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/mutex.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/mutex.h>
  33 #include <linux/slab.h>
  34 #include <asm/cputime.h>
  35
  36 #define CREATE_TRACE_POINTS
  37 #include <trace/events/cpufreq_interactive.h>
  38
  39 static atomic_t active_count = ATOMIC_INIT(0);
  40
  41 struct cpufreq_interactive_cpuinfo {
  42         struct timer_list cpu_timer;
  43         struct timer_list cpu_slack_timer;
  44         spinlock_t load_lock; /* protects the next 4 fields */
  45         u64 time_in_idle;
  46         u64 time_in_idle_timestamp;
  47         u64 cputime_speedadj;
  48         u64 cputime_speedadj_timestamp;
  49         struct cpufreq_policy *policy;
  50         struct cpufreq_frequency_table *freq_table;
  51         unsigned int target_freq;
  52         unsigned int floor_freq;
  53         u64 floor_validate_time;
  54         u64 hispeed_validate_time;
  55         int governor_enabled;
  56 };
  57
  58 static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo);
  59
  60 /* realtime thread handles frequency scaling */
  61 static struct task_struct *speedchange_task;
  62 static cpumask_t speedchange_cpumask;
  63 static spinlock_t speedchange_cpumask_lock;
  64
  65 /* Hi speed to bump to from lo speed when load burst (default max) */
  66 static unsigned int hispeed_freq;
  67
  68 /* Go to hi speed when CPU load at or above this value. */
  69 #define DEFAULT_GO_HISPEED_LOAD 85
  70 static unsigned long go_hispeed_load;
  71
  72 /* Target load.  Lower values result in higher CPU speeds. */
  73 #define DEFAULT_TARGET_LOAD 90
  74 static unsigned int default_target_loads[] = {DEFAULT_TARGET_LOAD};
  75 static spinlock_t target_loads_lock;
  76 static unsigned int *target_loads = default_target_loads;
  77 static int ntarget_loads = ARRAY_SIZE(default_target_loads);
  78
  79 /*
  80  * The minimum amount of time to spend at a frequency before we can ramp down.
  81  */
  82 #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC)
  83 static unsigned long min_sample_time;
  84
  85 /*
  86  * The sample rate of the timer used to increase frequency
  87  */
  88 #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC)
  89 static unsigned long timer_rate;
  90
  91 /*
  92  * Wait this long before raising speed above hispeed, by default a single
  93  * timer interval.
  94  */
  95 #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE
  96 static unsigned long above_hispeed_delay_val;
  97
  98 /* Non-zero means indefinite speed boost active */
  99 static int boost_val;
 100 /* Duration of a boot pulse in usecs */
 101 static int boostpulse_duration_val = DEFAULT_MIN_SAMPLE_TIME;
 102 /* End time of boost pulse in ktime converted to usecs */
 103 static u64 boostpulse_endtime;
 104
 105 /*
 106  * Max additional time to wait in idle, beyond timer_rate, at speeds above
 107  * minimum before wakeup to reduce speed, or -1 if unnecessary.
 108  */
 109 #define DEFAULT_TIMER_SLACK (4 * DEFAULT_TIMER_RATE)
 110 static int timer_slack_val = DEFAULT_TIMER_SLACK;
 111
 112 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
 113                 unsigned int event);
 114
 115 #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
 116 static
 117 #endif
 118 struct cpufreq_governor cpufreq_gov_interactive = {
 119         .name = "interactive",
 120         .governor = cpufreq_governor_interactive,
 121         .max_transition_latency = 10000000,
 122         .owner = THIS_MODULE,
 123 };
 124
 125 static void cpufreq_interactive_timer_resched(
 126         struct cpufreq_interactive_cpuinfo *pcpu)
 127 {
 128         unsigned long expires = jiffies + usecs_to_jiffies(timer_rate);
 129
 130         mod_timer_pinned(&pcpu->cpu_timer, expires);
 131         if (timer_slack_val >= 0 && pcpu->target_freq > pcpu->policy->min) {
 132                 expires += usecs_to_jiffies(timer_slack_val);
 133                 mod_timer_pinned(&pcpu->cpu_slack_timer, expires);
 134         }
 135
 136         spin_lock(&pcpu->load_lock);
 137         pcpu->time_in_idle =
 138                 get_cpu_idle_time_us(smp_processor_id(),
 139                                      &pcpu->time_in_idle_timestamp);
 140         pcpu->cputime_speedadj = 0;
 141         pcpu->cputime_speedadj_timestamp = pcpu->time_in_idle_timestamp;
 142         spin_unlock(&pcpu->load_lock);
 143 }
 144
 145 static unsigned int freq_to_targetload(unsigned int freq)
 146 {
 147         int i;
 148         unsigned int ret;
 149
 150         spin_lock(&target_loads_lock);
 151
 152         for (i = 0; i < ntarget_loads - 1 && freq >= target_loads[i+1]; i += 2)
 153                 ;
 154
 155         ret = target_loads[i];
 156         spin_unlock(&target_loads_lock);
 157         return ret;
 158 }
 159
 160 /*
 161  * If increasing frequencies never map to a lower target load then
 162  * choose_freq() will find the minimum frequency that does not exceed its
 163  * target load given the current load.
