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 #include <linux/kernel_stat.h>
  34 #include <asm/cputime.h>
  35
  36 #define CREATE_TRACE_POINTS
  37 #include <trace/events/cpufreq_interactive.h>
  38
  39 struct cpufreq_interactive_cpuinfo {
  40         struct timer_list cpu_timer;
  41         struct timer_list cpu_slack_timer;
  42         spinlock_t load_lock; /* protects the next 4 fields */
  43         u64 time_in_idle;
  44         u64 time_in_idle_timestamp;
  45         u64 cputime_speedadj;
  46         u64 cputime_speedadj_timestamp;
  47         struct cpufreq_policy *policy;
  48         struct cpufreq_frequency_table *freq_table;
  49         spinlock_t target_freq_lock; /*protects target freq */
  50         unsigned int target_freq;
  51         unsigned int floor_freq;
  52         unsigned int max_freq;
  53         u64 floor_validate_time;
  54         u64 hispeed_validate_time;
  55         struct rw_semaphore enable_sem;
  56         int governor_enabled;
  57 };
  58
  59 static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
  60
  61 /* realtime thread handles frequency scaling */
  62 static struct task_struct *speedchange_task;
  63 static cpumask_t speedchange_cpumask;
  64 static spinlock_t speedchange_cpumask_lock;
  65 static struct mutex gov_lock;
  66
  67 /* Target load.  Lower values result in higher CPU speeds. */
  68 #define DEFAULT_TARGET_LOAD 90
  69 static unsigned int default_target_loads[] = {DEFAULT_TARGET_LOAD};
  70
  71 #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC)
  72 #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE
  73 static unsigned int default_above_hispeed_delay[] = {
  74         DEFAULT_ABOVE_HISPEED_DELAY };
  75
  76 struct cpufreq_interactive_tunables {
  77         int usage_count;
  78         /* Hi speed to bump to from lo speed when load burst (default max) */
  79         unsigned int hispeed_freq;
  80         /* Go to hi speed when CPU load at or above this value. */
  81 #define DEFAULT_GO_HISPEED_LOAD 99
  82         unsigned long go_hispeed_load;
  83         /* Target load. Lower values result in higher CPU speeds. */
  84         spinlock_t target_loads_lock;
  85         unsigned int *target_loads;
  86         int ntarget_loads;
  87         /*
  88          * The minimum amount of time to spend at a frequency before we can ramp
  89          * down.
  90          */
  91 #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC)
  92         unsigned long min_sample_time;
  93         /*
  94          * The sample rate of the timer used to increase frequency
  95          */
  96         unsigned long timer_rate;
  97         /*
  98          * Wait this long before raising speed above hispeed, by default a
  99          * single timer interval.
 100          */
 101         spinlock_t above_hispeed_delay_lock;
 102         unsigned int *above_hispeed_delay;
 103         int nabove_hispeed_delay;
 104         /* Non-zero means indefinite speed boost active */
 105         int boost_val;
 106         /* Duration of a boot pulse in usecs */
 107         int boostpulse_duration_val;
 108         /* End time of boost pulse in ktime converted to usecs */
 109         u64 boostpulse_endtime;
 110         /*
 111          * Max additional time to wait in idle, beyond timer_rate, at speeds
 112          * above minimum before wakeup to reduce speed, or -1 if unnecessary.
 113          */
 114 #define DEFAULT_TIMER_SLACK (4 * DEFAULT_TIMER_RATE)
 115         int timer_slack_val;
 116         bool io_is_busy;
 117 };
 118
 119 /* For cases where we have single governor instance for system */
 120 static struct cpufreq_interactive_tunables *common_tunables;
 121
 122 static struct attribute_group *get_sysfs_attr(void);
 123
 124 static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
 125                                                   cputime64_t *wall)
 126 {
 127         u64 idle_time;
 128         u64 cur_wall_time;
 129         u64 busy_time;
 130
 131         cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
 132
 133         busy_time  = kcpustat_cpu(cpu).cpustat[CPUTIME_USER];
 134         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SYSTEM];
 135         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_IRQ];
 136         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_SOFTIRQ];
 137         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_STEAL];
 138         busy_time += kcpustat_cpu(cpu).cpustat[CPUTIME_NICE];
 139
 140         idle_time = cur_wall_time - busy_time;
 141         if (wall)
 142                 *wall = jiffies_to_usecs(cur_wall_time);
 143
 144         return jiffies_to_usecs(idle_time);
 145 }
 146
 147 static inline cputime64_t get_cpu_idle_time(
 148         unsigned int cpu,
 149         cputime64_t *wall,
 150         bool io_is_busy)
 151 {
 152         u64 idle_time = get_cpu_idle_time_us(cpu, wall);
 153
 154         if (idle_time == -1ULL)
 155                 idle_time = get_cpu_idle_time_jiffy(cpu, wall);
 156         else if (!io_is_busy)
 157                 idle_time += get_cpu_iowait_time_us(cpu, wall);
 158
 159         return idle_time;
 160 }
 161
 162 static void cpufreq_interactive_timer_resched(
 163         struct cpufreq_interactive_cpuinfo *pcpu)
 164 {
 165         struct cpufreq_interactive_tunables *tunables =
 166                 pcpu->policy->governor_data;
 167         unsigned long expires;
 168         unsigned long flags;
 169
 170         spin_lock_irqsave(&pcpu->load_lock, flags);
 171         pcpu->time_in_idle =
 172                 get_cpu_idle_time(smp_processor_id(),
 173                                   &pcpu->time_in_idle_timestamp,
 174                                   tunables->io_is_busy);
 175         pcpu->cputime_speedadj = 0;
 176         pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
 177         expires = jiffies + usecs_to_jiffies(tunables->timer_rate);
 178         mod_timer_pinned(&pcpu->cpu_timer, expires);
 179
 180         if (tunables->timer_slack_val >= 0 &&
 181             pcpu->target_freq > pcpu->policy->min) {
 182                 expires += usecs_to_jiffies(tunables->timer_slack_val);
 183                 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
 184         }
 185
 186         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 187 }
 188
 189 /* The caller shall take enable_sem write semaphore to avoid any timer race.
