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