Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 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 | |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 65 | static unsigned int max_load_scale_factor = 1024; |
| 66 | static unsigned int max_possible_capacity = 1024; |
| 67 | |
| 68 | /* Mask of all CPUs that have max_possible_capacity */ |
| 69 | static cpumask_t mpc_mask = CPU_MASK_ALL; |
| 70 | |
| 71 | /* Window size (in ns) */ |
| 72 | __read_mostly unsigned int walt_ravg_window = 20000000; |
| 73 | |
| 74 | /* Min window size (in ns) = 10ms */ |
Joonwoo Park | 578db5d | 2017-06-01 10:59:12 -0700 | [diff] [blame^] | 75 | #ifdef CONFIG_HZ_300 |
| 76 | /* |
| 77 | * Tick interval becomes to 3333333 due to |
| 78 | * rounding error when HZ=300. |
| 79 | */ |
| 80 | #define MIN_SCHED_RAVG_WINDOW (3333333 * 6) |
| 81 | #else |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 82 | #define MIN_SCHED_RAVG_WINDOW 10000000 |
Joonwoo Park | 578db5d | 2017-06-01 10:59:12 -0700 | [diff] [blame^] | 83 | #endif |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 84 | |
| 85 | /* Max window size (in ns) = 1s */ |
| 86 | #define MAX_SCHED_RAVG_WINDOW 1000000000 |
| 87 | |
| 88 | static unsigned int sync_cpu; |
| 89 | static ktime_t ktime_last; |
Todd Poynor | 932dcee | 2017-04-10 18:31:28 -0700 | [diff] [blame] | 90 | static __read_mostly bool walt_ktime_suspended; |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 91 | |
| 92 | static unsigned int task_load(struct task_struct *p) |
| 93 | { |
| 94 | return p->ravg.demand; |
| 95 | } |
| 96 | |
| 97 | void |
| 98 | walt_inc_cumulative_runnable_avg(struct rq *rq, |
| 99 | struct task_struct *p) |
| 100 | { |
| 101 | rq->cumulative_runnable_avg += p->ravg.demand; |
| 102 | } |
| 103 | |
| 104 | void |
| 105 | walt_dec_cumulative_runnable_avg(struct rq *rq, |
| 106 | struct task_struct *p) |
| 107 | { |
| 108 | rq->cumulative_runnable_avg -= p->ravg.demand; |
| 109 | BUG_ON((s64)rq->cumulative_runnable_avg < 0); |
| 110 | } |
| 111 | |
| 112 | static void |
| 113 | fixup_cumulative_runnable_avg(struct rq *rq, |
| 114 | struct task_struct *p, s64 task_load_delta) |
| 115 | { |
| 116 | rq->cumulative_runnable_avg += task_load_delta; |
| 117 | if ((s64)rq->cumulative_runnable_avg < 0) |
| 118 | panic("cra less than zero: tld: %lld, task_load(p) = %u\n", |
| 119 | task_load_delta, task_load(p)); |
| 120 | } |
| 121 | |
| 122 | u64 walt_ktime_clock(void) |
| 123 | { |
| 124 | if (unlikely(walt_ktime_suspended)) |
| 125 | return ktime_to_ns(ktime_last); |
| 126 | return ktime_get_ns(); |
| 127 | } |
| 128 | |
| 129 | static void walt_resume(void) |
| 130 | { |
| 131 | walt_ktime_suspended = false; |
| 132 | } |
| 133 | |
| 134 | static int walt_suspend(void) |
| 135 | { |
| 136 | ktime_last = ktime_get(); |
| 137 | walt_ktime_suspended = true; |
| 138 | return 0; |
| 139 | } |
| 140 | |
| 141 | static struct syscore_ops walt_syscore_ops = { |
| 142 | .resume = walt_resume, |
| 143 | .suspend = walt_suspend |
| 144 | }; |
| 145 | |
| 146 | static int __init walt_init_ops(void) |
| 147 | { |
| 148 | register_syscore_ops(&walt_syscore_ops); |
| 149 | return 0; |
| 150 | } |
| 151 | late_initcall(walt_init_ops); |
| 152 | |
| 153 | void walt_inc_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 | void walt_dec_cfs_cumulative_runnable_avg(struct cfs_rq *cfs_rq, |
| 160 | struct task_struct *p) |
| 161 | { |
| 162 | cfs_rq->cumulative_runnable_avg -= p->ravg.demand; |
| 163 | } |
| 164 | |
| 165 | static int exiting_task(struct task_struct *p) |
| 166 | { |
| 167 | if (p->flags & PF_EXITING) { |
| 168 | if (p->ravg.sum_history[0] != EXITING_TASK_MARKER) { |
| 169 | p->ravg.sum_history[0] = EXITING_TASK_MARKER; |
| 170 | } |
| 171 | return 1; |
| 172 | } |
| 173 | return 0; |
| 174 | } |
| 175 | |
| 176 | static int __init set_walt_ravg_window(char *str) |
| 177 | { |
| 178 | get_option(&str, &walt_ravg_window); |
| 179 | |
| 180 | walt_disabled = (walt_ravg_window < MIN_SCHED_RAVG_WINDOW || |
| 181 | walt_ravg_window > MAX_SCHED_RAVG_WINDOW); |
| 182 | return 0; |
| 183 | } |
| 184 | |
| 185 | early_param("walt_ravg_window", set_walt_ravg_window); |
| 186 | |
| 187 | static void |
| 188 | update_window_start(struct rq *rq, u64 wallclock) |
| 189 | { |
| 190 | s64 delta; |
| 191 | int nr_windows; |
| 192 | |
| 193 | delta = wallclock - rq->window_start; |
Chris Redpath | b5e1207 | 2016-07-25 15:13:58 +0100 | [diff] [blame] | 194 | /* If the MPM global timer is cleared, set delta as 0 to avoid kernel BUG happening */ |
| 195 | if (delta < 0) { |
Chris Redpath | 5ea9de8 | 2016-09-20 17:00:47 +0100 | [diff] [blame] | 196 | delta = 0; |
| 197 | WARN_ONCE(1, "WALT wallclock appears to have gone backwards or reset\n"); |
Chris Redpath | b5e1207 | 2016-07-25 15:13:58 +0100 | [diff] [blame] | 198 | } |
