kernel/sched/walt.c

   1 /*
   2  * Copyright (c) 2016, The Linux Foundation. All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 and
   6  * only version 2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  11  * GNU General Public License for more details.
  12  *
  13  *
  14  * Window Assisted Load Tracking (WALT) implementation credits:
  15  * Srivatsa Vaddagiri, Steve Muckle, Syed Rameez Mustafa, Joonwoo Park,
  16  * Pavan Kumar Kondeti, Olav Haugan
  17  *
  18  * 2016-03-06: Integration with EAS/refactoring by Vikram Mulukutla
  19  *             and Todd Kjos
  20  */
  21
  22 #include <linux/syscore_ops.h>
  23 #include <linux/cpufreq.h>
  24 #include <trace/events/sched.h>
  25 #include "sched.h"
  26 #include "walt.h"
  27
  28 #define WINDOW_STATS_RECENT             0
  29 #define WINDOW_STATS_MAX                1
  30 #define WINDOW_STATS_MAX_RECENT_AVG     2
  31 #define WINDOW_STATS_AVG                3
  32 #define WINDOW_STATS_INVALID_POLICY     4
  33
  34 #define EXITING_TASK_MARKER     0xdeaddead
  35
  36 static __read_mostly unsigned int walt_ravg_hist_size = 5;
  37 static __read_mostly unsigned int walt_window_stats_policy =
  38         WINDOW_STATS_MAX_RECENT_AVG;
  39 static __read_mostly unsigned int walt_account_wait_time = 1;
  40 static __read_mostly unsigned int walt_freq_account_wait_time = 0;
  41 static __read_mostly unsigned int walt_io_is_busy = 0;
  42
  43 unsigned int sysctl_sched_walt_init_task_load_pct = 15;
  44
  45 /* 1 -> use PELT based load stats, 0 -> use window-based load stats */
  46 unsigned int __read_mostly walt_disabled = 0;
  47
  48 static unsigned int max_possible_efficiency = 1024;
  49 static unsigned int min_possible_efficiency = 1024;
  50
  51 /*
  52  * Maximum possible frequency across all cpus. Task demand and cpu
  53  * capacity (cpu_power) metrics are scaled in reference to it.
  54  */
  55 static unsigned int max_possible_freq = 1;
  56
  57 /*
  58  * Minimum possible max_freq across all cpus. This will be same as
  59  * max_possible_freq on homogeneous systems and could be different from
  60  * max_possible_freq on heterogenous systems. min_max_freq is used to derive
  61  * capacity (cpu_power) of cpus.
  62  */
  63 static unsigned int min_max_freq = 1;
  64
  65 static unsigned int max_capacity = 1024;
  66 static unsigned int min_capacity = 1024;
  67 static unsigned int max_load_scale_factor = 1024;
  68 static unsigned int max_possible_capacity = 1024;
  69
  70 /* Mask of all CPUs that have  max_possible_capacity */
  71 static cpumask_t mpc_mask = CPU_MASK_ALL;
  72
  73 /* Window size (in ns) */
  74 __read_mostly unsigned int walt_ravg_window = 20000000;
  75
  76 /* Min window size (in ns) = 10ms */
  77 #define MIN_SCHED_RAVG_WINDOW 10000000
  78
  79 /* Max window size (in ns) = 1s */
  80 #define MAX_SCHED_RAVG_WINDOW 1000000000
  81
  82 static unsigned int sync_cpu;
  83 static ktime_t ktime_last;
  84 static bool walt_ktime_suspended;
  85
  86 static unsigned int task_load(struct task_struct *p)
  87 {
  88         return p->ravg.demand;
  89 }
  90
  91 void
  92 walt_inc_cumulative_runnable_avg(struct rq *rq,
  93                                  struct task_struct *p)
  94 {
  95         rq->cumulative_runnable_avg += p->ravg.demand;
  96 }
  97
  98 void
  99 walt_dec_cumulative_runnable_avg(struct rq *rq,
 100                                  struct task_struct *p)
 101 {
 102         rq->cumulative_runnable_avg -= p->ravg.demand;
 103         BUG_ON((s64)rq->cumulative_runnable_avg < 0);
 104 }
 105
 106 static void
 107 fixup_cumulative_runnable_avg(struct rq *rq,
 108                               struct task_struct *p, s64 task_load_delta)
 109 {
 110         rq->cumulative_runnable_avg += task_load_delta;
 111         if ((s64)rq->cumulative_runnable_avg < 0)
 112                 panic("cra less than zero: tld: %lld, task_load(p) = %u\n",
 113                         task_load_delta, task_load(p));
 114 }
 115
 116 u64 walt_ktime_clock(void)
 117 {
 118         if (unlikely(walt_ktime_suspended))
 119                 return ktime_to_ns(ktime_last);
 120         return ktime_get_ns();
 121 }
 122
 123 static void walt_resume(void)
 124 {
 125         walt_ktime_suspended = false;
 126 }
 127
 128 static int walt_suspend(void)
 129 {
 130         ktime_last = ktime_get();
 131         walt_ktime_suspended = true;
 132         return 0;
 133 }
 134
 135 static struct syscore_ops walt_syscore_ops = {
 136         .resume = walt_resume,
 137         .suspend = walt_suspend
 138 };
 139
 140 static int __init walt_init_ops(void)
 141 {
 142         register_syscore_ops(&walt_syscore_ops);
 143         return 0;
