drivers/cpufreq/cpufreq_interactive.c

   1 /*
   2  * drivers/cpufreq/cpufreq_interactive.c
   3  *
   4  * Copyright (C) 2010 Google, Inc.
   5  *
   6  * This software is licensed under the terms of the GNU General Public
   7  * License version 2, as published by the Free Software Foundation, and
   8  * may be copied, distributed, and modified under those terms.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * Author: Mike Chan (mike@android.com)
  16  *
  17  */
  18
  19 #include <linux/cpu.h>
  20 #include <linux/cpumask.h>
  21 #include <linux/cpufreq.h>
  22 #include <linux/module.h>
  23 #include <linux/moduleparam.h>
  24 #include <linux/rwsem.h>
  25 #include <linux/sched.h>
  26 #include <linux/sched/rt.h>
  27 #include <linux/tick.h>
  28 #include <linux/time.h>
  29 #include <linux/timer.h>
  30 #include <linux/workqueue.h>
  31 #include <linux/kthread.h>
  32 #include <linux/slab.h>
  33
  34 #define CREATE_TRACE_POINTS
  35 #include <trace/events/cpufreq_interactive.h>
  36
  37 struct cpufreq_interactive_cpuinfo {
  38         struct timer_list cpu_timer;
  39         struct timer_list cpu_slack_timer;
  40         spinlock_t load_lock; /* protects the next 4 fields */
  41         u64 time_in_idle;
  42         u64 time_in_idle_timestamp;
  43         u64 cputime_speedadj;
  44         u64 cputime_speedadj_timestamp;
  45         struct cpufreq_policy *policy;
  46         struct cpufreq_frequency_table *freq_table;
  47         spinlock_t target_freq_lock; /*protects target freq */
  48         unsigned int target_freq;
  49         unsigned int floor_freq;
  50         u64 pol_floor_val_time; /* policy floor_validate_time */
  51         u64 loc_floor_val_time; /* per-cpu floor_validate_time */
  52         u64 pol_hispeed_val_time; /* policy hispeed_validate_time */
  53         u64 loc_hispeed_val_time; /* per-cpu hispeed_validate_time */
  54         struct rw_semaphore enable_sem;
  55         int governor_enabled;
  56 };
  57
  58 static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
  59
  60 /* realtime thread handles frequency scaling */
  61 static struct task_struct *speedchange_task;
  62 static cpumask_t speedchange_cpumask;
  63 static spinlock_t speedchange_cpumask_lock;
  64 static struct mutex gov_lock;
  65
  66 /* Target load.  Lower values result in higher CPU speeds. */
  67 #define DEFAULT_TARGET_LOAD 90
  68 static unsigned int default_target_loads[] = {DEFAULT_TARGET_LOAD};
  69
  70 #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC)
  71 #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE
  72 static unsigned int default_above_hispeed_delay[] = {
  73         DEFAULT_ABOVE_HISPEED_DELAY };
  74
  75 struct cpufreq_interactive_tunables {
  76         int usage_count;
  77         /* Hi speed to bump to from lo speed when load burst (default max) */
  78         unsigned int hispeed_freq;
  79         /* Go to hi speed when CPU load at or above this value. */
  80 #define DEFAULT_GO_HISPEED_LOAD 99
  81         unsigned long go_hispeed_load;
  82         /* Target load. Lower values result in higher CPU speeds. */
  83         spinlock_t target_loads_lock;
  84         unsigned int *target_loads;
  85         int ntarget_loads;
  86         /*
  87          * The minimum amount of time to spend at a frequency before we can ramp
  88          * down.
  89          */
  90 #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC)
  91         unsigned long min_sample_time;
  92         /*
  93          * The sample rate of the timer used to increase frequency
  94          */
  95         unsigned long timer_rate;
  96         /*
  97          * Wait this long before raising speed above hispeed, by default a
  98          * single timer interval.
  99          */
 100         spinlock_t above_hispeed_delay_lock;
 101         unsigned int *above_hispeed_delay;
 102         int nabove_hispeed_delay;
 103         /* Non-zero means indefinite speed boost active */
 104         int boost_val;
 105         /* Duration of a boot pulse in usecs */
 106         int boostpulse_duration_val;
 107         /* End time of boost pulse in ktime converted to usecs */
 108         u64 boostpulse_endtime;
 109         bool boosted;
 110         /*
 111          * Max additional time to wait in idle, beyond timer_rate, at speeds
 112          * above minimum before wakeup to reduce speed, or -1 if unnecessary.
 113          */
 114 #define DEFAULT_TIMER_SLACK (4 * DEFAULT_TIMER_RATE)
 115         int timer_slack_val;
 116         bool io_is_busy;
 117 };
 118
 119 /*
 120  * HACK: FIXME: Bring back cpufreq_{get,put}_global_kobject()
 121  * definition removed by upstream commit 8eec1020f0c0 "cpufreq:
 122  * create cpu/cpufreq at boot time" to fix build failures.
 123  */
 124 static int cpufreq_global_kobject_usage;
 125
 126 int cpufreq_get_global_kobject(void)
 127 {
 128         if (!cpufreq_global_kobject_usage++)
 129                 return kobject_add(cpufreq_global_kobject,
 130                                 &cpu_subsys.dev_root->kobj, "%s", "cpufreq");
 131
 132         return 0;
 133 }
 134
 135 void cpufreq_put_global_kobject(void)
 136 {
 137         if (!--cpufreq_global_kobject_usage)
 138                 kobject_del(cpufreq_global_kobject);
 139 }
 140
 141 /* For cases where we have single governor instance for system */
 142 static struct cpufreq_interactive_tunables *common_tunables;
 143
 144 static struct attribute_group *get_sysfs_attr(void);
 145
 146 static void cpufreq_interactive_timer_resched(
 147         struct cpufreq_interactive_cpuinfo *pcpu)
 148 {
 149         struct cpufreq_interactive_tunables *tunables =
 150                 pcpu->policy->governor_data;
 151         unsigned long expires;
 152         unsigned long flags;
 153
 154         spin_lock_irqsave(&pcpu->load_lock, flags);
 155         pcpu->time_in_idle =
 156                 get_cpu_idle_time(smp_processor_id(),
 157                                   &pcpu->time_in_idle_timestamp,
 158                                   tunables->io_is_busy);
 159         pcpu->cputime_speedadj = 0;
 160         pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
 161         expires = jiffies + usecs_to_jiffies(tunables->timer_rate);
 162         mod_timer_pinned(&pcpu->cpu_timer, expires);
 163
 164         if (tunables->timer_slack_val >= 0 &&
 165             pcpu->target_freq > pcpu->policy->min) {
 166                 expires += usecs_to_jiffies(tunables->timer_slack_val);
 167                 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
 168         }
 169
 170         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 171 }
 172
 173 /* The caller shall take enable_sem write semaphore to avoid any timer race.