| 199 | |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 200 | if (delta < walt_ravg_window) |
| 201 | return; |
| 202 | |
| 203 | nr_windows = div64_u64(delta, walt_ravg_window); |
| 204 | rq->window_start += (u64)nr_windows * (u64)walt_ravg_window; |
| 205 | } |
| 206 | |
| 207 | static u64 scale_exec_time(u64 delta, struct rq *rq) |
| 208 | { |
| 209 | unsigned int cur_freq = rq->cur_freq; |
| 210 | int sf; |
| 211 | |
| 212 | if (unlikely(cur_freq > max_possible_freq)) |
| 213 | cur_freq = rq->max_possible_freq; |
| 214 | |
| 215 | /* round up div64 */ |
| 216 | delta = div64_u64(delta * cur_freq + max_possible_freq - 1, |
| 217 | max_possible_freq); |
| 218 | |
| 219 | sf = DIV_ROUND_UP(rq->efficiency * 1024, max_possible_efficiency); |
| 220 | |
| 221 | delta *= sf; |
| 222 | delta >>= 10; |
| 223 | |
| 224 | return delta; |
| 225 | } |
| 226 | |
| 227 | static int cpu_is_waiting_on_io(struct rq *rq) |
| 228 | { |
| 229 | if (!walt_io_is_busy) |
| 230 | return 0; |
| 231 | |
| 232 | return atomic_read(&rq->nr_iowait); |
| 233 | } |
| 234 | |
Srinath Sridharan | 3a73c96 | 2016-07-22 13:21:15 +0100 | [diff] [blame] | 235 | void walt_account_irqtime(int cpu, struct task_struct *curr, |
| 236 | u64 delta, u64 wallclock) |
| 237 | { |
| 238 | struct rq *rq = cpu_rq(cpu); |
| 239 | unsigned long flags, nr_windows; |
| 240 | u64 cur_jiffies_ts; |
| 241 | |
| 242 | raw_spin_lock_irqsave(&rq->lock, flags); |
| 243 | |
| 244 | /* |
| 245 | * cputime (wallclock) uses sched_clock so use the same here for |
| 246 | * consistency. |
| 247 | */ |
| 248 | delta += sched_clock() - wallclock; |
| 249 | cur_jiffies_ts = get_jiffies_64(); |
| 250 | |
| 251 | if (is_idle_task(curr)) |
| 252 | walt_update_task_ravg(curr, rq, IRQ_UPDATE, walt_ktime_clock(), |
| 253 | delta); |
| 254 | |
| 255 | nr_windows = cur_jiffies_ts - rq->irqload_ts; |
| 256 | |
| 257 | if (nr_windows) { |
| 258 | if (nr_windows < 10) { |
| 259 | /* Decay CPU's irqload by 3/4 for each window. */ |
| 260 | rq->avg_irqload *= (3 * nr_windows); |
| 261 | rq->avg_irqload = div64_u64(rq->avg_irqload, |
| 262 | 4 * nr_windows); |
| 263 | } else { |
| 264 | rq->avg_irqload = 0; |
| 265 | } |
| 266 | rq->avg_irqload += rq->cur_irqload; |
| 267 | rq->cur_irqload = 0; |
| 268 | } |
| 269 | |
| 270 | rq->cur_irqload += delta; |
| 271 | rq->irqload_ts = cur_jiffies_ts; |
| 272 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 273 | } |
| 274 | |
| 275 | |
| 276 | #define WALT_HIGH_IRQ_TIMEOUT 3 |
| 277 | |
| 278 | u64 walt_irqload(int cpu) { |
| 279 | struct rq *rq = cpu_rq(cpu); |
| 280 | s64 delta; |
| 281 | delta = get_jiffies_64() - rq->irqload_ts; |
| 282 | |
| 283 | /* |
| 284 | * Current context can be preempted by irq and rq->irqload_ts can be |
| 285 | * updated by irq context so that delta can be negative. |
| 286 | * But this is okay and we can safely return as this means there |
| 287 | * was recent irq occurrence. |
| 288 | */ |
| 289 | |
| 290 | if (delta < WALT_HIGH_IRQ_TIMEOUT) |
| 291 | return rq->avg_irqload; |
| 292 | else |
| 293 | return 0; |
| 294 | } |
| 295 | |
| 296 | int walt_cpu_high_irqload(int cpu) { |
| 297 | return walt_irqload(cpu) >= sysctl_sched_walt_cpu_high_irqload; |
| 298 | } |
| 299 | |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 300 | static int account_busy_for_cpu_time(struct rq *rq, struct task_struct *p, |
| 301 | u64 irqtime, int event) |
| 302 | { |
| 303 | if (is_idle_task(p)) { |
| 304 | /* TASK_WAKE && TASK_MIGRATE is not possible on idle task! */ |
| 305 | if (event == PICK_NEXT_TASK) |
| 306 | return 0; |
| 307 | |
| 308 | /* PUT_PREV_TASK, TASK_UPDATE && IRQ_UPDATE are left */ |
| 309 | return irqtime || cpu_is_waiting_on_io(rq); |
| 310 | } |
| 311 | |
| 312 | if (event == TASK_WAKE) |
| 313 | return 0; |
| 314 | |
| 315 | if (event == PUT_PREV_TASK || event == IRQ_UPDATE || |
| 316 | event == TASK_UPDATE) |
| 317 | return 1; |
| 318 | |
| 319 | /* Only TASK_MIGRATE && PICK_NEXT_TASK left */ |
| 320 | return walt_freq_account_wait_time; |
| 321 | } |
| 322 | |
| 323 | /* |
| 324 | * Account cpu activity in its busy time counters (rq->curr/prev_runnable_sum) |
| 325 | */ |
| 326 | static void update_cpu_busy_time(struct task_struct *p, struct rq *rq, |
| 327 | int event, u64 wallclock, u64 irqtime) |
| 328 | { |
| 329 | int new_window, nr_full_windows = 0; |
| 330 | int p_is_curr_task = (p == rq->curr); |
| 331 | u64 mark_start = p->ravg.mark_start; |
| 332 | u64 window_start = rq->window_start; |
| 333 | u32 window_size = walt_ravg_window; |
| 334 | u64 delta; |
| 335 | |
| 336 | new_window = mark_start < window_start; |
| 337 | if (new_window) { |
| 338 | nr_full_windows = div64_u64((window_start - mark_start), |
| 339 | window_size); |
| 340 | if (p->ravg.active_windows < USHRT_MAX) |
| 341 | p->ravg.active_windows++; |
| 342 | } |
| 343 | |
| 344 | /* Handle per-task window rollover. We don't care about the idle |
