1// SPDX-License-Identifier: GPL-2.0-only 2/* 3 * menu.c - the menu idle governor 4 * 5 * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com> 6 * Copyright (C) 2009 Intel Corporation 7 * Author: 8 * Arjan van de Ven <arjan@linux.intel.com> 9 */ 10 11#include <linux/kernel.h> 12#include <linux/cpuidle.h> 13#include <linux/time.h> 14#include <linux/ktime.h> 15#include <linux/hrtimer.h> 16#include <linux/tick.h> 17#include <linux/sched.h> 18#include <linux/sched/loadavg.h> 19#include <linux/sched/stat.h> 20#include <linux/math64.h> 21 22#define BUCKETS 12 23#define INTERVAL_SHIFT 3 24#define INTERVALS (1UL << INTERVAL_SHIFT) 25#define RESOLUTION 1024 26#define DECAY 8 27#define MAX_INTERESTING (50000 * NSEC_PER_USEC) 28 29/* 30 * Concepts and ideas behind the menu governor 31 * 32 * For the menu governor, there are 3 decision factors for picking a C 33 * state: 34 * 1) Energy break even point 35 * 2) Performance impact 36 * 3) Latency tolerance (from pmqos infrastructure) 37 * These these three factors are treated independently. 38 * 39 * Energy break even point 40 * ----------------------- 41 * C state entry and exit have an energy cost, and a certain amount of time in 42 * the C state is required to actually break even on this cost. CPUIDLE 43 * provides us this duration in the "target_residency" field. So all that we 44 * need is a good prediction of how long we'll be idle. Like the traditional 45 * menu governor, we start with the actual known "next timer event" time. 46 * 47 * Since there are other source of wakeups (interrupts for example) than 48 * the next timer event, this estimation is rather optimistic. To get a 49 * more realistic estimate, a correction factor is applied to the estimate, 50 * that is based on historic behavior. For example, if in the past the actual 51 * duration always was 50% of the next timer tick, the correction factor will 52 * be 0.5. 53 * 54 * menu uses a running average for this correction factor, however it uses a 55 * set of factors, not just a single factor. This stems from the realization 56 * that the ratio is dependent on the order of magnitude of the expected 57 * duration; if we expect 500 milliseconds of idle time the likelihood of 58 * getting an interrupt very early is much higher than if we expect 50 micro 59 * seconds of idle time. A second independent factor that has big impact on 60 * the actual factor is if there is (disk) IO outstanding or not. 61 * (as a special twist, we consider every sleep longer than 50 milliseconds 62 * as perfect; there are no power gains for sleeping longer than this) 63 * 64 * For these two reasons we keep an array of 12 independent factors, that gets 65 * indexed based on the magnitude of the expected duration as well as the 66 * "is IO outstanding" property. 67 * 68 * Repeatable-interval-detector 69 * ---------------------------- 70 * There are some cases where "next timer" is a completely unusable predictor: 71 * Those cases where the interval is fixed, for example due to hardware 72 * interrupt mitigation, but also due to fixed transfer rate devices such as 73 * mice. 74 * For this, we use a different predictor: We track the duration of the last 8 75 * intervals and if the stand deviation of these 8 intervals is below a 76 * threshold value, we use the average of these intervals as prediction. 77 * 78 * Limiting Performance Impact 79 * --------------------------- 80 * C states, especially those with large exit latencies, can have a real 81 * noticeable impact on workloads, which is not acceptable for most sysadmins, 82 * and in addition, less performance has a power price of its own. 83 * 84 * As a general rule of thumb, menu assumes that the following heuristic 85 * holds: 86 * The busier the system, the less impact of C states is acceptable 87 * 88 * This rule-of-thumb is implemented using a performance-multiplier: 89 * If the exit latency times the performance multiplier is longer than 90 * the predicted duration, the C state is not considered a candidate 91 * for selection due to a too high performance impact. So the higher 92 * this multiplier is, the longer we need to be idle to pick a deep C 93 * state, and thus the less likely a busy CPU will hit such a deep 94 * C state. 95 * 96 * Two factors are used in determing this multiplier: 97 * a value of 10 is added for each point of "per cpu load average" we have. 98 * a value of 5 points is added for each process that is waiting for 99 * IO on this CPU. 100 * (these values are experimentally determined) 101 * 102 * The load average factor gives a longer term (few seconds) input to the 103 * decision, while the iowait value gives a cpu local instantanious input. 