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 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         }
 421
 422         if (pcpu->target_freq >= tunables->hispeed_freq &&
 423             new_freq > pcpu->target_freq &&
 424             now - pcpu->hispeed_validate_time <
 425             freq_to_above_hispeed_delay(tunables, pcpu->target_freq)) {
 426                 trace_cpufreq_interactive_notyet(
 427                         data, cpu_load, pcpu->target_freq,
 428                         pcpu->policy->cur, new_freq);
 429                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 430                 goto rearm;
 431         }
 432
 433         pcpu->hispeed_validate_time = now;
 434
 435         if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
 436                                            new_freq, CPUFREQ_RELATION_L,
 437                                            &index)) {
 438                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 439                 goto rearm;
 440         }
 441
 442         new_freq = pcpu->freq_table[index].frequency;
 443
 444         /*
 445          * Do not scale below floor_freq unless we have been at or above the
 446          * floor frequency for the minimum sample time since last validated.
 447          */
 448         if (new_freq < pcpu->floor_freq) {
 449                 if (now - pcpu->floor_validate_time <
 450                                 tunables->min_sample_time) {
 451                         trace_cpufreq_interactive_notyet(
 452                                 data, cpu_load, pcpu->target_freq,
 453                                 pcpu->policy->cur, new_freq);
 454                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 455                         goto rearm;
 456                 }
 457         }
 458
 459         /*
 460          * Update the timestamp for checking whether speed has been held at
 461          * or above the selected frequency for a minimum of min_sample_time,
 462          * if not boosted to hispeed_freq.  If boosted to hispeed_freq then we
 463          * allow the speed to drop as soon as the boostpulse duration expires
 464          * (or the indefinite boost is turned off).
 465          */
 466
 467         if (!boosted || new_freq > tunables->hispeed_freq) {
 468                 pcpu->floor_freq = new_freq;
 469                 pcpu->floor_validate_time = now;
 470         }
 471
 472         if (pcpu->target_freq == new_freq) {
 473                 trace_cpufreq_interactive_already(
 474                         data, cpu_load, pcpu->target_freq,
 475                         pcpu->policy->cur, new_freq);
 476                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 477                 goto rearm_if_notmax;
 478         }
 479
 480         trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
 481                                          pcpu->policy->cur, new_freq);
 482
 483         pcpu->target_freq = new_freq;
 484         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
 485         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 486         cpumask_set_cpu(data, &speedchange_cpumask);
 487         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 488         wake_up_process(speedchange_task);
 489
 490 rearm_if_notmax:
 491         /*
 492          * Already set max speed and don't see a need to change that,
 493          * wait until next idle to re-evaluate, don't need timer.
 494          */
 495         if (pcpu->target_freq == pcpu->policy->max)
 496                 goto exit;
 497
 498 rearm:
 499         if (!timer_pending(&pcpu->cpu_timer))
 500                 cpufreq_interactive_timer_resched(pcpu);
 501
 502 exit:
 503         up_read(&pcpu->enable_sem);
 504         return;
 505 }
 506
 507 static void cpufreq_interactive_idle_start(void)
 508 {
 509         struct cpufreq_interactive_cpuinfo *pcpu =
 510                 &per_cpu(cpuinfo, smp_processor_id());
 511         int pending;
 512
 513         if (!down_read_trylock(&pcpu->enable_sem))
 514                 return;
 515         if (!pcpu->governor_enabled) {
 516                 up_read(&pcpu->enable_sem);
 517                 return;
 518         }
 519
 520         pending = timer_pending(&pcpu->cpu_timer);
 521
 522         if (pcpu->target_freq != pcpu->policy->min) {
 523                 /*
 524                  * Entering idle while not at lowest speed.  On some
 525                  * platforms this can hold the other CPU(s) at that speed
 526                  * even though the CPU is idle. Set a timer to re-evaluate
 527                  * speed so this idle CPU doesn't hold the other CPUs above
 528                  * min indefinitely.  This should probably be a quirk of
 529                  * the CPUFreq driver.