| 345 | * task or exiting tasks. */ |
| 346 | if (new_window && !is_idle_task(p) && !exiting_task(p)) { |
| 347 | u32 curr_window = 0; |
| 348 | |
| 349 | if (!nr_full_windows) |
| 350 | curr_window = p->ravg.curr_window; |
| 351 | |
| 352 | p->ravg.prev_window = curr_window; |
| 353 | p->ravg.curr_window = 0; |
| 354 | } |
| 355 | |
| 356 | if (!account_busy_for_cpu_time(rq, p, irqtime, event)) { |
| 357 | /* account_busy_for_cpu_time() = 0, so no update to the |
| 358 | * task's current window needs to be made. This could be |
| 359 | * for example |
| 360 | * |
| 361 | * - a wakeup event on a task within the current |
| 362 | * window (!new_window below, no action required), |
| 363 | * - switching to a new task from idle (PICK_NEXT_TASK) |
| 364 | * in a new window where irqtime is 0 and we aren't |
| 365 | * waiting on IO */ |
| 366 | |
| 367 | if (!new_window) |
| 368 | return; |
| 369 | |
| 370 | /* A new window has started. The RQ demand must be rolled |
| 371 | * over if p is the current task. */ |
| 372 | if (p_is_curr_task) { |
| 373 | u64 prev_sum = 0; |
| 374 | |
| 375 | /* p is either idle task or an exiting task */ |
| 376 | if (!nr_full_windows) { |
| 377 | prev_sum = rq->curr_runnable_sum; |
| 378 | } |
| 379 | |
| 380 | rq->prev_runnable_sum = prev_sum; |
| 381 | rq->curr_runnable_sum = 0; |
| 382 | } |
| 383 | |
| 384 | return; |
| 385 | } |
| 386 | |
| 387 | if (!new_window) { |
| 388 | /* account_busy_for_cpu_time() = 1 so busy time needs |
| 389 | * to be accounted to the current window. No rollover |
| 390 | * since we didn't start a new window. An example of this is |
| 391 | * when a task starts execution and then sleeps within the |
| 392 | * same window. */ |
| 393 | |
| 394 | if (!irqtime || !is_idle_task(p) || cpu_is_waiting_on_io(rq)) |
| 395 | delta = wallclock - mark_start; |
| 396 | else |
| 397 | delta = irqtime; |
| 398 | delta = scale_exec_time(delta, rq); |
| 399 | rq->curr_runnable_sum += delta; |
| 400 | if (!is_idle_task(p) && !exiting_task(p)) |
| 401 | p->ravg.curr_window += delta; |
| 402 | |
| 403 | return; |
| 404 | } |
| 405 | |
| 406 | if (!p_is_curr_task) { |
| 407 | /* account_busy_for_cpu_time() = 1 so busy time needs |
| 408 | * to be accounted to the current window. A new window |
| 409 | * has also started, but p is not the current task, so the |
| 410 | * window is not rolled over - just split up and account |
| 411 | * as necessary into curr and prev. The window is only |
| 412 | * rolled over when a new window is processed for the current |
| 413 | * task. |
| 414 | * |
| 415 | * Irqtime can't be accounted by a task that isn't the |
| 416 | * currently running task. */ |
| 417 | |
| 418 | if (!nr_full_windows) { |
| 419 | /* A full window hasn't elapsed, account partial |
| 420 | * contribution to previous completed window. */ |
| 421 | delta = scale_exec_time(window_start - mark_start, rq); |
| 422 | if (!exiting_task(p)) |
| 423 | p->ravg.prev_window += delta; |
| 424 | } else { |
| 425 | /* Since at least one full window has elapsed, |
| 426 | * the contribution to the previous window is the |
| 427 | * full window (window_size). */ |
| 428 | delta = scale_exec_time(window_size, rq); |
| 429 | if (!exiting_task(p)) |
| 430 | p->ravg.prev_window = delta; |
| 431 | } |
| 432 | rq->prev_runnable_sum += delta; |
| 433 | |
| 434 | /* Account piece of busy time in the current window. */ |
| 435 | delta = scale_exec_time(wallclock - window_start, rq); |
| 436 | rq->curr_runnable_sum += delta; |
| 437 | if (!exiting_task(p)) |
| 438 | p->ravg.curr_window = delta; |
| 439 | |
| 440 | return; |
| 441 | } |
| 442 | |
| 443 | if (!irqtime || !is_idle_task(p) || cpu_is_waiting_on_io(rq)) { |
| 444 | /* account_busy_for_cpu_time() = 1 so busy time needs |
| 445 | * to be accounted to the current window. A new window |
| 446 | * has started and p is the current task so rollover is |
| 447 | * needed. If any of these three above conditions are true |
| 448 | * then this busy time can't be accounted as irqtime. |
| 449 | * |
| 450 | * Busy time for the idle task or exiting tasks need not |
| 451 | * be accounted. |
| 452 | * |
| 453 | * An example of this would be a task that starts execution |
| 454 | * and then sleeps once a new window has begun. */ |
| 455 | |
| 456 | if (!nr_full_windows) { |
| 457 | /* A full window hasn't elapsed, account partial |
| 458 | * contribution to previous completed window. */ |
| 459 | delta = scale_exec_time(window_start - mark_start, rq); |
| 460 | if (!is_idle_task(p) && !exiting_task(p)) |
| 461 | p->ravg.prev_window += delta; |
| 462 | |
| 463 | delta += rq->curr_runnable_sum; |
| 464 | } else { |
| 465 | /* Since at least one full window has elapsed, |
| 466 | * the contribution to the previous window is the |
| 467 | * full window (window_size). */ |