 174  * The cpu_timer and cpu_slack_timer must be deactivated when calling this
 175  * function.
 176  */
 177 static void cpufreq_interactive_timer_start(
 178         struct cpufreq_interactive_tunables *tunables, int cpu)
 179 {
 180         struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
 181         unsigned long expires = jiffies +
 182                 usecs_to_jiffies(tunables->timer_rate);
 183         unsigned long flags;
 184
 185         pcpu->cpu_timer.expires = expires;
 186         add_timer_on(&pcpu->cpu_timer, cpu);
 187         if (tunables->timer_slack_val >= 0 &&
 188             pcpu->target_freq > pcpu->policy->min) {
 189                 expires += usecs_to_jiffies(tunables->timer_slack_val);
 190                 pcpu->cpu_slack_timer.expires = expires;
 191                 add_timer_on(&pcpu->cpu_slack_timer, cpu);
 192         }
 193
 194         spin_lock_irqsave(&pcpu->load_lock, flags);
 195         pcpu->time_in_idle =
 196                 get_cpu_idle_time(cpu, &pcpu->time_in_idle_timestamp,
 197                                   tunables->io_is_busy);
 198         pcpu->cputime_speedadj = 0;
 199         pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
 200         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 201 }
 202
 203 static unsigned int freq_to_above_hispeed_delay(
 204         struct cpufreq_interactive_tunables *tunables,
 205         unsigned int freq)
 206 {
 207         int i;
 208         unsigned int ret;
 209         unsigned long flags;
 210
 211         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 212
 213         for (i = 0; i < tunables->nabove_hispeed_delay - 1 &&
 214                         freq >= tunables->above_hispeed_delay[i+1]; i += 2)
 215                 ;
 216
 217         ret = tunables->above_hispeed_delay[i];
 218         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 219         return ret;
 220 }
 221
 222 static unsigned int freq_to_targetload(
 223         struct cpufreq_interactive_tunables *tunables, unsigned int freq)
 224 {
 225         int i;
 226         unsigned int ret;
 227         unsigned long flags;
 228
 229         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 230
 231         for (i = 0; i < tunables->ntarget_loads - 1 &&
 232                     freq >= tunables->target_loads[i+1]; i += 2)
 233                 ;
 234
 235         ret = tunables->target_loads[i];
 236         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 237         return ret;
 238 }
 239
 240 /*
 241  * If increasing frequencies never map to a lower target load then
 242  * choose_freq() will find the minimum frequency that does not exceed its
 243  * target load given the current load.
 244  */
 245 static unsigned int choose_freq(struct cpufreq_interactive_cpuinfo *pcpu,
 246                 unsigned int loadadjfreq)
 247 {
 248         unsigned int freq = pcpu->policy->cur;
 249         unsigned int prevfreq, freqmin, freqmax;
 250         unsigned int tl;
 251         int index;
 252
 253         freqmin = 0;
 254         freqmax = UINT_MAX;
 255
 256         do {
 257                 prevfreq = freq;
 258                 tl = freq_to_targetload(pcpu->policy->governor_data, freq);
 259
 260                 /*
 261                  * Find the lowest frequency where the computed load is less
 262                  * than or equal to the target load.
 263                  */
 264
 265                 if (cpufreq_frequency_table_target(
 266                             pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
 267                             CPUFREQ_RELATION_L, &index))
 268                         break;
 269                 freq = pcpu->freq_table[index].frequency;
 270
 271                 if (freq > prevfreq) {
 272                         /* The previous frequency is too low. */
 273                         freqmin = prevfreq;
 274
 275                         if (freq >= freqmax) {
 276                                 /*
 277                                  * Find the highest frequency that is less
 278                                  * than freqmax.
 279                                  */
 280                                 if (cpufreq_frequency_table_target(
 281                                             pcpu->policy, pcpu->freq_table,
 282                                             freqmax - 1, CPUFREQ_RELATION_H,
 283                                             &index))
 284                                         break;
 285                                 freq = pcpu->freq_table[index].frequency;
 286
 287                                 if (freq == freqmin) {
 288                                         /*
 289                                          * The first frequency below freqmax
 290                                          * has already been found to be too
 291                                          * low.  freqmax is the lowest speed
 292                                          * we found that is fast enough.
 293                                          */
 294                                         freq = freqmax;
 295                                         break;
 296                                 }
 297                         }
 298                 } else if (freq < prevfreq) {
 299                         /* The previous frequency is high enough. */
 300                         freqmax = prevfreq;
 301
 302                         if (freq <= freqmin) {
 303                                 /*
 304                                  * Find the lowest frequency that is higher
 305                                  * than freqmin.
 306                                  */
 307                                 if (cpufreq_frequency_table_target(
 308                                             pcpu->policy, pcpu->freq_table,
 309                                             freqmin + 1, CPUFREQ_RELATION_L,
 310                                             &index))
 311                                         break;
 312                                 freq = pcpu->freq_table[index].frequency;
 313
 314                                 /*
 315                                  * If freqmax is the first frequency above
 316                                  * freqmin then we have already found that
 317                                  * this speed is fast enough.
