drivers/cpufreq/cpufreq_interactive.c

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