 530                  */
 531                 if (!pending)
 532                         cpufreq_interactive_timer_resched(pcpu);
 533         }
 534
 535         up_read(&pcpu->enable_sem);
 536 }
 537
 538 static void cpufreq_interactive_idle_end(void)
 539 {
 540         struct cpufreq_interactive_cpuinfo *pcpu =
 541                 &per_cpu(cpuinfo, smp_processor_id());
 542
 543         if (!down_read_trylock(&pcpu->enable_sem))
 544                 return;
 545         if (!pcpu->governor_enabled) {
 546                 up_read(&pcpu->enable_sem);
 547                 return;
 548         }
 549
 550         /* Arm the timer for 1-2 ticks later if not already. */
 551         if (!timer_pending(&pcpu->cpu_timer)) {
 552                 cpufreq_interactive_timer_resched(pcpu);
 553         } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
 554                 del_timer(&pcpu->cpu_timer);
 555                 del_timer(&pcpu->cpu_slack_timer);
 556                 cpufreq_interactive_timer(smp_processor_id());
 557         }
 558
 559         up_read(&pcpu->enable_sem);
 560 }
 561
 562 static int cpufreq_interactive_speedchange_task(void *data)
 563 {
 564         unsigned int cpu;
 565         cpumask_t tmp_mask;
 566         unsigned long flags;
 567         struct cpufreq_interactive_cpuinfo *pcpu;
 568
 569         while (1) {
 570                 set_current_state(TASK_INTERRUPTIBLE);
 571                 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 572
 573                 if (cpumask_empty(&speedchange_cpumask)) {
 574                         spin_unlock_irqrestore(&speedchange_cpumask_lock,
 575                                                flags);
 576                         schedule();
 577
 578                         if (kthread_should_stop())
 579                                 break;
 580
 581                         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 582                 }
 583
 584                 set_current_state(TASK_RUNNING);
 585                 tmp_mask = speedchange_cpumask;
 586                 cpumask_clear(&speedchange_cpumask);
 587                 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 588
 589                 for_each_cpu(cpu, &tmp_mask) {
 590                         unsigned int j;
 591                         unsigned int max_freq = 0;
 592
 593                         pcpu = &per_cpu(cpuinfo, cpu);
 594                         if (!down_read_trylock(&pcpu->enable_sem))
 595                                 continue;
 596                         if (!pcpu->governor_enabled) {
 597                                 up_read(&pcpu->enable_sem);
 598                                 continue;
 599                         }
 600
 601                         for_each_cpu(j, pcpu->policy->cpus) {
 602                                 struct cpufreq_interactive_cpuinfo *pjcpu =
 603                                         &per_cpu(cpuinfo, j);
 604
 605                                 if (pjcpu->target_freq > max_freq)
 606                                         max_freq = pjcpu->target_freq;
 607                         }
 608
 609                         if (max_freq != pcpu->policy->cur)
 610                                 __cpufreq_driver_target(pcpu->policy,
 611                                                         max_freq,
 612                                                         CPUFREQ_RELATION_H);
 613                         trace_cpufreq_interactive_setspeed(cpu,
 614                                                      pcpu->target_freq,
 615                                                      pcpu->policy->cur);
 616
 617                         up_read(&pcpu->enable_sem);
 618                 }
 619         }
 620
 621         return 0;
 622 }
 623
 624 static void cpufreq_interactive_boost(void)
 625 {
 626         int i;
 627         int anyboost = 0;
 628         unsigned long flags[2];
 629         struct cpufreq_interactive_cpuinfo *pcpu;
 630         struct cpufreq_interactive_tunables *tunables;
 631
 632         spin_lock_irqsave(&speedchange_cpumask_lock, flags[0]);
 633
 634         for_each_online_cpu(i) {
 635                 pcpu = &per_cpu(cpuinfo, i);
 636                 tunables = pcpu->policy->governor_data;
 637
 638                 spin_lock_irqsave(&pcpu->target_freq_lock, flags[1]);
 639                 if (pcpu->target_freq < tunables->hispeed_freq) {
 640                         pcpu->target_freq = tunables->hispeed_freq;
 641                         cpumask_set_cpu(i, &speedchange_cpumask);
 642                         pcpu->hispeed_validate_time =
 643                                 ktime_to_us(ktime_get());
 644                         anyboost = 1;
 645                 }
 646
 647                 /*
 648                  * Set floor freq and (re)start timer for when last
 649                  * validated.
