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
 613                 /*
 614                  * Set floor freq and (re)start timer for when last
 615                  * validated.
 616                  */
 617
 618                 pcpu->floor_freq = tunables->hispeed_freq;
 619                 pcpu->floor_validate_time = ktime_to_us(ktime_get());
 620                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
 621         }
 622
 623         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
 624
 625         if (anyboost)
 626                 wake_up_process(speedchange_task);
 627 }
 628
 629 static int cpufreq_interactive_notifier(
 630         struct notifier_block *nb, unsigned long val, void *data)
 631 {
 632         struct cpufreq_freqs *freq = data;
 633         struct cpufreq_interactive_cpuinfo *pcpu;
 634         int cpu;
 635         unsigned long flags;
 636
 637         if (val == CPUFREQ_POSTCHANGE) {
 638                 pcpu = &per_cpu(cpuinfo, freq->cpu);
 639                 if (!down_read_trylock(&pcpu->enable_sem))
 640                         return 0;
 641                 if (!pcpu->governor_enabled) {
 642                         up_read(&pcpu->enable_sem);
 643                         return 0;
 644                 }
 645
 646                 for_each_cpu(cpu, pcpu->policy->cpus) {
 647                         struct cpufreq_interactive_cpuinfo *pjcpu =
 648                                 &per_cpu(cpuinfo, cpu);
 649                         if (cpu != freq->cpu) {
 650                                 if (!down_read_trylock(&pjcpu->enable_sem))
 651                                         continue;
 652                                 if (!pjcpu->governor_enabled) {
 653                                         up_read(&pjcpu->enable_sem);
 654                                         continue;
 655                                 }
 656                         }
 657                         spin_lock_irqsave(&pjcpu->load_lock, flags);
 658                         update_load(cpu);
 659                         spin_unlock_irqrestore(&pjcpu->load_lock, flags);
 660                         if (cpu != freq->cpu)
 661                                 up_read(&pjcpu->enable_sem);
 662                 }
 663
 664                 up_read(&pcpu->enable_sem);
 665         }
 666         return 0;
 667 }
 668
 669 static struct notifier_block cpufreq_notifier_block = {
 670         .notifier_call = cpufreq_interactive_notifier,
 671 };
 672
 673 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
 674 {
 675         const char *cp;
 676         int i;
 677         int ntokens = 1;
 678         unsigned int *tokenized_data;
 679         int err = -EINVAL;
 680
 681         cp = buf;
 682         while ((cp = strpbrk(cp + 1, " :")))
 683                 ntokens++;
 684
 685         if (!(ntokens & 0x1))
 686                 goto err;
 687
 688         tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
 689         if (!tokenized_data) {
 690                 err = -ENOMEM;
 691                 goto err;
 692         }
 693
 694         cp = buf;
 695         i = 0;
 696         while (i < ntokens) {
 697                 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
 698                         goto err_kfree;
 699
 700                 cp = strpbrk(cp, " :");
 701                 if (!cp)
 702                         break;
 703                 cp++;
 704         }
 705
 706         if (i != ntokens)
 707                 goto err_kfree;
 708
 709         *num_tokens = ntokens;
 710         return tokenized_data;
 711
 712 err_kfree:
 713         kfree(tokenized_data);
 714 err:
 715         return ERR_PTR(err);
 716 }
 717
 718 static ssize_t show_target_loads(
 719         struct cpufreq_interactive_tunables *tunables,
 720         char *buf)
 721 {
 722         int i;
 723         ssize_t ret = 0;
 724         unsigned long flags;
 725
 726         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 727
 728         for (i = 0; i < tunables->ntarget_loads; i++)
 729                 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
 730                                i & 0x1 ? ":" : " ");
 731
 732         sprintf(buf + ret - 1, "\n");
 733         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 734         return ret;
 735 }
 736
 737 static ssize_t store_target_loads(
 738         struct cpufreq_interactive_tunables *tunables,
 739         const char *buf, size_t count)
 740 {
 741         int ntokens;
 742         unsigned int *new_target_loads = NULL;
 743         unsigned long flags;
 744
 745         new_target_loads = get_tokenized_data(buf, &ntokens);
 746         if (IS_ERR(new_target_loads))
 747                 return PTR_RET(new_target_loads);
 748
 749         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 750         if (tunables->target_loads != default_target_loads)
 751                 kfree(tunables->target_loads);
 752         tunables->target_loads = new_target_loads;
 753         tunables->ntarget_loads = ntokens;
 754         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 755         return count;
 756 }
 757
 758 static ssize_t show_above_hispeed_delay(
 759         struct cpufreq_interactive_tunables *tunables, char *buf)
