224 static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf,
 225                 size_t count)
 226 {
 227         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 228         unsigned int input;
 229         int ret;
 230         unsigned int j;
 231
 232         ret = sscanf(buf, "%u", &input);
 233         if (ret != 1)
 234                 return -EINVAL;
 235         od_tuners->io_is_busy = !!input;
 236
 237         /* we need to re-evaluate prev_cpu_idle */
 238         for_each_online_cpu(j) {
 239                 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 240                                                                         j);
 241                 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 242                         &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
 243         }
 244         return count;
 245 }
 246
 247 static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf,
 248                 size_t count)
 249 {
 250         unsigned int input;
 251         int ret;
 252         ret = sscanf(buf, "%u", &input);
 253
 254         if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
 255                         input < MIN_FREQUENCY_UP_THRESHOLD) {
 256                 return -EINVAL;
 257         }
 258
 259         dbs_data->up_threshold = input;
 260         return count;
 261 }
 262
 263 static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data,
 264                 const char *buf, size_t count)
 265 {
 266         unsigned int input, j;
 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         dbs_data->sampling_down_factor = input;
 273
 274         /* Reset down sampling multiplier in case it was active */
 275         for_each_online_cpu(j) {
 276                 struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info,
 277                                 j);
 278                 dbs_info->rate_mult = 1;
 279         }
 280         return count;
 281 }
 282
 283 static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data,
 284                 const char *buf, size_t count)
 285 {
 286         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 287         unsigned int input;
 288         int ret;
 289
 290         unsigned int j;
 291
 292         ret = sscanf(buf, "%u", &input);
 293         if (ret != 1)
 294                 return -EINVAL;
 295
 296         if (input > 1)
 297                 input = 1;
 298
 299         if (input == dbs_data->ignore_nice_load) { /* nothing to do */
 300                 return count;
 301         }
 302         dbs_data->ignore_nice_load = input;
 303
 304         /* we need to re-evaluate prev_cpu_idle */
 305         for_each_online_cpu(j) {
 306                 struct od_cpu_dbs_info_s *dbs_info;
 307                 dbs_info = &per_cpu(od_cpu_dbs_info, j);
 308                 dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j,
 309                         &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy);
 310                 if (dbs_data->ignore_nice_load)
 311                         dbs_info->cdbs.prev_cpu_nice =
 312                                 kcpustat_cpu(j).cpustat[CPUTIME_NICE];
 313
 314         }
 315         return count;
 316 }
 317
 318 static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf,
 319                 size_t count)
 320 {
 321         struct od_dbs_tuners *od_tuners = dbs_data->tuners;
 322         unsigned int input;
 323         int ret;
 324         ret = sscanf(buf, "%u", &input);
 325
 326         if (ret != 1)
 327                 return -EINVAL;
 328
 329         if (input > 1000)
 330                 input = 1000;
 331
 332         od_tuners->powersave_bias = input;
 333         ondemand_powersave_bias_init();
 334         return count;
 335 }
 336
 337 gov_show_one_common(sampling_rate);
 338 gov_show_one_common(up_threshold);
 339 gov_show_one_common(sampling_down_factor);
 340 gov_show_one_common(ignore_nice_load);
 341 gov_show_one_common(min_sampling_rate);
 342 gov_show_one(od, io_is_busy);
 343 gov_show_one(od, powersave_bias);
 344
 345 gov_attr_rw(sampling_rate);
 346 gov_attr_rw(io_is_busy);
 347 gov_attr_rw(up_threshold);
 348 gov_attr_rw(sampling_down_factor);
 349 gov_attr_rw(ignore_nice_load);
 350 gov_attr_rw(powersave_bias);
 351 gov_attr_ro(min_sampling_rate);
 352
 353 static struct attribute *od_attributes[] = {
 354         &min_sampling_rate.attr,
 355         &sampling_rate.attr,
 356         &up_threshold.attr,
 357         &sampling_down_factor.attr,
 358         &ignore_nice_load.attr,
 359         &powersave_bias.attr,
 360         &io_is_busy.attr,
 361         NULL
 362 };
 363
 364 /************************** sysfs end ************************/
 365
 366 static int od_init(struct dbs_data *dbs_data, bool notify)
 367 {
 368         struct od_dbs_tuners *tuners;
 369         u64 idle_time;
 370         int cpu;
 371
 372         tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
 373         if (!tuners) {
 374                 pr_err("%s: kzalloc failed\n", __func__);
 375                 return -ENOMEM;
 376         }
 377
 378         cpu = get_cpu();
 379         idle_time = get_cpu_idle_time_us(cpu, NULL);
 380         put_cpu();
 381         if (idle_time != -1ULL) {
 382                 /* Idle micro accounting is supported. Use finer thresholds */
 383                 dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
 384                 /*
 385                  * In nohz/micro accounting case we set the minimum frequency
 386                  * not depending on HZ, but fixed (very low). The deferred
 387                  * timer might skip some samples if idle/sleeping as needed.