 318                                  */
 319                                 if (freq == freqmax)
 320                                         break;
 321                         }
 322                 }
 323
 324                 /* If same frequency chosen as previous then done. */
 325         } while (freq != prevfreq);
 326
 327         return freq;
 328 }
 329
 330 static u64 update_load(int cpu)
 331 {
 332         struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
 333         struct cpufreq_interactive_tunables *tunables =
 334                 pcpu->policy->governor_data;
 335         u64 now;
 336         u64 now_idle;
 337         unsigned int delta_idle;
 338         unsigned int delta_time;
 339         u64 active_time;
 340
 341         now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
 342         delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
 343         delta_time = (unsigned int)(now - pcpu->time_in_idle_timestamp);
 344
 345         if (delta_time <= delta_idle)
 346                 active_time = 0;
 347         else
 348                 active_time = delta_time - delta_idle;
 349
 350         pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
 351
 352         pcpu->time_in_idle = now_idle;
 353         pcpu->time_in_idle_timestamp = now;
 354         return now;
 355 }
 356
 357 static void cpufreq_interactive_timer(unsigned long data)
 358 {
 359         u64 now;
 360         unsigned int delta_time;
 361         u64 cputime_speedadj;
 362         int cpu_load;
 363         struct cpufreq_interactive_cpuinfo *pcpu =
 364                 &per_cpu(cpuinfo, data);
 365         struct cpufreq_interactive_tunables *tunables =
 366                 pcpu->policy->governor_data;
 367         unsigned int new_freq;
 368         unsigned int loadadjfreq;
 369         unsigned int index;
 370         unsigned long flags;
 371         u64 max_fvtime;
 372
 373         if (!down_read_trylock(&pcpu->enable_sem))
 374                 return;
 375         if (!pcpu->governor_enabled)
 376                 goto exit;
 377
 378         spin_lock_irqsave(&pcpu->load_lock, flags);
 379         now = update_load(data);
 380         delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
 381         cputime_speedadj = pcpu->cputime_speedadj;
 382         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 383
 384         if (WARN_ON_ONCE(!delta_time))
 385                 goto rearm;
 386
 387         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
 388         do_div(cputime_speedadj, delta_time);
 389         loadadjfreq = (unsigned int)cputime_speedadj * 100;
 390         cpu_load = loadadjfreq / pcpu->policy->cur;
 391         tunables->boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
 392
 393         if (cpu_load >= tunables->go_hispeed_load || tunables->boosted) {
 394                 if (pcpu->policy->cur < tunables->hispeed_freq) {
 395                         new_freq = tunables->hispeed_freq;
 396                 } else {
 397                         new_freq = choose_freq(pcpu, loadadjfreq);
 398
 399                         if (new_freq < tunables->hispeed_freq)
 400                                 new_freq = tunables->hispeed_freq;
 401                 }
 402         } else {
 403                 new_freq = choose_freq(pcpu, loadadjfreq);
 404                 if (new_freq > tunables->hispeed_freq &&
 405                                 pcpu->policy->cur < tunables->hispeed_freq)
 406                         new_freq = tunables->hispeed_freq;
 407         }
 408
 409         if (pcpu->policy->cur >= tunables->hispeed_freq &&
 410             new_freq > pcpu->policy->cur &&
 411             now - pcpu->pol_hispeed_val_time <
 412             freq_to_above_hispeed_delay(tunables, pcpu->policy->cur)) {
 413                 trace_cpufreq_interactive_notyet(
 414                         data, cpu_load, pcpu->target_freq,
 415                         pcpu->policy->cur, new_freq);
 416                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 417                 goto rearm;
 418         }
 419
 420         pcpu->loc_hispeed_val_time = now;
 421
 422         if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
 423                                            new_freq, CPUFREQ_RELATION_L,
 424                                            &index)) {
 425                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 426                 goto rearm;
 427         }
 428
 429         new_freq = pcpu->freq_table[index].frequency;
 430
 431         /*
 432          * Do not scale below floor_freq unless we have been at or above the
 433          * floor frequency for the minimum sample time since last validated.
 434          */
 435         max_fvtime = max(pcpu->pol_floor_val_time, pcpu->loc_floor_val_time);
 436         if (new_freq < pcpu->floor_freq &&
 437             pcpu->target_freq >= pcpu->policy->cur) {
 438                 if (now - max_fvtime < tunables->min_sample_time) {
 439                         trace_cpufreq_interactive_notyet(
 440                                 data, cpu_load, pcpu->target_freq,
 441                                 pcpu->policy->cur, new_freq);
 442                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 443                         goto rearm;
 444                 }
 445         }
 446
 447         /*
 448          * Update the timestamp for checking whether speed has been held at
 449          * or above the selected frequency for a minimum of min_sample_time,
 450          * if not boosted to hispeed_freq.  If boosted to hispeed_freq then we
 451          * allow the speed to drop as soon as the boostpulse duration expires
 452          * (or the indefinite boost is turned off).
