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