 190  * The cpu_timer and cpu_slack_timer must be deactivated when calling this
 191  * function.
 192  */
 193 static void cpufreq_interactive_timer_start(
 194         struct cpufreq_interactive_tunables *tunables, int cpu)
 195 {
 196         struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
 197         unsigned long expires = jiffies +
 198                 usecs_to_jiffies(tunables->timer_rate);
 199         unsigned long flags;
 200
 201         pcpu->cpu_timer.expires = expires;
 202         add_timer_on(&pcpu->cpu_timer, cpu);
 203         if (tunables->timer_slack_val >= 0 &&
 204             pcpu->target_freq > pcpu->policy->min) {
 205                 expires += usecs_to_jiffies(tunables->timer_slack_val);
 206                 pcpu->cpu_slack_timer.expires = expires;
 207                 add_timer_on(&pcpu->cpu_slack_timer, cpu);
 208         }
 209
 210         spin_lock_irqsave(&pcpu->load_lock, flags);
 211         pcpu->time_in_idle =
 212                 get_cpu_idle_time(cpu, &pcpu->time_in_idle_timestamp,
 213                                   tunables->io_is_busy);
 214         pcpu->cputime_speedadj = 0;
 215         pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
 216         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 217 }
 218
 219 static unsigned int freq_to_above_hispeed_delay(
 220         struct cpufreq_interactive_tunables *tunables,
 221         unsigned int freq)
 222 {
 223         int i;
 224         unsigned int ret;
 225         unsigned long flags;
 226
 227         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 228
 229         for (i = 0; i < tunables->nabove_hispeed_delay - 1 &&
 230                         freq >= tunables->above_hispeed_delay[i+1]; i += 2)
 231                 ;
 232
 233         ret = tunables->above_hispeed_delay[i];
 234         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 235         return ret;
 236 }
 237
 238 static unsigned int freq_to_targetload(
 239         struct cpufreq_interactive_tunables *tunables, unsigned int freq)
 240 {
 241         int i;
 242         unsigned int ret;
 243         unsigned long flags;
 244
 245         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 246
 247         for (i = 0; i < tunables->ntarget_loads - 1 &&
 248                     freq >= tunables->target_loads[i+1]; i += 2)
 249                 ;
 250
 251         ret = tunables->target_loads[i];
 252         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 253         return ret;
 254 }
 255
 256 /*
 257  * If increasing frequencies never map to a lower target load then
 258  * choose_freq() will find the minimum frequency that does not exceed its
 259  * target load given the current load.
 260  */
 261 static unsigned int choose_freq(struct cpufreq_interactive_cpuinfo *pcpu,
 262                 unsigned int loadadjfreq)
 263 {
 264         unsigned int freq = pcpu->policy->cur;
 265         unsigned int prevfreq, freqmin, freqmax;
 266         unsigned int tl;
 267         int index;
 268
 269         freqmin = 0;
 270         freqmax = UINT_MAX;
 271
 272         do {
 273                 prevfreq = freq;
 274                 tl = freq_to_targetload(pcpu->policy->governor_data, freq);
 275
 276                 /*
 277                  * Find the lowest frequency where the computed load is less
 278                  * than or equal to the target load.
 279                  */
 280
 281                 if (cpufreq_frequency_table_target(
 282                             pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
 283                             CPUFREQ_RELATION_L, &index))
 284                         break;
 285                 freq = pcpu->freq_table[index].frequency;
 286
 287                 if (freq > prevfreq) {
 288                         /* The previous frequency is too low. */
 289                         freqmin = prevfreq;
 290
 291                         if (freq >= freqmax) {
 292                                 /*
 293                                  * Find the highest frequency that is less
 294                                  * than freqmax.
 295                                  */
 296                                 if (cpufreq_frequency_table_target(
 297                                             pcpu->policy, pcpu->freq_table,
 298                                             freqmax - 1, CPUFREQ_RELATION_H,
 299                                             &index))
 300                                         break;
 301                                 freq = pcpu->freq_table[index].frequency;
 302
 303                                 if (freq == freqmin) {
 304                                         /*
 305                                          * The first frequency below freqmax
 306                                          * has already been found to be too
 307                                          * low.  freqmax is the lowest speed
 308                                          * we found that is fast enough.
 309                                          */
 310                                         freq = freqmax;
 311                                         break;
 312                                 }
 313                         }
 314                 } else if (freq < prevfreq) {
 315                         /* The previous frequency is high enough. */
 316                         freqmax = prevfreq;
 317
 318                         if (freq <= freqmin) {
 319                                 /*
 320                                  * Find the lowest frequency that is higher
 321                                  * than freqmin.
 322                                  */
 323                                 if (cpufreq_frequency_table_target(
 324                                             pcpu->policy, pcpu->freq_table,
 325                                             freqmin + 1, CPUFREQ_RELATION_L,
 326                                             &index))
 327                                         break;
 328                                 freq = pcpu->freq_table[index].frequency;
 329
 330                                 /*
 331                                  * If freqmax is the first frequency above
 332                                  * freqmin then we have already found that
 333                                  * this speed is fast enough.