 144 }
 145 late_initcall(walt_init_ops);
 146
 147 void walt_inc_cfs_cumulative_runnable_avg(struct cfs_rq *cfs_rq,
 148                 struct task_struct *p)
 149 {
 150         cfs_rq->cumulative_runnable_avg += p->ravg.demand;
 151 }
 152
 153 void walt_dec_cfs_cumulative_runnable_avg(struct cfs_rq *cfs_rq,
 154                 struct task_struct *p)
 155 {
 156         cfs_rq->cumulative_runnable_avg -= p->ravg.demand;
 157 }
 158
 159 static int exiting_task(struct task_struct *p)
 160 {
 161         if (p->flags & PF_EXITING) {
 162                 if (p->ravg.sum_history[0] != EXITING_TASK_MARKER) {
 163                         p->ravg.sum_history[0] = EXITING_TASK_MARKER;
 164                 }
 165                 return 1;
 166         }
 167         return 0;
 168 }
 169
 170 static int __init set_walt_ravg_window(char *str)
 171 {
 172         get_option(&str, &walt_ravg_window);
 173
 174         walt_disabled = (walt_ravg_window < MIN_SCHED_RAVG_WINDOW ||
 175                                 walt_ravg_window > MAX_SCHED_RAVG_WINDOW);
 176         return 0;
 177 }
 178
 179 early_param("walt_ravg_window", set_walt_ravg_window);
 180
 181 static void
 182 update_window_start(struct rq *rq, u64 wallclock)
 183 {
 184         s64 delta;
 185         int nr_windows;
 186
 187         delta = wallclock - rq->window_start;
 188         /* If the MPM global timer is cleared, set delta as 0 to avoid kernel BUG happening */
 189         if (delta < 0) {
 190                 delta = 0;
 191                 WARN_ONCE(1, "WALT wallclock appears to have gone backwards or reset\n");
 192         }
 193
 194         if (delta < walt_ravg_window)
 195                 return;
 196
 197         nr_windows = div64_u64(delta, walt_ravg_window);
 198         rq->window_start += (u64)nr_windows * (u64)walt_ravg_window;
 199 }
 200
 201 static u64 scale_exec_time(u64 delta, struct rq *rq)
 202 {
 203         unsigned int cur_freq = rq->cur_freq;
 204         int sf;
 205
 206         if (unlikely(cur_freq > max_possible_freq))
 207                 cur_freq = rq->max_possible_freq;
 208
 209         /* round up div64 */
 210         delta = div64_u64(delta * cur_freq + max_possible_freq - 1,
 211                           max_possible_freq);
 212
 213         sf = DIV_ROUND_UP(rq->efficiency * 1024, max_possible_efficiency);
 214
 215         delta *= sf;
 216         delta >>= 10;
 217
 218         return delta;
 219 }
 220
 221 static int cpu_is_waiting_on_io(struct rq *rq)
 222 {
 223         if (!walt_io_is_busy)
 224                 return 0;
 225
 226         return atomic_read(&rq->nr_iowait);
 227 }
 228
 229 void walt_account_irqtime(int cpu, struct task_struct *curr,
 230                                  u64 delta, u64 wallclock)
 231 {
 232         struct rq *rq = cpu_rq(cpu);
 233         unsigned long flags, nr_windows;
 234         u64 cur_jiffies_ts;
 235
 236         raw_spin_lock_irqsave(&rq->lock, flags);
 237
 238         /*
 239          * cputime (wallclock) uses sched_clock so use the same here for
 240          * consistency.
 241          */
 242         delta += sched_clock() - wallclock;
 243         cur_jiffies_ts = get_jiffies_64();
 244
 245         if (is_idle_task(curr))
 246                 walt_update_task_ravg(curr, rq, IRQ_UPDATE, walt_ktime_clock(),
 247                                  delta);
 248
 249         nr_windows = cur_jiffies_ts - rq->irqload_ts;
 250
 251         if (nr_windows) {
 252                 if (nr_windows < 10) {
 253                         /* Decay CPU's irqload by 3/4 for each window. */
 254                         rq->avg_irqload *= (3 * nr_windows);
 255                         rq->avg_irqload = div64_u64(rq->avg_irqload,
 256                                                     4 * nr_windows);
 257                 } else {
 258                         rq->avg_irqload = 0;
 259                 }
 260                 rq->avg_irqload += rq->cur_irqload;
 261                 rq->cur_irqload = 0;
 262         }
 263
 264         rq->cur_irqload += delta;
 265         rq->irqload_ts = cur_jiffies_ts;
 266         raw_spin_unlock_irqrestore(&rq->lock, flags);
 267 }
 268
 269
 270 #define WALT_HIGH_IRQ_TIMEOUT 3
 271
 272 u64 walt_irqload(int cpu) {
 273         struct rq *rq = cpu_rq(cpu);
 274         s64 delta;
 275         delta = get_jiffies_64() - rq->irqload_ts;
 276
 277         /*
 278          * Current context can be preempted by irq and rq->irqload_ts can be
 279          * updated by irq context so that delta can be negative.
 280          * But this is okay and we can safely return as this means there
 281          * was recent irq occurrence.