 164  */
 165
 166 static unsigned int choose_freq(
 167         struct cpufreq_interactive_cpuinfo *pcpu, unsigned int loadadjfreq)
 168 {
 169         unsigned int freq = pcpu->policy->cur;
 170         unsigned int prevfreq, freqmin, freqmax;
 171         unsigned int tl;
 172         int index;
 173
 174         freqmin = 0;
 175         freqmax = UINT_MAX;
 176
 177         do {
 178                 prevfreq = freq;
 179                 tl = freq_to_targetload(freq);
 180
 181                 /*
 182                  * Find the lowest frequency where the computed load is less
 183                  * than or equal to the target load.
 184                  */
 185
 186                 cpufreq_frequency_table_target(
 187                         pcpu->policy, pcpu->freq_table, loadadjfreq / tl,
 188                         CPUFREQ_RELATION_L, &index);
 189                 freq = pcpu->freq_table[index].frequency;
 190
 191                 if (freq > prevfreq) {
 192                         /* The previous frequency is too low. */
 193                         freqmin = prevfreq;
 194
 195                         if (freq >= freqmax) {
 196                                 /*
 197                                  * Find the highest frequency that is less
 198                                  * than freqmax.
 199                                  */
 200                                 cpufreq_frequency_table_target(
 201                                         pcpu->policy, pcpu->freq_table,
 202                                         freqmax - 1, CPUFREQ_RELATION_H,
 203                                         &index);
 204                                 freq = pcpu->freq_table[index].frequency;
 205
 206                                 if (freq == freqmin) {
 207                                         /*
 208                                          * The first frequency below freqmax
 209                                          * has already been found to be too
 210                                          * low.  freqmax is the lowest speed
 211                                          * we found that is fast enough.
 212                                          */
 213                                         freq = freqmax;
 214                                         break;
 215                                 }
 216                         }
 217                 } else if (freq < prevfreq) {
 218                         /* The previous frequency is high enough. */
 219                         freqmax = prevfreq;
 220
 221                         if (freq <= freqmin) {
 222                                 /*
 223                                  * Find the lowest frequency that is higher
 224                                  * than freqmin.
 225                                  */
 226                                 cpufreq_frequency_table_target(
 227                                         pcpu->policy, pcpu->freq_table,
 228                                         freqmin + 1, CPUFREQ_RELATION_L,
 229                                         &index);
 230                                 freq = pcpu->freq_table[index].frequency;
 231
 232                                 /*
 233                                  * If freqmax is the first frequency above
 234                                  * freqmin then we have already found that
 235                                  * this speed is fast enough.