 334                                  */
 335                                 if (freq == freqmax)
 336                                         break;
 337                         }
 338                 }
 339
 340                 /* If same frequency chosen as previous then done. */
 341         } while (freq != prevfreq);
 342
 343         return freq;
 344 }
 345
 346 static u64 update_load(int cpu)
 347 {
 348         struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
 349         struct cpufreq_interactive_tunables *tunables =
 350                 pcpu->policy->governor_data;
 351         u64 now;
 352         u64 now_idle;
 353         unsigned int delta_idle;
 354         unsigned int delta_time;
 355         u64 active_time;
 356
 357         now_idle = get_cpu_idle_time(cpu, &now, tunables->io_is_busy);
 358         delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
 359         delta_time = (unsigned int)(now - pcpu->time_in_idle_timestamp);
 360
 361         if (delta_time <= delta_idle)
 362                 active_time = 0;
 363         else
 364                 active_time = delta_time - delta_idle;
 365
 366         pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
 367
 368         pcpu->time_in_idle = now_idle;
 369         pcpu->time_in_idle_timestamp = now;
 370         return now;
 371 }
 372
 373 static void cpufreq_interactive_timer(unsigned long data)
 374 {
 375         u64 now;
 376         unsigned int delta_time;
 377         u64 cputime_speedadj;
 378         int cpu_load;
 379         struct cpufreq_interactive_cpuinfo *pcpu =
 380                 &per_cpu(cpuinfo, data);
 381         struct cpufreq_interactive_tunables *tunables =
 382                 pcpu->policy->governor_data;
 383         unsigned int new_freq;
 384         unsigned int loadadjfreq;
 385         unsigned int index;
 386         unsigned long flags;
 387         bool boosted;
 388
 389         if (!down_read_trylock(&pcpu->enable_sem))
 390                 return;
 391         if (!pcpu->governor_enabled)
 392                 goto exit;
 393
 394         spin_lock_irqsave(&pcpu->load_lock, flags);
 395         now = update_load(data);
 396         delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
 397         cputime_speedadj = pcpu->cputime_speedadj;
 398         spin_unlock_irqrestore(&pcpu->load_lock, flags);
 399
 400         if (WARN_ON_ONCE(!delta_time))
 401                 goto rearm;
 402
 403         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
 404         do_div(cputime_speedadj, delta_time);
 405         loadadjfreq = (unsigned int)cputime_speedadj * 100;
 406         cpu_load = loadadjfreq / pcpu->target_freq;
 407         boosted = tunables->boost_val || now < tunables->boostpulse_endtime;
 408
 409         if (cpu_load >= tunables->go_hispeed_load || boosted) {
 410                 if (pcpu->target_freq < tunables->hispeed_freq) {
 411                         new_freq = tunables->hispeed_freq;
 412                 } else {
 413                         new_freq = choose_freq(pcpu, loadadjfreq);
 414
 415                         if (new_freq < tunables->hispeed_freq)
 416                                 new_freq = tunables->hispeed_freq;
 417                 }
 418         } else {
 419                 new_freq = choose_freq(pcpu, loadadjfreq);
 420                 if (new_freq > tunables->hispeed_freq &&
 421                                 pcpu->target_freq < tunables->hispeed_freq)
 422                         new_freq = tunables->hispeed_freq;
 423         }
 424
 425         if (pcpu->target_freq >= tunables->hispeed_freq &&
 426             new_freq > pcpu->target_freq &&
 427             now - pcpu->hispeed_validate_time <
 428             freq_to_above_hispeed_delay(tunables, pcpu->target_freq)) {
 429                 trace_cpufreq_interactive_notyet(
 430                         data, cpu_load, pcpu->target_freq,
 431                         pcpu->policy->cur, new_freq);
 432                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 433                 goto rearm;
 434         }
 435
 436         pcpu->hispeed_validate_time = now;
 437
 438         if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
 439                                            new_freq, CPUFREQ_RELATION_L,
 440                                            &index)) {
 441                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 442                 goto rearm;
 443         }
 444
 445         new_freq = pcpu->freq_table[index].frequency;
 446
 447         /*
 448          * Do not scale below floor_freq unless we have been at or above the
 449          * floor frequency for the minimum sample time since last validated.
 450          */
 451         if (new_freq < pcpu->floor_freq) {
 452                 if (now - pcpu->floor_validate_time <
 453                                 tunables->min_sample_time) {
 454                         trace_cpufreq_interactive_notyet(
 455                                 data, cpu_load, pcpu->target_freq,
 456                                 pcpu->policy->cur, new_freq);
 457                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 458                         goto rearm;
 459                 }
 460         }
 461
 462         /*
 463          * Update the timestamp for checking whether speed has been held at
 464          * or above the selected frequency for a minimum of min_sample_time,
 465          * if not boosted to hispeed_freq.  If boosted to hispeed_freq then we
 466          * allow the speed to drop as soon as the boostpulse duration expires
 467          * (or the indefinite boost is turned off).
 468          */
 469
 470         if (!boosted || new_freq > tunables->hispeed_freq) {
 471                 pcpu->floor_freq = new_freq;
 472                 pcpu->floor_validate_time = now;
 473         }
 474
 475         if (pcpu->target_freq == new_freq) {
 476                 trace_cpufreq_interactive_already(
 477                         data, cpu_load, pcpu->target_freq,
 478                         pcpu->policy->cur, new_freq);
 479                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 480                 goto rearm_if_notmax;
 481         }
 482
 483         trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
 484                                          pcpu->policy->cur, new_freq);
 485
 486         pcpu->target_freq = new_freq;
 487         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 488         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 489         cpumask_set_cpu(data, &speedchange_cpumask);
 490         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 491         wake_up_process(speedchange_task);
 492
 493 rearm_if_notmax:
 494         /*
 495          * Already set max speed and don't see a need to change that,
 496          * wait until next idle to re-evaluate, don't need timer.
 497          */
 498         if (pcpu->target_freq == pcpu->policy->max)
 499                 goto exit;
 500
 501 rearm:
 502         if (!timer_pending(&pcpu->cpu_timer))
 503                 cpufreq_interactive_timer_resched(pcpu);
 504
 505 exit:
 506         up_read(&pcpu->enable_sem);
 507         return;
 508 }
 509
 510 static void cpufreq_interactive_idle_start(void)
 511 {
 512         struct cpufreq_interactive_cpuinfo *pcpu =
 513                 &per_cpu(cpuinfo, smp_processor_id());
 514         int pending;
 515
 516         if (!down_read_trylock(&pcpu->enable_sem))
 517                 return;
 518         if (!pcpu->governor_enabled) {
 519                 up_read(&pcpu->enable_sem);
 520                 return;
 521         }
 522
 523         pending = timer_pending(&pcpu->cpu_timer);
 524
 525         if (pcpu->target_freq != pcpu->policy->min) {
 526                 /*
 527                  * Entering idle while not at lowest speed.  On some
 528                  * platforms this can hold the other CPU(s) at that speed
 529                  * even though the CPU is idle. Set a timer to re-evaluate
 530                  * speed so this idle CPU doesn't hold the other CPUs above
 531                  * min indefinitely.  This should probably be a quirk of
 532                  * the CPUFreq driver.