 282          */
 283
 284         if (delta < WALT_HIGH_IRQ_TIMEOUT)
 285                 return rq->avg_irqload;
 286         else
 287                 return 0;
 288 }
 289
 290 int walt_cpu_high_irqload(int cpu) {
 291         return walt_irqload(cpu) >= sysctl_sched_walt_cpu_high_irqload;
 292 }
 293
 294 static int account_busy_for_cpu_time(struct rq *rq, struct task_struct *p,
 295                                      u64 irqtime, int event)
 296 {
 297         if (is_idle_task(p)) {
 298                 /* TASK_WAKE && TASK_MIGRATE is not possible on idle task! */
 299                 if (event == PICK_NEXT_TASK)
 300                         return 0;
 301
 302                 /* PUT_PREV_TASK, TASK_UPDATE && IRQ_UPDATE are left */
 303                 return irqtime || cpu_is_waiting_on_io(rq);
 304         }
 305
 306         if (event == TASK_WAKE)
 307                 return 0;
 308
 309         if (event == PUT_PREV_TASK || event == IRQ_UPDATE ||
 310                                          event == TASK_UPDATE)
 311                 return 1;
 312
 313         /* Only TASK_MIGRATE && PICK_NEXT_TASK left */
 314         return walt_freq_account_wait_time;
 315 }
 316
 317 /*
 318  * Account cpu activity in its busy time counters (rq->curr/prev_runnable_sum)
 319  */
 320 static void update_cpu_busy_time(struct task_struct *p, struct rq *rq,
 321              int event, u64 wallclock, u64 irqtime)
 322 {
 323         int new_window, nr_full_windows = 0;
 324         int p_is_curr_task = (p == rq->curr);
 325         u64 mark_start = p->ravg.mark_start;
 326         u64 window_start = rq->window_start;
 327         u32 window_size = walt_ravg_window;
 328         u64 delta;
 329
 330         new_window = mark_start < window_start;
 331         if (new_window) {
 332                 nr_full_windows = div64_u64((window_start - mark_start),
 333                                                 window_size);
 334                 if (p->ravg.active_windows < USHRT_MAX)
 335                         p->ravg.active_windows++;
 336         }
 337
 338         /* Handle per-task window rollover. We don't care about the idle
 339          * task or exiting tasks. */
 340         if (new_window && !is_idle_task(p) && !exiting_task(p)) {
 341                 u32 curr_window = 0;
 342
 343                 if (!nr_full_windows)
 344                         curr_window = p->ravg.curr_window;
 345
 346                 p->ravg.prev_window = curr_window;
 347                 p->ravg.curr_window = 0;
 348         }
 349
 350         if (!account_busy_for_cpu_time(rq, p, irqtime, event)) {
 351                 /* account_busy_for_cpu_time() = 0, so no update to the
 352                  * task's current window needs to be made. This could be
 353                  * for example
 354                  *
 355                  *   - a wakeup event on a task within the current
 356                  *     window (!new_window below, no action required),
 357                  *   - switching to a new task from idle (PICK_NEXT_TASK)
 358                  *     in a new window where irqtime is 0 and we aren't
 359                  *     waiting on IO */
 360
 361                 if (!new_window)
 362                         return;
 363
 364                 /* A new window has started. The RQ demand must be rolled
 365                  * over if p is the current task. */
 366                 if (p_is_curr_task) {
 367                         u64 prev_sum = 0;
 368
 369                         /* p is either idle task or an exiting task */
 370                         if (!nr_full_windows) {
 371                                 prev_sum = rq->curr_runnable_sum;
 372                         }
 373
 374                         rq->prev_runnable_sum = prev_sum;
 375                         rq->curr_runnable_sum = 0;
 376                 }
 377
 378                 return;
 379         }
 380
 381         if (!new_window) {
 382                 /* account_busy_for_cpu_time() = 1 so busy time needs
 383                  * to be accounted to the current window. No rollover
 384                  * since we didn't start a new window. An example of this is
 385                  * when a task starts execution and then sleeps within the
 386                  * same window. */
 387
 388                 if (!irqtime || !is_idle_task(p) || cpu_is_waiting_on_io(rq))
 389                         delta = wallclock - mark_start;
 390                 else
 391                         delta = irqtime;
 392                 delta = scale_exec_time(delta, rq);
 393                 rq->curr_runnable_sum += delta;
 394                 if (!is_idle_task(p) && !exiting_task(p))
 395                         p->ravg.curr_window += delta;
 396
 397                 return;
 398         }
 399
 400         if (!p_is_curr_task) {
 401                 /* account_busy_for_cpu_time() = 1 so busy time needs
 402                  * to be accounted to the current window. A new window
 403                  * has also started, but p is not the current task, so the
 404                  * window is not rolled over - just split up and account
 405                  * as necessary into curr and prev. The window is only
 406                  * rolled over when a new window is processed for the current
 407                  * task.
 408                  *
 409                  * Irqtime can't be accounted by a task that isn't the
 410                  * currently running task. */
 411
 412                 if (!nr_full_windows) {
 413                         /* A full window hasn't elapsed, account partial
 414                          * contribution to previous completed window. */
 415                         delta = scale_exec_time(window_start - mark_start, rq);
 416                         if (!exiting_task(p))
 417                                 p->ravg.prev_window += delta;
 418                 } else {
 419                         /* Since at least one full window has elapsed,
 420                          * the contribution to the previous window is the
 421                          * full window (window_size). */
 422                         delta = scale_exec_time(window_size, rq);
 423                         if (!exiting_task(p))
 424                                 p->ravg.prev_window = delta;
 425                 }
 426                 rq->prev_runnable_sum += delta;
 427
 428                 /* Account piece of busy time in the current window. */
 429                 delta = scale_exec_time(wallclock - window_start, rq);
 430                 rq->curr_runnable_sum += delta;
 431                 if (!exiting_task(p))
 432                         p->ravg.curr_window = delta;
 433
 434                 return;
 435         }
 436
 437         if (!irqtime || !is_idle_task(p) || cpu_is_waiting_on_io(rq)) {
 438                 /* account_busy_for_cpu_time() = 1 so busy time needs
 439                  * to be accounted to the current window. A new window
 440                  * has started and p is the current task so rollover is
 441                  * needed. If any of these three above conditions are true
 442                  * then this busy time can't be accounted as irqtime.