| 468 | delta = scale_exec_time(window_size, rq); |
| 469 | if (!is_idle_task(p) && !exiting_task(p)) |
| 470 | p->ravg.prev_window = delta; |
| 471 | |
| 472 | } |
| 473 | /* |
| 474 | * Rollover for normal runnable sum is done here by overwriting |
| 475 | * the values in prev_runnable_sum and curr_runnable_sum. |
| 476 | * Rollover for new task runnable sum has completed by previous |
| 477 | * if-else statement. |
| 478 | */ |
| 479 | rq->prev_runnable_sum = delta; |
| 480 | |
| 481 | /* Account piece of busy time in the current window. */ |
| 482 | delta = scale_exec_time(wallclock - window_start, rq); |
| 483 | rq->curr_runnable_sum = delta; |
| 484 | if (!is_idle_task(p) && !exiting_task(p)) |
| 485 | p->ravg.curr_window = delta; |
| 486 | |
| 487 | return; |
| 488 | } |
| 489 | |
| 490 | if (irqtime) { |
| 491 | /* account_busy_for_cpu_time() = 1 so busy time needs |
| 492 | * to be accounted to the current window. A new window |
| 493 | * has started and p is the current task so rollover is |
| 494 | * needed. The current task must be the idle task because |
| 495 | * irqtime is not accounted for any other task. |
| 496 | * |
| 497 | * Irqtime will be accounted each time we process IRQ activity |
| 498 | * after a period of idleness, so we know the IRQ busy time |
| 499 | * started at wallclock - irqtime. */ |
| 500 | |
| 501 | BUG_ON(!is_idle_task(p)); |
| 502 | mark_start = wallclock - irqtime; |
| 503 | |
| 504 | /* Roll window over. If IRQ busy time was just in the current |
| 505 | * window then that is all that need be accounted. */ |
| 506 | rq->prev_runnable_sum = rq->curr_runnable_sum; |
| 507 | if (mark_start > window_start) { |
| 508 | rq->curr_runnable_sum = scale_exec_time(irqtime, rq); |
| 509 | return; |
| 510 | } |
| 511 | |
| 512 | /* The IRQ busy time spanned multiple windows. Process the |
| 513 | * busy time preceding the current window start first. */ |
| 514 | delta = window_start - mark_start; |
| 515 | if (delta > window_size) |
| 516 | delta = window_size; |
| 517 | delta = scale_exec_time(delta, rq); |
| 518 | rq->prev_runnable_sum += delta; |
| 519 | |
| 520 | /* Process the remaining IRQ busy time in the current window. */ |
| 521 | delta = wallclock - window_start; |
| 522 | rq->curr_runnable_sum = scale_exec_time(delta, rq); |
| 523 | |
| 524 | return; |
| 525 | } |
| 526 | |
| 527 | BUG(); |
| 528 | } |
| 529 | |
| 530 | static int account_busy_for_task_demand(struct task_struct *p, int event) |
| 531 | { |
| 532 | /* No need to bother updating task demand for exiting tasks |
| 533 | * or the idle task. */ |
| 534 | if (exiting_task(p) || is_idle_task(p)) |
| 535 | return 0; |
| 536 | |
| 537 | /* When a task is waking up it is completing a segment of non-busy |
| 538 | * time. Likewise, if wait time is not treated as busy time, then |
| 539 | * when a task begins to run or is migrated, it is not running and |
| 540 | * is completing a segment of non-busy time. */ |
| 541 | if (event == TASK_WAKE || (!walt_account_wait_time && |
| 542 | (event == PICK_NEXT_TASK || event == TASK_MIGRATE))) |
| 543 | return 0; |
| 544 | |
| 545 | return 1; |
| 546 | } |
| 547 | |
| 548 | /* |
| 549 | * Called when new window is starting for a task, to record cpu usage over |
| 550 | * recently concluded window(s). Normally 'samples' should be 1. It can be > 1 |
| 551 | * when, say, a real-time task runs without preemption for several windows at a |
| 552 | * stretch. |
| 553 | */ |
| 554 | static void update_history(struct rq *rq, struct task_struct *p, |
| 555 | u32 runtime, int samples, int event) |
| 556 | { |
| 557 | u32 *hist = &p->ravg.sum_history[0]; |
| 558 | int ridx, widx; |
| 559 | u32 max = 0, avg, demand; |
| 560 | u64 sum = 0; |
| 561 | |
| 562 | /* Ignore windows where task had no activity */ |
| 563 | if (!runtime || is_idle_task(p) || exiting_task(p) || !samples) |
| 564 | goto done; |
| 565 | |
| 566 | /* Push new 'runtime' value onto stack */ |
| 567 | widx = walt_ravg_hist_size - 1; |
| 568 | ridx = widx - samples; |
| 569 | for (; ridx >= 0; --widx, --ridx) { |
| 570 | hist[widx] = hist[ridx]; |
| 571 | sum += hist[widx]; |
| 572 | if (hist[widx] > max) |
| 573 | max = hist[widx]; |
| 574 | } |
| 575 | |
| 576 | for (widx = 0; widx < samples && widx < walt_ravg_hist_size; widx++) { |
| 577 | hist[widx] = runtime; |
| 578 | sum += hist[widx]; |
| 579 | if (hist[widx] > max) |
| 580 | max = hist[widx]; |
| 581 | } |
| 582 | |
| 583 | p->ravg.sum = 0; |
| 584 | |
| 585 | if (walt_window_stats_policy == WINDOW_STATS_RECENT) { |
| 586 | demand = runtime; |
| 587 | } else if (walt_window_stats_policy == WINDOW_STATS_MAX) { |
| 588 | demand = max; |
| 589 | } else { |
| 590 | avg = div64_u64(sum, walt_ravg_hist_size); |
| 591 | if (walt_window_stats_policy == WINDOW_STATS_AVG) |
| 592 | demand = avg; |
| 593 | else |
| 594 | demand = max(avg, runtime); |
