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 /*
 120  * HACK: FIXME: Bring back cpufreq_{get,put}_global_kobject()
 121  * definition removed by upstream commit 8eec1020f0c0 "cpufreq:
 122  * create cpu/cpufreq at boot time" to fix build failures.
 123  */
 124 static int cpufreq_global_kobject_usage;
 125
 126 int cpufreq_get_global_kobject(void)
 127 {
 128         if (!cpufreq_global_kobject_usage++)
 129                 return kobject_add(cpufreq_global_kobject,
 130                                 &cpu_subsys.dev_root->kobj, "%s", "cpufreq");
 131
 132         return 0;
 133 }
 134
 135 void cpufreq_put_global_kobject(void)
 136 {
 137         if (!--cpufreq_global_kobject_usage)
 138                 kobject_del(cpufreq_global_kobject);
 139 }
 140
 141 /* For cases where we have single governor instance for system */
 142 static struct cpufreq_interactive_tunables *common_tunables;
 143
 144 static struct attribute_group *get_sysfs_attr(void);
 145
 146 static void cpufreq_interactive_timer_resched(
 147         struct cpufreq_interactive_cpuinfo *pcpu)
 148 {
 149         struct cpufreq_interactive_tunables *tunables =
 150                 pcpu->policy->governor_data;
 151         unsigned long expires;
 152         unsigned long flags;
 153
 154         spin_lock_irqsave(&pcpu->load_lock, flags);
 155         pcpu->time_in_idle =
 156                 get_cpu_idle_time(smp_processor_id(),
 157                                   &pcpu->time_in_idle_timestamp,
 158                                   tunables->io_is_busy);
 159         pcpu->cputime_speedadj = 0;
 160         pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
 161         expires = jiffies + usecs_to_jiffies(tunables->timer_rate);
 162         mod_timer_pinned(&pcpu->cpu_timer, expires);
 163
 164         if (tunables->timer_slack_val >= 0 &&
 165             pcpu->target_freq > pcpu->policy->min) {
 166                 expires += usecs_to_jiffies(tunables->timer_slack_val);
 167                 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
 168         }
 169
 170         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 171 }
 172
 173 /* The caller shall take enable_sem write semaphore to avoid any timer race.
 174  * The cpu_timer and cpu_slack_timer must be deactivated when calling this
 175  * function.
 176  */
 177 static void cpufreq_interactive_timer_start(
 178         struct cpufreq_interactive_tunables *tunables, int cpu)
 179 {
 180         struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
 181         unsigned long expires = jiffies +
 182                 usecs_to_jiffies(tunables->timer_rate);
 183         unsigned long flags;
 184
 185         pcpu->cpu_timer.expires = expires;
 186         add_timer_on(&pcpu->cpu_timer, cpu);
 187         if (tunables->timer_slack_val >= 0 &&
 188             pcpu->target_freq > pcpu->policy->min) {
 189                 expires += usecs_to_jiffies(tunables->timer_slack_val);
 190                 pcpu->cpu_slack_timer.expires = expires;
 191                 add_timer_on(&pcpu->cpu_slack_timer, cpu);
 192         }
 193
 194         spin_lock_irqsave(&pcpu->load_lock, flags);
 195         pcpu->time_in_idle =
 196                 get_cpu_idle_time(cpu, &pcpu->time_in_idle_timestamp,
 197                                   tunables->io_is_busy);
 198         pcpu->cputime_speedadj = 0;
 199         pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
 200         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 201 }
 202
 203 static unsigned int freq_to_above_hispeed_delay(
 204         struct cpufreq_interactive_tunables *tunables,
 205         unsigned int freq)
 206 {
 207         int i;
 208         unsigned int ret;
 209         unsigned long flags;
 210
 211         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 212
 213         for (i = 0; i < tunables->nabove_hispeed_delay - 1 &&
 214                         freq >= tunables->above_hispeed_delay[i+1]; i += 2)
 215                 ;
 216
 217         ret = tunables->above_hispeed_delay[i];
 218         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 219         return ret;
 220 }
 221
 222 static unsigned int freq_to_targetload(
 223         struct cpufreq_interactive_tunables *tunables, unsigned int freq)
 224 {
 225         int i;
 226         unsigned int ret;
 227         unsigned long flags;
 228
 229         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 230
 231         for (i = 0; i < tunables->ntarget_loads - 1 &&
 232                     freq >= tunables->target_loads[i+1]; i += 2)
 233                 ;
 234
 235         ret = tunables->target_loads[i];
 236         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 237         return ret;
 238 }
 239
 240 /*
 241  * If increasing frequencies never map to a lower target load then
 242  * choose_freq() will find the minimum frequency that does not exceed its
 243  * target load given the current load.
 244  */
 245 static unsigned int choose_freq(struct cpufreq_interactive_cpuinfo *pcpu,
 246                 unsigned int loadadjfreq)
 247 {
 248         unsigned int freq = pcpu->policy->cur;
 249         unsigned int prevfreq, freqmin, freqmax;
 250         unsigned int tl;
 251         int index;
 252
 253         freqmin = 0;
 254         freqmax = UINT_MAX;
 255
 256         do {
 257                 prevfreq = freq;
 258                 tl = freq_to_targetload(pcpu->policy->governor_data, freq);
 259
 260                 /*
 261                  * Find the lowest frequency where the computed load is less
 262                  * than or equal to the target load.