 236                                  */
 237                                 if (freq == freqmax)
 238                                         break;
 239                         }
 240                 }
 241
 242                 /* If same frequency chosen as previous then done. */
 243         } while (freq != prevfreq);
 244
 245         return freq;
 246 }
 247
 248 static u64 update_load(int cpu)
 249 {
 250         struct cpufreq_interactive_cpuinfo *pcpu = &per_cpu(cpuinfo, cpu);
 251         u64 now;
 252         u64 now_idle;
 253         unsigned int delta_idle;
 254         unsigned int delta_time;
 255         u64 active_time;
 256
 257         now_idle = get_cpu_idle_time_us(cpu, &now);
 258         delta_idle = (unsigned int)(now_idle - pcpu->time_in_idle);
 259         delta_time = (unsigned int)(now - pcpu->time_in_idle_timestamp);
 260         active_time = delta_time - delta_idle;
 261         pcpu->cputime_speedadj += active_time * pcpu->policy->cur;
 262
 263         pcpu->time_in_idle = now_idle;
 264         pcpu->time_in_idle_timestamp = now;
 265         return now;
 266 }
 267
 268 static void cpufreq_interactive_timer(unsigned long data)
 269 {
 270         u64 now;
 271         unsigned int delta_time;
 272         u64 cputime_speedadj;
 273         int cpu_load;
 274         struct cpufreq_interactive_cpuinfo *pcpu =
 275                 &per_cpu(cpuinfo, data);
 276         unsigned int new_freq;
 277         unsigned int loadadjfreq;
 278         unsigned int index;
 279         unsigned long flags;
 280         bool boosted;
 281
 282         smp_rmb();
 283
 284         if (!pcpu->governor_enabled)
 285                 goto exit;
 286
 287         spin_lock(&pcpu->load_lock);
 288         now = update_load(data);
 289         delta_time = (unsigned int)(now - pcpu->cputime_speedadj_timestamp);
 290         cputime_speedadj = pcpu->cputime_speedadj;
 291         spin_unlock(&pcpu->load_lock);
 292
 293         if (WARN_ON_ONCE(!delta_time))
 294                 goto rearm;
 295
 296         do_div(cputime_speedadj, delta_time);
 297         loadadjfreq = (unsigned int)cputime_speedadj * 100;
 298         cpu_load = loadadjfreq / pcpu->target_freq;
 299         boosted = boost_val || now < boostpulse_endtime;
 300
 301         if (cpu_load >= go_hispeed_load || boosted) {
 302                 if (pcpu->target_freq < hispeed_freq) {
 303                         new_freq = hispeed_freq;
 304                 } else {
 305                         new_freq = choose_freq(pcpu, loadadjfreq);
 306
 307                         if (new_freq < hispeed_freq)
 308                                 new_freq = hispeed_freq;
 309                 }
 310         } else {
 311                 new_freq = choose_freq(pcpu, loadadjfreq);
 312         }
 313
 314         if (pcpu->target_freq >= hispeed_freq &&
 315             new_freq > pcpu->target_freq &&
 316             now - pcpu->hispeed_validate_time < above_hispeed_delay_val) {
 317                 trace_cpufreq_interactive_notyet(
 318                         data, cpu_load, pcpu->target_freq,
 319                         pcpu->policy->cur, new_freq);
 320                 goto rearm;
 321         }
 322
 323         pcpu->hispeed_validate_time = now;
 324
 325         if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table,
 326                                            new_freq, CPUFREQ_RELATION_L,
 327                                            &index)) {
 328                 pr_warn_once("timer %d: cpufreq_frequency_table_target error\n",
 329                              (int) data);
 330                 goto rearm;
 331         }
 332
 333         new_freq = pcpu->freq_table[index].frequency;
 334
 335         /*
 336          * Do not scale below floor_freq unless we have been at or above the
 337          * floor frequency for the minimum sample time since last validated.
 338          */
 339         if (new_freq < pcpu->floor_freq) {
 340                 if (now - pcpu->floor_validate_time < min_sample_time) {
 341                         trace_cpufreq_interactive_notyet(
 342                                 data, cpu_load, pcpu->target_freq,
 343                                 pcpu->policy->cur, new_freq);
 344                         goto rearm;
 345                 }
 346         }
 347
 348         /*
 349          * Update the timestamp for checking whether speed has been held at
 350          * or above the selected frequency for a minimum of min_sample_time,
 351          * if not boosted to hispeed_freq.  If boosted to hispeed_freq then we
 352          * allow the speed to drop as soon as the boostpulse duration expires
 353          * (or the indefinite boost is turned off).
 354          */
 355
 356         if (!boosted || new_freq > hispeed_freq) {
 357                 pcpu->floor_freq = new_freq;
 358                 pcpu->floor_validate_time = now;
 359         }
 360
 361         if (pcpu->target_freq == new_freq) {
 362                 trace_cpufreq_interactive_already(
 363                         data, cpu_load, pcpu->target_freq,
 364                         pcpu->policy->cur, new_freq);
 365                 goto rearm_if_notmax;
 366         }
 367
 368         trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq,
 369                                          pcpu->policy->cur, new_freq);
 370
 371         pcpu->target_freq = new_freq;
 372         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 373         cpumask_set_cpu(data, &speedchange_cpumask);
 374         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 375         wake_up_process(speedchange_task);
 376
 377 rearm_if_notmax:
 378         /*
 379          * Already set max speed and don't see a need to change that,
 380          * wait until next idle to re-evaluate, don't need timer.