 443                  *
 444                  * Busy time for the idle task or exiting tasks need not
 445                  * be accounted.
 446                  *
 447                  * An example of this would be a task that starts execution
 448                  * and then sleeps once a new window has begun. */
 449
 450                 if (!nr_full_windows) {
 451                         /* A full window hasn't elapsed, account partial
 452                          * contribution to previous completed window. */
 453                         delta = scale_exec_time(window_start - mark_start, rq);
 454                         if (!is_idle_task(p) && !exiting_task(p))
 455                                 p->ravg.prev_window += delta;
 456
 457                         delta += rq->curr_runnable_sum;
 458                 } else {
 459                         /* Since at least one full window has elapsed,
 460                          * the contribution to the previous window is the
 461                          * full window (window_size). */
 462                         delta = scale_exec_time(window_size, rq);
 463                         if (!is_idle_task(p) && !exiting_task(p))
 464                                 p->ravg.prev_window = delta;
 465
 466                 }
 467                 /*
 468                  * Rollover for normal runnable sum is done here by overwriting
 469                  * the values in prev_runnable_sum and curr_runnable_sum.
 470                  * Rollover for new task runnable sum has completed by previous
 471                  * if-else statement.
 472                  */
 473                 rq->prev_runnable_sum = delta;
 474
 475                 /* Account piece of busy time in the current window. */
 476                 delta = scale_exec_time(wallclock - window_start, rq);
 477                 rq->curr_runnable_sum = delta;
 478                 if (!is_idle_task(p) && !exiting_task(p))
 479                         p->ravg.curr_window = delta;
 480
 481                 return;
 482         }
 483
 484         if (irqtime) {
 485                 /* account_busy_for_cpu_time() = 1 so busy time needs
 486                  * to be accounted to the current window. A new window
 487                  * has started and p is the current task so rollover is
 488                  * needed. The current task must be the idle task because
 489                  * irqtime is not accounted for any other task.
 490                  *
 491                  * Irqtime will be accounted each time we process IRQ activity
 492                  * after a period of idleness, so we know the IRQ busy time
 493                  * started at wallclock - irqtime. */
 494
 495                 BUG_ON(!is_idle_task(p));
 496                 mark_start = wallclock - irqtime;
 497
 498                 /* Roll window over. If IRQ busy time was just in the current
 499                  * window then that is all that need be accounted. */
 500                 rq->prev_runnable_sum = rq->curr_runnable_sum;
 501                 if (mark_start > window_start) {
 502                         rq->curr_runnable_sum = scale_exec_time(irqtime, rq);
 503                         return;
 504                 }
 505
 506                 /* The IRQ busy time spanned multiple windows. Process the
 507                  * busy time preceding the current window start first. */
 508                 delta = window_start - mark_start;
 509                 if (delta > window_size)
 510                         delta = window_size;
 511                 delta = scale_exec_time(delta, rq);
 512                 rq->prev_runnable_sum += delta;
 513
 514                 /* Process the remaining IRQ busy time in the current window. */
 515                 delta = wallclock - window_start;
 516                 rq->curr_runnable_sum = scale_exec_time(delta, rq);
 517
 518                 return;
 519         }
 520
 521         BUG();
 522 }
 523
 524 static int account_busy_for_task_demand(struct task_struct *p, int event)
 525 {
 526         /* No need to bother updating task demand for exiting tasks
 527          * or the idle task. */
 528         if (exiting_task(p) || is_idle_task(p))
 529                 return 0;
 530
 531         /* When a task is waking up it is completing a segment of non-busy
 532          * time. Likewise, if wait time is not treated as busy time, then
 533          * when a task begins to run or is migrated, it is not running and
 534          * is completing a segment of non-busy time. */
 535         if (event == TASK_WAKE || (!walt_account_wait_time &&
 536                          (event == PICK_NEXT_TASK || event == TASK_MIGRATE)))
 537                 return 0;
 538
 539         return 1;
 540 }
 541
 542 /*
 543  * Called when new window is starting for a task, to record cpu usage over
 544  * recently concluded window(s). Normally 'samples' should be 1. It can be > 1
 545  * when, say, a real-time task runs without preemption for several windows at a
 546  * stretch.
 547  */
 548 static void update_history(struct rq *rq, struct task_struct *p,
 549                          u32 runtime, int samples, int event)
 550 {
 551         u32 *hist = &p->ravg.sum_history[0];
 552         int ridx, widx;
 553         u32 max = 0, avg, demand;
 554         u64 sum = 0;
 555
 556         /* Ignore windows where task had no activity */
 557         if (!runtime || is_idle_task(p) || exiting_task(p) || !samples)
 558                         goto done;
 559
 560         /* Push new 'runtime' value onto stack */
 561         widx = walt_ravg_hist_size - 1;
 562         ridx = widx - samples;
 563         for (; ridx >= 0; --widx, --ridx) {
 564                 hist[widx] = hist[ridx];
 565                 sum += hist[widx];
 566                 if (hist[widx] > max)
 567                         max = hist[widx];
 568         }
 569
 570         for (widx = 0; widx < samples && widx < walt_ravg_hist_size; widx++) {
 571                 hist[widx] = runtime;
 572                 sum += hist[widx];
 573                 if (hist[widx] > max)
 574                         max = hist[widx];
 575         }
 576
 577         p->ravg.sum = 0;
 578
 579         if (walt_window_stats_policy == WINDOW_STATS_RECENT) {
 580                 demand = runtime;
 581         } else if (walt_window_stats_policy == WINDOW_STATS_MAX) {
 582                 demand = max;
 583         } else {
 584                 avg = div64_u64(sum, walt_ravg_hist_size);
 585                 if (walt_window_stats_policy == WINDOW_STATS_AVG)
 586                         demand = avg;
 587                 else
 588                         demand = max(avg, runtime);
 589         }
 590
 591         /*
 592          * A throttled deadline sched class task gets dequeued without
 593          * changing p->on_rq. Since the dequeue decrements hmp stats
 594          * avoid decrementing it here again.