 650                  */
 651
 652                 pcpu->floor_freq = tunables->hispeed_freq;
 653                 pcpu->floor_validate_time = ktime_to_us(ktime_get());
 654                 spin_unlock_irqrestore(&pcpu->target_freq_lock, flags[1]);
 655         }
 656
 657         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags[0]);
 658
 659         if (anyboost)
 660                 wake_up_process(speedchange_task);
 661 }
 662
 663 static int cpufreq_interactive_notifier(
 664         struct notifier_block *nb, unsigned long val, void *data)
 665 {
 666         struct cpufreq_freqs *freq = data;
 667         struct cpufreq_interactive_cpuinfo *pcpu;
 668         int cpu;
 669         unsigned long flags;
 670
 671         if (val == CPUFREQ_POSTCHANGE) {
 672                 pcpu = &per_cpu(cpuinfo, freq->cpu);
 673                 if (!down_read_trylock(&pcpu->enable_sem))
 674                         return 0;
 675                 if (!pcpu->governor_enabled) {
 676                         up_read(&pcpu->enable_sem);
 677                         return 0;
 678                 }
 679
 680                 for_each_cpu(cpu, pcpu->policy->cpus) {
 681                         struct cpufreq_interactive_cpuinfo *pjcpu =
 682                                 &per_cpu(cpuinfo, cpu);
 683                         if (cpu != freq->cpu) {
 684                                 if (!down_read_trylock(&pjcpu->enable_sem))
 685                                         continue;
 686                                 if (!pjcpu->governor_enabled) {
 687                                         up_read(&pjcpu->enable_sem);
 688                                         continue;
 689                                 }
 690                         }
 691                         spin_lock_irqsave(&pjcpu->load_lock, flags);
 692                         update_load(cpu);
 693                         spin_unlock_irqrestore(&pjcpu->load_lock, flags);
 694                         if (cpu != freq->cpu)
 695                                 up_read(&pjcpu->enable_sem);
 696                 }
 697
 698                 up_read(&pcpu->enable_sem);
 699         }
 700         return 0;
 701 }
 702
 703 static struct notifier_block cpufreq_notifier_block = {
 704         .notifier_call = cpufreq_interactive_notifier,
 705 };
 706
 707 static unsigned int *get_tokenized_data(const char *buf, int *num_tokens)
 708 {
 709         const char *cp;
 710         int i;
 711         int ntokens = 1;
 712         unsigned int *tokenized_data;
 713         int err = -EINVAL;
 714
 715         cp = buf;
 716         while ((cp = strpbrk(cp + 1, " :")))
 717                 ntokens++;
 718
 719         if (!(ntokens & 0x1))
 720                 goto err;
 721
 722         tokenized_data = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
 723         if (!tokenized_data) {
 724                 err = -ENOMEM;
 725                 goto err;
 726         }
 727
 728         cp = buf;
 729         i = 0;
 730         while (i < ntokens) {
 731                 if (sscanf(cp, "%u", &tokenized_data[i++]) != 1)
 732                         goto err_kfree;
 733
 734                 cp = strpbrk(cp, " :");
 735                 if (!cp)
 736                         break;
 737                 cp++;
 738         }
 739
 740         if (i != ntokens)
 741                 goto err_kfree;
 742
 743         *num_tokens = ntokens;
 744         return tokenized_data;
 745
 746 err_kfree:
 747         kfree(tokenized_data);
 748 err:
 749         return ERR_PTR(err);
 750 }
 751
 752 static ssize_t show_target_loads(
 753         struct cpufreq_interactive_tunables *tunables,
 754         char *buf)
 755 {
 756         int i;
 757         ssize_t ret = 0;
 758         unsigned long flags;
 759
 760         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 761
 762         for (i = 0; i < tunables->ntarget_loads; i++)
 763                 ret += sprintf(buf + ret, "%u%s", tunables->target_loads[i],
 764                                i & 0x1 ? ":" : " ");
 765
 766         sprintf(buf + ret - 1, "\n");
 767         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 768         return ret;
 769 }
 770
 771 static ssize_t store_target_loads(
 772         struct cpufreq_interactive_tunables *tunables,
 773         const char *buf, size_t count)
 774 {
 775         int ntokens;
 776         unsigned int *new_target_loads = NULL;
 777         unsigned long flags;
 778
 779         new_target_loads = get_tokenized_data(buf, &ntokens);
 780         if (IS_ERR(new_target_loads))
 781                 return PTR_RET(new_target_loads);
 782
 783         spin_lock_irqsave(&tunables->target_loads_lock, flags);
 784         if (tunables->target_loads != default_target_loads)
 785                 kfree(tunables->target_loads);
 786         tunables->target_loads = new_target_loads;
 787         tunables->ntarget_loads = ntokens;
 788         spin_unlock_irqrestore(&tunables->target_loads_lock, flags);