 533                  */
 534                 if (!pending)
 535                         cpufreq_interactive_timer_resched(pcpu);
 536         }
 537
 538         up_read(&pcpu->enable_sem);
 539 }
 540
 541 static void cpufreq_interactive_idle_end(void)
 542 {
 543         struct cpufreq_interactive_cpuinfo *pcpu =
 544                 &per_cpu(cpuinfo, smp_processor_id());
 545
 546         if (!down_read_trylock(&pcpu->enable_sem))
 547                 return;
 548         if (!pcpu->governor_enabled) {
 549                 up_read(&pcpu->enable_sem);
 550                 return;
 551         }
 552
 553         /* Arm the timer for 1-2 ticks later if not already. */
 554         if (!timer_pending(&pcpu->cpu_timer)) {
 555                 cpufreq_interactive_timer_resched(pcpu);
 556         } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
 557                 del_timer(&pcpu->cpu_timer);
 558                 del_timer(&pcpu->cpu_slack_timer);
 559                 cpufreq_interactive_timer(smp_processor_id());
 560         }
 561
 562         up_read(&pcpu->enable_sem);
 563 }
 564
 565 static int cpufreq_interactive_speedchange_task(void *data)
 566 {
 567         unsigned int cpu;
 568         cpumask_t tmp_mask;
 569         unsigned long flags;
 570         struct cpufreq_interactive_cpuinfo *pcpu;
 571
 572         while (1) {
 573                 set_current_state(TASK_INTERRUPTIBLE);
 574                 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 575
 576                 if (cpumask_empty(&speedchange_cpumask)) {
 577                         spin_unlock_irqrestore(&speedchange_cpumask_lock,
 578                                                flags);
 579                         schedule();
 580
 581                         if (kthread_should_stop())
 582                                 break;
 583
 584                         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 585                 }
 586
 587                 set_current_state(TASK_RUNNING);
 588                 tmp_mask = speedchange_cpumask;
 589                 cpumask_clear(&speedchange_cpumask);
 590                 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 591
 592                 for_each_cpu(cpu, &tmp_mask) {
 593                         unsigned int j;
 594                         unsigned int max_freq = 0;
 595
 596                         pcpu = &per_cpu(cpuinfo, cpu);
 597                         if (!down_read_trylock(&pcpu->enable_sem))
 598                                 continue;
 599                         if (!pcpu->governor_enabled) {
 600                                 up_read(&pcpu->enable_sem);
 601                                 continue;
 602                         }
 603
 604                         for_each_cpu(j, pcpu->policy->cpus) {
 605                                 struct cpufreq_interactive_cpuinfo *pjcpu =
 606                                         &per_cpu(cpuinfo, j);
 607
 608                                 if (pjcpu->target_freq > max_freq)
 609                                         max_freq = pjcpu->target_freq;
 610                         }
 611
 612                         if (max_freq != pcpu->policy->cur)
 613                                 __cpufreq_driver_target(pcpu->policy,
 614                                                         max_freq,
 615                                                         CPUFREQ_RELATION_H);
 616                         trace_cpufreq_interactive_setspeed(cpu,
 617                                                      pcpu->target_freq,
 618                                                      pcpu->policy->cur);
 619
 620                         up_read(&pcpu->enable_sem);
 621                 }
 622         }
 623
 624         return 0;
 625 }
 626
 627 static void cpufreq_interactive_boost(void)
 628 {
 629         int i;
 630         int anyboost = 0;
 631         unsigned long flags[2];
 632         struct cpufreq_interactive_cpuinfo *pcpu;
 633         struct cpufreq_interactive_tunables *tunables;
 634
 635         spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
 636
 637         for_each_online_cpu(i) {
 638                 pcpu = &per_cpu(cpuinfo, i);
 639                 tunables = pcpu->policy->governor_data;
 640
 641                 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
 642                 if (pcpu->target_freq < tunables->hispeed_freq) {
 643                         pcpu->target_freq = tunables->hispeed_freq;
 644                         cpumask_set_cpu(i, &speedchange_cpumask);
 645                         pcpu->hispeed_validate_time =
 646                                 ktime_to_us(ktime_get());
 647                         anyboost = 1;
 648                 }
 649
 650                 /*
 651                  * Set floor freq and (re)start timer for when last
 652                  * validated.