 453          */
 454
 455         if (!tunables->boosted || new_freq > tunables->hispeed_freq) {
 456                 pcpu->floor_freq = new_freq;
 457                 if (pcpu->target_freq >= pcpu->policy->cur ||
 458                     new_freq >= pcpu->policy->cur)
 459                         pcpu->loc_floor_val_time = now;
 460         }
 461
 462         if (pcpu->target_freq == new_freq &&
 463                         pcpu->target_freq <= pcpu->policy->cur) {
 464                 trace_cpufreq_interactive_already(
 465                         data, cpu_load, pcpu->target_freq,
 466                         pcpu->policy->cur, new_freq);
 467                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 468                 goto rearm;
 469         }
 470
 471         trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
 472                                          pcpu->policy->cur, new_freq);
 473
 474         pcpu->target_freq = new_freq;
 475         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 476         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 477         cpumask_set_cpu(data, &speedchange_cpumask);
 478         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 479         wake_up_process(speedchange_task);
 480
 481 rearm:
 482         if (!timer_pending(&pcpu->cpu_timer))
 483                 cpufreq_interactive_timer_resched(pcpu);
 484
 485 exit:
 486         up_read(&pcpu->enable_sem);
 487         return;
 488 }
 489
 490 static void cpufreq_interactive_idle_end(void)
 491 {
 492         struct cpufreq_interactive_cpuinfo *pcpu =
 493                 &per_cpu(cpuinfo, smp_processor_id());
 494
 495         if (!down_read_trylock(&pcpu->enable_sem))
 496                 return;
 497         if (!pcpu->governor_enabled) {
 498                 up_read(&pcpu->enable_sem);
 499                 return;
 500         }
 501
 502         /* Arm the timer for 1-2 ticks later if not already. */
 503         if (!timer_pending(&pcpu->cpu_timer)) {
 504                 cpufreq_interactive_timer_resched(pcpu);
 505         } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
 506                 del_timer(&pcpu->cpu_timer);
 507                 del_timer(&pcpu->cpu_slack_timer);
 508                 cpufreq_interactive_timer(smp_processor_id());
 509         }
 510
 511         up_read(&pcpu->enable_sem);
 512 }
 513
 514 static void cpufreq_interactive_get_policy_info(struct cpufreq_policy *policy,
 515                                                 unsigned int *pmax_freq,
 516                                                 u64 *phvt, u64 *pfvt)
 517 {
 518         struct cpufreq_interactive_cpuinfo *pcpu;
 519         unsigned int max_freq = 0;
 520         u64 hvt = ~0ULL, fvt = 0;
 521         unsigned int i;
 522
 523         for_each_cpu(i, policy->cpus) {
 524                 pcpu = &per_cpu(cpuinfo, i);
 525
 526                 fvt = max(fvt, pcpu->loc_floor_val_time);
 527                 if (pcpu->target_freq > max_freq) {
 528                         max_freq = pcpu->target_freq;
 529                         hvt = pcpu->loc_hispeed_val_time;
 530                 } else if (pcpu->target_freq == max_freq) {
 531                         hvt = min(hvt, pcpu->loc_hispeed_val_time);
 532                 }
 533         }
 534
 535         *pmax_freq = max_freq;
 536         *phvt = hvt;
 537         *pfvt = fvt;
 538 }
 539
 540 static void cpufreq_interactive_adjust_cpu(unsigned int cpu,
 541                                            struct cpufreq_policy *policy)
 542 {
 543         struct cpufreq_interactive_cpuinfo *pcpu;
 544         u64 hvt, fvt;
 545         unsigned int max_freq;
 546         int i;
 547
 548         cpufreq_interactive_get_policy_info(policy, &max_freq, &hvt, &fvt);
 549
 550         for_each_cpu(i, policy->cpus) {
 551                 pcpu = &per_cpu(cpuinfo, i);
 552                 pcpu->pol_floor_val_time = fvt;
 553         }
 554
 555         if (max_freq != policy->cur) {
 556                 __cpufreq_driver_target(policy, max_freq, CPUFREQ_RELATION_H);
 557                 for_each_cpu(i, policy->cpus) {
 558                         pcpu = &per_cpu(cpuinfo, i);
 559                         pcpu->pol_hispeed_val_time = hvt;
 560                 }
 561         }
 562
 563         trace_cpufreq_interactive_setspeed(cpu, max_freq, policy->cur);
 564 }
 565
 566 static int cpufreq_interactive_speedchange_task(void *data)
 567 {
 568         unsigned int cpu;
 569         cpumask_t tmp_mask;
 570         unsigned long flags;
 571         struct cpufreq_interactive_cpuinfo *pcpu;
 572
 573         while (1) {
 574                 set_current_state(TASK_INTERRUPTIBLE);
 575                 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 576
 577                 if (cpumask_empty(&speedchange_cpumask)) {
 578                         spin_unlock_irqrestore(&speedchange_cpumask_lock,
 579                                                flags);
 580                         schedule();
 581
 582                         if (kthread_should_stop())
 583                                 break;
 584
 585                         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 586                 }
 587
 588                 set_current_state(TASK_RUNNING);
 589                 tmp_mask = speedchange_cpumask;
 590                 cpumask_clear(&speedchange_cpumask);
 591                 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 592
 593                 for_each_cpu(cpu, &tmp_mask) {
 594                         pcpu = &per_cpu(cpuinfo, cpu);
 595
 596                         down_write(&pcpu->policy->rwsem);
 597
 598                         if (likely(down_read_trylock(&pcpu->enable_sem))) {
 599                                 if (likely(pcpu->governor_enabled))
 600                                         cpufreq_interactive_adjust_cpu(cpu,
 601                                                         pcpu->policy);
 602                                 up_read(&pcpu->enable_sem);
 603                         }
 604
 605                         up_write(&pcpu->policy->rwsem);
 606                 }
 607         }
 608
 609         return 0;
 610 }
 611
 612 static void cpufreq_interactive_boost(struct cpufreq_interactive_tunables *tunables)
 613 {
 614         int i;
 615         int anyboost = 0;
 616         unsigned long flags[2];
 617         struct cpufreq_interactive_cpuinfo *pcpu;
 618
 619         tunables->boosted = true;
 620
 621         spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
 622
 623         for_each_online_cpu(i) {
 624                 pcpu = &per_cpu(cpuinfo, i);
 625
 626                 if (!down_read_trylock(&pcpu->enable_sem))
 627                         continue;
 628
 629                 if (!pcpu->governor_enabled) {
 630                         up_read(&pcpu->enable_sem);
 631                         continue;
 632                 }
 633
 634                 if (tunables != pcpu->policy->governor_data) {
 635                         up_read(&pcpu->enable_sem);
 636                         continue;
 637                 }
 638
 639                 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
 640                 if (pcpu->target_freq < tunables->hispeed_freq) {