 789         return count;
 790 }
 791
 792 static ssize_t show_above_hispeed_delay(
 793         struct cpufreq_interactive_tunables *tunables, char *buf)
 794 {
 795         int i;
 796         ssize_t ret = 0;
 797         unsigned long flags;
 798
 799         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 800
 801         for (i = 0; i < tunables->nabove_hispeed_delay; i++)
 802                 ret += sprintf(buf + ret, "%u%s",
 803                                tunables->above_hispeed_delay[i],
 804                                i & 0x1 ? ":" : " ");
 805
 806         sprintf(buf + ret - 1, "\n");
 807         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 808         return ret;
 809 }
 810
 811 static ssize_t store_above_hispeed_delay(
 812         struct cpufreq_interactive_tunables *tunables,
 813         const char *buf, size_t count)
 814 {
 815         int ntokens;
 816         unsigned int *new_above_hispeed_delay = NULL;
 817         unsigned long flags;
 818
 819         new_above_hispeed_delay = get_tokenized_data(buf, &ntokens);
 820         if (IS_ERR(new_above_hispeed_delay))
 821                 return PTR_RET(new_above_hispeed_delay);
 822
 823         spin_lock_irqsave(&tunables->above_hispeed_delay_lock, flags);
 824         if (tunables->above_hispeed_delay != default_above_hispeed_delay)
 825                 kfree(tunables->above_hispeed_delay);
 826         tunables->above_hispeed_delay = new_above_hispeed_delay;
 827         tunables->nabove_hispeed_delay = ntokens;
 828         spin_unlock_irqrestore(&tunables->above_hispeed_delay_lock, flags);
 829         return count;
 830
 831 }
 832
 833 static ssize_t show_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 834                 char *buf)
 835 {
 836         return sprintf(buf, "%u\n", tunables->hispeed_freq);
 837 }
 838
 839 static ssize_t store_hispeed_freq(struct cpufreq_interactive_tunables *tunables,
 840                 const char *buf, size_t count)
 841 {
 842         int ret;
 843         long unsigned int val;
 844
 845         ret = strict_strtoul(buf, 0, &val);
 846         if (ret < 0)
 847                 return ret;
 848         tunables->hispeed_freq = val;
 849         return count;
 850 }
 851
 852 static ssize_t show_go_hispeed_load(struct cpufreq_interactive_tunables
 853                 *tunables, char *buf)
 854 {
 855         return sprintf(buf, "%lu\n", tunables->go_hispeed_load);
 856 }
 857
 858 static ssize_t store_go_hispeed_load(struct cpufreq_interactive_tunables
 859                 *tunables, const char *buf, size_t count)
 860 {
 861         int ret;
 862         unsigned long val;
 863
 864         ret = strict_strtoul(buf, 0, &val);
 865         if (ret < 0)
 866                 return ret;
 867         tunables->go_hispeed_load = val;
 868         return count;
 869 }
 870
 871 static ssize_t show_min_sample_time(struct cpufreq_interactive_tunables
 872                 *tunables, char *buf)
 873 {
 874         return sprintf(buf, "%lu\n", tunables->min_sample_time);
 875 }
 876
 877 static ssize_t store_min_sample_time(struct cpufreq_interactive_tunables
 878                 *tunables, const char *buf, size_t count)
 879 {
 880         int ret;
 881         unsigned long val;
 882
 883         ret = strict_strtoul(buf, 0, &val);
 884         if (ret < 0)
 885                 return ret;
 886         tunables->min_sample_time = val;
 887         return count;
 888 }
 889
 890 static ssize_t show_timer_rate(struct cpufreq_interactive_tunables *tunables,
 891                 char *buf)
 892 {
 893         return sprintf(buf, "%lu\n", tunables->timer_rate);
 894 }
 895
 896 static ssize_t store_timer_rate(struct cpufreq_interactive_tunables *tunables,
 897                 const char *buf, size_t count)
 898 {
 899         int ret;
 900         unsigned long val;
 901
 902         ret = strict_strtoul(buf, 0, &val);
 903         if (ret < 0)
 904                 return ret;
 905         tunables->timer_rate = val;
 906         return count;
 907 }
 908
 909 static ssize_t show_timer_slack(struct cpufreq_interactive_tunables *tunables,
 910                 char *buf)
 911 {
 912         return sprintf(buf, "%d\n", tunables->timer_slack_val);
 913 }
 914
 915 static ssize_t store_timer_slack(struct cpufreq_interactive_tunables *tunables,
 916                 const char *buf, size_t count)
 917 {
 918         int ret;
 919         unsigned long val;
 920
 921         ret = kstrtol(buf, 10, &val);
 922         if (ret < 0)
 923                 return ret;
 924
 925         tunables->timer_slack_val = val;
 926         return count;
 927 }
 928
 929 static ssize_t show_boost(struct cpufreq_interactive_tunables *tunables,
 930                           char *buf)
 931 {
 932         return sprintf(buf, "%d\n", tunables->boost_val);
 933 }
 934
 935 static ssize_t store_boost(struct cpufreq_interactive_tunables *tunables,