104 * The iowait factor may look low, but realize that this is also already 105 * represented in the system load average. 106 * 107 */ 108 109struct menu_device { 110 int needs_update; 111 int tick_wakeup; 112 113 u64 next_timer_ns; 114 unsigned int bucket; 115 unsigned int correction_factor[BUCKETS]; 116 unsigned int intervals[INTERVALS]; 117 int interval_ptr; 118}; 119 120static inline int which_bucket(u64 duration_ns, unsigned long nr_iowaiters) 121{ 122 int bucket = 0; 123 124 /* 125 * We keep two groups of stats; one with no 126 * IO pending, one without. 127 * This allows us to calculate 128 * E(duration)|iowait 129 */ 130 if (nr_iowaiters) 131 bucket = BUCKETS/2; 132 133 if (duration_ns < 10ULL * NSEC_PER_USEC) 134 return bucket; 135 if (duration_ns < 100ULL * NSEC_PER_USEC) 136 return bucket + 1; 137 if (duration_ns < 1000ULL * NSEC_PER_USEC) 138 return bucket + 2; 139 if (duration_ns < 10000ULL * NSEC_PER_USEC) 140 return bucket + 3; 141 if (duration_ns < 100000ULL * NSEC_PER_USEC) 142 return bucket + 4; 143 return bucket + 5; 144} 145 146/* 147 * Return a multiplier for the exit latency that is intended 148 * to take performance requirements into account. 149 * The more performance critical we estimate the system 150 * to be, the higher this multiplier, and thus the higher 151 * the barrier to go to an expensive C state. 152 */ 153static inline int performance_multiplier(unsigned long nr_iowaiters) 154{ 155 /* for IO wait tasks (per cpu!) we add 10x each */ 156 return 1 + 10 * nr_iowaiters; 157} 158 159static DEFINE_PER_CPU(struct menu_device, menu_devices); 160 161static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev); 162 163/* 164 * Try detecting repeating patterns by keeping track of the last 8 165 * intervals, and checking if the standard deviation of that set 166 * of points is below a threshold. If it is... then use the 167 * average of these 8 points as the estimated value. 168 */ 169static unsigned int get_typical_interval(struct menu_device *data, 170 unsigned int predicted_us) 171{ 172 int i, divisor; 173 unsigned int min, max, thresh, avg; 174 uint64_t sum, variance; 175 176 thresh = INT_MAX; /* Discard outliers above this value */ 177 178again: 179 180 /* First calculate the average of past intervals */ 181 min = UINT_MAX; 182 max = 0; 183 sum = 0; 184 divisor = 0; 185 for (i = 0; i < INTERVALS; i++) { 186 unsigned int value = data->intervals[i]; 187 if (value <= thresh) { 188 sum += value; 189 divisor++; 190 if (value > max) 191 max = value; 192 193 if (value < min) 194 min = value; 195 } 196 } 197 198 /* 199 * If the result of the computation is going to be discarded anyway, 200 * avoid the computation altogether. 201 */ 202 if (min >= predicted_us) 203 return UINT_MAX; 204 205 if (divisor == INTERVALS) 206 avg = sum >> INTERVAL_SHIFT; 207 else 208 avg = div_u64(sum, divisor); 209 210 /* Then try to determine variance */ 211 variance = 0; 212 for (i = 0; i < INTERVALS; i++) { 213 unsigned int value = data->intervals[i]; 214 if (value <= thresh) { 215 int64_t diff = (int64_t)value - avg; 216 variance += diff * diff; 217 } 218 } 219 if (divisor == INTERVALS) 220 variance >>= INTERVAL_SHIFT; 221 else 222 do_div(variance, divisor); 223 224 /* 225 * The typical interval is obtained when standard deviation is 226 * small (stddev <= 20 us, variance <= 400 us^2) or standard 227 * deviation is small compared to the average interval (avg > 228 * 6*stddev, avg^2 > 36*variance). The average is smaller than 229 * UINT_MAX aka U32_MAX, so computing its square does not 230 * overflow a u64. We simply reject this candidate average if 231 * the standard deviation is greater than 715 s (which is 232 * rather unlikely). 233 * 234 * Use this result only if there is no timer to wake us up sooner. 235 */ 236 if (likely(variance <= U64_MAX/36)) { 237 if ((((u64)avg*avg > variance*36) && (divisor * 4 >= INTERVALS * 3)) 238 || variance <= 400) { 239 return avg; 240 } 241 } 242 243 /* 244 * If we have outliers to the upside in our distribution, discard 245 * those by setting the threshold to exclude these outliers, then 246 * calculate the average and standard deviation again. Once we get 247 * down to the bottom 3/4 of our samples, stop excluding samples. 