 653                  */
 654
 655                 pcpu->floor_freq = tunables->hispeed_freq;
 656                 pcpu->floor_validate_time = ktime_to_us(ktime_get());
 657                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
 658         }
 659
 660         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
 661
 662         if (anyboost)
 663                 wake_up_process(speedchange_task);
 664 }
 665
 666 static int cpufreq_interactive_notifier(
 667         struct notifier_block *nb, unsigned long val, void *data)
 668 {
 669         struct cpufreq_freqs *freq = data;
 670         struct cpufreq_interactive_cpuinfo *pcpu;
 671         int cpu;
 672         unsigned long flags;
 673
 674         if (val == CPUFREQ_POSTCHANGE) {
 675                 pcpu = &per_cpu(cpuinfo, freq->cpu);
 676                 if (!down_read_trylock(&pcpu->enable_sem))
 677                         return 0;
 678                 if (!pcpu->governor_enabled) {
 679                         up_read(&pcpu->enable_sem);
 680                         return 0;
 681                 }
 682
 683                 for_each_cpu(cpu, pcpu->policy->cpus) {
 684                         struct cpufreq_interactive_cpuinfo *pjcpu =
 685                                 &per_cpu(cpuinfo, cpu);
 686                         if (cpu != freq->cpu) {
 687                                 if (!down_read_trylock(&pjcpu->enable_sem))
 688                                         continue;
 689                                 if (!pjcpu->governor_enabled) {
 690                                         up_read(&pjcpu->enable_sem);
 691                                         continue;
 692                                 }
 693                         }
 694                         spin_lock_irqsave(&pjcpu->load_lock, flags);
 695                         update_load(cpu);
 696                         spin_unlock_irqrestore(&pjcpu->load_lock, flags);
 697                         if (cpu != freq->cpu)
 698                                 up_read(&pjcpu->enable_sem);
 699                 }
 700
 701                 up_read(&pcpu->enable_sem);
 702         }
 703         return 0;
 704 }
 705
 706 static struct notifier_block cpufreq_notifier_block = {
 707         .notifier_call = cpufreq_interactive_notifier,
 708 };
 709
 710 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
 711 {
 712         const char *cp;
 713         int i;
 714         int ntokens = 1;
 715         unsigned int *tokenized_data;
 716         int err = -EINVAL;
 717
 718         cp = buf;
 719         while ((cp = strpbrk(cp + 1, " :")))
 720                 ntokens++;
 721
 722         if (!(ntokens & 0x1))
 723                 goto err;
 724
 725         tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
 726         if (!tokenized_data) {
 727                 err = -ENOMEM;
 728                 goto err;
 729         }
 730
 731         cp = buf;
 732         i = 0;
 733         while (i < ntokens) {
 734                 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
 735                         goto err_kfree;
 736
 737                 cp = strpbrk(cp, " :");
 738                 if (!cp)
 739                         break;
 740                 cp++;
 741         }
 742
 743         if (i != ntokens)
 744                 goto err_kfree;
 745
 746         *num_tokens = ntokens;
 747         return tokenized_data;
 748
 749 err_kfree:
 750         kfree(tokenized_data);
 751 err:
 752         return ERR_PTR(err);
 753 }
 754
 755 static ssize_t show_target_loads(
 756         struct cpufreq_interactive_tunables *tunables,
 757         char *buf)
 758 {
 759         int i;
 760         ssize_t ret = 0;
 761         unsigned long flags;
 762
 763         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 764
 765         for (i = 0; i < tunables->ntarget_loads; i++)
 766                 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
 767                                i & 0x1 ? ":" : " ");
 768
 769         sprintf(buf + ret - 1, "\n");
 770         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 771         return ret;
 772 }
 773
 774 static ssize_t store_target_loads(
 775         struct cpufreq_interactive_tunables *tunables,
 776         const char *buf, size_t count)
 777 {
 778         int ntokens;
 779         unsigned int *new_target_loads = NULL;
 780         unsigned long flags;
 781
 782         new_target_loads = get_tokenized_data(buf, &ntokens);
 783         if (IS_ERR(new_target_loads))
 784                 return PTR_RET(new_target_loads);
 785
 786         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 787         if (tunables->target_loads != default_target_loads)
 788                 kfree(tunables->target_loads);
 789         tunables->target_loads = new_target_loads;
 790         tunables->ntarget_loads = ntokens;
 791         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 792         return count;
 793 }
 794
 795 static ssize_t show_above_hispeed_delay(
 796         struct cpufreq_interactive_tunables *tunables, char *buf)
 797 {
 798         int i;
 799         ssize_t ret = 0;
 800         unsigned long flags;
 801
 802         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 803
 804         for (i = 0; i < tunables->nabove_hispeed_delay; i++)
 805                 ret += sprintf(buf + ret, "%u%s",
 806                                tunables->above_hispeed_delay[i],
 807                                i & 0x1 ? ":" : " ");
 808
 809         sprintf(buf + ret - 1, "\n");
 810         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 811         return ret;
 812 }
 813
 814 static ssize_t store_above_hispeed_delay(
 815         struct cpufreq_interactive_tunables *tunables,
 816         const char *buf, size_t count)
 817 {
 818         int ntokens;
 819         unsigned int *new_above_hispeed_delay = NULL;
 820         unsigned long flags;
 821
 822         new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
 823         if (IS_ERR(new_above_hispeed_delay))
 824                 return PTR_RET(new_above_hispeed_delay);
 825
 826         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 827         if (tunables->above_hispeed_delay != default_above_hispeed_delay)
 828                 kfree(tunables->above_hispeed_delay);
 829         tunables->above_hispeed_delay = new_above_hispeed_delay;
 830         tunables->nabove_hispeed_delay = ntokens;
 831         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 832         return count;
 833
 834 }
 835
 836 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 837                 char *buf)
 838 {
 839         return sprintf(buf, "%u\n", tunables->hispeed_freq);
 840 }
 841
 842 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 843                 const char *buf, size_t count)
 844 {
 845         int ret;
 846         long unsigned int val;
 847
 848         ret = strict_strtoul(buf, 0, &val);
 849         if (ret < 0)
 850                 return ret;
 851         tunables->hispeed_freq = val;
 852         return count;
 853 }
 854
 855 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
 856                 *tunables, char *buf)
 857 {
 858         return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
 859 }
 860