 381          */
 382         if (pcpu->target_freq == pcpu->policy->max)
 383                 goto exit;
 384
 385 rearm:
 386         if (!timer_pending(&pcpu->cpu_timer))
 387                 cpufreq_interactive_timer_resched(pcpu);
 388
 389 exit:
 390         return;
 391 }
 392
 393 static void cpufreq_interactive_idle_start(void)
 394 {
 395         struct cpufreq_interactive_cpuinfo *pcpu =
 396                 &per_cpu(cpuinfo, smp_processor_id());
 397         int pending;
 398
 399         if (!pcpu->governor_enabled)
 400                 return;
 401
 402         pending = timer_pending(&pcpu->cpu_timer);
 403
 404         if (pcpu->target_freq != pcpu->policy->min) {
 405                 /*
 406                  * Entering idle while not at lowest speed.  On some
 407                  * platforms this can hold the other CPU(s) at that speed
 408                  * even though the CPU is idle. Set a timer to re-evaluate
 409                  * speed so this idle CPU doesn't hold the other CPUs above
 410                  * min indefinitely.  This should probably be a quirk of
 411                  * the CPUFreq driver.
 412                  */
 413                 if (!pending)
 414                         cpufreq_interactive_timer_resched(pcpu);
 415         }
 416
 417 }
 418
 419 static void cpufreq_interactive_idle_end(void)
 420 {
 421         struct cpufreq_interactive_cpuinfo *pcpu =
 422                 &per_cpu(cpuinfo, smp_processor_id());
 423
 424         if (!pcpu->governor_enabled)
 425                 return;
 426
 427         /* Arm the timer for 1-2 ticks later if not already. */
 428         if (!timer_pending(&pcpu->cpu_timer)) {
 429                 cpufreq_interactive_timer_resched(pcpu);
 430         } else if (time_after_eq(jiffies, pcpu->cpu_timer.expires)) {
 431                 del_timer(&pcpu->cpu_timer);
 432                 del_timer(&pcpu->cpu_slack_timer);
 433                 cpufreq_interactive_timer(smp_processor_id());
 434         }
 435 }
 436
 437 static int cpufreq_interactive_speedchange_task(void *data)
 438 {
 439         unsigned int cpu;
 440         cpumask_t tmp_mask;
 441         unsigned long flags;
 442         struct cpufreq_interactive_cpuinfo *pcpu;
 443
 444         while (1) {
 445                 set_current_state(TASK_INTERRUPTIBLE);
 446                 spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 447
 448                 if (cpumask_empty(&speedchange_cpumask)) {
 449                         spin_unlock_irqrestore(&speedchange_cpumask_lock,
 450                                                flags);
 451                         schedule();
 452
 453                         if (kthread_should_stop())
 454                                 break;
 455
 456                         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 457                 }
 458
 459                 set_current_state(TASK_RUNNING);
 460                 tmp_mask = speedchange_cpumask;
 461                 cpumask_clear(&speedchange_cpumask);
 462                 spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 463
 464                 for_each_cpu(cpu, &tmp_mask) {
 465                         unsigned int j;
 466                         unsigned int max_freq = 0;
 467
 468                         pcpu = &per_cpu(cpuinfo, cpu);
 469                         smp_rmb();
 470
 471                         if (!pcpu->governor_enabled)
 472                                 continue;
 473
 474                         for_each_cpu(j, pcpu->policy->cpus) {
 475                                 struct cpufreq_interactive_cpuinfo *pjcpu =
 476                                         &per_cpu(cpuinfo, j);
 477
 478                                 if (pjcpu->target_freq > max_freq)
 479                                         max_freq = pjcpu->target_freq;
 480                         }
 481
 482                         if (max_freq != pcpu->policy->cur)
 483                                 __cpufreq_driver_target(pcpu->policy,
 484                                                         max_freq,
 485                                                         CPUFREQ_RELATION_H);
 486                         trace_cpufreq_interactive_setspeed(cpu,
 487                                                      pcpu->target_freq,
 488                                                      pcpu->policy->cur);
 489                 }
 490         }
 491
 492         return 0;
 493 }
 494
 495 static void cpufreq_interactive_boost(void)
 496 {
 497         int i;
 498         int anyboost = 0;
 499         unsigned long flags;
 500         struct cpufreq_interactive_cpuinfo *pcpu;
 501
 502         spin_lock_irqsave(&speedchange_cpumask_lock, flags);
 503
 504         for_each_online_cpu(i) {
 505                 pcpu = &per_cpu(cpuinfo, i);
 506
 507                 if (pcpu->target_freq < hispeed_freq) {
 508                         pcpu->target_freq = hispeed_freq;
 509                         cpumask_set_cpu(i, &speedchange_cpumask);
 510                         pcpu->hispeed_validate_time =
 511                                 ktime_to_us(ktime_get());
 512                         anyboost = 1;
 513                 }
 514
 515                 /*
 516                  * Set floor freq and (re)start timer for when last
 517                  * validated.