 263                  */
 264
 265                 if (cpufreq_frequency_table_target(
 266                             pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
 267                             CPUFREQ_RELATION_L, &index))
 268                         break;
 269                 freq = pcpu->freq_table[index].frequency;
 270
 271                 if (freq > prevfreq) {
 272                         /* The previous frequency is too low. */
 273                         freqmin = prevfreq;
 274
 275                         if (freq >= freqmax) {
 276                                 /*
 277                                  * Find the highest frequency that is less
 278                                  * than freqmax.
 279                                  */
 280                                 if (cpufreq_frequency_table_target(
 281                                             pcpu->policy, pcpu->freq_table,
 282                                             freqmax - 1, CPUFREQ_RELATION_H,
 283                                             &index))
 284                                         break;
 285                                 freq = pcpu->freq_table[index].frequency;
 286
 287                                 if (freq == freqmin) {
 288                                         /*
 289                                          * The first frequency below freqmax
 290                                          * has already been found to be too
 291                                          * low.  freqmax is the lowest speed
 292                                          * we found that is fast enough.
 293                                          */
 294                                         freq = freqmax;
 295                                         break;
 296                                 }
 297                         }
 298                 } else if (freq < prevfreq) {
 299                         /* The previous frequency is high enough. */
 300                         freqmax = prevfreq;
 301
 302                         if (freq <= freqmin) {
 303                                 /*
 304                                  * Find the lowest frequency that is higher
 305                                  * than freqmin.
 306                                  */
 307                                 if (cpufreq_frequency_table_target(
 308                                             pcpu->policy, pcpu->freq_table,
 309                                             freqmin + 1, CPUFREQ_RELATION_L,
 310                                             &index))
 311                                         break;
 312                                 freq = pcpu->freq_table[index].frequency;
 313
 314                                 /*
 315                                  * If freqmax is the first frequency above
 316                                  * freqmin then we have already found that
 317                                  * this speed is fast enough.
 318                                  */
 319                                 if (freq == freqmax)
 320                                         break;
 321                         }
 322                 }
 323
 324                 /* If same frequency chosen as previous then done. */
 325         } while (freq != prevfreq);
 326
 327         return freq;
 328 }
 329
 330 static u64 update_load(int cpu)
 331 {
 332         struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
 333         struct cpufreq_interactive_tunables *tunables =
 334                 pcpu->policy->governor_data;
 335         u64 now;
 336         u64 now_idle;
 337         unsigned int delta_idle;
 338         unsigned int delta_time;
 339         u64 active_time;
 340
 341         now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
 342         delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
 343         delta_time = (unsigned int)(now - pcpu->time_in_idle_timestamp);
 344
 345         if (delta_time <= delta_idle)
 346                 active_time = 0;
 347         else
 348                 active_time = delta_time - delta_idle;
 349
 350         pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
 351
 352         pcpu->time_in_idle = now_idle;
 353         pcpu->time_in_idle_timestamp = now;
 354         return now;
 355 }
 356
 357 static void cpufreq_interactive_timer(unsigned long data)
 358 {
 359         u64 now;
 360         unsigned int delta_time;
 361         u64 cputime_speedadj;
 362         int cpu_load;
 363         struct cpufreq_interactive_cpuinfo *pcpu =
 364                 &per_cpu(cpuinfo, data);
 365         struct cpufreq_interactive_tunables *tunables =
 366                 pcpu->policy->governor_data;
 367         unsigned int new_freq;
 368         unsigned int loadadjfreq;
 369         unsigned int index;
 370         unsigned long flags;
 371         u64 max_fvtime;
 372
 373         if (!down_read_trylock(&pcpu->enable_sem))
 374                 return;
 375         if (!pcpu->governor_enabled)
 376                 goto exit;
 377
 378         spin_lock_irqsave(&pcpu->load_lock, flags);
 379         now = update_load(data);
 380         delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
 381         cputime_speedadj = pcpu->cputime_speedadj;
 382         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 383
 384         if (WARN_ON_ONCE(!delta_time))
 385                 goto rearm;
 386
 387         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
 388         do_div(cputime_speedadj, delta_time);
 389         loadadjfreq = (unsigned int)cputime_speedadj * 100;
 390         cpu_load = loadadjfreq / pcpu->policy->cur;
 391         tunables->boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
 392
 393         if (cpu_load >= tunables->go_hispeed_load || tunables->boosted) {
 394                 if (pcpu->policy->cur < tunables->hispeed_freq) {
 395                         new_freq = tunables->hispeed_freq;
 396                 } else {
 397                         new_freq = choose_freq(pcpu, loadadjfreq);
 398
 399                         if (new_freq < tunables->hispeed_freq)
 400                                 new_freq = tunables->hispeed_freq;
 401                 }
 402         } else {
 403                 new_freq = choose_freq(pcpu, loadadjfreq);
 404                 if (new_freq > tunables->hispeed_freq &&
 405                                 pcpu->policy->cur < tunables->hispeed_freq)
 406                         new_freq = tunables->hispeed_freq;
 407         }
 408
 409         if (pcpu->policy->cur >= tunables->hispeed_freq &&
 410             new_freq > pcpu->policy->cur &&
 411             now - pcpu->pol_hispeed_val_time <
 412             freq_to_above_hispeed_delay(tunables, pcpu->policy->cur)) {
 413                 trace_cpufreq_interactive_notyet(
 414                         data, cpu_load, pcpu->target_freq,
 415                         pcpu->policy->cur, new_freq);
 416                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 417                 goto rearm;
 418         }
 419
 420         pcpu->loc_hispeed_val_time = now;
 421
 422         if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
 423                                            new_freq, CPUFREQ_RELATION_L,
 424                                            &index)) {
 425                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 426                 goto rearm;
 427         }
 428
 429         new_freq = pcpu->freq_table[index].frequency;
 430
 431         /*
 432          * Do not scale below floor_freq unless we have been at or above the
 433          * floor frequency for the minimum sample time since last validated.