 861 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
 862                 *tunables, const char *buf, size_t count)
 863 {
 864         int ret;
 865         unsigned long val;
 866
 867         ret = strict_strtoul(buf, 0, &val);
 868         if (ret < 0)
 869                 return ret;
 870         tunables->go_hispeed_load = val;
 871         return count;
 872 }
 873
 874 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
 875                 *tunables, char *buf)
 876 {
 877         return sprintf(buf, "%lu\n", tunables->min_sample_time);
 878 }
 879
 880 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
 881                 *tunables, const char *buf, size_t count)
 882 {
 883         int ret;
 884         unsigned long val;
 885
 886         ret = strict_strtoul(buf, 0, &val);
 887         if (ret < 0)
 888                 return ret;
 889         tunables->min_sample_time = val;
 890         return count;
 891 }
 892
 893 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
 894                 char *buf)
 895 {
 896         return sprintf(buf, "%lu\n", tunables->timer_rate);
 897 }
 898
 899 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
 900                 const char *buf, size_t count)
 901 {
 902         int ret;
 903         unsigned long val;
 904
 905         ret = strict_strtoul(buf, 0, &val);
 906         if (ret < 0)
 907                 return ret;
 908         tunables->timer_rate = val;
 909         return count;
 910 }
 911
 912 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
 913                 char *buf)
 914 {
 915         return sprintf(buf, "%d\n", tunables->timer_slack_val);
 916 }
 917
 918 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
 919                 const char *buf, size_t count)
 920 {
 921         int ret;
 922         unsigned long val;
 923
 924         ret = kstrtol(buf, 10, &val);
 925         if (ret < 0)
 926                 return ret;
 927
 928         tunables->timer_slack_val = val;
 929         return count;
 930 }
 931
 932 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
 933                           char *buf)
 934 {
 935         return sprintf(buf, "%d\n", tunables->boost_val);
 936 }
 937
 938 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
 939                            const char *buf, size_t count)
 940 {
 941         int ret;
 942         unsigned long val;
 943
 944         ret = kstrtoul(buf, 0, &val);
 945         if (ret < 0)
 946                 return ret;
 947
 948         tunables->boost_val = val;
 949
 950         if (tunables->boost_val) {
 951                 trace_cpufreq_interactive_boost("on");
 952                 cpufreq_interactive_boost();
 953         } else {
 954                 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
 955                 trace_cpufreq_interactive_unboost("off");
 956         }
 957
 958         return count;
 959 }
 960
 961 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
 962                                 const char *buf, size_t count)
 963 {
 964         int ret;
 965         unsigned long val;
 966
 967         ret = kstrtoul(buf, 0, &val);
 968         if (ret < 0)
 969                 return ret;
 970
 971         tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
 972                 tunables->boostpulse_duration_val;
 973         trace_cpufreq_interactive_boost("pulse");
 974         cpufreq_interactive_boost();
 975         return count;
 976 }
 977
 978 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
 979                 *tunables, char *buf)
 980 {
 981         return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
 982 }
 983
 984 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
 985                 *tunables, const char *buf, size_t count)
 986 {
 987         int ret;
 988         unsigned long val;
 989
 990         ret = kstrtoul(buf, 0, &val);
 991         if (ret < 0)
 992                 return ret;
 993
 994         tunables->boostpulse_duration_val = val;
 995         return count;
 996 }
 997
 998 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
 999                 char *buf)
1000 {
1001         return sprintf(buf, "%u\n", tunables->io_is_busy);
1002 }
1003
1004 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
1005                 const char *buf, size_t count)
1006 {
1007         int ret;
1008         unsigned long val;
1009
1010         ret = kstrtoul(buf, 0, &val);
1011         if (ret < 0)
1012                 return ret;
1013         tunables->io_is_busy = val;
1014         return count;
1015 }
1016
1017 /*
1018  * Create show/store routines
1019  * - sys: One governor instance for complete SYSTEM
1020  * - pol: One governor instance per struct cpufreq_policy
1021  */
1022 #define show_gov_pol_sys(file_name)                                     \
1023 static ssize_t show_##file_name##_gov_sys                               \
1024 (struct kobject *kobj, struct attribute *attr, char *buf)               \
1025 {                                                                       \
1026         return show_##file_name(common_tunables, buf);                  \
1027 }                                                                       \
1028                                                                         \
1029 static ssize_t show_##file_name##_gov_pol                               \
1030 (struct cpufreq_policy *policy, char *buf)                              \
1031 {                                                                       \
1032         return show_##file_name(policy->governor_data, buf);            \
1033 }
1034
1035 #define store_gov_pol_sys(file_name)                                    \
1036 static ssize_t store_##file_name##_gov_sys                              \
1037 (struct kobject *kobj, struct attribute *attr, const char *buf,         \
1038         size_t count)                                                   \
1039 {                                                                       \
1040         return store_##file_name(common_tunables, buf, count);          \
1041 }                                                                       \
1042                                                                         \
1043 static ssize_t store_##file_name##_gov_pol                              \
1044 (struct cpufreq_policy *policy, const char *buf, size_t count)          \
1045 {                                                                       \
1046         return store_##file_name(policy->governor_data, buf, count);    \
1047 }
1048
1049 #define show_store_gov_pol_sys(file_name)                               \
1050 show_gov_pol_sys(file_name);                                            \
1051 store_gov_pol_sys(file_name)
1052
1053 show_store_gov_pol_sys(target_loads);
1054 show_store_gov_pol_sys(above_hispeed_delay);
1055 show_store_gov_pol_sys(hispeed_freq);
1056 show_store_gov_pol_sys(go_hispeed_load);
1057 show_store_gov_pol_sys(min_sample_time);
1058 show_store_gov_pol_sys(timer_rate);
1059 show_store_gov_pol_sys(timer_slack);
1060 show_store_gov_pol_sys(boost);
1061 store_gov_pol_sys(boostpulse);
1062 show_store_gov_pol_sys(boostpulse_duration);
1063 show_store_gov_pol_sys(io_is_busy);
1064
1065 #define gov_sys_attr_rw(_name)                                          \