| 595 | } |
| 596 | |
| 597 | /* |
| 598 | * A throttled deadline sched class task gets dequeued without |
| 599 | * changing p->on_rq. Since the dequeue decrements hmp stats |
| 600 | * avoid decrementing it here again. |
| 601 | */ |
| 602 | if (task_on_rq_queued(p) && (!task_has_dl_policy(p) || |
| 603 | !p->dl.dl_throttled)) |
| 604 | fixup_cumulative_runnable_avg(rq, p, demand); |
| 605 | |
| 606 | p->ravg.demand = demand; |
| 607 | |
| 608 | done: |
| 609 | trace_walt_update_history(rq, p, runtime, samples, event); |
| 610 | return; |
| 611 | } |
| 612 | |
| 613 | static void add_to_task_demand(struct rq *rq, struct task_struct *p, |
| 614 | u64 delta) |
| 615 | { |
| 616 | delta = scale_exec_time(delta, rq); |
| 617 | p->ravg.sum += delta; |
| 618 | if (unlikely(p->ravg.sum > walt_ravg_window)) |
| 619 | p->ravg.sum = walt_ravg_window; |
| 620 | } |
| 621 | |
| 622 | /* |
| 623 | * Account cpu demand of task and/or update task's cpu demand history |
| 624 | * |
| 625 | * ms = p->ravg.mark_start; |
| 626 | * wc = wallclock |
| 627 | * ws = rq->window_start |
| 628 | * |
| 629 | * Three possibilities: |
| 630 | * |
| 631 | * a) Task event is contained within one window. |
| 632 | * window_start < mark_start < wallclock |
| 633 | * |
| 634 | * ws ms wc |
| 635 | * | | | |
| 636 | * V V V |
| 637 | * |---------------| |
| 638 | * |
| 639 | * In this case, p->ravg.sum is updated *iff* event is appropriate |
| 640 | * (ex: event == PUT_PREV_TASK) |
| 641 | * |
| 642 | * b) Task event spans two windows. |
| 643 | * mark_start < window_start < wallclock |
| 644 | * |
| 645 | * ms ws wc |
| 646 | * | | | |
| 647 | * V V V |
| 648 | * -----|------------------- |
| 649 | * |
| 650 | * In this case, p->ravg.sum is updated with (ws - ms) *iff* event |
| 651 | * is appropriate, then a new window sample is recorded followed |
| 652 | * by p->ravg.sum being set to (wc - ws) *iff* event is appropriate. |
| 653 | * |
| 654 | * c) Task event spans more than two windows. |
| 655 | * |
| 656 | * ms ws_tmp ws wc |
| 657 | * | | | | |
| 658 | * V V V V |
| 659 | * ---|-------|-------|-------|-------|------ |
| 660 | * | | |
| 661 | * |<------ nr_full_windows ------>| |
| 662 | * |
| 663 | * In this case, p->ravg.sum is updated with (ws_tmp - ms) first *iff* |
| 664 | * event is appropriate, window sample of p->ravg.sum is recorded, |
| 665 | * 'nr_full_window' samples of window_size is also recorded *iff* |
| 666 | * event is appropriate and finally p->ravg.sum is set to (wc - ws) |
| 667 | * *iff* event is appropriate. |
| 668 | * |
| 669 | * IMPORTANT : Leave p->ravg.mark_start unchanged, as update_cpu_busy_time() |
| 670 | * depends on it! |
| 671 | */ |
| 672 | static void update_task_demand(struct task_struct *p, struct rq *rq, |
| 673 | int event, u64 wallclock) |
| 674 | { |
| 675 | u64 mark_start = p->ravg.mark_start; |
| 676 | u64 delta, window_start = rq->window_start; |
| 677 | int new_window, nr_full_windows; |
| 678 | u32 window_size = walt_ravg_window; |
| 679 | |
| 680 | new_window = mark_start < window_start; |
| 681 | if (!account_busy_for_task_demand(p, event)) { |
| 682 | if (new_window) |
| 683 | /* If the time accounted isn't being accounted as |
| 684 | * busy time, and a new window started, only the |
| 685 | * previous window need be closed out with the |
| 686 | * pre-existing demand. Multiple windows may have |
| 687 | * elapsed, but since empty windows are dropped, |
| 688 | * it is not necessary to account those. */ |
| 689 | update_history(rq, p, p->ravg.sum, 1, event); |
| 690 | return; |
| 691 | } |
| 692 | |
| 693 | if (!new_window) { |
| 694 | /* The simple case - busy time contained within the existing |
| 695 | * window. */ |
| 696 | add_to_task_demand(rq, p, wallclock - mark_start); |
| 697 | return; |
| 698 | } |
| 699 | |
| 700 | /* Busy time spans at least two windows. Temporarily rewind |
| 701 | * window_start to first window boundary after mark_start. */ |
| 702 | delta = window_start - mark_start; |
| 703 | nr_full_windows = div64_u64(delta, window_size); |
| 704 | window_start -= (u64)nr_full_windows * (u64)window_size; |
| 705 | |
| 706 | /* Process (window_start - mark_start) first */ |
| 707 | add_to_task_demand(rq, p, window_start - mark_start); |
| 708 | |
| 709 | /* Push new sample(s) into task's demand history */ |
| 710 | update_history(rq, p, p->ravg.sum, 1, event); |
| 711 | if (nr_full_windows) |
| 712 | update_history(rq, p, scale_exec_time(window_size, rq), |
| 713 | nr_full_windows, event); |
| 714 | |
| 715 | /* Roll window_start back to current to process any remainder |
| 716 | * in current window. */ |
| 717 | window_start += (u64)nr_full_windows * (u64)window_size; |
| 718 | |
| 719 | /* Process (wallclock - window_start) next */ |
| 720 | mark_start = window_start; |
| 721 | add_to_task_demand(rq, p, wallclock - mark_start); |