 434          */
 435         max_fvtime = max(pcpu->pol_floor_val_time, pcpu->loc_floor_val_time);
 436         if (new_freq < pcpu->floor_freq &&
 437             pcpu->target_freq >= pcpu->policy->cur) {
 438                 if (now - max_fvtime < tunables->min_sample_time) {
 439                         trace_cpufreq_interactive_notyet(
 440                                 data, cpu_load, pcpu->target_freq,
 441                                 pcpu->policy->cur, new_freq);
 442                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 443                         goto rearm;
 444                 }
 445         }
 446
 447         /*
 448          * Update the timestamp for checking whether speed has been held at
 449          * or above the selected frequency for a minimum of min_sample_time,
 450          * if not boosted to hispeed_freq.  If boosted to hispeed_freq then we
 451          * allow the speed to drop as soon as the boostpulse duration expires
 452          * (or the indefinite boost is turned off).
 453          */
 454
 455         if (!tunables->boosted || new_freq > tunables->hispeed_freq) {
 456                 pcpu->floor_freq = new_freq;
 457                 if (pcpu->target_freq >= pcpu->policy->cur ||
 458                     new_freq >= pcpu->policy->cur)
 459                         pcpu->loc_floor_val_time = now;
 460         }
 461
 462         if (pcpu->target_freq == new_freq &&
 463                         pcpu->target_freq <= pcpu->policy->cur) {
 464                 trace_cpufreq_interactive_already(
 465                         data, cpu_load, pcpu->target_freq,
 466                         pcpu->policy->cur, new_freq);
 467                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 468                 goto rearm;
 469         }
 470
 471         trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
 472                                          pcpu->policy->cur, new_freq);
 473
 474         pcpu->target_freq = new_freq;
 475         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 476         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 477         cpumask_set_cpu(data, &speedchange_cpumask);
 478         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 479         wake_up_process(speedchange_task);
 480
 481 rearm:
 482         if (!timer_pending(&pcpu->cpu_timer))
 483                 cpufreq_interactive_timer_resched(pcpu);
 484
 485 exit:
 486         up_read(&pcpu->enable_sem);
 487         return;
 488 }
 489
 490 static void cpufreq_interactive_idle_end(void)
 491 {
 492         struct cpufreq_interactive_cpuinfo *pcpu =
 493                 &per_cpu(cpuinfo, smp_processor_id());
 494
 495         if (!down_read_trylock(&pcpu->enable_sem))
 496                 return;
 497         if (!pcpu->governor_enabled) {
 498                 up_read(&pcpu->enable_sem);
 499                 return;
 500         }
 501
 502         /* Arm the timer for 1-2 ticks later if not already. */
 503         if (!timer_pending(&pcpu->cpu_timer)) {
 504                 cpufreq_interactive_timer_resched(pcpu);
 505         } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
 506                 del_timer(&pcpu->cpu_timer);
 507                 del_timer(&pcpu->cpu_slack_timer);
 508                 cpufreq_interactive_timer(smp_processor_id());
 509         }
 510
 511         up_read(&pcpu->enable_sem);
 512 }
 513
 514 static void cpufreq_interactive_get_policy_info(struct cpufreq_policy *policy,
 515                                                 unsigned int *pmax_freq,
 516                                                 u64 *phvt, u64 *pfvt)
 517 {
 518         struct cpufreq_interactive_cpuinfo *pcpu;
 519         unsigned int max_freq = 0;
 520         u64 hvt = ~0ULL, fvt = 0;
 521         unsigned int i;
 522
 523         for_each_cpu(i, policy->cpus) {
 524                 pcpu = &per_cpu(cpuinfo, i);
 525
 526                 fvt = max(fvt, pcpu->loc_floor_val_time);
 527                 if (pcpu->target_freq > max_freq) {
 528                         max_freq = pcpu->target_freq;
 529                         hvt = pcpu->loc_hispeed_val_time;
 530                 } else if (pcpu->target_freq == max_freq) {
 531                         hvt = min(hvt, pcpu->loc_hispeed_val_time);
 532                 }
 533         }
 534
 535         *pmax_freq = max_freq;
 536         *phvt = hvt;
 537         *pfvt = fvt;
 538 }
 539
 540 static void cpufreq_interactive_adjust_cpu(unsigned int cpu,
 541                                            struct cpufreq_policy *policy)
 542 {
 543         struct cpufreq_interactive_cpuinfo *pcpu;
 544         u64 hvt, fvt;
 545         unsigned int max_freq;
 546         int i;
 547
 548         cpufreq_interactive_get_policy_info(policy, &max_freq, &hvt, &fvt);
 549
 550         for_each_cpu(i, policy->cpus) {
 551                 pcpu = &per_cpu(cpuinfo, i);
 552                 pcpu->pol_floor_val_time = fvt;
 553         }
 554
 555         if (max_freq != policy->cur) {
 556                 __cpufreq_driver_target(policy, max_freq, CPUFREQ_RELATION_H);
 557                 for_each_cpu(i, policy->cpus) {
 558                         pcpu = &per_cpu(cpuinfo, i);
 559                         pcpu->pol_hispeed_val_time = hvt;
 560                 }
 561         }
 562
 563         trace_cpufreq_interactive_setspeed(cpu, max_freq, policy->cur);
 564 }
 565
 566 static int cpufreq_interactive_speedchange_task(void *data)
 567 {
 568         unsigned int cpu;
 569         cpumask_t tmp_mask;
 570         unsigned long flags;
 571         struct cpufreq_interactive_cpuinfo *pcpu;
 572
 573         while (1) {
 574                 set_current_state(TASK_INTERRUPTIBLE);
 575                 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 576
 577                 if (cpumask_empty(&speedchange_cpumask)) {
 578                         spin_unlock_irqrestore(&speedchange_cpumask_lock,
 579                                                flags);
 580                         schedule();
 581
 582                         if (kthread_should_stop())
 583                                 break;