 936                            const char *buf, size_t count)
 937 {
 938         int ret;
 939         unsigned long val;
 940
 941         ret = kstrtoul(buf, 0, &val);
 942         if (ret < 0)
 943                 return ret;
 944
 945         tunables->boost_val = val;
 946
 947         if (tunables->boost_val) {
 948                 trace_cpufreq_interactive_boost("on");
 949                 cpufreq_interactive_boost();
 950         } else {
 951                 tunables->boostpulse_endtime = ktime_to_us(ktime_get());
 952                 trace_cpufreq_interactive_unboost("off");
 953         }
 954
 955         return count;
 956 }
 957
 958 static ssize_t store_boostpulse(struct cpufreq_interactive_tunables *tunables,
 959                                 const char *buf, size_t count)
 960 {
 961         int ret;
 962         unsigned long val;
 963
 964         ret = kstrtoul(buf, 0, &val);
 965         if (ret < 0)
 966                 return ret;
 967
 968         tunables->boostpulse_endtime = ktime_to_us(ktime_get()) +
 969                 tunables->boostpulse_duration_val;
 970         trace_cpufreq_interactive_boost("pulse");
 971         cpufreq_interactive_boost();
 972         return count;
 973 }
 974
 975 static ssize_t show_boostpulse_duration(struct cpufreq_interactive_tunables
 976                 *tunables, char *buf)
 977 {
 978         return sprintf(buf, "%d\n", tunables->boostpulse_duration_val);
 979 }
 980
 981 static ssize_t store_boostpulse_duration(struct cpufreq_interactive_tunables
 982                 *tunables, const char *buf, size_t count)
 983 {
 984         int ret;
 985         unsigned long val;
 986
 987         ret = kstrtoul(buf, 0, &val);
 988         if (ret < 0)
 989                 return ret;
 990
 991         tunables->boostpulse_duration_val = val;
 992         return count;
 993 }
 994
 995 static ssize_t show_io_is_busy(struct cpufreq_interactive_tunables *tunables,
 996                 char *buf)
 997 {
 998         return sprintf(buf, "%u\n", tunables->io_is_busy);
 999 }
1000
1001 static ssize_t store_io_is_busy(struct cpufreq_interactive_tunables *tunables,
1002                 const char *buf, size_t count)
1003 {
1004         int ret;
1005         unsigned long val;
1006
1007         ret = kstrtoul(buf, 0, &val);
1008         if (ret < 0)
1009                 return ret;
1010         tunables->io_is_busy = val;
1011         return count;
1012 }
1013
1014 /*
1015  * Create show/store routines
1016  * - sys: One governor instance for complete SYSTEM
1017  * - pol: One governor instance per struct cpufreq_policy
1018  */
1019 #define show_gov_pol_sys(file_name)                                     \
1020 static ssize_t show_##file_name##_gov_sys                               \
1021 (struct kobject *kobj, struct attribute *attr, char *buf)               \
1022 {                                                                       \
1023         return show_##file_name(common_tunables, buf);                  \
1024 }                                                                       \
1025                                                                         \
1026 static ssize_t show_##file_name##_gov_pol                               \
1027 (struct cpufreq_policy *policy, char *buf)                              \
1028 {                                                                       \
1029         return show_##file_name(policy->governor_data, buf);            \
1030 }
1031
1032 #define store_gov_pol_sys(file_name)                                    \
1033 static ssize_t store_##file_name##_gov_sys                              \
1034 (struct kobject *kobj, struct attribute *attr, const char *buf,         \
1035         size_t count)                                                   \
1036 {                                                                       \
1037         return store_##file_name(common_tunables, buf, count);          \
1038 }                                                                       \
1039                                                                         \
1040 static ssize_t store_##file_name##_gov_pol                              \
1041 (struct cpufreq_policy *policy, const char *buf, size_t count)          \
1042 {                                                                       \
1043         return store_##file_name(policy->governor_data, buf, count);    \
1044 }
1045
1046 #define show_store_gov_pol_sys(file_name)                               \
1047 show_gov_pol_sys(file_name);                                            \
1048 store_gov_pol_sys(file_name)
1049
1050 show_store_gov_pol_sys(target_loads);
1051 show_store_gov_pol_sys(above_hispeed_delay);
1052 show_store_gov_pol_sys(hispeed_freq);
1053 show_store_gov_pol_sys(go_hispeed_load);
1054 show_store_gov_pol_sys(min_sample_time);
1055 show_store_gov_pol_sys(timer_rate);
1056 show_store_gov_pol_sys(timer_slack);
1057 show_store_gov_pol_sys(boost);