| 722 | } |
| 723 | |
| 724 | /* Reflect task activity on its demand and cpu's busy time statistics */ |
| 725 | void walt_update_task_ravg(struct task_struct *p, struct rq *rq, |
| 726 | int event, u64 wallclock, u64 irqtime) |
| 727 | { |
| 728 | if (walt_disabled || !rq->window_start) |
| 729 | return; |
| 730 | |
| 731 | lockdep_assert_held(&rq->lock); |
| 732 | |
| 733 | update_window_start(rq, wallclock); |
| 734 | |
| 735 | if (!p->ravg.mark_start) |
| 736 | goto done; |
| 737 | |
| 738 | update_task_demand(p, rq, event, wallclock); |
| 739 | update_cpu_busy_time(p, rq, event, wallclock, irqtime); |
| 740 | |
| 741 | done: |
| 742 | trace_walt_update_task_ravg(p, rq, event, wallclock, irqtime); |
| 743 | |
| 744 | p->ravg.mark_start = wallclock; |
| 745 | } |
| 746 | |
| 747 | unsigned long __weak arch_get_cpu_efficiency(int cpu) |
| 748 | { |
| 749 | return SCHED_CAPACITY_SCALE; |
| 750 | } |
| 751 | |
| 752 | void walt_init_cpu_efficiency(void) |
| 753 | { |
| 754 | int i, efficiency; |
| 755 | unsigned int max = 0, min = UINT_MAX; |
| 756 | |
| 757 | for_each_possible_cpu(i) { |
| 758 | efficiency = arch_get_cpu_efficiency(i); |
| 759 | cpu_rq(i)->efficiency = efficiency; |
| 760 | |
| 761 | if (efficiency > max) |
| 762 | max = efficiency; |
| 763 | if (efficiency < min) |
| 764 | min = efficiency; |
| 765 | } |
| 766 | |
| 767 | if (max) |
| 768 | max_possible_efficiency = max; |
| 769 | |
| 770 | if (min) |
| 771 | min_possible_efficiency = min; |
| 772 | } |
| 773 | |
| 774 | static void reset_task_stats(struct task_struct *p) |
| 775 | { |
| 776 | u32 sum = 0; |
| 777 | |
| 778 | if (exiting_task(p)) |
| 779 | sum = EXITING_TASK_MARKER; |
| 780 | |
| 781 | memset(&p->ravg, 0, sizeof(struct ravg)); |
| 782 | /* Retain EXITING_TASK marker */ |
| 783 | p->ravg.sum_history[0] = sum; |
| 784 | } |
| 785 | |
| 786 | void walt_mark_task_starting(struct task_struct *p) |
| 787 | { |
| 788 | u64 wallclock; |
| 789 | struct rq *rq = task_rq(p); |
| 790 | |
| 791 | if (!rq->window_start) { |
| 792 | reset_task_stats(p); |
| 793 | return; |
| 794 | } |
| 795 | |
| 796 | wallclock = walt_ktime_clock(); |
| 797 | p->ravg.mark_start = wallclock; |
| 798 | } |
| 799 | |
| 800 | void walt_set_window_start(struct rq *rq) |
| 801 | { |
| 802 | int cpu = cpu_of(rq); |
| 803 | struct rq *sync_rq = cpu_rq(sync_cpu); |
| 804 | |
| 805 | if (rq->window_start) |
| 806 | return; |
| 807 | |
| 808 | if (cpu == sync_cpu) { |
| 809 | rq->window_start = walt_ktime_clock(); |
| 810 | } else { |
| 811 | raw_spin_unlock(&rq->lock); |
| 812 | double_rq_lock(rq, sync_rq); |
| 813 | rq->window_start = cpu_rq(sync_cpu)->window_start; |
| 814 | rq->curr_runnable_sum = rq->prev_runnable_sum = 0; |
| 815 | raw_spin_unlock(&sync_rq->lock); |
| 816 | } |
| 817 | |
| 818 | rq->curr->ravg.mark_start = rq->window_start; |
| 819 | } |
| 820 | |
| 821 | void walt_migrate_sync_cpu(int cpu) |
| 822 | { |
| 823 | if (cpu == sync_cpu) |
| 824 | sync_cpu = smp_processor_id(); |
| 825 | } |
| 826 | |
| 827 | void walt_fixup_busy_time(struct task_struct *p, int new_cpu) |
| 828 | { |
| 829 | struct rq *src_rq = task_rq(p); |
| 830 | struct rq *dest_rq = cpu_rq(new_cpu); |
| 831 | u64 wallclock; |
| 832 | |
| 833 | if (!p->on_rq && p->state != TASK_WAKING) |
| 834 | return; |
| 835 | |
| 836 | if (exiting_task(p)) { |
| 837 | return; |
| 838 | } |
| 839 | |
| 840 | if (p->state == TASK_WAKING) |
| 841 | double_rq_lock(src_rq, dest_rq); |
| 842 | |
| 843 | wallclock = walt_ktime_clock(); |
| 844 | |
| 845 | walt_update_task_ravg(task_rq(p)->curr, task_rq(p), |
| 846 | TASK_UPDATE, wallclock, 0); |
| 847 | walt_update_task_ravg(dest_rq->curr, dest_rq, |
| 848 | TASK_UPDATE, wallclock, 0); |
| 849 | |
| 850 | walt_update_task_ravg(p, task_rq(p), TASK_MIGRATE, wallclock, 0); |
| 851 | |
| 852 | if (p->ravg.curr_window) { |
| 853 | src_rq->curr_runnable_sum -= p->ravg.curr_window; |
| 854 | dest_rq->curr_runnable_sum += p->ravg.curr_window; |
| 855 | } |
| 856 | |
| 857 | if (p->ravg.prev_window) { |
| 858 | src_rq->prev_runnable_sum -= p->ravg.prev_window; |
| 859 | dest_rq->prev_runnable_sum += p->ravg.prev_window; |
| 860 | } |
| 861 | |
| 862 | if ((s64)src_rq->prev_runnable_sum < 0) { |
| 863 | src_rq->prev_runnable_sum = 0; |
| 864 | WARN_ON(1); |
| 865 | } |
| 866 | if ((s64)src_rq->curr_runnable_sum < 0) { |
| 867 | src_rq->curr_runnable_sum = 0; |
| 868 | WARN_ON(1); |
| 869 | } |
| 870 | |
| 871 | trace_walt_migration_update_sum(src_rq, p); |
| 872 | trace_walt_migration_update_sum(dest_rq, p); |
| 873 | |
| 874 | if (p->state == TASK_WAKING) |
| 875 | double_rq_unlock(src_rq, dest_rq); |
| 876 | } |
| 877 | |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 878 | /* |
| 879 | * Return 'capacity' of a cpu in reference to "least" efficient cpu, such that |
| 880 | * least efficient cpu gets capacity of 1024 |