 595          */
 596         if (task_on_rq_queued(p) && (!task_has_dl_policy(p) ||
 597                                                 !p->dl.dl_throttled))
 598                 fixup_cumulative_runnable_avg(rq, p, demand);
 599
 600         p->ravg.demand = demand;
 601
 602 done:
 603         trace_walt_update_history(rq, p, runtime, samples, event);
 604         return;
 605 }
 606
 607 static void add_to_task_demand(struct rq *rq, struct task_struct *p,
 608                                 u64 delta)
 609 {
 610         delta = scale_exec_time(delta, rq);
 611         p->ravg.sum += delta;
 612         if (unlikely(p->ravg.sum > walt_ravg_window))
 613                 p->ravg.sum = walt_ravg_window;
 614 }
 615
 616 /*
 617  * Account cpu demand of task and/or update task's cpu demand history
 618  *
 619  * ms = p->ravg.mark_start;
 620  * wc = wallclock
 621  * ws = rq->window_start
 622  *
 623  * Three possibilities:
 624  *
 625  *      a) Task event is contained within one window.
 626  *              window_start < mark_start < wallclock
 627  *
 628  *              ws   ms  wc
 629  *              |    |   |
 630  *              V    V   V
 631  *              |---------------|
 632  *
 633  *      In this case, p->ravg.sum is updated *iff* event is appropriate
 634  *      (ex: event == PUT_PREV_TASK)
 635  *
 636  *      b) Task event spans two windows.
 637  *              mark_start < window_start < wallclock
 638  *
 639  *              ms   ws   wc
 640  *              |    |    |
 641  *              V    V    V
 642  *              -----|-------------------
 643  *
 644  *      In this case, p->ravg.sum is updated with (ws - ms) *iff* event
 645  *      is appropriate, then a new window sample is recorded followed
 646  *      by p->ravg.sum being set to (wc - ws) *iff* event is appropriate.
 647  *
 648  *      c) Task event spans more than two windows.
 649  *
 650  *              ms ws_tmp                          ws  wc
 651  *              |  |                               |   |
 652  *              V  V                               V   V
 653  *              ---|-------|-------|-------|-------|------
 654  *                 |                               |
 655  *                 |<------ nr_full_windows ------>|
 656  *
 657  *      In this case, p->ravg.sum is updated with (ws_tmp - ms) first *iff*
 658  *      event is appropriate, window sample of p->ravg.sum is recorded,
 659  *      'nr_full_window' samples of window_size is also recorded *iff*
 660  *      event is appropriate and finally p->ravg.sum is set to (wc - ws)
 661  *      *iff* event is appropriate.
 662  *
 663  * IMPORTANT : Leave p->ravg.mark_start unchanged, as update_cpu_busy_time()
 664  * depends on it!
 665  */
 666 static void update_task_demand(struct task_struct *p, struct rq *rq,
 667              int event, u64 wallclock)
 668 {
 669         u64 mark_start = p->ravg.mark_start;
 670         u64 delta, window_start = rq->window_start;
 671         int new_window, nr_full_windows;
 672         u32 window_size = walt_ravg_window;
 673
 674         new_window = mark_start < window_start;
 675         if (!account_busy_for_task_demand(p, event)) {
 676                 if (new_window)
 677                         /* If the time accounted isn't being accounted as
 678                          * busy time, and a new window started, only the
 679                          * previous window need be closed out with the
 680                          * pre-existing demand. Multiple windows may have
 681                          * elapsed, but since empty windows are dropped,
 682                          * it is not necessary to account those. */
 683                         update_history(rq, p, p->ravg.sum, 1, event);
 684                 return;
 685         }
 686
 687         if (!new_window) {
 688                 /* The simple case - busy time contained within the existing
 689                  * window. */
 690                 add_to_task_demand(rq, p, wallclock - mark_start);
 691                 return;
 692         }
 693
 694         /* Busy time spans at least two windows. Temporarily rewind
 695          * window_start to first window boundary after mark_start. */
 696         delta = window_start - mark_start;
 697         nr_full_windows = div64_u64(delta, window_size);
 698         window_start -= (u64)nr_full_windows * (u64)window_size;
 699
 700         /* Process (window_start - mark_start) first */
 701         add_to_task_demand(rq, p, window_start - mark_start);
 702
 703         /* Push new sample(s) into task's demand history */
 704         update_history(rq, p, p->ravg.sum, 1, event);
 705         if (nr_full_windows)
 706                 update_history(rq, p, scale_exec_time(window_size, rq),
 707                                nr_full_windows, event);
 708
 709         /* Roll window_start back to current to process any remainder
 710          * in current window. */
 711         window_start += (u64)nr_full_windows * (u64)window_size;
 712
 713         /* Process (wallclock - window_start) next */
 714         mark_start = window_start;
 715         add_to_task_demand(rq, p, wallclock - mark_start);
 716 }
 717
 718 /* Reflect task activity on its demand and cpu's busy time statistics */
 719 void walt_update_task_ravg(struct task_struct *p, struct rq *rq,
 720              int event, u64 wallclock, u64 irqtime)
 721 {
 722         if (walt_disabled || !rq->window_start)
 723                 return;
 724
 725         lockdep_assert_held(&rq->lock);
 726
 727         update_window_start(rq, wallclock);
 728
 729         if (!p->ravg.mark_start)
 730                 goto done;
 731
 732         update_task_demand(p, rq, event, wallclock);
 733         update_cpu_busy_time(p, rq, event, wallclock, irqtime);