1058 store_gov_pol_sys(boostpulse);
1059 show_store_gov_pol_sys(boostpulse_duration);
1060 show_store_gov_pol_sys(io_is_busy);
1061
1062 #define gov_sys_attr_rw(_name)                                          \
1063 static struct global_attr _name##_gov_sys =                             \
1064 __ATTR(_name, 0644, show_##_name##_gov_sys, store_##_name##_gov_sys)
1065
1066 #define gov_pol_attr_rw(_name)                                          \
1067 static struct freq_attr _name##_gov_pol =                               \
1068 __ATTR(_name, 0644, show_##_name##_gov_pol, store_##_name##_gov_pol)
1069
1070 #define gov_sys_pol_attr_rw(_name)                                      \
1071         gov_sys_attr_rw(_name);                                         \
1072         gov_pol_attr_rw(_name)
1073
1074 gov_sys_pol_attr_rw(target_loads);
1075 gov_sys_pol_attr_rw(above_hispeed_delay);
1076 gov_sys_pol_attr_rw(hispeed_freq);
1077 gov_sys_pol_attr_rw(go_hispeed_load);
1078 gov_sys_pol_attr_rw(min_sample_time);
1079 gov_sys_pol_attr_rw(timer_rate);
1080 gov_sys_pol_attr_rw(timer_slack);
1081 gov_sys_pol_attr_rw(boost);
1082 gov_sys_pol_attr_rw(boostpulse_duration);
1083 gov_sys_pol_attr_rw(io_is_busy);
1084
1085 static struct global_attr boostpulse_gov_sys =
1086         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_sys);
1087
1088 static struct freq_attr boostpulse_gov_pol =
1089         __ATTR(boostpulse, 0200, NULL, store_boostpulse_gov_pol);
1090
1091 /* One Governor instance for entire system */
1092 static struct attribute *interactive_attributes_gov_sys[] = {
1093         &target_loads_gov_sys.attr,
1094         &above_hispeed_delay_gov_sys.attr,
1095         &hispeed_freq_gov_sys.attr,
1096         &go_hispeed_load_gov_sys.attr,
1097         &min_sample_time_gov_sys.attr,
1098         &timer_rate_gov_sys.attr,
1099         &timer_slack_gov_sys.attr,
1100         &boost_gov_sys.attr,
1101         &boostpulse_gov_sys.attr,
1102         &boostpulse_duration_gov_sys.attr,
1103         &io_is_busy_gov_sys.attr,
1104         NULL,
1105 };
1106
1107 static struct attribute_group interactive_attr_group_gov_sys = {
1108         .attrs = interactive_attributes_gov_sys,
1109         .name = "interactive",
1110 };
1111
1112 /* Per policy governor instance */
1113 static struct attribute *interactive_attributes_gov_pol[] = {
1114         &target_loads_gov_pol.attr,
1115         &above_hispeed_delay_gov_pol.attr,
1116         &hispeed_freq_gov_pol.attr,
1117         &go_hispeed_load_gov_pol.attr,
1118         &min_sample_time_gov_pol.attr,
1119         &timer_rate_gov_pol.attr,
1120         &timer_slack_gov_pol.attr,
1121         &boost_gov_pol.attr,
1122         &boostpulse_gov_pol.attr,
1123         &boostpulse_duration_gov_pol.attr,
1124         &io_is_busy_gov_pol.attr,
1125         NULL,
1126 };
1127
1128 static struct attribute_group interactive_attr_group_gov_pol = {
1129         .attrs = interactive_attributes_gov_pol,
1130         .name = "interactive",
1131 };
1132
1133 static struct attribute_group *get_sysfs_attr(void)
1134 {
1135         if (have_governor_per_policy())
1136                 return &interactive_attr_group_gov_pol;
1137         else
1138                 return &interactive_attr_group_gov_sys;
1139 }
1140
1141 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
1142                                              unsigned long val,
1143                                              void *data)
1144 {
1145         switch (val) {
1146         case IDLE_START:
1147                 cpufreq_interactive_idle_start();
1148                 break;
1149         case IDLE_END:
1150                 cpufreq_interactive_idle_end();
1151                 break;
1152         }
1153
1154         return 0;
1155 }
1156
1157 static struct notifier_block cpufreq_interactive_idle_nb = {
1158         .notifier_call = cpufreq_interactive_idle_notifier,
1159 };
1160
1161 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
1162                 unsigned int event)
1163 {
1164         int rc;
1165         unsigned int j;
1166         struct cpufreq_interactive_cpuinfo *pcpu;
1167         struct cpufreq_frequency_table *freq_table;
1168         struct cpufreq_interactive_tunables *tunables;
1169         unsigned long flags;
1170
1171         if (have_governor_per_policy())
1172                 tunables = policy->governor_data;
1173         else
1174                 tunables = common_tunables;
1175
1176         WARN_ON(!tunables && (event != CPUFREQ_GOV_POLICY_INIT));
1177
1178         switch (event) {
1179         case CPUFREQ_GOV_POLICY_INIT:
1180                 if (have_governor_per_policy()) {
1181                         WARN_ON(tunables);
1182                 } else if (tunables) {
1183                         tunables->usage_count++;
1184                         policy->governor_data = tunables;
1185                         return 0;
1186                 }
1187
1188                 tunables = kzalloc(sizeof(*tunables), GFP_KERNEL);