 584
 585                         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 586                 }
 587
 588                 set_current_state(TASK_RUNNING);
 589                 tmp_mask = speedchange_cpumask;
 590                 cpumask_clear(&speedchange_cpumask);
 591                 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 592
 593                 for_each_cpu(cpu, &tmp_mask) {
 594                         pcpu = &per_cpu(cpuinfo, cpu);
 595
 596                         down_write(&pcpu->policy->rwsem);
 597
 598                         if (likely(down_read_trylock(&pcpu->enable_sem))) {
 599                                 if (likely(pcpu->governor_enabled))
 600                                         cpufreq_interactive_adjust_cpu(cpu,
 601                                                         pcpu->policy);
 602                                 up_read(&pcpu->enable_sem);
 603                         }
 604
 605                         up_write(&pcpu->policy->rwsem);
 606                 }
 607         }
 608
 609         return 0;
 610 }
 611
 612 static void cpufreq_interactive_boost(struct cpufreq_interactive_tunables *tunables)
 613 {
 614         int i;
 615         int anyboost = 0;
 616         unsigned long flags[2];
 617         struct cpufreq_interactive_cpuinfo *pcpu;
 618
 619         tunables->boosted = true;
 620
 621         spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
 622
 623         for_each_online_cpu(i) {
 624                 pcpu = &per_cpu(cpuinfo, i);
 625                 if (tunables != pcpu->policy->governor_data)
 626                         continue;
 627
 628                 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
 629                 if (pcpu->target_freq < tunables->hispeed_freq) {
 630                         pcpu->target_freq = tunables->hispeed_freq;
 631                         cpumask_set_cpu(i, &speedchange_cpumask);
 632                         pcpu->pol_hispeed_val_time =
 633                                 ktime_to_us(ktime_get());
 634                         anyboost = 1;
 635                 }
 636                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
 637         }
 638
 639         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
 640
 641         if (anyboost)
 642                 wake_up_process(speedchange_task);
 643 }
 644
 645 static int cpufreq_interactive_notifier(
 646         struct notifier_block *nb, unsigned long val, void *data)
 647 {
 648         struct cpufreq_freqs *freq = data;
 649         struct cpufreq_interactive_cpuinfo *pcpu;
 650         int cpu;
 651         unsigned long flags;
 652
 653         if (val == CPUFREQ_POSTCHANGE) {
 654                 pcpu = &per_cpu(cpuinfo, freq->cpu);
 655                 if (!down_read_trylock(&pcpu->enable_sem))
 656                         return 0;
 657                 if (!pcpu->governor_enabled) {
 658                         up_read(&pcpu->enable_sem);
 659                         return 0;
 660                 }
 661
 662                 for_each_cpu(cpu, pcpu->policy->cpus) {
 663                         struct cpufreq_interactive_cpuinfo *pjcpu =
 664                                 &per_cpu(cpuinfo, cpu);
 665                         if (cpu != freq->cpu) {
 666                                 if (!down_read_trylock(&pjcpu->enable_sem))
 667                                         continue;
 668                                 if (!pjcpu->governor_enabled) {
 669                                         up_read(&pjcpu->enable_sem);
 670                                         continue;
 671                                 }
 672                         }
 673                         spin_lock_irqsave(&pjcpu->load_lock, flags);
 674                         update_load(cpu);
 675                         spin_unlock_irqrestore(&pjcpu->load_lock, flags);
 676                         if (cpu != freq->cpu)
 677                                 up_read(&pjcpu->enable_sem);
 678                 }
 679
 680                 up_read(&pcpu->enable_sem);
 681         }
 682         return 0;
 683 }
 684
 685 static struct notifier_block cpufreq_notifier_block = {
 686         .notifier_call = cpufreq_interactive_notifier,
 687 };
 688
 689 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
 690 {
 691         const char *cp;
 692         int i;
 693         int ntokens = 1;
 694         unsigned int *tokenized_data;
 695         int err = -EINVAL;
 696
 697         cp = buf;
 698         while ((cp = strpbrk(cp + 1, " :")))
 699                 ntokens++;
 700
 701         if (!(ntokens & 0x1))
 702                 goto err;
 703
 704         tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
 705         if (!tokenized_data) {
 706                 err = -ENOMEM;
 707                 goto err;
 708         }
 709
 710         cp = buf;
 711         i = 0;
 712         while (i < ntokens) {
 713                 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
 714                         goto err_kfree;
 715
 716                 cp = strpbrk(cp, " :");
 717                 if (!cp)
 718                         break;
 719                 cp++;
 720         }
 721
 722         if (i != ntokens)
 723                 goto err_kfree;
 724
 725         *num_tokens = ntokens;
 726         return tokenized_data;
 727
 728 err_kfree:
 729         kfree(tokenized_data);
 730 err:
 731         return ERR_PTR(err);
 732 }
 733
 734 static ssize_t show_target_loads(
 735         struct cpufreq_interactive_tunables *tunables,
 736         char *buf)
 737 {
 738         int i;
 739         ssize_t ret = 0;
 740         unsigned long flags;
 741
 742         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 743
 744         for (i = 0; i < tunables->ntarget_loads; i++)
 745                 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
 746                                i & 0x1 ? ":" : " ");
 747
 748         sprintf(buf + ret - 1, "\n");
 749         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 750         return ret;
 751 }
 752
 753 static ssize_t store_target_loads(