 518                  */
 519
 520                 pcpu->floor_freq = hispeed_freq;
 521                 pcpu->floor_validate_time = ktime_to_us(ktime_get());
 522         }
 523
 524         spin_unlock_irqrestore(&speedchange_cpumask_lock, flags);
 525
 526         if (anyboost)
 527                 wake_up_process(speedchange_task);
 528 }
 529
 530 static int cpufreq_interactive_notifier(
 531         struct notifier_block *nb, unsigned long val, void *data)
 532 {
 533         struct cpufreq_freqs *freq = data;
 534         struct cpufreq_interactive_cpuinfo *pcpu;
 535         int cpu;
 536
 537         if (val == CPUFREQ_POSTCHANGE) {
 538                 pcpu = &per_cpu(cpuinfo, freq->cpu);
 539
 540                 for_each_cpu(cpu, pcpu->policy->cpus) {
 541                         struct cpufreq_interactive_cpuinfo *pjcpu =
 542                                 &per_cpu(cpuinfo, cpu);
 543                         spin_lock(&pjcpu->load_lock);
 544                         update_load(cpu);
 545                         spin_unlock(&pjcpu->load_lock);
 546                 }
 547         }
 548
 549         return 0;
 550 }
 551
 552 static struct notifier_block cpufreq_notifier_block = {
 553         .notifier_call = cpufreq_interactive_notifier,
 554 };
 555
 556 static ssize_t show_target_loads(
 557         struct kobject *kobj, struct attribute *attr, char *buf)
 558 {
 559         int i;
 560         ssize_t ret = 0;
 561
 562         spin_lock(&target_loads_lock);
 563
 564         for (i = 0; i < ntarget_loads; i++)
 565                 ret += sprintf(buf + ret, "%u%s", target_loads[i],
 566                                i & 0x1 ? ":" : " ");
 567
 568         ret += sprintf(buf + ret, "\n");
 569         spin_unlock(&target_loads_lock);
 570         return ret;
 571 }
 572
 573 static ssize_t store_target_loads(
 574         struct kobject *kobj, struct attribute *attr, const char *buf,
 575         size_t count)
 576 {
 577         int ret;
 578         const char *cp;
 579         unsigned int *new_target_loads = NULL;
 580         int ntokens = 1;
 581         int i;
 582
 583         cp = buf;
 584         while ((cp = strpbrk(cp + 1, " :")))
 585                 ntokens++;
 586
 587         if (!(ntokens & 0x1))
 588                 goto err_inval;
 589
 590         new_target_loads = kmalloc(ntokens * sizeof(unsigned int), GFP_KERNEL);
 591         if (!new_target_loads) {
 592                 ret = -ENOMEM;
 593                 goto err;
 594         }
 595
 596         cp = buf;
 597         i = 0;
 598         while (i < ntokens) {
 599                 if (sscanf(cp, "%u", &new_target_loads[i++]) != 1)
 600                         goto err_inval;
 601
 602                 cp = strpbrk(cp, " :");
 603                 if (!cp)
 604                         break;
 605                 cp++;
 606         }
 607
 608         if (i != ntokens)
 609                 goto err_inval;
 610
 611         spin_lock(&target_loads_lock);
 612         if (target_loads != default_target_loads)
 613                 kfree(target_loads);
 614         target_loads = new_target_loads;
 615         ntarget_loads = ntokens;
 616         spin_unlock(&target_loads_lock);
 617         return count;
 618
 619 err_inval:
 620         ret = -EINVAL;
 621 err:
 622         kfree(new_target_loads);
 623         return ret;
 624 }
 625
 626 static struct global_attr target_loads_attr =
 627         __ATTR(target_loads, S_IRUGO | S_IWUSR,
 628                 show_target_loads, store_target_loads);
 629
 630 static ssize_t show_hispeed_freq(struct kobject *kobj,
 631                                  struct attribute *attr, char *buf)
 632 {
 633         return sprintf(buf, "%u\n", hispeed_freq);
 634 }
 635
 636 static ssize_t store_hispeed_freq(struct kobject *kobj,
 637                                   struct attribute *attr, const char *buf,
 638                                   size_t count)
 639 {
 640         int ret;
 641         long unsigned int val;
 642
 643         ret = strict_strtoul(buf, 0, &val);
 644         if (ret < 0)
 645                 return ret;
 646         hispeed_freq = val;
 647         return count;
 648 }
 649
 650 static struct global_attr hispeed_freq_attr = __ATTR(hispeed_freq, 0644,
 651                 show_hispeed_freq, store_hispeed_freq);
 652
 653
 654 static ssize_t show_go_hispeed_load(struct kobject *kobj,
 655                                      struct attribute *attr, char *buf)
 656 {