1189                 if (!tunables) {
1190                         pr_err("%s: POLICY_INIT: kzalloc failed\n", __func__);
1191                         return -ENOMEM;
1192                 }
1193
1194                 tunables->usage_count = 1;
1195                 tunables->above_hispeed_delay = default_above_hispeed_delay;
1196                 tunables->nabove_hispeed_delay =
1197                         ARRAY_SIZE(default_above_hispeed_delay);
1198                 tunables->go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
1199                 tunables->target_loads = default_target_loads;
1200                 tunables->ntarget_loads = ARRAY_SIZE(default_target_loads);
1201                 tunables->min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
1202                 tunables->timer_rate = DEFAULT_TIMER_RATE;
1203                 tunables->boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
1204                 tunables->timer_slack_val = DEFAULT_TIMER_SLACK;
1205
1206                 spin_lock_init(&tunables->target_loads_lock);
1207                 spin_lock_init(&tunables->above_hispeed_delay_lock);
1208
1209                 policy->governor_data = tunables;
1210                 if (!have_governor_per_policy())
1211                         common_tunables = tunables;
1212
1213                 rc = sysfs_create_group(get_governor_parent_kobj(policy),
1214                                 get_sysfs_attr());
1215                 if (rc) {
1216                         kfree(tunables);
1217                         policy->governor_data = NULL;
1218                         if (!have_governor_per_policy())
1219                                 common_tunables = NULL;
1220                         return rc;
1221                 }
1222
1223                 if (!policy->governor->initialized) {
1224                         idle_notifier_register(&cpufreq_interactive_idle_nb);
1225                         cpufreq_register_notifier(&cpufreq_notifier_block,
1226                                         CPUFREQ_TRANSITION_NOTIFIER);
1227                 }
1228
1229                 break;
1230
1231         case CPUFREQ_GOV_POLICY_EXIT:
1232                 if (!--tunables->usage_count) {
1233                         if (policy->governor->initialized == 1) {
1234                                 cpufreq_unregister_notifier(&cpufreq_notifier_block,
1235                                                 CPUFREQ_TRANSITION_NOTIFIER);
1236                                 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
1237                         }
1238
1239                         sysfs_remove_group(get_governor_parent_kobj(policy),
1240                                         get_sysfs_attr());
1241                         kfree(tunables);
1242                         common_tunables = NULL;
1243                 }
1244
1245                 policy->governor_data = NULL;
1246                 break;
1247
1248         case CPUFREQ_GOV_START:
1249                 mutex_lock(&gov_lock);
1250
1251                 freq_table = cpufreq_frequency_get_table(policy->cpu);
1252                 if (!tunables->hispeed_freq)
1253                         tunables->hispeed_freq = policy->max;
1254
1255                 for_each_cpu(j, policy->cpus) {
1256                         pcpu = &per_cpu(cpuinfo, j);
1257                         pcpu->policy = policy;
1258                         pcpu->target_freq = policy->cur;
1259                         pcpu->freq_table = freq_table;
1260                         pcpu->floor_freq = pcpu->target_freq;
1261                         pcpu->floor_validate_time =
1262                                 ktime_to_us(ktime_get());
1263                         pcpu->hispeed_validate_time =
1264                                 pcpu->floor_validate_time;
1265                         pcpu->max_freq = policy->max;
1266                         down_write(&pcpu->enable_sem);
1267                         del_timer_sync(&pcpu->cpu_timer);
1268                         del_timer_sync(&pcpu->cpu_slack_timer);
1269                         cpufreq_interactive_timer_start(tunables, j);
1270                         pcpu->governor_enabled = 1;
1271                         up_write(&pcpu->enable_sem);
1272                 }
1273
1274                 mutex_unlock(&gov_lock);
1275                 break;
1276
1277         case CPUFREQ_GOV_STOP:
1278                 mutex_lock(&gov_lock);
1279                 for_each_cpu(j, policy->cpus) {
1280                         pcpu = &per_cpu(cpuinfo, j);
1281                         down_write(&pcpu->enable_sem);
1282                         pcpu->governor_enabled = 0;
1283                         del_timer_sync(&pcpu->cpu_timer);
1284                         del_timer_sync(&pcpu->cpu_slack_timer);
1285                         up_write(&pcpu->enable_sem);
1286                 }
1287
1288                 mutex_unlock(&gov_lock);
1289                 break;