 754         struct cpufreq_interactive_tunables *tunables,
 755         const char *buf, size_t count)
 756 {
 757         int ntokens;
 758         unsigned int *new_target_loads = NULL;
 759         unsigned long flags;
 760
 761         new_target_loads = get_tokenized_data(buf, &ntokens);
 762         if (IS_ERR(new_target_loads))
 763                 return PTR_RET(new_target_loads);
 764
 765         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 766         if (tunables->target_loads != default_target_loads)
 767                 kfree(tunables->target_loads);
 768         tunables->target_loads = new_target_loads;
 769         tunables->ntarget_loads = ntokens;
 770         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 771         return count;
 772 }
 773
 774 static ssize_t show_above_hispeed_delay(
 775         struct cpufreq_interactive_tunables *tunables, char *buf)
 776 {
 777         int i;
 778         ssize_t ret = 0;
 779         unsigned long flags;
 780
 781         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 782
 783         for (i = 0; i < tunables->nabove_hispeed_delay; i++)
 784                 ret += sprintf(buf + ret, "%u%s",
 785                                tunables->above_hispeed_delay[i],
 786                                i & 0x1 ? ":" : " ");
 787
 788         sprintf(buf + ret - 1, "\n");
 789         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 790         return ret;
 791 }
 792
 793 static ssize_t store_above_hispeed_delay(
 794         struct cpufreq_interactive_tunables *tunables,
 795         const char *buf, size_t count)
 796 {
 797         int ntokens;
 798         unsigned int *new_above_hispeed_delay = NULL;
 799         unsigned long flags;
 800
 801         new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
 802         if (IS_ERR(new_above_hispeed_delay))
 803                 return PTR_RET(new_above_hispeed_delay);
 804
 805         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 806         if (tunables->above_hispeed_delay != default_above_hispeed_delay)
 807                 kfree(tunables->above_hispeed_delay);
 808         tunables->above_hispeed_delay = new_above_hispeed_delay;
 809         tunables->nabove_hispeed_delay = ntokens;
 810         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 811         return count;
 812
 813 }
 814
 815 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 816                 char *buf)
 817 {
 818         return sprintf(buf, "%u\n", tunables->hispeed_freq);
 819 }
 820
 821 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 822                 const char *buf, size_t count)
 823 {
 824         int ret;
 825         long unsigned int val;
 826
 827         ret = kstrtoul(buf, 0, &val);
 828         if (ret < 0)
 829                 return ret;
 830         tunables->hispeed_freq = val;
 831         return count;
 832 }
 833
 834 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
 835                 *tunables, char *buf)
 836 {
 837         return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
 838 }
 839
 840 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
 841                 *tunables, const char *buf, size_t count)
 842 {
 843         int ret;
 844         unsigned long val;
 845
 846         ret = kstrtoul(buf, 0, &val);
 847         if (ret < 0)
 848                 return ret;
 849         tunables->go_hispeed_load = val;
 850         return count;
 851 }
 852
 853 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
 854                 *tunables, char *buf)
 855 {
 856         return sprintf(buf, "%lu\n", tunables->min_sample_time);
 857 }
 858
 859 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
 860                 *tunables, const char *buf, size_t count)
 861 {
 862         int ret;
 863         unsigned long val;
 864
 865         ret = kstrtoul(buf, 0, &val);
 866         if (ret < 0)
 867                 return ret;
 868         tunables->min_sample_time = val;
 869         return count;
 870 }
 871
 872 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
 873                 char *buf)
 874 {
 875         return sprintf(buf, "%lu\n", tunables->timer_rate);
 876 }
 877
 878 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
 879                 const char *buf, size_t count)
 880 {
 881         int ret;
 882         unsigned long val, val_round;
 883
 884         ret = kstrtoul(buf, 0, &val);
 885         if (ret < 0)
 886                 return ret;
 887
 888         val_round = jiffies_to_usecs(usecs_to_jiffies(val));
 889         if (val != val_round)
 890                 pr_warn("timer_rate not aligned to jiffy. Rounded up to %lu\n",
 891                         val_round);
 892
 893         tunables->timer_rate = val_round;
 894         return count;
 895 }
 896
 897 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
 898                 char *buf)
 899 {
 900         return sprintf(buf, "%d\n", tunables->timer_slack_val);
 901 }
 902
 903 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
 904                 const char *buf, size_t count)
 905 {
 906         int ret;
 907         unsigned long val;
 908
 909         ret = kstrtol(buf, 10, &val);
 910         if (ret < 0)
 911                 return ret;
 912
 913         tunables->timer_slack_val = val;
 914         return count;
 915 }
 916
 917 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
 918                           char *buf)
 919 {
 920         return sprintf(buf, "%d\n", tunables->boost_val);
 921 }
 922
 923 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
 924                            const char *buf, size_t count)
 925 {
 926         int ret;
 927         unsigned long val;
 928
 929         ret = kstrtoul(buf, 0, &val);
 930         if (ret < 0)
 931                 return ret;
 932
 933         tunables->boost_val = val;
 934
 935         if (tunables->boost_val) {
 936                 trace_cpufreq_interactive_boost("on");
 937                 if (!tunables->boosted)
 938                         cpufreq_interactive_boost(tunables);
 939         } else {
 940                 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