 388                 */
 389                 dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
 390         } else {
 391                 dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
 392
 393                 /* For correct statistics, we need 10 ticks for each measure */
 394                 dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO *
 395                         jiffies_to_usecs(10);
 396         }
 397
 398         dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
 399         dbs_data->ignore_nice_load = 0;
 400         tuners->powersave_bias = default_powersave_bias;
 401         tuners->io_is_busy = should_io_be_busy();
 402
 403         dbs_data->tuners = tuners;
 404         return 0;
 405 }
 406
 407 static void od_exit(struct dbs_data *dbs_data, bool notify)
 408 {
 409         kfree(dbs_data->tuners);
 410 }
 411
 412 define_get_cpu_dbs_routines(od_cpu_dbs_info);
 413
 414 static struct od_ops od_ops = {
 415         .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu,
 416         .powersave_bias_target = generic_powersave_bias_target,
 417         .freq_increase = dbs_freq_increase,
 418 };
 419
 420 static struct dbs_governor od_dbs_gov = {
 421         .gov = {
 422                 .name = "ondemand",
 423                 .governor = cpufreq_governor_dbs,
 424                 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
 425                 .owner = THIS_MODULE,
 426         },
 427         .governor = GOV_ONDEMAND,
 428         .kobj_type = { .default_attrs = od_attributes },
 429         .get_cpu_cdbs = get_cpu_cdbs,
 430         .get_cpu_dbs_info_s = get_cpu_dbs_info_s,
 431         .gov_dbs_timer = od_dbs_timer,
 432         .gov_check_cpu = od_check_cpu,
 433         .gov_ops = &od_ops,
 434         .init = od_init,
 435         .exit = od_exit,
 436 };
 437
 438 #define CPU_FREQ_GOV_ONDEMAND   (&od_dbs_gov.gov)
 439
 440 static void od_set_powersave_bias(unsigned int powersave_bias)
 441 {
 442         struct cpufreq_policy *policy;
 443         struct dbs_data *dbs_data;
 444         struct od_dbs_tuners *od_tuners;
 445         unsigned int cpu;
 446         cpumask_t done;
 447
 448         default_powersave_bias = powersave_bias;
 449         cpumask_clear(&done);
 450
 451         get_online_cpus();
 452         for_each_online_cpu(cpu) {
 453                 struct policy_dbs_info *policy_dbs;
 454
 455                 if (cpumask_test_cpu(cpu, &done))
 456                         continue;
 457
 458                 policy_dbs = per_cpu(od_cpu_dbs_info, cpu).cdbs.policy_dbs;
 459                 if (!policy_dbs)
 460                         continue;
 461
 462                 policy = policy_dbs->policy;
 463                 cpumask_or(&done, &done, policy->cpus);
 464
 465                 if (policy->governor != CPU_FREQ_GOV_ONDEMAND)
 466                         continue;
 467
 468                 dbs_data = policy_dbs->dbs_data;
 469                 od_tuners = dbs_data->tuners;
 470                 od_tuners->powersave_bias = default_powersave_bias;
 471         }
 472         put_online_cpus();
 473 }
 474
 475 void od_register_powersave_bias_handler(unsigned int (*f)
 476                 (struct cpufreq_policy *, unsigned int, unsigned int),
 477                 unsigned int powersave_bias)
 478 {
 479         od_ops.powersave_bias_target = f;
 480         od_set_powersave_bias(powersave_bias);
 481 }
 482 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
 483
 484 void od_unregister_powersave_bias_handler(void)
 485 {
 486         od_ops.powersave_bias_target = generic_powersave_bias_target;
 487         od_set_powersave_bias(0);
 488 }
 489 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
 490
 491 static int __init cpufreq_gov_dbs_init(void)
 492 {
 493         return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
 494 }
 495
 496 static void __exit cpufreq_gov_dbs_exit(void)
 497 {
 498         cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
 499 }
 500
 501 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
 502 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
 503 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
 504         "Low Latency Frequency Transition capable processors");
 505 MODULE_LICENSE("GPL");
 506
 507 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
 508 struct cpufreq_governor *cpufreq_default_governor(void)
 509 {
 510         return CPU_FREQ_GOV_ONDEMAND;
 511 }
 512
 513 fs_initcall(cpufreq_gov_dbs_init);
 514 #else
 515 module_init(cpufreq_gov_dbs_init);
 516 #endif
 517 module_exit(cpufreq_gov_dbs_exit);