1290
1291         case CPUFREQ_GOV_LIMITS:
1292                 if (policy->max < policy->cur)
1293                         __cpufreq_driver_target(policy,
1294                                         policy->max, CPUFREQ_RELATION_H);
1295                 else if (policy->min > policy->cur)
1296                         __cpufreq_driver_target(policy,
1297                                         policy->min, CPUFREQ_RELATION_L);
1298                 for_each_cpu(j, policy->cpus) {
1299                         pcpu = &per_cpu(cpuinfo, j);
1300
1301                         down_read(&pcpu->enable_sem);
1302                         if (pcpu->governor_enabled == 0) {
1303                                 up_read(&pcpu->enable_sem);
1304                                 continue;
1305                         }
1306
1307                         spin_lock_irqsave(&pcpu->target_freq_lock, flags);
1308                         if (policy->max < pcpu->target_freq)
1309                                 pcpu->target_freq = policy->max;
1310                         else if (policy->min > pcpu->target_freq)
1311                                 pcpu->target_freq = policy->min;
1312
1313                         spin_unlock_irqrestore(&pcpu->target_freq_lock, flags);
1314                         up_read(&pcpu->enable_sem);
1315
1316                         /* Reschedule timer only if policy->max is raised.
1317                          * Delete the timers, else the timer callback may
1318                          * return without re-arm the timer when failed
1319                          * acquire the semaphore. This race may cause timer
1320                          * stopped unexpectedly.
1321                          */
1322
1323                         if (policy->max > pcpu->max_freq) {
1324                                 down_write(&pcpu->enable_sem);
1325                                 del_timer_sync(&pcpu->cpu_timer);
1326                                 del_timer_sync(&pcpu->cpu_slack_timer);
1327                                 cpufreq_interactive_timer_start(tunables, j);
1328                                 up_write(&pcpu->enable_sem);
1329                         }
1330
1331                         pcpu->max_freq = policy->max;
1332                 }
1333                 break;
1334         }
1335         return 0;
1336 }
1337
1338 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1339 static
1340 #endif
1341 struct cpufreq_governor cpufreq_gov_interactive = {
1342         .name = "interactive",
1343         .governor = cpufreq_governor_interactive,
1344         .max_transition_latency = 10000000,
1345         .owner = THIS_MODULE,
1346 };
1347
1348 static void cpufreq_interactive_nop_timer(unsigned long data)
1349 {
1350 }
1351
1352 static int __init cpufreq_interactive_init(void)
1353 {
1354         unsigned int i;
1355         struct cpufreq_interactive_cpuinfo *pcpu;
1356         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
1357
1358         /* Initalize per-cpu timers */
1359         for_each_possible_cpu(i) {
1360                 pcpu = &per_cpu(cpuinfo, i);
1361                 init_timer_deferrable(&pcpu->cpu_timer);
1362                 pcpu->cpu_timer.function = cpufreq_interactive_timer;
1363                 pcpu->cpu_timer.data = i;
1364                 init_timer(&pcpu->cpu_slack_timer);
1365                 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
1366                 spin_lock_init(&pcpu->load_lock);
1367                 spin_lock_init(&pcpu->target_freq_lock);
1368                 init_rwsem(&pcpu->enable_sem);
1369         }
1370
1371         spin_lock_init(&speedchange_cpumask_lock);
1372         mutex_init(&gov_lock);
1373         speedchange_task =
1374                 kthread_create(cpufreq_interactive_speedchange_task, NULL,
1375                                "cfinteractive");
1376         if (IS_ERR(speedchange_task))
1377                 return PTR_ERR(speedchange_task);
1378
1379         sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, &param);
1380         get_task_struct(speedchange_task);
1381
1382         /* NB: wake up so the thread does not look hung to the freezer */
1383         wake_up_process(speedchange_task);
1384
1385         return cpufreq_register_governor(&cpufreq_gov_interactive);
1386 }
1387
1388 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1389 fs_initcall(cpufreq_interactive_init);
1390 #else
1391 module_init(cpufreq_interactive_init);
1392 #endif
1393
1394 static void __exit cpufreq_interactive_exit(void)
1395 {
1396         cpufreq_unregister_governor(&cpufreq_gov_interactive);
1397         kthread_stop(speedchange_task);
1398         put_task_struct(speedchange_task);
1399 }
1400
1401 module_exit(cpufreq_interactive_exit);
1402
1403 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1404 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1405         "Latency sensitive workloads");
1406 MODULE_LICENSE("GPL");