248 * 249 * This can deal with workloads that have long pauses interspersed 250 * with sporadic activity with a bunch of short pauses. 251 */ 252 if ((divisor * 4) <= INTERVALS * 3) 253 return UINT_MAX; 254 255 thresh = max - 1; 256 goto again; 257} 258 259/** 260 * menu_select - selects the next idle state to enter 261 * @drv: cpuidle driver containing state data 262 * @dev: the CPU 263 * @stop_tick: indication on whether or not to stop the tick 264 */ 265static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, 266 bool *stop_tick) 267{ 268 struct menu_device *data = this_cpu_ptr(&menu_devices); 269 s64 latency_req = cpuidle_governor_latency_req(dev->cpu); 270 unsigned int predicted_us; 271 u64 predicted_ns; 272 u64 interactivity_req; 273 unsigned long nr_iowaiters; 274 ktime_t delta_next; 275 int i, idx; 276 277 if (data->needs_update) { 278 menu_update(drv, dev); 279 data->needs_update = 0; 280 } 281 282 /* determine the expected residency time, round up */ 283 data->next_timer_ns = tick_nohz_get_sleep_length(&delta_next); 284 285 nr_iowaiters = nr_iowait_cpu(dev->cpu); 286 data->bucket = which_bucket(data->next_timer_ns, nr_iowaiters); 287 288 if (unlikely(drv->state_count <= 1 || latency_req == 0) || 289 ((data->next_timer_ns < drv->states[1].target_residency_ns || 290 latency_req < drv->states[1].exit_latency_ns) && 291 !dev->states_usage[0].disable)) { 292 /* 293 * In this case state[0] will be used no matter what, so return 294 * it right away and keep the tick running if state[0] is a 295 * polling one. 296 */ 297 *stop_tick = !(drv->states[0].flags & CPUIDLE_FLAG_POLLING); 298 return 0; 299 } 300 301 /* Round up the result for half microseconds. */ 302 predicted_us = div_u64(data->next_timer_ns * 303 data->correction_factor[data->bucket] + 304 (RESOLUTION * DECAY * NSEC_PER_USEC) / 2, 305 RESOLUTION * DECAY * NSEC_PER_USEC); 306 /* Use the lowest expected idle interval to pick the idle state. */ 307 predicted_ns = (u64)min(predicted_us, 308 get_typical_interval(data, predicted_us)) * 309 NSEC_PER_USEC; 310 311 if (tick_nohz_tick_stopped()) { 312 /* 313 * If the tick is already stopped, the cost of possible short 314 * idle duration misprediction is much higher, because the CPU 315 * may be stuck in a shallow idle state for a long time as a 316 * result of it. In that case say we might mispredict and use 317 * the known time till the closest timer event for the idle 318 * state selection. 319 */ 320 if (predicted_ns < TICK_NSEC) 321 predicted_ns = delta_next; 322 } else { 323 /* 324 * Use the performance multiplier and the user-configurable 325 * latency_req to determine the maximum exit latency. 326 */ 327 interactivity_req = div64_u64(predicted_ns, 328 performance_multiplier(nr_iowaiters)); 329 if (latency_req > interactivity_req) 330 latency_req = interactivity_req; 331 } 332 333 /* 334 * Find the idle state with the lowest power while satisfying 335 * our constraints. 336 */ 337 idx = -1; 338 for (i = 0; i < drv->state_count; i++) { 339 struct cpuidle_state *s = &drv->states[i]; 340 341 if (dev->states_usage[i].disable) 342 continue; 343 344 if (idx == -1) 345 idx = i; /* first enabled state */ 346 347 if (s->target_residency_ns > predicted_ns) { 348 /* 349 * Use a physical idle state, not busy polling, unless 350 * a timer is going to trigger soon enough. 351 */ 352 if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) && 353 s->exit_latency_ns <= latency_req && 354 s->target_residency_ns <= data->next_timer_ns) { 355 predicted_ns = s->target_residency_ns; 356 idx = i; 357 break; 358 } 359 if (predicted_ns < TICK_NSEC) 360 break; 361 362 if (!tick_nohz_tick_stopped()) { 363 /* 364 * If the state selected so far is shallow, 365 * waking up early won't hurt, so retain the 366 * tick in that case and let the governor run 367 * again in the next iteration of the loop. 368 */ 369 predicted_ns = drv->states[idx].target_residency_ns; 370 break; 371 } 372 373 /* 374 * If the state selected so far is shallow and this 375 * state's target residency matches the time till the 376 * closest timer event, select this one to avoid getting 377 * stuck in the shallow one for too long. 378 */ 379 if (drv->states[idx].target_residency_ns < TICK_NSEC && 380 s->target_residency_ns <= delta_next) 381 idx = i; 382 383 return idx; 384 } 385 if (s->exit_latency_ns > latency_req) 386 break; 387 388 idx = i; 389 } 390 391 if (idx == -1) 392 idx = 0; /* No states enabled. Must use 0. */ 393 394 /* 395 * Don't stop the tick if the selected state is a polling one or if the 396 * expected idle duration is shorter than the tick period length. 