 760 {
 761         int i;
 762         ssize_t ret = 0;
 763         unsigned long flags;
 764
 765         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 766
 767         for (i = 0; i < tunables->nabove_hispeed_delay; i++)
 768                 ret += sprintf(buf + ret, "%u%s",
 769                                tunables->above_hispeed_delay[i],
 770                                i & 0x1 ? ":" : " ");
 771
 772         sprintf(buf + ret - 1, "\n");
 773         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 774         return ret;
 775 }
 776
 777 static ssize_t store_above_hispeed_delay(
 778         struct cpufreq_interactive_tunables *tunables,
 779         const char *buf, size_t count)
 780 {
 781         int ntokens;
 782         unsigned int *new_above_hispeed_delay = NULL;
 783         unsigned long flags;
 784
 785         new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
 786         if (IS_ERR(new_above_hispeed_delay))
 787                 return PTR_RET(new_above_hispeed_delay);
 788
 789         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 790         if (tunables->above_hispeed_delay != default_above_hispeed_delay)
 791                 kfree(tunables->above_hispeed_delay);
 792         tunables->above_hispeed_delay = new_above_hispeed_delay;
 793         tunables->nabove_hispeed_delay = ntokens;
 794         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 795         return count;
 796
 797 }
 798
 799 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 800                 char *buf)
 801 {
 802         return sprintf(buf, "%u\n", tunables->hispeed_freq);
 803 }
 804
 805 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 806                 const char *buf, size_t count)
 807 {
 808         int ret;
 809         long unsigned int val;
 810
 811         ret = strict_strtoul(buf, 0, &val);
 812         if (ret < 0)
 813                 return ret;
 814         tunables->hispeed_freq = val;
 815         return count;
 816 }
 817
 818 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
 819                 *tunables, char *buf)
 820 {
 821         return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
 822 }
 823
 824 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
 825                 *tunables, const char *buf, size_t count)
 826 {
 827         int ret;
 828         unsigned long val;
 829
 830         ret = strict_strtoul(buf, 0, &val);
 831         if (ret < 0)
 832                 return ret;
 833         tunables->go_hispeed_load = val;
 834         return count;
 835 }
 836
 837 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
 838                 *tunables, char *buf)
 839 {
 840         return sprintf(buf, "%lu\n", tunables->min_sample_time);
 841 }
 842
 843 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
 844                 *tunables, const char *buf, size_t count)
 845 {
 846         int ret;
 847         unsigned long val;
 848
 849         ret = strict_strtoul(buf, 0, &val);
 850         if (ret < 0)
 851                 return ret;
 852         tunables->min_sample_time = val;
 853         return count;
 854 }
 855
 856 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
 857                 char *buf)
 858 {
 859         return sprintf(buf, "%lu\n", tunables->timer_rate);
 860 }
 861
 862 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
 863                 const char *buf, size_t count)
 864 {
 865         int ret;
 866         unsigned long val;
 867
 868         ret = strict_strtoul(buf, 0, &val);
 869         if (ret < 0)
 870                 return ret;
 871         tunables->timer_rate = val;
 872         return count;
 873 }
 874
 875 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
 876                 char *buf)
 877 {
 878         return sprintf(buf, "%d\n", tunables->timer_slack_val);
 879 }
 880
 881 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
 882                 const char *buf, size_t count)
 883 {
 884         int ret;
 885         unsigned long val;
 886
 887         ret = kstrtol(buf, 10, &val);
 888         if (ret < 0)
 889                 return ret;
 890
 891         tunables->timer_slack_val = val;
 892         return count;
 893 }
 894
 895 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
 896                           char *buf)
 897 {
 898         return sprintf(buf, "%d\n", tunables->boost_val);
 899 }
 900
 901 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
 902                            const char *buf, size_t count)
 903 {
 904         int ret;
 905         unsigned long val;
 906
 907         ret = kstrtoul(buf, 0, &val);
 908         if (ret < 0)
 909                 return ret;
 910
 911         tunables->boost_val = val;
 912
 913         if (tunables->boost_val) {
 914                 trace_cpufreq_interactive_boost("on");
 915                 if (!tunables->boosted)
 916                         cpufreq_interactive_boost(tunables);
 917         } else {
 918                 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
 919                 trace_cpufreq_interactive_unboost("off");
 920         }
 921
 922         return count;
 923 }
 924
 925 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