1066 static struct global_attr _name##_gov_sys =                             \
1067 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1068
1069 #define gov_pol_attr_rw(_name)                                          \
1070 static struct freq_attr _name##_gov_pol =                               \
1071 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1072
1073 #define gov_sys_pol_attr_rw(_name)                                      \
1074         gov_sys_attr_rw(_name);                                         \
1075         gov_pol_attr_rw(_name)
1076
1077 gov_sys_pol_attr_rw(target_loads);
1078 gov_sys_pol_attr_rw(above_hispeed_delay);
1079 gov_sys_pol_attr_rw(hispeed_freq);
1080 gov_sys_pol_attr_rw(go_hispeed_load);
1081 gov_sys_pol_attr_rw(min_sample_time);
1082 gov_sys_pol_attr_rw(timer_rate);
1083 gov_sys_pol_attr_rw(timer_slack);
1084 gov_sys_pol_attr_rw(boost);
1085 gov_sys_pol_attr_rw(boostpulse_duration);
1086 gov_sys_pol_attr_rw(io_is_busy);
1087
1088 static struct global_attr boostpulse_gov_sys =
1089         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1090
1091 static struct freq_attr boostpulse_gov_pol =
1092         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1093
1094 /* One Governor instance for entire system */
1095 static struct attribute *interactive_attributes_gov_sys[] = {
1096         &target_loads_gov_sys.attr,
1097         &above_hispeed_delay_gov_sys.attr,
1098         &hispeed_freq_gov_sys.attr,
1099         &go_hispeed_load_gov_sys.attr,
1100         &min_sample_time_gov_sys.attr,
1101         &timer_rate_gov_sys.attr,
1102         &timer_slack_gov_sys.attr,
1103         &boost_gov_sys.attr,
1104         &boostpulse_gov_sys.attr,
1105         &boostpulse_duration_gov_sys.attr,
1106         &io_is_busy_gov_sys.attr,
1107         NULL,
1108 };
1109
1110 static struct attribute_group interactive_attr_group_gov_sys = {
1111         .attrs = interactive_attributes_gov_sys,
1112         .name = "interactive",
1113 };
1114
1115 /* Per policy governor instance */
1116 static struct attribute *interactive_attributes_gov_pol[] = {
1117         &target_loads_gov_pol.attr,
1118         &above_hispeed_delay_gov_pol.attr,
1119         &hispeed_freq_gov_pol.attr,
1120         &go_hispeed_load_gov_pol.attr,
1121         &min_sample_time_gov_pol.attr,
1122         &timer_rate_gov_pol.attr,
1123         &timer_slack_gov_pol.attr,
1124         &boost_gov_pol.attr,
1125         &boostpulse_gov_pol.attr,
1126         &boostpulse_duration_gov_pol.attr,
1127         &io_is_busy_gov_pol.attr,
1128         NULL,
1129 };
1130
1131 static struct attribute_group interactive_attr_group_gov_pol = {
1132         .attrs = interactive_attributes_gov_pol,
1133         .name = "interactive",
1134 };
1135
1136 static struct attribute_group *get_sysfs_attr(void)
1137 {
1138         if (have_governor_per_policy())
1139                 return &interactive_attr_group_gov_pol;
1140         else
1141                 return &interactive_attr_group_gov_sys;
1142 }
1143
1144 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1145                                              unsigned long val,
1146                                              void *data)
1147 {
1148         switch (val) {
1149         case IDLE_START:
1150                 cpufreq_interactive_idle_start();
1151                 break;
1152         case IDLE_END:
1153                 cpufreq_interactive_idle_end();
1154                 break;
1155         }
1156
1157         return 0;
1158 }
1159
1160 static struct notifier_block cpufreq_interactive_idle_nb = {
1161         .notifier_call = cpufreq_interactive_idle_notifier,
1162 };
1163
1164 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1165                 unsigned int event)
1166 {
1167         int rc;
1168         unsigned int j;
1169         struct cpufreq_interactive_cpuinfo *pcpu;
1170         struct cpufreq_frequency_table *freq_table;
1171         struct cpufreq_interactive_tunables *tunables;
1172         unsigned long flags;
1173
1174         if (have_governor_per_policy())
1175                 tunables = policy->governor_data;
1176         else
1177                 tunables = common_tunables;
1178
1179         WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1180
1181         switch (event) {
1182         case CPUFREQ_GOV_POLICY_INIT:
1183                 if (have_governor_per_policy()) {
1184                         WARN_ON(tunables);
1185                 } else if (tunables) {
1186                         tunables->usage_count++;
1187                         policy->governor_data = tunables;
1188                         return 0;
1189                 }
1190
1191                 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1192                 if (!tunables) {
1193                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1194                         return -ENOMEM;
1195                 }
1196
1197                 tunables->usage_count = 1;
1198                 tunables->above_hispeed_delay = default_above_hispeed_delay;
1199                 tunables->nabove_hispeed_delay =
1200                         ARRAY_SIZE(default_above_hispeed_delay);
1201                 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1202                 tunables->target_loads = default_target_loads;
1203                 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1204                 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1205                 tunables->timer_rate = DEFAULT_TIMER_RATE;
1206                 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1207                 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1208
1209                 spin_lock_init(&tunables->target_loads_lock);
1210                 spin_lock_init(&tunables->above_hispeed_delay_lock);
1211
1212                 policy->governor_data = tunables;
1213                 if (!have_governor_per_policy())
1214                         common_tunables = tunables;
1215
1216                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1217                                 get_sysfs_attr());
1218                 if (rc) {
1219                         kfree(tunables);
1220                         policy->governor_data = NULL;
1221                         if (!have_governor_per_policy())
1222                                 common_tunables = NULL;
1223                         return rc;
1224                 }
1225
1226                 if (!policy->governor->initialized) {
1227                         idle_notifier_register(&cpufreq_interactive_idle_nb);
1228                         cpufreq_register_notifier(&cpufreq_notifier_block,
1229                                         CPUFREQ_TRANSITION_NOTIFIER);
1230                 }
1231
1232                 break;
1233
1234         case CPUFREQ_GOV_POLICY_EXIT:
1235                 if (!--tunables->usage_count) {
1236                         if (policy->governor->initialized == 1) {
1237                                 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1238                                                 CPUFREQ_TRANSITION_NOTIFIER);
1239                                 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1240                         }
1241