 657         return sprintf(buf, "%lu\n", go_hispeed_load);
 658 }
 659
 660 static ssize_t store_go_hispeed_load(struct kobject *kobj,
 661                         struct attribute *attr, const char *buf, size_t count)
 662 {
 663         int ret;
 664         unsigned long val;
 665
 666         ret = strict_strtoul(buf, 0, &val);
 667         if (ret < 0)
 668                 return ret;
 669         go_hispeed_load = val;
 670         return count;
 671 }
 672
 673 static struct global_attr go_hispeed_load_attr = __ATTR(go_hispeed_load, 0644,
 674                 show_go_hispeed_load, store_go_hispeed_load);
 675
 676 static ssize_t show_min_sample_time(struct kobject *kobj,
 677                                 struct attribute *attr, char *buf)
 678 {
 679         return sprintf(buf, "%lu\n", min_sample_time);
 680 }
 681
 682 static ssize_t store_min_sample_time(struct kobject *kobj,
 683                         struct attribute *attr, const char *buf, size_t count)
 684 {
 685         int ret;
 686         unsigned long val;
 687
 688         ret = strict_strtoul(buf, 0, &val);
 689         if (ret < 0)
 690                 return ret;
 691         min_sample_time = val;
 692         return count;
 693 }
 694
 695 static struct global_attr min_sample_time_attr = __ATTR(min_sample_time, 0644,
 696                 show_min_sample_time, store_min_sample_time);
 697
 698 static ssize_t show_above_hispeed_delay(struct kobject *kobj,
 699                                         struct attribute *attr, char *buf)
 700 {
 701         return sprintf(buf, "%lu\n", above_hispeed_delay_val);
 702 }
 703
 704 static ssize_t store_above_hispeed_delay(struct kobject *kobj,
 705                                          struct attribute *attr,
 706                                          const char *buf, size_t count)
 707 {
 708         int ret;
 709         unsigned long val;
 710
 711         ret = strict_strtoul(buf, 0, &val);
 712         if (ret < 0)
 713                 return ret;
 714         above_hispeed_delay_val = val;
 715         return count;
 716 }
 717
 718 define_one_global_rw(above_hispeed_delay);
 719
 720 static ssize_t show_timer_rate(struct kobject *kobj,
 721                         struct attribute *attr, char *buf)
 722 {
 723         return sprintf(buf, "%lu\n", timer_rate);
 724 }
 725
 726 static ssize_t store_timer_rate(struct kobject *kobj,
 727                         struct attribute *attr, const char *buf, size_t count)
 728 {
 729         int ret;
 730         unsigned long val;
 731
 732         ret = strict_strtoul(buf, 0, &val);
 733         if (ret < 0)
 734                 return ret;
 735         timer_rate = val;
 736         return count;
 737 }
 738
 739 static struct global_attr timer_rate_attr = __ATTR(timer_rate, 0644,
 740                 show_timer_rate, store_timer_rate);
 741
 742 static ssize_t show_timer_slack(
 743         struct kobject *kobj, struct attribute *attr, char *buf)
 744 {
 745         return sprintf(buf, "%d\n", timer_slack_val);
 746 }
 747
 748 static ssize_t store_timer_slack(
 749         struct kobject *kobj, struct attribute *attr, const char *buf,
 750         size_t count)
 751 {
 752         int ret;
 753         unsigned long val;
 754
 755         ret = kstrtol(buf, 10, &val);
 756         if (ret < 0)
 757                 return ret;
 758
 759         timer_slack_val = val;
 760         return count;
 761 }
 762
 763 define_one_global_rw(timer_slack);
 764
 765 static ssize_t show_boost(struct kobject *kobj, struct attribute *attr,
 766                           char *buf)
 767 {
 768         return sprintf(buf, "%d\n", boost_val);
 769 }
 770
 771 static ssize_t store_boost(struct kobject *kobj, struct attribute *attr,
 772                            const char *buf, size_t count)
 773 {
 774         int ret;
 775         unsigned long val;
 776
 777         ret = kstrtoul(buf, 0, &val);
 778         if (ret < 0)
 779                 return ret;
 780
 781         boost_val = val;
 782
 783         if (boost_val) {
 784                 trace_cpufreq_interactive_boost("on");
 785                 cpufreq_interactive_boost();
 786         } else {
 787                 trace_cpufreq_interactive_unboost("off");
 788         }
 789
 790         return count;
 791 }
 792
 793 define_one_global_rw(boost);
 794
 795 static ssize_t store_boostpulse(struct kobject *kobj, struct attribute *attr,
 796                                 const char *buf, size_t count)
 797 {
 798         int ret;
 799         unsigned long val;