 926                                 const char *buf, size_t count)
 927 {
 928         int ret;
 929         unsigned long val;
 930
 931         ret = kstrtoul(buf, 0, &val);
 932         if (ret < 0)
 933                 return ret;
 934
 935         tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
 936                 tunables->boostpulse_duration_val;
 937         trace_cpufreq_interactive_boost("pulse");
 938         if (!tunables->boosted)
 939                 cpufreq_interactive_boost(tunables);
 940         return count;
 941 }
 942
 943 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
 944                 *tunables, char *buf)
 945 {
 946         return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
 947 }
 948
 949 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
 950                 *tunables, const char *buf, size_t count)
 951 {
 952         int ret;
 953         unsigned long val;
 954
 955         ret = kstrtoul(buf, 0, &val);
 956         if (ret < 0)
 957                 return ret;
 958
 959         tunables->boostpulse_duration_val = val;
 960         return count;
 961 }
 962
 963 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
 964                 char *buf)
 965 {
 966         return sprintf(buf, "%u\n", tunables->io_is_busy);
 967 }
 968
 969 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
 970                 const char *buf, size_t count)
 971 {
 972         int ret;
 973         unsigned long val;
 974
 975         ret = kstrtoul(buf, 0, &val);
 976         if (ret < 0)
 977                 return ret;
 978         tunables->io_is_busy = val;
 979         return count;
 980 }
 981
 982 /*
 983  * Create show/store routines
 984  * - sys: One governor instance for complete SYSTEM
 985  * - pol: One governor instance per struct cpufreq_policy
 986  */
 987 #define show_gov_pol_sys(file_name)                                     \
 988 static ssize_t show_##file_name##_gov_sys                               \
 989 (struct kobject *kobj, struct attribute *attr, char *buf)               \
 990 {                                                                       \
 991         return show_##file_name(common_tunables, buf);                  \
 992 }                                                                       \
 993                                                                         \
 994 static ssize_t show_##file_name##_gov_pol                               \
 995 (struct cpufreq_policy *policy, char *buf)                              \
 996 {                                                                       \
 997         return show_##file_name(policy->governor_data, buf);            \
 998 }
 999
1000 #define store_gov_pol_sys(file_name)                                    \
1001 static ssize_t store_##file_name##_gov_sys                              \
1002 (struct kobject *kobj, struct attribute *attr, const char *buf,         \
1003         size_t count)                                                   \
1004 {                                                                       \
1005         return store_##file_name(common_tunables, buf, count);          \
1006 }                                                                       \
1007                                                                         \
1008 static ssize_t store_##file_name##_gov_pol                              \
1009 (struct cpufreq_policy *policy, const char *buf, size_t count)          \
1010 {                                                                       \
1011         return store_##file_name(policy->governor_data, buf, count);    \
1012 }
1013
1014 #define show_store_gov_pol_sys(file_name)                               \
1015 show_gov_pol_sys(file_name);                                            \
1016 store_gov_pol_sys(file_name)
1017
1018 show_store_gov_pol_sys(target_loads);
1019 show_store_gov_pol_sys(above_hispeed_delay);
1020 show_store_gov_pol_sys(hispeed_freq);
1021 show_store_gov_pol_sys(go_hispeed_load);
1022 show_store_gov_pol_sys(min_sample_time);
1023 show_store_gov_pol_sys(timer_rate);
1024 show_store_gov_pol_sys(timer_slack);
1025 show_store_gov_pol_sys(boost);
1026 store_gov_pol_sys(boostpulse);
1027 show_store_gov_pol_sys(boostpulse_duration);
1028 show_store_gov_pol_sys(io_is_busy);
1029
1030 #define gov_sys_attr_rw(_name)                                          \
1031 static struct global_attr _name##_gov_sys =                             \
1032 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1033
1034 #define gov_pol_attr_rw(_name)                                          \
1035 static struct freq_attr _name##_gov_pol =                               \
1036 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1037
1038 #define gov_sys_pol_attr_rw(_name)                                      \
1039         gov_sys_attr_rw(_name);                                         \
1040         gov_pol_attr_rw(_name)
1041
1042 gov_sys_pol_attr_rw(target_loads);
1043 gov_sys_pol_attr_rw(above_hispeed_delay);
1044 gov_sys_pol_attr_rw(hispeed_freq);
1045 gov_sys_pol_attr_rw(go_hispeed_load);
1046 gov_sys_pol_attr_rw(min_sample_time);
1047 gov_sys_pol_attr_rw(timer_rate);
1048 gov_sys_pol_attr_rw(timer_slack);
