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