1242                         sysfs_remove_group(get_governor_parent_kobj(policy),
1243                                         get_sysfs_attr());
1244                         kfree(tunables);
1245                         common_tunables = NULL;
1246                 }
1247
1248                 policy->governor_data = NULL;
1249                 break;
1250
1251         case CPUFREQ_GOV_START:
1252                 mutex_lock(&gov_lock);
1253
1254                 freq_table = cpufreq_frequency_get_table(policy->cpu);
1255                 if (!tunables->hispeed_freq)
1256                         tunables->hispeed_freq = policy->max;
1257
1258                 for_each_cpu(j, policy->cpus) {
1259                         pcpu = &per_cpu(cpuinfo, j);
1260                         pcpu->policy = policy;
1261                         pcpu->target_freq = policy->cur;
1262                         pcpu->freq_table = freq_table;
1263                         pcpu->floor_freq = pcpu->target_freq;
1264                         pcpu->floor_validate_time =
1265                                 ktime_to_us(ktime_get());
1266                         pcpu->hispeed_validate_time =
1267                                 pcpu->floor_validate_time;
1268                         pcpu->max_freq = policy->max;
1269                         down_write(&pcpu->enable_sem);
1270                         del_timer_sync(&pcpu->cpu_timer);
1271                         del_timer_sync(&pcpu->cpu_slack_timer);
1272                         cpufreq_interactive_timer_start(tunables, j);
1273                         pcpu->governor_enabled = 1;
1274                         up_write(&pcpu->enable_sem);
1275                 }
1276
1277                 mutex_unlock(&gov_lock);
1278                 break;
1279
1280         case CPUFREQ_GOV_STOP:
1281                 mutex_lock(&gov_lock);
1282                 for_each_cpu(j, policy->cpus) {
1283                         pcpu = &per_cpu(cpuinfo, j);
1284                         down_write(&pcpu->enable_sem);
1285                         pcpu->governor_enabled = 0;
1286                         del_timer_sync(&pcpu->cpu_timer);
1287                         del_timer_sync(&pcpu->cpu_slack_timer);
1288                         up_write(&pcpu->enable_sem);
1289                 }
1290
1291                 mutex_unlock(&gov_lock);
1292                 break;
1293
1294         case CPUFREQ_GOV_LIMITS:
1295                 if (policy->max < policy->cur)
1296                         __cpufreq_driver_target(policy,
1297                                         policy->max, CPUFREQ_RELATION_H);
1298                 else if (policy->min > policy->cur)
1299                         __cpufreq_driver_target(policy,
1300                                         policy->min, CPUFREQ_RELATION_L);
1301                 for_each_cpu(j, policy->cpus) {
1302                         pcpu = &per_cpu(cpuinfo, j);
1303
1304                         down_read(&pcpu->enable_sem);
1305                         if (pcpu->governor_enabled == 0) {
1306                                 up_read(&pcpu->enable_sem);
1307                                 continue;
1308                         }
1309
1310                         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1311                         if (policy->max < pcpu->target_freq)
1312                                 pcpu->target_freq = policy->max;
1313                         else if (policy->min > pcpu->target_freq)
1314                                 pcpu->target_freq = policy->min;
1315
1316                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1317                         up_read(&pcpu->enable_sem);
1318
1319                         /* Reschedule timer only if policy->max is raised.
1320                          * Delete the timers, else the timer callback may
1321                          * return without re-arm the timer when failed
1322                          * acquire the semaphore. This race may cause timer
1323                          * stopped unexpectedly.
1324                          */
1325
1326                         if (policy->max > pcpu->max_freq) {
1327                                 down_write(&pcpu->enable_sem);
1328                                 del_timer_sync(&pcpu->cpu_timer);
1329                                 del_timer_sync(&pcpu->cpu_slack_timer);
1330                                 cpufreq_interactive_timer_start(tunables, j);
1331                                 up_write(&pcpu->enable_sem);
1332                         }
1333
1334                         pcpu->max_freq = policy->max;
1335                 }
1336                 break;
1337         }
1338         return 0;
1339 }
1340
1341 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1342 static
1343 #endif
1344 struct cpufreq_governor cpufreq_gov_interactive = {
1345         .name = "interactive",
1346         .governor = cpufreq_governor_interactive,
1347         .max_transition_latency = 10000000,
1348         .owner = THIS_MODULE,
1349 };
1350
1351 static void cpufreq_interactive_nop_timer(unsigned long data)
1352 {
1353 }
1354
1355 static int __init cpufreq_interactive_init(void)
1356 {
1357         unsigned int i;
1358         struct cpufreq_interactive_cpuinfo *pcpu;
1359         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1360
1361         /* Initalize per-cpu timers */
1362         for_each_possible_cpu(i) {
1363                 pcpu = &per_cpu(cpuinfo, i);
1364                 init_timer_deferrable(&pcpu->cpu_timer);
1365                 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1366                 pcpu->cpu_timer.data = i;
1367                 init_timer(&pcpu->cpu_slack_timer);
1368                 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1369                 spin_lock_init(&pcpu->load_lock);
1370                 spin_lock_init(&pcpu->target_freq_lock);
1371                 init_rwsem(&pcpu->enable_sem);
1372         }
1373
1374         spin_lock_init(&speedchange_cpumask_lock);
1375         mutex_init(&gov_lock);
1376         speedchange_task =
1377                 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1378                                "cfinteractive");
1379         if (IS_ERR(speedchange_task))
1380                 return PTR_ERR(speedchange_task);
1381
1382         sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, &param);
1383         get_task_struct(speedchange_task);
1384
1385         /* NB: wake up so the thread does not look hung to the freezer */
1386         wake_up_process(speedchange_task);
1387
1388         return cpufreq_register_governor(&cpufreq_gov_interactive);
1389 }
1390
1391 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1392 fs_initcall(cpufreq_interactive_init);
1393 #else
1394 module_init(cpufreq_interactive_init);
1395 #endif
1396
1397 static void __exit cpufreq_interactive_exit(void)
1398 {
1399         cpufreq_unregister_governor(&cpufreq_gov_interactive);
1400         kthread_stop(speedchange_task);
1401         put_task_struct(speedchange_task);
1402 }
1403
1404 module_exit(cpufreq_interactive_exit);
1405
1406 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1407 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1408         "Latency sensitive workloads");
1409 MODULE_LICENSE("GPL");