 641                         pcpu->target_freq = tunables->hispeed_freq;
 642                         cpumask_set_cpu(i, &speedchange_cpumask);
 643                         pcpu->pol_hispeed_val_time =
 644                                 ktime_to_us(ktime_get());
 645                         anyboost = 1;
 646                 }
 647                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
 648
 649                 up_read(&pcpu->enable_sem);
 650         }
 651
 652         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
 653
 654         if (anyboost)
 655                 wake_up_process(speedchange_task);
 656 }
 657
 658 static int cpufreq_interactive_notifier(
 659         struct notifier_block *nb, unsigned long val, void *data)
 660 {
 661         struct cpufreq_freqs *freq = data;
 662         struct cpufreq_interactive_cpuinfo *pcpu;
 663         int cpu;
 664         unsigned long flags;
 665
 666         if (val == CPUFREQ_POSTCHANGE) {
 667                 pcpu = &per_cpu(cpuinfo, freq->cpu);
 668                 if (!down_read_trylock(&pcpu->enable_sem))
 669                         return 0;
 670                 if (!pcpu->governor_enabled) {
 671                         up_read(&pcpu->enable_sem);
 672                         return 0;
 673                 }
 674
 675                 for_each_cpu(cpu, pcpu->policy->cpus) {
 676                         struct cpufreq_interactive_cpuinfo *pjcpu =
 677                                 &per_cpu(cpuinfo, cpu);
 678                         if (cpu != freq->cpu) {
 679                                 if (!down_read_trylock(&pjcpu->enable_sem))
 680                                         continue;
 681                                 if (!pjcpu->governor_enabled) {
 682                                         up_read(&pjcpu->enable_sem);
 683                                         continue;
 684                                 }
 685                         }
 686                         spin_lock_irqsave(&pjcpu->load_lock, flags);
 687                         update_load(cpu);
 688                         spin_unlock_irqrestore(&pjcpu->load_lock, flags);
 689                         if (cpu != freq->cpu)
 690                                 up_read(&pjcpu->enable_sem);
 691                 }
 692
 693                 up_read(&pcpu->enable_sem);
 694         }
 695         return 0;
 696 }
 697
 698 static struct notifier_block cpufreq_notifier_block = {
 699         .notifier_call = cpufreq_interactive_notifier,
 700 };
 701
 702 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
 703 {
 704         const char *cp;
 705         int i;
 706         int ntokens = 1;
 707         unsigned int *tokenized_data;
 708         int err = -EINVAL;
 709
 710         cp = buf;
 711         while ((cp = strpbrk(cp + 1, " :")))
 712                 ntokens++;
 713
 714         if (!(ntokens & 0x1))
 715                 goto err;
 716
 717         tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
 718         if (!tokenized_data) {
 719                 err = -ENOMEM;
 720                 goto err;
 721         }
 722
 723         cp = buf;
 724         i = 0;
 725         while (i < ntokens) {
 726                 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
 727                         goto err_kfree;
 728
 729                 cp = strpbrk(cp, " :");
 730                 if (!cp)
 731                         break;
 732                 cp++;
 733         }
 734
 735         if (i != ntokens)
 736                 goto err_kfree;
 737
 738         *num_tokens = ntokens;
 739         return tokenized_data;
 740
 741 err_kfree:
 742         kfree(tokenized_data);
 743 err:
 744         return ERR_PTR(err);
 745 }
 746
 747 static ssize_t show_target_loads(
 748         struct cpufreq_interactive_tunables *tunables,
 749         char *buf)
 750 {
 751         int i;
 752         ssize_t ret = 0;
 753         unsigned long flags;
 754
 755         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 756
 757         for (i = 0; i < tunables->ntarget_loads; i++)
 758                 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
 759                                i & 0x1 ? ":" : " ");
 760
 761         sprintf(buf + ret - 1, "\n");
 762         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 763         return ret;
 764 }
 765
 766 static ssize_t store_target_loads(
 767         struct cpufreq_interactive_tunables *tunables,
 768         const char *buf, size_t count)
 769 {
 770         int ntokens;
 771         unsigned int *new_target_loads = NULL;
 772         unsigned long flags;
 773
 774         new_target_loads = get_tokenized_data(buf, &ntokens);
 775         if (IS_ERR(new_target_loads))
 776                 return PTR_RET(new_target_loads);
 777
 778         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 779         if (tunables->target_loads != default_target_loads)
 780                 kfree(tunables->target_loads);
 781         tunables->target_loads = new_target_loads;
 782         tunables->ntarget_loads = ntokens;
 783         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 784         return count;
 785 }
 786
 787 static ssize_t show_above_hispeed_delay(
 788         struct cpufreq_interactive_tunables *tunables, char *buf)
 789 {
 790         int i;
 791         ssize_t ret = 0;
 792         unsigned long flags;
 793
 794         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 795
 796         for (i = 0; i < tunables->nabove_hispeed_delay; i++)
 797                 ret += sprintf(buf + ret, "%u%s",
 798                                tunables->above_hispeed_delay[i],
 799                                i & 0x1 ? ":" : " ");
 800
 801         sprintf(buf + ret - 1, "\n");
 802         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 803         return ret;
 804 }
 805
 806 static ssize_t store_above_hispeed_delay(
 807         struct cpufreq_interactive_tunables *tunables,
 808         const char *buf, size_t count)
 809 {
 810         int ntokens;
 811         unsigned int *new_above_hispeed_delay = NULL;
 812         unsigned long flags;
 813
 814         new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
 815         if (IS_ERR(new_above_hispeed_delay))
 816                 return PTR_RET(new_above_hispeed_delay);
 817
 818         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 819         if (tunables->above_hispeed_delay != default_above_hispeed_delay)
 820                 kfree(tunables->above_hispeed_delay);
 821         tunables->above_hispeed_delay = new_above_hispeed_delay;
 822         tunables->nabove_hispeed_delay = ntokens;
 823         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 824         return count;
 825
 826 }
 827
 828 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 829                 char *buf)
 830 {
 831         return sprintf(buf, "%u\n", tunables->hispeed_freq);
 832 }
 833
 834 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 835                 const char *buf, size_t count)
 836 {
 837         int ret;
 838         long unsigned int val;
 839
 840         ret = kstrtoul(buf, 0, &val);
 841         if (ret < 0)
 842                 return ret;