 734
 735 done:
 736         trace_walt_update_task_ravg(p, rq, event, wallclock, irqtime);
 737
 738         p->ravg.mark_start = wallclock;
 739 }
 740
 741 unsigned long __weak arch_get_cpu_efficiency(int cpu)
 742 {
 743         return SCHED_LOAD_SCALE;
 744 }
 745
 746 void walt_init_cpu_efficiency(void)
 747 {
 748         int i, efficiency;
 749         unsigned int max = 0, min = UINT_MAX;
 750
 751         for_each_possible_cpu(i) {
 752                 efficiency = arch_get_cpu_efficiency(i);
 753                 cpu_rq(i)->efficiency = efficiency;
 754
 755                 if (efficiency > max)
 756                         max = efficiency;
 757                 if (efficiency < min)
 758                         min = efficiency;
 759         }
 760
 761         if (max)
 762                 max_possible_efficiency = max;
 763
 764         if (min)
 765                 min_possible_efficiency = min;
 766 }
 767
 768 static void reset_task_stats(struct task_struct *p)
 769 {
 770         u32 sum = 0;
 771
 772         if (exiting_task(p))
 773                 sum = EXITING_TASK_MARKER;
 774
 775         memset(&p->ravg, 0, sizeof(struct ravg));
 776         /* Retain EXITING_TASK marker */
 777         p->ravg.sum_history[0] = sum;
 778 }
 779
 780 void walt_mark_task_starting(struct task_struct *p)
 781 {
 782         u64 wallclock;
 783         struct rq *rq = task_rq(p);
 784
 785         if (!rq->window_start) {
 786                 reset_task_stats(p);
 787                 return;
 788         }
 789
 790         wallclock = walt_ktime_clock();
 791         p->ravg.mark_start = wallclock;
 792 }
 793
 794 void walt_set_window_start(struct rq *rq)
 795 {
 796         int cpu = cpu_of(rq);
 797         struct rq *sync_rq = cpu_rq(sync_cpu);
 798
 799         if (rq->window_start)
 800                 return;
 801
 802         if (cpu == sync_cpu) {
 803                 rq->window_start = walt_ktime_clock();
 804         } else {
 805                 raw_spin_unlock(&rq->lock);
 806                 double_rq_lock(rq, sync_rq);
 807                 rq->window_start = cpu_rq(sync_cpu)->window_start;
 808                 rq->curr_runnable_sum = rq->prev_runnable_sum = 0;
 809                 raw_spin_unlock(&sync_rq->lock);
 810         }
 811
 812         rq->curr->ravg.mark_start = rq->window_start;
 813 }
 814
 815 void walt_migrate_sync_cpu(int cpu)
 816 {
 817         if (cpu == sync_cpu)
 818                 sync_cpu = smp_processor_id();
 819 }
 820
 821 void walt_fixup_busy_time(struct task_struct *p, int new_cpu)
 822 {
 823         struct rq *src_rq = task_rq(p);
 824         struct rq *dest_rq = cpu_rq(new_cpu);
 825         u64 wallclock;
 826
 827         if (!p->on_rq && p->state != TASK_WAKING)
 828                 return;
 829
 830         if (exiting_task(p)) {
 831                 return;
 832         }
 833
 834         if (p->state == TASK_WAKING)
 835                 double_rq_lock(src_rq, dest_rq);
 836
 837         wallclock = walt_ktime_clock();
 838
 839         walt_update_task_ravg(task_rq(p)->curr, task_rq(p),
 840                         TASK_UPDATE, wallclock, 0);
 841         walt_update_task_ravg(dest_rq->curr, dest_rq,
 842                         TASK_UPDATE, wallclock, 0);
 843
 844         walt_update_task_ravg(p, task_rq(p), TASK_MIGRATE, wallclock, 0);
 845
 846         if (p->ravg.curr_window) {
 847                 src_rq->curr_runnable_sum -= p->ravg.curr_window;
 848                 dest_rq->curr_runnable_sum += p->ravg.curr_window;
 849         }
 850
 851         if (p->ravg.prev_window) {
 852                 src_rq->prev_runnable_sum -= p->ravg.prev_window;
 853                 dest_rq->prev_runnable_sum += p->ravg.prev_window;
 854         }
 855
 856         if ((s64)src_rq->prev_runnable_sum < 0) {
 857                 src_rq->prev_runnable_sum = 0;
 858                 WARN_ON(1);
 859         }
 860         if ((s64)src_rq->curr_runnable_sum < 0) {
 861                 src_rq->curr_runnable_sum = 0;
 862                 WARN_ON(1);
 863         }
 864
 865         trace_walt_migration_update_sum(src_rq, p);
 866         trace_walt_migration_update_sum(dest_rq, p);
 867
 868         if (p->state == TASK_WAKING)
 869                 double_rq_unlock(src_rq, dest_rq);
 870 }
 871
 872 /* Keep track of max/min capacity possible across CPUs "currently" */
 873 static void __update_min_max_capacity(void)
 874 {
 875         int i;
 876         int max = 0, min = INT_MAX;
 877
 878         for_each_online_cpu(i) {
 879                 if (cpu_rq(i)->capacity > max)
 880                         max = cpu_rq(i)->capacity;
 881                 if (cpu_rq(i)->capacity < min)
 882                         min = cpu_rq(i)->capacity;
 883         }
 884
 885         max_capacity = max;
 886         min_capacity = min;
 887 }
 888
 889 static void update_min_max_capacity(void)
 890 {
 891         unsigned long flags;
 892         int i;
 893
 894         local_irq_save(flags);
 895         for_each_possible_cpu(i)
 896                 raw_spin_lock(&cpu_rq(i)->lock);
 897
 898         __update_min_max_capacity();
 899
 900         for_each_possible_cpu(i)
 901                 raw_spin_unlock(&cpu_rq(i)->lock);
 902         local_irq_restore(flags);
 903 }
 904
 905 /*
 906  * Return 'capacity' of a cpu in reference to "least" efficient cpu, such that
 907  * least efficient cpu gets capacity of 1024
 908  */
 909 static unsigned long capacity_scale_cpu_efficiency(int cpu)
 910 {
 911         return (1024 * cpu_rq(cpu)->efficiency) / min_possible_efficiency;
 912 }
 913
 914 /*
 915  * Return 'capacity' of a cpu in reference to cpu with lowest max_freq
 916  * (min_max_freq), such that one with lowest max_freq gets capacity of 1024.