 800
 801         ret = kstrtoul(buf, 0, &val);
 802         if (ret < 0)
 803                 return ret;
 804
 805         boostpulse_endtime = ktime_to_us(ktime_get()) + boostpulse_duration_val;
 806         trace_cpufreq_interactive_boost("pulse");
 807         cpufreq_interactive_boost();
 808         return count;
 809 }
 810
 811 static struct global_attr boostpulse =
 812         __ATTR(boostpulse, 0200, NULL, store_boostpulse);
 813
 814 static ssize_t show_boostpulse_duration(
 815         struct kobject *kobj, struct attribute *attr, char *buf)
 816 {
 817         return sprintf(buf, "%d\n", boostpulse_duration_val);
 818 }
 819
 820 static ssize_t store_boostpulse_duration(
 821         struct kobject *kobj, struct attribute *attr, const char *buf,
 822         size_t count)
 823 {
 824         int ret;
 825         unsigned long val;
 826
 827         ret = kstrtoul(buf, 0, &val);
 828         if (ret < 0)
 829                 return ret;
 830
 831         boostpulse_duration_val = val;
 832         return count;
 833 }
 834
 835 define_one_global_rw(boostpulse_duration);
 836
 837 static struct attribute *interactive_attributes[] = {
 838         &target_loads_attr.attr,
 839         &hispeed_freq_attr.attr,
 840         &go_hispeed_load_attr.attr,
 841         &above_hispeed_delay.attr,
 842         &min_sample_time_attr.attr,
 843         &timer_rate_attr.attr,
 844         &timer_slack.attr,
 845         &boost.attr,
 846         &boostpulse.attr,
 847         &boostpulse_duration.attr,
 848         NULL,
 849 };
 850
 851 static struct attribute_group interactive_attr_group = {
 852         .attrs = interactive_attributes,
 853         .name = "interactive",
 854 };
 855
 856 static int cpufreq_interactive_idle_notifier(struct notifier_block *nb,
 857                                              unsigned long val,
 858                                              void *data)
 859 {
 860         switch (val) {
 861         case IDLE_START:
 862                 cpufreq_interactive_idle_start();
 863                 break;
 864         case IDLE_END:
 865                 cpufreq_interactive_idle_end();
 866                 break;
 867         }
 868
 869         return 0;
 870 }
 871
 872 static struct notifier_block cpufreq_interactive_idle_nb = {
 873         .notifier_call = cpufreq_interactive_idle_notifier,
 874 };
 875
 876 static int cpufreq_governor_interactive(struct cpufreq_policy *policy,
 877                 unsigned int event)
 878 {
 879         int rc;
 880         unsigned int j;
 881         struct cpufreq_interactive_cpuinfo *pcpu;
 882         struct cpufreq_frequency_table *freq_table;
 883
 884         switch (event) {
 885         case CPUFREQ_GOV_START:
 886                 if (!cpu_online(policy->cpu))
 887                         return -EINVAL;
 888
 889                 freq_table =
 890                         cpufreq_frequency_get_table(policy->cpu);
 891                 if (!hispeed_freq)
 892                         hispeed_freq = policy->max;
 893
 894                 for_each_cpu(j, policy->cpus) {
 895                         unsigned long expires;
 896
 897                         pcpu = &per_cpu(cpuinfo, j);
 898                         pcpu->policy = policy;
 899                         pcpu->target_freq = policy->cur;
 900                         pcpu->freq_table = freq_table;
 901                         pcpu->floor_freq = pcpu->target_freq;
 902                         pcpu->floor_validate_time =
 903                                 ktime_to_us(ktime_get());
 904                         pcpu->hispeed_validate_time =
 905                                 pcpu->floor_validate_time;
 906                         pcpu->governor_enabled = 1;
 907                         smp_wmb();
 908                         expires = jiffies + usecs_to_jiffies(timer_rate);
 909                         pcpu->cpu_timer.expires = expires;
 910                         add_timer_on(&pcpu->cpu_timer, j);
 911
 912                         if (timer_slack_val >= 0) {
 913                                 expires += usecs_to_jiffies(timer_slack_val);
 914                                 pcpu->cpu_slack_timer.expires = expires;
 915                                 add_timer_on(&pcpu->cpu_slack_timer, j);
 916                         }
 917                 }
 918
 919                 /*
 920                  * Do not register the idle hook and create sysfs
 921                  * entries if we have already done so.
