drivers/cpufreq/cpufreq_interactive.c

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