 941                 trace_cpufreq_interactive_unboost("off");
 942         }
 943
 944         return count;
 945 }
 946
 947 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
 948                                 const char *buf, size_t count)
 949 {
 950         int ret;
 951         unsigned long val;
 952
 953         ret = kstrtoul(buf, 0, &val);
 954         if (ret < 0)
 955                 return ret;
 956
 957         tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
 958                 tunables->boostpulse_duration_val;
 959         trace_cpufreq_interactive_boost("pulse");
 960         if (!tunables->boosted)
 961                 cpufreq_interactive_boost(tunables);
 962         return count;
 963 }
 964
 965 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
 966                 *tunables, char *buf)
 967 {
 968         return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
 969 }
 970
 971 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
 972                 *tunables, const char *buf, size_t count)
 973 {
 974         int ret;
 975         unsigned long val;
 976
 977         ret = kstrtoul(buf, 0, &val);
 978         if (ret < 0)
 979                 return ret;
 980
 981         tunables->boostpulse_duration_val = val;
 982         return count;
 983 }
 984
 985 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
 986                 char *buf)
 987 {
 988         return sprintf(buf, "%u\n", tunables->io_is_busy);
 989 }
 990
 991 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
 992                 const char *buf, size_t count)
 993 {
 994         int ret;
 995         unsigned long val;
 996
 997         ret = kstrtoul(buf, 0, &val);
 998         if (ret < 0)
 999                 return ret;
1000         tunables->io_is_busy = val;
1001         return count;
1002 }
1003
1004 /*
1005  * Create show/store routines
1006  * - sys: One governor instance for complete SYSTEM
1007  * - pol: One governor instance per struct cpufreq_policy
1008  */
1009 #define show_gov_pol_sys(file_name)                                     \
1010 static ssize_t show_##file_name##_gov_sys                               \
1011 (struct kobject *kobj, struct attribute *attr, char *buf)               \
1012 {                                                                       \
1013         return show_##file_name(common_tunables, buf);                  \
1014 }                                                                       \
1015                                                                         \
1016 static ssize_t show_##file_name##_gov_pol                               \
1017 (struct cpufreq_policy *policy, char *buf)                              \
1018 {                                                                       \
1019         return show_##file_name(policy->governor_data, buf);            \
1020 }
1021
1022 #define store_gov_pol_sys(file_name)                                    \
1023 static ssize_t store_##file_name##_gov_sys                              \
1024 (struct kobject *kobj, struct attribute *attr, const char *buf,         \
1025         size_t count)                                                   \
1026 {                                                                       \
1027         return store_##file_name(common_tunables, buf, count);          \
1028 }                                                                       \
1029                                                                         \
1030 static ssize_t store_##file_name##_gov_pol                              \
1031 (struct cpufreq_policy *policy, const char *buf, size_t count)          \
1032 {                                                                       \
1033         return store_##file_name(policy->governor_data, buf, count);    \
1034 }
1035
1036 #define show_store_gov_pol_sys(file_name)                               \
1037 show_gov_pol_sys(file_name);                                            \
1038 store_gov_pol_sys(file_name)
1039
1040 show_store_gov_pol_sys(target_loads);
1041 show_store_gov_pol_sys(above_hispeed_delay);
1042 show_store_gov_pol_sys(hispeed_freq);
1043 show_store_gov_pol_sys(go_hispeed_load);
1044 show_store_gov_pol_sys(min_sample_time);
1045 show_store_gov_pol_sys(timer_rate);
1046 show_store_gov_pol_sys(timer_slack);
1047 show_store_gov_pol_sys(boost);
1048 store_gov_pol_sys(boostpulse);
1049 show_store_gov_pol_sys(boostpulse_duration);
1050 show_store_gov_pol_sys(io_is_busy);
1051
1052 #define gov_sys_attr_rw(_name)                                          \
1053 static struct global_attr _name##_gov_sys =                             \
1054 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1055
1056 #define gov_pol_attr_rw(_name)                                          \
1057 static struct freq_attr _name##_gov_pol =                               \
1058 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1059
1060 #define gov_sys_pol_attr_rw(_name)                                      \
1061         gov_sys_attr_rw(_name);                                         \
1062         gov_pol_attr_rw(_name)
1063
1064 gov_sys_pol_attr_rw(target_loads);
1065 gov_sys_pol_attr_rw(above_hispeed_delay);
1066 gov_sys_pol_attr_rw(hispeed_freq);
1067 gov_sys_pol_attr_rw(go_hispeed_load);
1068 gov_sys_pol_attr_rw(min_sample_time);
1069 gov_sys_pol_attr_rw(timer_rate);
1070 gov_sys_pol_attr_rw(timer_slack);
1071 gov_sys_pol_attr_rw(boost);
1072 gov_sys_pol_attr_rw(boostpulse_duration);
1073 gov_sys_pol_attr_rw(io_is_busy);
1074
1075 static struct global_attr boostpulse_gov_sys =
1076         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1077
1078 static struct freq_attr boostpulse_gov_pol =
1079         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1080
1081 /* One Governor instance for entire system */
1082 static struct attribute *interactive_attributes_gov_sys[] = {
1083         &target_loads_gov_sys.attr,
1084         &above_hispeed_delay_gov_sys.attr,
1085         &hispeed_freq_gov_sys.attr,
1086         &go_hispeed_load_gov_sys.attr,
1087         &min_sample_time_gov_sys.attr,