| 881 | */ |
| 882 | static unsigned long capacity_scale_cpu_efficiency(int cpu) |
| 883 | { |
| 884 | return (1024 * cpu_rq(cpu)->efficiency) / min_possible_efficiency; |
| 885 | } |
| 886 | |
| 887 | /* |
| 888 | * Return 'capacity' of a cpu in reference to cpu with lowest max_freq |
| 889 | * (min_max_freq), such that one with lowest max_freq gets capacity of 1024. |
| 890 | */ |
| 891 | static unsigned long capacity_scale_cpu_freq(int cpu) |
| 892 | { |
| 893 | return (1024 * cpu_rq(cpu)->max_freq) / min_max_freq; |
| 894 | } |
| 895 | |
| 896 | /* |
| 897 | * Return load_scale_factor of a cpu in reference to "most" efficient cpu, so |
| 898 | * that "most" efficient cpu gets a load_scale_factor of 1 |
| 899 | */ |
| 900 | static unsigned long load_scale_cpu_efficiency(int cpu) |
| 901 | { |
| 902 | return DIV_ROUND_UP(1024 * max_possible_efficiency, |
| 903 | cpu_rq(cpu)->efficiency); |
| 904 | } |
| 905 | |
| 906 | /* |
| 907 | * Return load_scale_factor of a cpu in reference to cpu with best max_freq |
| 908 | * (max_possible_freq), so that one with best max_freq gets a load_scale_factor |
| 909 | * of 1. |
| 910 | */ |
| 911 | static unsigned long load_scale_cpu_freq(int cpu) |
| 912 | { |
| 913 | return DIV_ROUND_UP(1024 * max_possible_freq, cpu_rq(cpu)->max_freq); |
| 914 | } |
| 915 | |
| 916 | static int compute_capacity(int cpu) |
| 917 | { |
| 918 | int capacity = 1024; |
| 919 | |
| 920 | capacity *= capacity_scale_cpu_efficiency(cpu); |
| 921 | capacity >>= 10; |
| 922 | |
| 923 | capacity *= capacity_scale_cpu_freq(cpu); |
| 924 | capacity >>= 10; |
| 925 | |
| 926 | return capacity; |
| 927 | } |
| 928 | |
| 929 | static int compute_load_scale_factor(int cpu) |
| 930 | { |
| 931 | int load_scale = 1024; |
| 932 | |
| 933 | /* |
| 934 | * load_scale_factor accounts for the fact that task load |
| 935 | * is in reference to "best" performing cpu. Task's load will need to be |
| 936 | * scaled (up) by a factor to determine suitability to be placed on a |
| 937 | * (little) cpu. |
| 938 | */ |
| 939 | load_scale *= load_scale_cpu_efficiency(cpu); |
| 940 | load_scale >>= 10; |
| 941 | |
| 942 | load_scale *= load_scale_cpu_freq(cpu); |
| 943 | load_scale >>= 10; |
| 944 | |
| 945 | return load_scale; |
| 946 | } |
| 947 | |
| 948 | static int cpufreq_notifier_policy(struct notifier_block *nb, |
| 949 | unsigned long val, void *data) |
| 950 | { |
| 951 | struct cpufreq_policy *policy = (struct cpufreq_policy *)data; |
| 952 | int i, update_max = 0; |
| 953 | u64 highest_mpc = 0, highest_mplsf = 0; |
| 954 | const struct cpumask *cpus = policy->related_cpus; |
| 955 | unsigned int orig_min_max_freq = min_max_freq; |
| 956 | unsigned int orig_max_possible_freq = max_possible_freq; |
| 957 | /* Initialized to policy->max in case policy->related_cpus is empty! */ |
| 958 | unsigned int orig_max_freq = policy->max; |
| 959 | |
Juri Lelli | 3d89857 | 2016-12-06 11:50:53 +0000 | [diff] [blame] | 960 | if (val != CPUFREQ_NOTIFY) |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 961 | return 0; |
| 962 | |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 963 | for_each_cpu(i, policy->related_cpus) { |
| 964 | cpumask_copy(&cpu_rq(i)->freq_domain_cpumask, |
| 965 | policy->related_cpus); |
| 966 | orig_max_freq = cpu_rq(i)->max_freq; |
| 967 | cpu_rq(i)->min_freq = policy->min; |
| 968 | cpu_rq(i)->max_freq = policy->max; |
| 969 | cpu_rq(i)->cur_freq = policy->cur; |
| 970 | cpu_rq(i)->max_possible_freq = policy->cpuinfo.max_freq; |
| 971 | } |
| 972 | |
| 973 | max_possible_freq = max(max_possible_freq, policy->cpuinfo.max_freq); |
| 974 | if (min_max_freq == 1) |
| 975 | min_max_freq = UINT_MAX; |
| 976 | min_max_freq = min(min_max_freq, policy->cpuinfo.max_freq); |
| 977 | BUG_ON(!min_max_freq); |
| 978 | BUG_ON(!policy->max); |
| 979 | |
| 980 | /* Changes to policy other than max_freq don't require any updates */ |
| 981 | if (orig_max_freq == policy->max) |
| 982 | return 0; |
| 983 | |
| 984 | /* |
| 985 | * A changed min_max_freq or max_possible_freq (possible during bootup) |
| 986 | * needs to trigger re-computation of load_scale_factor and capacity for |
| 987 | * all possible cpus (even those offline). It also needs to trigger |
| 988 | * re-computation of nr_big_task count on all online cpus. |
| 989 | * |
| 990 | * A changed rq->max_freq otoh needs to trigger re-computation of |
| 991 | * load_scale_factor and capacity for just the cluster of cpus involved. |
| 992 | * Since small task definition depends on max_load_scale_factor, a |
| 993 | * changed load_scale_factor of one cluster could influence |
| 994 | * classification of tasks in another cluster. Hence a changed |
| 995 | * rq->max_freq will need to trigger re-computation of nr_big_task |
| 996 | * count on all online cpus. |
| 997 | * |