 917  */
 918 static unsigned long capacity_scale_cpu_freq(int cpu)
 919 {
 920         return (1024 * cpu_rq(cpu)->max_freq) / min_max_freq;
 921 }
 922
 923 /*
 924  * Return load_scale_factor of a cpu in reference to "most" efficient cpu, so
 925  * that "most" efficient cpu gets a load_scale_factor of 1
 926  */
 927 static unsigned long load_scale_cpu_efficiency(int cpu)
 928 {
 929         return DIV_ROUND_UP(1024 * max_possible_efficiency,
 930                             cpu_rq(cpu)->efficiency);
 931 }
 932
 933 /*
 934  * Return load_scale_factor of a cpu in reference to cpu with best max_freq
 935  * (max_possible_freq), so that one with best max_freq gets a load_scale_factor
 936  * of 1.
 937  */
 938 static unsigned long load_scale_cpu_freq(int cpu)
 939 {
 940         return DIV_ROUND_UP(1024 * max_possible_freq, cpu_rq(cpu)->max_freq);
 941 }
 942
 943 static int compute_capacity(int cpu)
 944 {
 945         int capacity = 1024;
 946
 947         capacity *= capacity_scale_cpu_efficiency(cpu);
 948         capacity >>= 10;
 949
 950         capacity *= capacity_scale_cpu_freq(cpu);
 951         capacity >>= 10;
 952
 953         return capacity;
 954 }
 955
 956 static int compute_load_scale_factor(int cpu)
 957 {
 958         int load_scale = 1024;
 959
 960         /*
 961          * load_scale_factor accounts for the fact that task load
 962          * is in reference to "best" performing cpu. Task's load will need to be
 963          * scaled (up) by a factor to determine suitability to be placed on a
 964          * (little) cpu.
 965          */
 966         load_scale *= load_scale_cpu_efficiency(cpu);
 967         load_scale >>= 10;
 968
 969         load_scale *= load_scale_cpu_freq(cpu);
 970         load_scale >>= 10;
 971
 972         return load_scale;
 973 }
 974
 975 static int cpufreq_notifier_policy(struct notifier_block *nb,
 976                 unsigned long val, void *data)
 977 {
 978         struct cpufreq_policy *policy = (struct cpufreq_policy *)data;
 979         int i, update_max = 0;
 980         u64 highest_mpc = 0, highest_mplsf = 0;
 981         const struct cpumask *cpus = policy->related_cpus;
 982         unsigned int orig_min_max_freq = min_max_freq;
 983         unsigned int orig_max_possible_freq = max_possible_freq;
 984         /* Initialized to policy->max in case policy->related_cpus is empty! */
 985         unsigned int orig_max_freq = policy->max;
 986
 987         if (val != CPUFREQ_NOTIFY && val != CPUFREQ_REMOVE_POLICY &&
 988                                                 val != CPUFREQ_CREATE_POLICY)
 989                 return 0;
 990
 991         if (val == CPUFREQ_REMOVE_POLICY || val == CPUFREQ_CREATE_POLICY) {
 992                 update_min_max_capacity();
 993                 return 0;
 994         }
 995
 996         for_each_cpu(i, policy->related_cpus) {
 997                 cpumask_copy(&cpu_rq(i)->freq_domain_cpumask,
 998                              policy->related_cpus);
 999                 orig_max_freq = cpu_rq(i)->max_freq;
1000                 cpu_rq(i)->min_freq = policy->min;
1001                 cpu_rq(i)->max_freq = policy->max;
1002                 cpu_rq(i)->cur_freq = policy->cur;
1003                 cpu_rq(i)->max_possible_freq = policy->cpuinfo.max_freq;
1004         }
1005
1006         max_possible_freq = max(max_possible_freq, policy->cpuinfo.max_freq);
1007         if (min_max_freq == 1)
1008                 min_max_freq = UINT_MAX;
1009         min_max_freq = min(min_max_freq, policy->cpuinfo.max_freq);
1010         BUG_ON(!min_max_freq);
1011         BUG_ON(!policy->max);
1012
1013         /* Changes to policy other than max_freq don't require any updates */
1014         if (orig_max_freq == policy->max)
1015                 return 0;
1016
1017         /*
1018          * A changed min_max_freq or max_possible_freq (possible during bootup)
1019          * needs to trigger re-computation of load_scale_factor and capacity for
1020          * all possible cpus (even those offline). It also needs to trigger
1021          * re-computation of nr_big_task count on all online cpus.