1088         &timer_rate_gov_sys.attr,
1089         &timer_slack_gov_sys.attr,
1090         &boost_gov_sys.attr,
1091         &boostpulse_gov_sys.attr,
1092         &boostpulse_duration_gov_sys.attr,
1093         &io_is_busy_gov_sys.attr,
1094         NULL,
1095 };
1096
1097 static struct attribute_group interactive_attr_group_gov_sys = {
1098         .attrs = interactive_attributes_gov_sys,
1099         .name = "interactive",
1100 };
1101
1102 /* Per policy governor instance */
1103 static struct attribute *interactive_attributes_gov_pol[] = {
1104         &target_loads_gov_pol.attr,
1105         &above_hispeed_delay_gov_pol.attr,
1106         &hispeed_freq_gov_pol.attr,
1107         &go_hispeed_load_gov_pol.attr,
1108         &min_sample_time_gov_pol.attr,
1109         &timer_rate_gov_pol.attr,
1110         &timer_slack_gov_pol.attr,
1111         &boost_gov_pol.attr,
1112         &boostpulse_gov_pol.attr,
1113         &boostpulse_duration_gov_pol.attr,
1114         &io_is_busy_gov_pol.attr,
1115         NULL,
1116 };
1117
1118 static struct attribute_group interactive_attr_group_gov_pol = {
1119         .attrs = interactive_attributes_gov_pol,
1120         .name = "interactive",
1121 };
1122
1123 static struct attribute_group *get_sysfs_attr(void)
1124 {
1125         if (have_governor_per_policy())
1126                 return &interactive_attr_group_gov_pol;
1127         else
1128                 return &interactive_attr_group_gov_sys;
1129 }
1130
1131 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1132                                              unsigned long val,
1133                                              void *data)
1134 {
1135         if (val == IDLE_END)
1136                 cpufreq_interactive_idle_end();
1137
1138         return 0;
1139 }
1140
1141 static struct notifier_block cpufreq_interactive_idle_nb = {
1142         .notifier_call = cpufreq_interactive_idle_notifier,
1143 };
1144
1145 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1146                 unsigned int event)
1147 {
1148         int rc;
1149         unsigned int j;
1150         struct cpufreq_interactive_cpuinfo *pcpu;
1151         struct cpufreq_frequency_table *freq_table;
1152         struct cpufreq_interactive_tunables *tunables;
1153         unsigned long flags;
1154
1155         if (have_governor_per_policy())
1156                 tunables = policy->governor_data;
1157         else
1158                 tunables = common_tunables;
1159
1160         WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1161
1162         switch (event) {
1163         case CPUFREQ_GOV_POLICY_INIT:
1164                 if (have_governor_per_policy()) {
1165                         WARN_ON(tunables);
1166                 } else if (tunables) {
1167                         tunables->usage_count++;
1168                         policy->governor_data = tunables;
1169                         return 0;
1170                 }
1171
1172                 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1173                 if (!tunables) {
1174                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1175                         return -ENOMEM;
1176                 }
1177
1178                 tunables->usage_count = 1;
1179                 tunables->above_hispeed_delay = default_above_hispeed_delay;
1180                 tunables->nabove_hispeed_delay =
1181                         ARRAY_SIZE(default_above_hispeed_delay);
1182                 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1183                 tunables->target_loads = default_target_loads;
1184                 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1185                 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1186                 tunables->timer_rate = DEFAULT_TIMER_RATE;
1187                 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1188                 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1189
1190                 spin_lock_init(&tunables->target_loads_lock);
1191                 spin_lock_init(&tunables->above_hispeed_delay_lock);
1192
1193                 policy->governor_data = tunables;
1194                 if (!have_governor_per_policy()) {
1195                         common_tunables = tunables;
1196                         WARN_ON(cpufreq_get_global_kobject());
1197                 }
1198
1199                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1200                                 get_sysfs_attr());
1201                 if (rc) {
1202                         kfree(tunables);
1203                         policy->governor_data = NULL;
1204                         if (!have_governor_per_policy()) {
1205                                 common_tunables = NULL;
1206                                 cpufreq_put_global_kobject();
1207                         }
1208                         return rc;
1209                 }
1210
1211                 if (!policy->governor->initialized) {
1212                         idle_notifier_register(&cpufreq_interactive_idle_nb);
1213                         cpufreq_register_notifier(&cpufreq_notifier_block,
1214                                         CPUFREQ_TRANSITION_NOTIFIER);
1215                 }
1216
1217                 break;
1218
1219         case CPUFREQ_GOV_POLICY_EXIT:
1220                 if (!--tunables->usage_count) {
1221                         if (policy->governor->initialized == 1) {
1222                                 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1223                                                 CPUFREQ_TRANSITION_NOTIFIER);
1224                                 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1225                         }
1226
1227                         sysfs_remove_group(get_governor_parent_kobj(policy),
1228                                         get_sysfs_attr());
1229
1230                         if (!have_governor_per_policy())
1231                                 cpufreq_put_global_kobject();
1232
1233                         kfree(tunables);
1234                         common_tunables = NULL;
1235                 }
1236