1049 gov_sys_pol_attr_rw(boost);
1050 gov_sys_pol_attr_rw(boostpulse_duration);
1051 gov_sys_pol_attr_rw(io_is_busy);
1052
1053 static struct global_attr boostpulse_gov_sys =
1054         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1055
1056 static struct freq_attr boostpulse_gov_pol =
1057         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1058
1059 /* One Governor instance for entire system */
1060 static struct attribute *interactive_attributes_gov_sys[] = {
1061         &target_loads_gov_sys.attr,
1062         &above_hispeed_delay_gov_sys.attr,
1063         &hispeed_freq_gov_sys.attr,
1064         &go_hispeed_load_gov_sys.attr,
1065         &min_sample_time_gov_sys.attr,
1066         &timer_rate_gov_sys.attr,
1067         &timer_slack_gov_sys.attr,
1068         &boost_gov_sys.attr,
1069         &boostpulse_gov_sys.attr,
1070         &boostpulse_duration_gov_sys.attr,
1071         &io_is_busy_gov_sys.attr,
1072         NULL,
1073 };
1074
1075 static struct attribute_group interactive_attr_group_gov_sys = {
1076         .attrs = interactive_attributes_gov_sys,
1077         .name = "interactive",
1078 };
1079
1080 /* Per policy governor instance */
1081 static struct attribute *interactive_attributes_gov_pol[] = {
1082         &target_loads_gov_pol.attr,
1083         &above_hispeed_delay_gov_pol.attr,
1084         &hispeed_freq_gov_pol.attr,
1085         &go_hispeed_load_gov_pol.attr,
1086         &min_sample_time_gov_pol.attr,
1087         &timer_rate_gov_pol.attr,
1088         &timer_slack_gov_pol.attr,
1089         &boost_gov_pol.attr,
1090         &boostpulse_gov_pol.attr,
1091         &boostpulse_duration_gov_pol.attr,
1092         &io_is_busy_gov_pol.attr,
1093         NULL,
1094 };
1095
1096 static struct attribute_group interactive_attr_group_gov_pol = {
1097         .attrs = interactive_attributes_gov_pol,
1098         .name = "interactive",
1099 };
1100
1101 static struct attribute_group *get_sysfs_attr(void)
1102 {
1103         if (have_governor_per_policy())
1104                 return &interactive_attr_group_gov_pol;
1105         else
1106                 return &interactive_attr_group_gov_sys;
1107 }
1108
1109 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1110                                              unsigned long val,
1111                                              void *data)
1112 {
1113         switch (val) {
1114         case IDLE_START:
1115                 cpufreq_interactive_idle_start();
1116                 break;
1117         case IDLE_END:
1118                 cpufreq_interactive_idle_end();
1119                 break;
1120         }
1121
1122         return 0;
1123 }
1124
1125 static struct notifier_block cpufreq_interactive_idle_nb = {
1126         .notifier_call = cpufreq_interactive_idle_notifier,
1127 };
1128
1129 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1130                 unsigned int event)
1131 {
1132         int rc;
1133         unsigned int j;
1134         struct cpufreq_interactive_cpuinfo *pcpu;
1135         struct cpufreq_frequency_table *freq_table;
1136         struct cpufreq_interactive_tunables *tunables;
1137         unsigned long flags;
1138
1139         if (have_governor_per_policy())
1140                 tunables = policy->governor_data;
1141         else
1142                 tunables = common_tunables;
1143
1144         WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1145
1146         switch (event) {
1147         case CPUFREQ_GOV_POLICY_INIT:
1148                 if (have_governor_per_policy()) {
1149                         WARN_ON(tunables);
1150                 } else if (tunables) {
1151                         tunables->usage_count++;
1152                         policy->governor_data = tunables;
1153                         return 0;
1154                 }
1155
1156                 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1157                 if (!tunables) {
1158                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1159                         return -ENOMEM;
1160                 }
1161
1162                 tunables->usage_count = 1;
1163                 tunables->above_hispeed_delay = default_above_hispeed_delay;
1164                 tunables->nabove_hispeed_delay =
1165                         ARRAY_SIZE(default_above_hispeed_delay);
1166                 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1167                 tunables->target_loads = default_target_loads;
1168                 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1169                 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1170                 tunables->timer_rate = DEFAULT_TIMER_RATE;
1171                 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1172                 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1173
1174                 spin_lock_init(&tunables->target_loads_lock);
1175                 spin_lock_init(&tunables->above_hispeed_delay_lock);
1176
1177                 policy->governor_data = tunables;
1178                 if (!have_governor_per_policy()) {
1179                         common_tunables = tunables;
1180                         WARN_ON(cpufreq_get_global_kobject());
1181                 }
1182
