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