 843         tunables->hispeed_freq = val;
 844         return count;
 845 }
 846
 847 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
 848                 *tunables, char *buf)
 849 {
 850         return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
 851 }
 852
 853 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
 854                 *tunables, const char *buf, size_t count)
 855 {
 856         int ret;
 857         unsigned long val;
 858
 859         ret = kstrtoul(buf, 0, &val);
 860         if (ret < 0)
 861                 return ret;
 862         tunables->go_hispeed_load = val;
 863         return count;
 864 }
 865
 866 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
 867                 *tunables, char *buf)
 868 {
 869         return sprintf(buf, "%lu\n", tunables->min_sample_time);
 870 }
 871
 872 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
 873                 *tunables, const char *buf, size_t count)
 874 {
 875         int ret;
 876         unsigned long val;
 877
 878         ret = kstrtoul(buf, 0, &val);
 879         if (ret < 0)
 880                 return ret;
 881         tunables->min_sample_time = val;
 882         return count;
 883 }
 884
 885 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
 886                 char *buf)
 887 {
 888         return sprintf(buf, "%lu\n", tunables->timer_rate);
 889 }
 890
 891 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
 892                 const char *buf, size_t count)
 893 {
 894         int ret;
 895         unsigned long val, val_round;
 896
 897         ret = kstrtoul(buf, 0, &val);
 898         if (ret < 0)
 899                 return ret;
 900
 901         val_round = jiffies_to_usecs(usecs_to_jiffies(val));
 902         if (val != val_round)
 903                 pr_warn("timer_rate not aligned to jiffy. Rounded up to %lu\n",
 904                         val_round);
 905
 906         tunables->timer_rate = val_round;
 907         return count;
 908 }
 909
 910 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
 911                 char *buf)
 912 {
 913         return sprintf(buf, "%d\n", tunables->timer_slack_val);
 914 }
 915
 916 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
 917                 const char *buf, size_t count)
 918 {
 919         int ret;
 920         unsigned long val;
 921
 922         ret = kstrtol(buf, 10, &val);
 923         if (ret < 0)
 924                 return ret;
 925
 926         tunables->timer_slack_val = val;
 927         return count;
 928 }
 929
 930 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
 931                           char *buf)
 932 {
 933         return sprintf(buf, "%d\n", tunables->boost_val);
 934 }
 935
 936 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
 937                            const char *buf, size_t count)
 938 {
 939         int ret;
 940         unsigned long val;
 941
 942         ret = kstrtoul(buf, 0, &val);
 943         if (ret < 0)
 944                 return ret;
 945
 946         tunables->boost_val = val;
 947
 948         if (tunables->boost_val) {
 949                 trace_cpufreq_interactive_boost("on");
 950                 if (!tunables->boosted)
 951                         cpufreq_interactive_boost(tunables);
 952         } else {
 953                 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
 954                 trace_cpufreq_interactive_unboost("off");
 955         }
 956
 957         return count;
 958 }
 959
 960 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
 961                                 const char *buf, size_t count)
 962 {
 963         int ret;
 964         unsigned long val;
 965
 966         ret = kstrtoul(buf, 0, &val);
 967         if (ret < 0)
 968                 return ret;
 969
 970         tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
 971                 tunables->boostpulse_duration_val;
 972         trace_cpufreq_interactive_boost("pulse");
 973         if (!tunables->boosted)
 974                 cpufreq_interactive_boost(tunables);
 975         return count;
 976 }
 977
 978 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
 979                 *tunables, char *buf)
 980 {
 981         return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
 982 }
 983
 984 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
 985                 *tunables, const char *buf, size_t count)
 986 {
 987         int ret;
 988         unsigned long val;
 989
 990         ret = kstrtoul(buf, 0, &val);
 991         if (ret < 0)
 992                 return ret;
 993
 994         tunables->boostpulse_duration_val = val;
 995         return count;
 996 }
 997
 998 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
 999                 char *buf)
1000 {
1001         return sprintf(buf, "%u\n", tunables->io_is_busy);
1002 }
1003
1004 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
1005                 const char *buf, size_t count)
1006 {
1007         int ret;
1008         unsigned long val;
1009
1010         ret = kstrtoul(buf, 0, &val);
1011         if (ret < 0)
1012                 return ret;
1013         tunables->io_is_busy = val;
1014         return count;
1015 }
1016
1017 /*
1018  * Create show/store routines
1019  * - sys: One governor instance for complete SYSTEM
1020  * - pol: One governor instance per struct cpufreq_policy
1021  */
1022 #define show_gov_pol_sys(file_name)                                     \
1023 static ssize_t show_##file_name##_gov_sys                               \
1024 (struct kobject *kobj, struct attribute *attr, char *buf)               \
1025 {                                                                       \
1026         return show_##file_name(common_tunables, buf);                  \
1027 }                                                                       \
1028                                                                         \
1029 static ssize_t show_##file_name##_gov_pol                               \
1030 (struct cpufreq_policy *policy, char *buf)                              \
1031 {                                                                       \
1032         return show_##file_name(policy->governor_data, buf);            \
1033 }
1034
1035 #define store_gov_pol_sys(file_name)                                    \
1036 static ssize_t store_##file_name##_gov_sys                              \
1037 (struct kobject *kobj, struct attribute *attr, const char *buf,         \
1038         size_t count)                                                   \
1039 {                                                                       \
1040         return store_##file_name(common_tunables, buf, count);          \
1041 }                                                                       \
1042                                                                         \
1043 static ssize_t store_##file_name##_gov_pol                              \
1044 (struct cpufreq_policy *policy, const char *buf, size_t count)          \
1045 {                                                                       \
1046         return store_##file_name(policy->governor_data, buf, count);    \
1047 }
1048
1049 #define show_store_gov_pol_sys(file_name)                               \