1022          *
1023          * A changed rq->max_freq otoh needs to trigger re-computation of
1024          * load_scale_factor and capacity for just the cluster of cpus involved.
1025          * Since small task definition depends on max_load_scale_factor, a
1026          * changed load_scale_factor of one cluster could influence
1027          * classification of tasks in another cluster. Hence a changed
1028          * rq->max_freq will need to trigger re-computation of nr_big_task
1029          * count on all online cpus.
1030          *
1031          * While it should be sufficient for nr_big_tasks to be
1032          * re-computed for only online cpus, we have inadequate context
1033          * information here (in policy notifier) with regard to hotplug-safety
1034          * context in which notification is issued. As a result, we can't use
1035          * get_online_cpus() here, as it can lead to deadlock. Until cpufreq is
1036          * fixed up to issue notification always in hotplug-safe context,
1037          * re-compute nr_big_task for all possible cpus.
1038          */
1039
1040         if (orig_min_max_freq != min_max_freq ||
1041                 orig_max_possible_freq != max_possible_freq) {
1042                         cpus = cpu_possible_mask;
1043                         update_max = 1;
1044         }
1045
1046         /*
1047          * Changed load_scale_factor can trigger reclassification of tasks as
1048          * big or small. Make this change "atomic" so that tasks are accounted
1049          * properly due to changed load_scale_factor
1050          */
1051         for_each_cpu(i, cpus) {
1052                 struct rq *rq = cpu_rq(i);
1053
1054                 rq->capacity = compute_capacity(i);
1055                 rq->load_scale_factor = compute_load_scale_factor(i);
1056
1057                 if (update_max) {
1058                         u64 mpc, mplsf;
1059
1060                         mpc = div_u64(((u64) rq->capacity) *
1061                                 rq->max_possible_freq, rq->max_freq);
1062                         rq->max_possible_capacity = (int) mpc;
1063
1064                         mplsf = div_u64(((u64) rq->load_scale_factor) *
1065                                 rq->max_possible_freq, rq->max_freq);
1066
1067                         if (mpc > highest_mpc) {
1068                                 highest_mpc = mpc;
1069                                 cpumask_clear(&mpc_mask);
1070                                 cpumask_set_cpu(i, &mpc_mask);
1071                         } else if (mpc == highest_mpc) {
1072                                 cpumask_set_cpu(i, &mpc_mask);
1073                         }
1074
1075                         if (mplsf > highest_mplsf)
1076                                 highest_mplsf = mplsf;
1077                 }
1078         }
1079
1080         if (update_max) {
1081                 max_possible_capacity = highest_mpc;
1082                 max_load_scale_factor = highest_mplsf;
1083         }
1084
1085         __update_min_max_capacity();
1086
1087         return 0;
1088 }
1089
1090 static int cpufreq_notifier_trans(struct notifier_block *nb,
1091                 unsigned long val, void *data)
1092 {
1093         struct cpufreq_freqs *freq = (struct cpufreq_freqs *)data;
1094         unsigned int cpu = freq->cpu, new_freq = freq->new;
1095         unsigned long flags;
1096         int i;
1097
1098         if (val != CPUFREQ_POSTCHANGE)
1099                 return 0;
1100
1101         BUG_ON(!new_freq);
1102
1103         if (cpu_rq(cpu)->cur_freq == new_freq)
1104                 return 0;
1105
1106         for_each_cpu(i, &cpu_rq(cpu)->freq_domain_cpumask) {
1107                 struct rq *rq = cpu_rq(i);
1108
1109                 raw_spin_lock_irqsave(&rq->lock, flags);
1110                 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
1111                                       walt_ktime_clock(), 0);
1112                 rq->cur_freq = new_freq;
1113                 raw_spin_unlock_irqrestore(&rq->lock, flags);
1114         }
1115
1116         return 0;
1117 }
1118
1119 static struct notifier_block notifier_policy_block = {
1120         .notifier_call = cpufreq_notifier_policy
1121 };
1122
1123 static struct notifier_block notifier_trans_block = {
1124         .notifier_call = cpufreq_notifier_trans
1125 };
1126
1127 static int register_sched_callback(void)
1128 {
1129         int ret;
1130
1131         ret = cpufreq_register_notifier(&notifier_policy_block,
1132                                                 CPUFREQ_POLICY_NOTIFIER);
1133
1134         if (!ret)
1135                 ret = cpufreq_register_notifier(&notifier_trans_block,
1136                                                 CPUFREQ_TRANSITION_NOTIFIER);
1137
1138         return 0;
1139 }
1140
1141 /*
1142  * cpufreq callbacks can be registered at core_initcall or later time.
1143  * Any registration done prior to that is "forgotten" by cpufreq. See
1144  * initialization of variable init_cpufreq_transition_notifier_list_called
1145  * for further information.
1146  */
1147 core_initcall(register_sched_callback);
1148
1149 void walt_init_new_task_load(struct task_struct *p)
1150 {
1151         int i;
1152         u32 init_load_windows =
1153                         div64_u64((u64)sysctl_sched_walt_init_task_load_pct *
1154                           (u64)walt_ravg_window, 100);
1155         u32 init_load_pct = current->init_load_pct;
1156
1157         p->init_load_pct = 0;
1158         memset(&p->ravg, 0, sizeof(struct ravg));
1159
1160         if (init_load_pct) {
1161                 init_load_windows = div64_u64((u64)init_load_pct *
1162                           (u64)walt_ravg_window, 100);
1163         }
1164
1165         p->ravg.demand = init_load_windows;
1166         for (i = 0; i < RAVG_HIST_SIZE_MAX; ++i)
1167                 p->ravg.sum_history[i] = init_load_windows;
1168 }