1183                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1184                                 get_sysfs_attr());
1185                 if (rc) {
1186                         kfree(tunables);
1187                         policy->governor_data = NULL;
1188                         if (!have_governor_per_policy()) {
1189                                 common_tunables = NULL;
1190                                 cpufreq_put_global_kobject();
1191                         }
1192                         return rc;
1193                 }
1194
1195                 if (!policy->governor->initialized) {
1196                         idle_notifier_register(&cpufreq_interactive_idle_nb);
1197                         cpufreq_register_notifier(&cpufreq_notifier_block,
1198                                         CPUFREQ_TRANSITION_NOTIFIER);
1199                 }
1200
1201                 break;
1202
1203         case CPUFREQ_GOV_POLICY_EXIT:
1204                 if (!--tunables->usage_count) {
1205                         if (policy->governor->initialized == 1) {
1206                                 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1207                                                 CPUFREQ_TRANSITION_NOTIFIER);
1208                                 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1209                         }
1210
1211                         sysfs_remove_group(get_governor_parent_kobj(policy),
1212                                         get_sysfs_attr());
1213
1214                         if (!have_governor_per_policy())
1215                                 cpufreq_put_global_kobject();
1216
1217                         kfree(tunables);
1218                         common_tunables = NULL;
1219                 }
1220
1221                 policy->governor_data = NULL;
1222                 break;
1223
1224         case CPUFREQ_GOV_START:
1225                 mutex_lock(&gov_lock);
1226
1227                 freq_table = cpufreq_frequency_get_table(policy->cpu);
1228                 if (!tunables->hispeed_freq)
1229                         tunables->hispeed_freq = policy->max;
1230
1231                 for_each_cpu(j, policy->cpus) {
1232                         pcpu = &per_cpu(cpuinfo, j);
1233                         pcpu->policy = policy;
1234                         pcpu->target_freq = policy->cur;
1235                         pcpu->freq_table = freq_table;
1236                         pcpu->floor_freq = pcpu->target_freq;
1237                         pcpu->floor_validate_time =
1238                                 ktime_to_us(ktime_get());
1239                         pcpu->hispeed_validate_time =
1240                                 pcpu->floor_validate_time;
1241                         pcpu->max_freq = policy->max;
1242                         down_write(&pcpu->enable_sem);
1243                         del_timer_sync(&pcpu->cpu_timer);
1244                         del_timer_sync(&pcpu->cpu_slack_timer);
1245                         cpufreq_interactive_timer_start(tunables, j);
1246                         pcpu->governor_enabled = 1;
1247                         up_write(&pcpu->enable_sem);
1248                 }
1249
1250                 mutex_unlock(&gov_lock);
1251                 break;
1252
1253         case CPUFREQ_GOV_STOP:
1254                 mutex_lock(&gov_lock);
1255                 for_each_cpu(j, policy->cpus) {
1256                         pcpu = &per_cpu(cpuinfo, j);
1257                         down_write(&pcpu->enable_sem);
1258                         pcpu->governor_enabled = 0;
1259                         del_timer_sync(&pcpu->cpu_timer);
1260                         del_timer_sync(&pcpu->cpu_slack_timer);
1261                         up_write(&pcpu->enable_sem);
1262                 }
1263
1264                 mutex_unlock(&gov_lock);
1265                 break;
1266
1267         case CPUFREQ_GOV_LIMITS:
1268                 if (policy->max < policy->cur)
1269                         __cpufreq_driver_target(policy,
1270                                         policy->max, CPUFREQ_RELATION_H);
1271                 else if (policy->min > policy->cur)
1272                         __cpufreq_driver_target(policy,
1273                                         policy->min, CPUFREQ_RELATION_L);
1274                 for_each_cpu(j, policy->cpus) {
1275                         pcpu = &per_cpu(cpuinfo, j);
1276
1277                         down_read(&pcpu->enable_sem);
1278                         if (pcpu->governor_enabled == 0) {
1279                                 up_read(&pcpu->enable_sem);
1280                                 continue;
1281                         }
1282
1283                         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1284                         if (policy->max < pcpu->target_freq)
1285                                 pcpu->target_freq = policy->max;
1286                         else if (policy->min > pcpu->target_freq)
1287                                 pcpu->target_freq = policy->min;
1288
1289                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1290                         up_read(&pcpu->enable_sem);
1291
1292                         /* Reschedule timer only if policy->max is raised.