1050 show_gov_pol_sys(file_name);                                            \
1051 store_gov_pol_sys(file_name)
1052
1053 show_store_gov_pol_sys(target_loads);
1054 show_store_gov_pol_sys(above_hispeed_delay);
1055 show_store_gov_pol_sys(hispeed_freq);
1056 show_store_gov_pol_sys(go_hispeed_load);
1057 show_store_gov_pol_sys(min_sample_time);
1058 show_store_gov_pol_sys(timer_rate);
1059 show_store_gov_pol_sys(timer_slack);
1060 show_store_gov_pol_sys(boost);
1061 store_gov_pol_sys(boostpulse);
1062 show_store_gov_pol_sys(boostpulse_duration);
1063 show_store_gov_pol_sys(io_is_busy);
1064
1065 #define gov_sys_attr_rw(_name)                                          \
1066 static struct global_attr _name##_gov_sys =                             \
1067 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1068
1069 #define gov_pol_attr_rw(_name)                                          \
1070 static struct freq_attr _name##_gov_pol =                               \
1071 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1072
1073 #define gov_sys_pol_attr_rw(_name)                                      \
1074         gov_sys_attr_rw(_name);                                         \
1075         gov_pol_attr_rw(_name)
1076
1077 gov_sys_pol_attr_rw(target_loads);
1078 gov_sys_pol_attr_rw(above_hispeed_delay);
1079 gov_sys_pol_attr_rw(hispeed_freq);
1080 gov_sys_pol_attr_rw(go_hispeed_load);
1081 gov_sys_pol_attr_rw(min_sample_time);
1082 gov_sys_pol_attr_rw(timer_rate);
1083 gov_sys_pol_attr_rw(timer_slack);
1084 gov_sys_pol_attr_rw(boost);
1085 gov_sys_pol_attr_rw(boostpulse_duration);
1086 gov_sys_pol_attr_rw(io_is_busy);
1087
1088 static struct global_attr boostpulse_gov_sys =
1089         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1090
1091 static struct freq_attr boostpulse_gov_pol =
1092         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1093
1094 /* One Governor instance for entire system */
1095 static struct attribute *interactive_attributes_gov_sys[] = {
1096         &target_loads_gov_sys.attr,
1097         &above_hispeed_delay_gov_sys.attr,
1098         &hispeed_freq_gov_sys.attr,
1099         &go_hispeed_load_gov_sys.attr,
1100         &min_sample_time_gov_sys.attr,
1101         &timer_rate_gov_sys.attr,
1102         &timer_slack_gov_sys.attr,
1103         &boost_gov_sys.attr,
1104         &boostpulse_gov_sys.attr,
1105         &boostpulse_duration_gov_sys.attr,
1106         &io_is_busy_gov_sys.attr,
1107         NULL,
1108 };
1109
1110 static struct attribute_group interactive_attr_group_gov_sys = {
1111         .attrs = interactive_attributes_gov_sys,
1112         .name = "interactive",
1113 };
1114
1115 /* Per policy governor instance */
1116 static struct attribute *interactive_attributes_gov_pol[] = {
1117         &target_loads_gov_pol.attr,
1118         &above_hispeed_delay_gov_pol.attr,
1119         &hispeed_freq_gov_pol.attr,
1120         &go_hispeed_load_gov_pol.attr,
1121         &min_sample_time_gov_pol.attr,
1122         &timer_rate_gov_pol.attr,
1123         &timer_slack_gov_pol.attr,
1124         &boost_gov_pol.attr,
1125         &boostpulse_gov_pol.attr,
1126         &boostpulse_duration_gov_pol.attr,
1127         &io_is_busy_gov_pol.attr,
1128         NULL,
1129 };
1130
1131 static struct attribute_group interactive_attr_group_gov_pol = {
1132         .attrs = interactive_attributes_gov_pol,
1133         .name = "interactive",
1134 };
1135
1136 static struct attribute_group *get_sysfs_attr(void)
1137 {
1138         if (have_governor_per_policy())
1139                 return &interactive_attr_group_gov_pol;
1140         else
1141                 return &interactive_attr_group_gov_sys;
1142 }
1143
1144 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1145                                              unsigned long val,
1146                                              void *data)
1147 {
1148         if (val == IDLE_END)
1149                 cpufreq_interactive_idle_end();
1150
1151         return 0;
1152 }
1153
1154 static struct notifier_block cpufreq_interactive_idle_nb = {
1155         .notifier_call = cpufreq_interactive_idle_notifier,
1156 };
1157
1158 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1159                 unsigned int event)
1160 {
1161         int rc;
1162         unsigned int j;
1163         struct cpufreq_interactive_cpuinfo *pcpu;
1164         struct cpufreq_frequency_table *freq_table;
1165         struct cpufreq_interactive_tunables *tunables;
1166         unsigned long flags;
1167
1168         if (have_governor_per_policy())
1169                 tunables = policy->governor_data;
1170         else
1171                 tunables = common_tunables;
1172
1173         WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1174
1175         switch (event) {
1176         case CPUFREQ_GOV_POLICY_INIT:
1177                 if (have_governor_per_policy()) {
1178                         WARN_ON(tunables);
1179                 } else if (tunables) {
1180                         tunables->usage_count++;
1181                         policy->governor_data = tunables;
1182                         return 0;
1183                 }
1184
1185                 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1186                 if (!tunables) {
1187                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1188                         return -ENOMEM;
1189                 }
1190
1191                 tunables->usage_count = 1;
1192                 tunables->above_hispeed_delay = default_above_hispeed_delay;
1193                 tunables->nabove_hispeed_delay =
1194                         ARRAY_SIZE(default_above_hispeed_delay);
1195                 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1196                 tunables->target_loads = default_target_loads;
1197                 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1198                 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1199                 tunables->timer_rate = DEFAULT_TIMER_RATE;
1200                 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1201                 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1202
1203                 spin_lock_init(&tunables->target_loads_lock);
1204                 spin_lock_init(&tunables->above_hispeed_delay_lock);
1205
1206                 policy->governor_data = tunables;
1207                 if (!have_governor_per_policy()) {
1208                         common_tunables = tunables;
1209                         WARN_ON(cpufreq_get_global_kobject());
1210                 }
1211
1212                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1213                                 get_sysfs_attr());
1214                 if (rc) {
1215                         kfree(tunables);
1216                         policy->governor_data = NULL;
1217                         if (!have_governor_per_policy()) {
1218                                 common_tunables = NULL;
1219                                 cpufreq_put_global_kobject();
1220                         }
1221                         return rc;
1222                 }
1223
1224                 if (!policy->governor->initialized) {
1225                         idle_notifier_register(&cpufreq_interactive_idle_nb);