 922                  */
 923                 if (atomic_inc_return(&active_count) > 1)
 924                         return 0;
 925
 926                 rc = sysfs_create_group(cpufreq_global_kobject,
 927                                 &interactive_attr_group);
 928                 if (rc)
 929                         return rc;
 930
 931                 idle_notifier_register(&cpufreq_interactive_idle_nb);
 932                 cpufreq_register_notifier(
 933                         &cpufreq_notifier_block, CPUFREQ_TRANSITION_NOTIFIER);
 934                 break;
 935
 936         case CPUFREQ_GOV_STOP:
 937                 for_each_cpu(j, policy->cpus) {
 938                         pcpu = &per_cpu(cpuinfo, j);
 939                         pcpu->governor_enabled = 0;
 940                         smp_wmb();
 941                         del_timer_sync(&pcpu->cpu_timer);
 942                         del_timer_sync(&pcpu->cpu_slack_timer);
 943                 }
 944
 945                 if (atomic_dec_return(&active_count) > 0)
 946                         return 0;
 947
 948                 cpufreq_unregister_notifier(
 949                         &cpufreq_notifier_block, CPUFREQ_TRANSITION_NOTIFIER);
 950                 idle_notifier_unregister(&cpufreq_interactive_idle_nb);
 951                 sysfs_remove_group(cpufreq_global_kobject,
 952                                 &interactive_attr_group);
 953
 954                 break;
 955
 956         case CPUFREQ_GOV_LIMITS:
 957                 if (policy->max < policy->cur)
 958                         __cpufreq_driver_target(policy,
 959                                         policy->max, CPUFREQ_RELATION_H);
 960                 else if (policy->min > policy->cur)
 961                         __cpufreq_driver_target(policy,
 962                                         policy->min, CPUFREQ_RELATION_L);
 963                 break;
 964         }
 965         return 0;
 966 }
 967
 968 static void cpufreq_interactive_nop_timer(unsigned long data)
 969 {
 970 }
 971
 972 static int __init cpufreq_interactive_init(void)
 973 {
 974         unsigned int i;
 975         struct cpufreq_interactive_cpuinfo *pcpu;
 976         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
 977
 978         go_hispeed_load = DEFAULT_GO_HISPEED_LOAD;
 979         min_sample_time = DEFAULT_MIN_SAMPLE_TIME;
 980         above_hispeed_delay_val = DEFAULT_ABOVE_HISPEED_DELAY;
 981         timer_rate = DEFAULT_TIMER_RATE;
 982
 983         /* Initalize per-cpu timers */
 984         for_each_possible_cpu(i) {
 985                 pcpu = &per_cpu(cpuinfo, i);
 986                 init_timer_deferrable(&pcpu->cpu_timer);
 987                 pcpu->cpu_timer.function = cpufreq_interactive_timer;
 988                 pcpu->cpu_timer.data = i;
 989                 init_timer(&pcpu->cpu_slack_timer);
 990                 pcpu->cpu_slack_timer.function = cpufreq_interactive_nop_timer;
 991                 spin_lock_init(&pcpu->load_lock);
 992         }
 993
 994         spin_lock_init(&target_loads_lock);
 995         spin_lock_init(&speedchange_cpumask_lock);
 996         speedchange_task =
 997                 kthread_create(cpufreq_interactive_speedchange_task, NULL,
 998                                "cfinteractive");
 999         if (IS_ERR(speedchange_task))
1000                 return PTR_ERR(speedchange_task);
1001
1002         sched_setscheduler_nocheck(speedchange_task, SCHED_FIFO, &param);
1003         get_task_struct(speedchange_task);
1004
1005         /* NB: wake up so the thread does not look hung to the freezer */
1006         wake_up_process(speedchange_task);
1007
1008         return cpufreq_register_governor(&cpufreq_gov_interactive);
1009 }
1010
1011 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE
1012 fs_initcall(cpufreq_interactive_init);
1013 #else
1014 module_init(cpufreq_interactive_init);
1015 #endif
1016
1017 static void __exit cpufreq_interactive_exit(void)
1018 {
1019         cpufreq_unregister_governor(&cpufreq_gov_interactive);
1020         kthread_stop(speedchange_task);
1021         put_task_struct(speedchange_task);
1022 }
1023
1024 module_exit(cpufreq_interactive_exit);
1025
1026 MODULE_AUTHOR("Mike Chan <mike@android.com>");
1027 MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for "
1028         "Latency sensitive workloads");
1029 MODULE_LICENSE("GPL");