1293                          * Delete the timers, else the timer callback may
1294                          * return without re-arm the timer when failed
1295                          * acquire the semaphore. This race may cause timer
1296                          * stopped unexpectedly.
1297                          */
1298
1299                         if (policy->max > pcpu->max_freq) {
1300                                 down_write(&pcpu->enable_sem);
1301                                 del_timer_sync(&pcpu->cpu_timer);
1302                                 del_timer_sync(&pcpu->cpu_slack_timer);
1303                                 cpufreq_interactive_timer_start(tunables, j);
1304                                 up_write(&pcpu->enable_sem);
1305                         }
1306
1307                         pcpu->max_freq = policy->max;
1308                 }
1309                 break;
1310         }
1311         return 0;
1312 }
1313
1314 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1315 static
1316 #endif
1317 struct cpufreq_governor cpufreq_gov_interactive = {
1318         .name = "interactive",
1319         .governor = cpufreq_governor_interactive,
1320         .max_transition_latency = 10000000,
1321         .owner = THIS_MODULE,
1322 };
1323
1324 static void cpufreq_interactive_nop_timer(unsigned long data)
1325 {
1326 }
1327
1328 static int __init cpufreq_interactive_init(void)
1329 {
1330         unsigned int i;
1331         struct cpufreq_interactive_cpuinfo *pcpu;
1332         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1333
1334         /* Initalize per-cpu timers */
1335         for_each_possible_cpu(i) {
1336                 pcpu = &per_cpu(cpuinfo, i);
1337                 init_timer_deferrable(&pcpu->cpu_timer);
1338                 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1339                 pcpu->cpu_timer.data = i;
1340                 init_timer(&pcpu->cpu_slack_timer);
1341                 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1342                 spin_lock_init(&pcpu->load_lock);
1343                 spin_lock_init(&pcpu->target_freq_lock);
1344                 init_rwsem(&pcpu->enable_sem);
1345         }
1346
1347         spin_lock_init(&speedchange_cpumask_lock);
1348         mutex_init(&gov_lock);
1349         speedchange_task =
1350                 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1351                                "cfinteractive");
1352         if (IS_ERR(speedchange_task))
1353                 return PTR_ERR(speedchange_task);
1354
1355         sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, &param);
1356         get_task_struct(speedchange_task);
1357
1358         /* NB: wake up so the thread does not look hung to the freezer */
1359         wake_up_process(speedchange_task);
1360
1361         return cpufreq_register_governor(&cpufreq_gov_interactive);
1362 }
1363
1364 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1365 fs_initcall(cpufreq_interactive_init);
1366 #else
1367 module_init(cpufreq_interactive_init);
1368 #endif
1369
1370 static void __exit cpufreq_interactive_exit(void)
1371 {
1372         cpufreq_unregister_governor(&cpufreq_gov_interactive);
1373         kthread_stop(speedchange_task);
1374         put_task_struct(speedchange_task);
1375 }
1376
1377 module_exit(cpufreq_interactive_exit);
1378
1379 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1380 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1381         "Latency sensitive workloads");
1382 MODULE_LICENSE("GPL");