1226                         cpufreq_register_notifier(&cpufreq_notifier_block,
1227                                         CPUFREQ_TRANSITION_NOTIFIER);
1228                 }
1229
1230                 break;
1231
1232         case CPUFREQ_GOV_POLICY_EXIT:
1233                 if (!--tunables->usage_count) {
1234                         if (policy->governor->initialized == 1) {
1235                                 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1236                                                 CPUFREQ_TRANSITION_NOTIFIER);
1237                                 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1238                         }
1239
1240                         sysfs_remove_group(get_governor_parent_kobj(policy),
1241                                         get_sysfs_attr());
1242
1243                         if (!have_governor_per_policy())
1244                                 cpufreq_put_global_kobject();
1245
1246                         kfree(tunables);
1247                         common_tunables = NULL;
1248                 }
1249
1250                 policy->governor_data = NULL;
1251                 break;
1252
1253         case CPUFREQ_GOV_START:
1254                 mutex_lock(&gov_lock);
1255
1256                 freq_table = cpufreq_frequency_get_table(policy->cpu);
1257                 if (!tunables->hispeed_freq)
1258                         tunables->hispeed_freq = policy->max;
1259
1260                 for_each_cpu(j, policy->cpus) {
1261                         pcpu = &per_cpu(cpuinfo, j);
1262                         pcpu->policy = policy;
1263                         pcpu->target_freq = policy->cur;
1264                         pcpu->freq_table = freq_table;
1265                         pcpu->floor_freq = pcpu->target_freq;
1266                         pcpu->pol_floor_val_time =
1267                                 ktime_to_us(ktime_get());
1268                         pcpu->loc_floor_val_time = pcpu->pol_floor_val_time;
1269                         pcpu->pol_hispeed_val_time = pcpu->pol_floor_val_time;
1270                         pcpu->loc_hispeed_val_time = pcpu->pol_floor_val_time;
1271                         down_write(&pcpu->enable_sem);
1272                         del_timer_sync(&pcpu->cpu_timer);
1273                         del_timer_sync(&pcpu->cpu_slack_timer);
1274                         cpufreq_interactive_timer_start(tunables, j);
1275                         pcpu->governor_enabled = 1;
1276                         up_write(&pcpu->enable_sem);
1277                 }
1278
1279                 mutex_unlock(&gov_lock);
1280                 break;
1281
1282         case CPUFREQ_GOV_STOP:
1283                 mutex_lock(&gov_lock);
1284                 for_each_cpu(j, policy->cpus) {
1285                         pcpu = &per_cpu(cpuinfo, j);
1286                         down_write(&pcpu->enable_sem);
1287                         pcpu->governor_enabled = 0;
1288                         del_timer_sync(&pcpu->cpu_timer);
1289                         del_timer_sync(&pcpu->cpu_slack_timer);
1290                         up_write(&pcpu->enable_sem);
1291                 }
1292
1293                 mutex_unlock(&gov_lock);
1294                 break;
1295
1296         case CPUFREQ_GOV_LIMITS:
1297                 if (policy->max < policy->cur)
1298                         __cpufreq_driver_target(policy,
1299                                         policy->max, CPUFREQ_RELATION_H);
1300                 else if (policy->min > policy->cur)
1301                         __cpufreq_driver_target(policy,
1302                                         policy->min, CPUFREQ_RELATION_L);
1303                 for_each_cpu(j, policy->cpus) {
1304                         pcpu = &per_cpu(cpuinfo, j);
1305
1306                         down_read(&pcpu->enable_sem);
1307                         if (pcpu->governor_enabled == 0) {
1308                                 up_read(&pcpu->enable_sem);
1309                                 continue;
1310                         }
1311
1312                         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1313                         if (policy->max < pcpu->target_freq)
1314                                 pcpu->target_freq = policy->max;
1315                         else if (policy->min > pcpu->target_freq)
1316                                 pcpu->target_freq = policy->min;
1317
1318                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1319                         up_read(&pcpu->enable_sem);
1320                 }
1321                 break;
1322         }
1323         return 0;
1324 }
1325
1326 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1327 static
1328 #endif
1329 struct cpufreq_governor cpufreq_gov_interactive = {
1330         .name = "interactive",
1331         .governor = cpufreq_governor_interactive,
1332         .max_transition_latency = 10000000,
1333         .owner = THIS_MODULE,
1334 };
1335
1336 static void cpufreq_interactive_nop_timer(unsigned long data)
1337 {
1338 }
1339
1340 static int __init cpufreq_interactive_init(void)
1341 {
1342         unsigned int i;
1343         struct cpufreq_interactive_cpuinfo *pcpu;
1344         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1345
1346         /* Initalize per-cpu timers */
1347         for_each_possible_cpu(i) {
1348                 pcpu = &per_cpu(cpuinfo, i);
1349                 init_timer_deferrable(&pcpu->cpu_timer);
1350                 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1351                 pcpu->cpu_timer.data = i;
1352                 init_timer(&pcpu->cpu_slack_timer);
1353                 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1354                 spin_lock_init(&pcpu->load_lock);
1355                 spin_lock_init(&pcpu->target_freq_lock);
1356                 init_rwsem(&pcpu->enable_sem);
1357         }
1358
1359         spin_lock_init(&speedchange_cpumask_lock);
1360         mutex_init(&gov_lock);
1361         speedchange_task =
1362                 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1363                                "cfinteractive");
1364         if (IS_ERR(speedchange_task))
1365                 return PTR_ERR(speedchange_task);
1366
1367         sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, &param);
1368         get_task_struct(speedchange_task);
1369
1370         /* NB: wake up so the thread does not look hung to the freezer */
1371         wake_up_process(speedchange_task);
1372
1373         return cpufreq_register_governor(&cpufreq_gov_interactive);
1374 }
1375
1376 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1377 fs_initcall(cpufreq_interactive_init);
1378 #else
1379 module_init(cpufreq_interactive_init);
1380 #endif
1381
1382 static void __exit cpufreq_interactive_exit(void)
1383 {
1384         cpufreq_unregister_governor(&cpufreq_gov_interactive);
1385         kthread_stop(speedchange_task);
1386         put_task_struct(speedchange_task);
1387 }
1388
1389 module_exit(cpufreq_interactive_exit);
1390
1391 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1392 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1393         "Latency sensitive workloads");
1394 MODULE_LICENSE("GPL");