| 998 | * While it should be sufficient for nr_big_tasks to be |
| 999 | * re-computed for only online cpus, we have inadequate context |
| 1000 | * information here (in policy notifier) with regard to hotplug-safety |
| 1001 | * context in which notification is issued. As a result, we can't use |
| 1002 | * get_online_cpus() here, as it can lead to deadlock. Until cpufreq is |
| 1003 | * fixed up to issue notification always in hotplug-safe context, |
| 1004 | * re-compute nr_big_task for all possible cpus. |
| 1005 | */ |
| 1006 | |
| 1007 | if (orig_min_max_freq != min_max_freq || |
| 1008 | orig_max_possible_freq != max_possible_freq) { |
| 1009 | cpus = cpu_possible_mask; |
| 1010 | update_max = 1; |
| 1011 | } |
| 1012 | |
| 1013 | /* |
| 1014 | * Changed load_scale_factor can trigger reclassification of tasks as |
| 1015 | * big or small. Make this change "atomic" so that tasks are accounted |
| 1016 | * properly due to changed load_scale_factor |
| 1017 | */ |
| 1018 | for_each_cpu(i, cpus) { |
| 1019 | struct rq *rq = cpu_rq(i); |
| 1020 | |
| 1021 | rq->capacity = compute_capacity(i); |
| 1022 | rq->load_scale_factor = compute_load_scale_factor(i); |
| 1023 | |
| 1024 | if (update_max) { |
| 1025 | u64 mpc, mplsf; |
| 1026 | |
| 1027 | mpc = div_u64(((u64) rq->capacity) * |
| 1028 | rq->max_possible_freq, rq->max_freq); |
| 1029 | rq->max_possible_capacity = (int) mpc; |
| 1030 | |
| 1031 | mplsf = div_u64(((u64) rq->load_scale_factor) * |
| 1032 | rq->max_possible_freq, rq->max_freq); |
| 1033 | |
| 1034 | if (mpc > highest_mpc) { |
| 1035 | highest_mpc = mpc; |
| 1036 | cpumask_clear(&mpc_mask); |
| 1037 | cpumask_set_cpu(i, &mpc_mask); |
| 1038 | } else if (mpc == highest_mpc) { |
| 1039 | cpumask_set_cpu(i, &mpc_mask); |
| 1040 | } |
| 1041 | |
| 1042 | if (mplsf > highest_mplsf) |
| 1043 | highest_mplsf = mplsf; |
| 1044 | } |
| 1045 | } |
| 1046 | |
| 1047 | if (update_max) { |
| 1048 | max_possible_capacity = highest_mpc; |
| 1049 | max_load_scale_factor = highest_mplsf; |
| 1050 | } |
| 1051 | |
Srivatsa Vaddagiri | 26c2154 | 2016-05-31 09:08:38 -0700 | [diff] [blame] | 1052 | return 0; |
| 1053 | } |
| 1054 | |
| 1055 | static int cpufreq_notifier_trans(struct notifier_block *nb, |
| 1056 | unsigned long val, void *data) |
| 1057 | { |
| 1058 | struct cpufreq_freqs *freq = (struct cpufreq_freqs *)data; |
| 1059 | unsigned int cpu = freq->cpu, new_freq = freq->new; |
| 1060 | unsigned long flags; |
| 1061 | int i; |
| 1062 | |
| 1063 | if (val != CPUFREQ_POSTCHANGE) |
| 1064 | return 0; |
| 1065 | |
| 1066 | BUG_ON(!new_freq); |
| 1067 | |
| 1068 | if (cpu_rq(cpu)->cur_freq == new_freq) |
| 1069 | return 0; |
| 1070 | |
| 1071 | for_each_cpu(i, &cpu_rq(cpu)->freq_domain_cpumask) { |
| 1072 | struct rq *rq = cpu_rq(i); |
| 1073 | |
| 1074 | raw_spin_lock_irqsave(&rq->lock, flags); |
| 1075 | walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, |
| 1076 | walt_ktime_clock(), 0); |
| 1077 | rq->cur_freq = new_freq; |
| 1078 | raw_spin_unlock_irqrestore(&rq->lock, flags); |
| 1079 | } |
| 1080 | |
| 1081 | return 0; |
| 1082 | } |
| 1083 | |
| 1084 | static struct notifier_block notifier_policy_block = { |
| 1085 | .notifier_call = cpufreq_notifier_policy |
| 1086 | }; |
| 1087 | |
| 1088 | static struct notifier_block notifier_trans_block = { |
| 1089 | .notifier_call = cpufreq_notifier_trans |
| 1090 | }; |
| 1091 | |
| 1092 | static int register_sched_callback(void) |
| 1093 | { |
| 1094 | int ret; |
| 1095 | |
| 1096 | ret = cpufreq_register_notifier(¬ifier_policy_block, |
| 1097 | CPUFREQ_POLICY_NOTIFIER); |
| 1098 | |
| 1099 | if (!ret) |
| 1100 | ret = cpufreq_register_notifier(¬ifier_trans_block, |
| 1101 | CPUFREQ_TRANSITION_NOTIFIER); |
| 1102 | |
| 1103 | return 0; |
| 1104 | } |
| 1105 | |
| 1106 | /* |
| 1107 | * cpufreq callbacks can be registered at core_initcall or later time. |
| 1108 | * Any registration done prior to that is "forgotten" by cpufreq. See |
| 1109 | * initialization of variable init_cpufreq_transition_notifier_list_called |
| 1110 | * for further information. |
| 1111 | */ |
| 1112 | core_initcall(register_sched_callback); |
| 1113 | |
| 1114 | void walt_init_new_task_load(struct task_struct *p) |
| 1115 | { |
| 1116 | int i; |
| 1117 | u32 init_load_windows = |
| 1118 | div64_u64((u64)sysctl_sched_walt_init_task_load_pct * |
| 1119 | (u64)walt_ravg_window, 100); |
| 1120 | u32 init_load_pct = current->init_load_pct; |
| 1121 | |
| 1122 | p->init_load_pct = 0; |
| 1123 | memset(&p->ravg, 0, sizeof(struct ravg)); |
| 1124 | |
| 1125 | if (init_load_pct) { |
| 1126 | init_load_windows = div64_u64((u64)init_load_pct * |
| 1127 | (u64)walt_ravg_window, 100); |
| 1128 | } |
| 1129 | |
| 1130 | p->ravg.demand = init_load_windows; |
| 1131 | for (i = 0; i < RAVG_HIST_SIZE_MAX; ++i) |
| 1132 | p->ravg.sum_history[i] = init_load_windows; |
| 1133 | } |