1237                 policy->governor_data = NULL;
1238                 break;
1239
1240         case CPUFREQ_GOV_START:
1241                 mutex_lock(&gov_lock);
1242
1243                 freq_table = cpufreq_frequency_get_table(policy->cpu);
1244                 if (!tunables->hispeed_freq)
1245                         tunables->hispeed_freq = policy->max;
1246
1247                 for_each_cpu(j, policy->cpus) {
1248                         pcpu = &per_cpu(cpuinfo, j);
1249                         pcpu->policy = policy;
1250                         pcpu->target_freq = policy->cur;
1251                         pcpu->freq_table = freq_table;
1252                         pcpu->floor_freq = pcpu->target_freq;
1253                         pcpu->pol_floor_val_time =
1254                                 ktime_to_us(ktime_get());
1255                         pcpu->loc_floor_val_time = pcpu->pol_floor_val_time;
1256                         pcpu->pol_hispeed_val_time = pcpu->pol_floor_val_time;
1257                         pcpu->loc_hispeed_val_time = pcpu->pol_floor_val_time;
1258                         down_write(&pcpu->enable_sem);
1259                         del_timer_sync(&pcpu->cpu_timer);
1260                         del_timer_sync(&pcpu->cpu_slack_timer);
1261                         cpufreq_interactive_timer_start(tunables, j);
1262                         pcpu->governor_enabled = 1;
1263                         up_write(&pcpu->enable_sem);
1264                 }
1265
1266                 mutex_unlock(&gov_lock);
1267                 break;
1268
1269         case CPUFREQ_GOV_STOP:
1270                 mutex_lock(&gov_lock);
1271                 for_each_cpu(j, policy->cpus) {
1272                         pcpu = &per_cpu(cpuinfo, j);
1273                         down_write(&pcpu->enable_sem);
1274                         pcpu->governor_enabled = 0;
1275                         del_timer_sync(&pcpu->cpu_timer);
1276                         del_timer_sync(&pcpu->cpu_slack_timer);
1277                         up_write(&pcpu->enable_sem);
1278                 }
1279
1280                 mutex_unlock(&gov_lock);
1281                 break;
1282
1283         case CPUFREQ_GOV_LIMITS:
1284                 if (policy->max < policy->cur)
1285                         __cpufreq_driver_target(policy,
1286                                         policy->max, CPUFREQ_RELATION_H);
1287                 else if (policy->min > policy->cur)
1288                         __cpufreq_driver_target(policy,
1289                                         policy->min, CPUFREQ_RELATION_L);
1290                 for_each_cpu(j, policy->cpus) {
1291                         pcpu = &per_cpu(cpuinfo, j);
1292
1293                         down_read(&pcpu->enable_sem);
1294                         if (pcpu->governor_enabled == 0) {
1295                                 up_read(&pcpu->enable_sem);
1296                                 continue;
1297                         }
1298
1299                         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1300                         if (policy->max < pcpu->target_freq)
1301                                 pcpu->target_freq = policy->max;
1302                         else if (policy->min > pcpu->target_freq)
1303                                 pcpu->target_freq = policy->min;
1304
1305                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1306                         up_read(&pcpu->enable_sem);
1307                 }
1308                 break;
1309         }
1310         return 0;
1311 }
1312
1313 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1314 static
1315 #endif
1316 struct cpufreq_governor cpufreq_gov_interactive = {
1317         .name = "interactive",
1318         .governor = cpufreq_governor_interactive,
1319         .max_transition_latency = 10000000,
1320         .owner = THIS_MODULE,
1321 };
1322
1323 static void cpufreq_interactive_nop_timer(unsigned long data)
1324 {
1325 }
1326
1327 static int __init cpufreq_interactive_init(void)
1328 {
1329         unsigned int i;
1330         struct cpufreq_interactive_cpuinfo *pcpu;
1331         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1332
1333         /* Initalize per-cpu timers */
1334         for_each_possible_cpu(i) {
1335                 pcpu = &per_cpu(cpuinfo, i);
1336                 init_timer_deferrable(&pcpu->cpu_timer);
1337                 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1338                 pcpu->cpu_timer.data = i;
1339                 init_timer(&pcpu->cpu_slack_timer);
1340                 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1341                 spin_lock_init(&pcpu->load_lock);
1342                 spin_lock_init(&pcpu->target_freq_lock);
1343                 init_rwsem(&pcpu->enable_sem);
1344         }
1345
1346         spin_lock_init(&speedchange_cpumask_lock);
1347         mutex_init(&gov_lock);
1348         speedchange_task =
1349                 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1350                                "cfinteractive");
1351         if (IS_ERR(speedchange_task))
1352                 return PTR_ERR(speedchange_task);
1353
1354         sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, &param);
1355         get_task_struct(speedchange_task);
1356
1357         /* NB: wake up so the thread does not look hung to the freezer */
1358         wake_up_process(speedchange_task);
1359
1360         return cpufreq_register_governor(&cpufreq_gov_interactive);
1361 }
1362
1363 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1364 fs_initcall(cpufreq_interactive_init);
1365 #else
1366 module_init(cpufreq_interactive_init);
1367 #endif
1368
1369 static void __exit cpufreq_interactive_exit(void)
1370 {
1371         cpufreq_unregister_governor(&cpufreq_gov_interactive);
1372         kthread_stop(speedchange_task);
1373         put_task_struct(speedchange_task);
1374 }
1375
1376 module_exit(cpufreq_interactive_exit);
1377
1378 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1379 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1380         "Latency sensitive workloads");
1381 MODULE_LICENSE("GPL");