397 */ 398 if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) || 399 predicted_ns < TICK_NSEC) && !tick_nohz_tick_stopped()) { 400 *stop_tick = false; 401 402 if (idx > 0 && drv->states[idx].target_residency_ns > delta_next) { 403 /* 404 * The tick is not going to be stopped and the target 405 * residency of the state to be returned is not within 406 * the time until the next timer event including the 407 * tick, so try to correct that. 408 */ 409 for (i = idx - 1; i >= 0; i--) { 410 if (dev->states_usage[i].disable) 411 continue; 412 413 idx = i; 414 if (drv->states[i].target_residency_ns <= delta_next) 415 break; 416 } 417 } 418 } 419 420 return idx; 421} 422 423/** 424 * menu_reflect - records that data structures need update 425 * @dev: the CPU 426 * @index: the index of actual entered state 427 * 428 * NOTE: it's important to be fast here because this operation will add to 429 * the overall exit latency. 430 */ 431static void menu_reflect(struct cpuidle_device *dev, int index) 432{ 433 struct menu_device *data = this_cpu_ptr(&menu_devices); 434 435 dev->last_state_idx = index; 436 data->needs_update = 1; 437 data->tick_wakeup = tick_nohz_idle_got_tick(); 438} 439 440/** 441 * menu_update - attempts to guess what happened after entry 442 * @drv: cpuidle driver containing state data 443 * @dev: the CPU 444 */ 445static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev) 446{ 447 struct menu_device *data = this_cpu_ptr(&menu_devices); 448 int last_idx = dev->last_state_idx; 449 struct cpuidle_state *target = &drv->states[last_idx]; 450 u64 measured_ns; 451 unsigned int new_factor; 452 453 /* 454 * Try to figure out how much time passed between entry to low 455 * power state and occurrence of the wakeup event. 456 * 457 * If the entered idle state didn't support residency measurements, 458 * we use them anyway if they are short, and if long, 459 * truncate to the whole expected time. 460 * 461 * Any measured amount of time will include the exit latency. 462 * Since we are interested in when the wakeup begun, not when it 463 * was completed, we must subtract the exit latency. However, if 464 * the measured amount of time is less than the exit latency, 465 * assume the state was never reached and the exit latency is 0. 466 */ 467 468 if (data->tick_wakeup && data->next_timer_ns > TICK_NSEC) { 469 /* 470 * The nohz code said that there wouldn't be any events within 471 * the tick boundary (if the tick was stopped), but the idle 472 * duration predictor had a differing opinion. Since the CPU 473 * was woken up by a tick (that wasn't stopped after all), the 474 * predictor was not quite right, so assume that the CPU could 475 * have been idle long (but not forever) to help the idle 476 * duration predictor do a better job next time. 477 */ 478 measured_ns = 9 * MAX_INTERESTING / 10; 479 } else if ((drv->states[last_idx].flags & CPUIDLE_FLAG_POLLING) && 480 dev->poll_time_limit) { 481 /* 482 * The CPU exited the "polling" state due to a time limit, so 483 * the idle duration prediction leading to the selection of that 484 * state was inaccurate. If a better prediction had been made, 485 * the CPU might have been woken up from idle by the next timer. 486 * Assume that to be the case. 487 */ 488 measured_ns = data->next_timer_ns; 489 } else { 490 /* measured value */ 491 measured_ns = dev->last_residency_ns; 492 493 /* Deduct exit latency */ 494 if (measured_ns > 2 * target->exit_latency_ns) 495 measured_ns -= target->exit_latency_ns; 496 else 497 measured_ns /= 2; 498 } 499 500 /* Make sure our coefficients do not exceed unity */ 501 if (measured_ns > data->next_timer_ns) 502 measured_ns = data->next_timer_ns; 503 504 /* Update our correction ratio */ 505 new_factor = data->correction_factor[data->bucket]; 506 new_factor -= new_factor / DECAY; 507 508 if (data->next_timer_ns > 0 && measured_ns < MAX_INTERESTING) 509 new_factor += div64_u64(RESOLUTION * measured_ns, 510 data->next_timer_ns); 511 else 512 /* 513 * we were idle so long that we count it as a perfect 514 * prediction 515 */ 516 new_factor += RESOLUTION; 517 518 /* 519 * We don't want 0 as factor; we always want at least 520 * a tiny bit of estimated time. Fortunately, due to rounding, 521 * new_factor will stay nonzero regardless of measured_us values 522 * and the compiler can eliminate this test as long as DECAY > 1. 523 */ 524 if (DECAY == 1 && unlikely(new_factor == 0)) 525 new_factor = 1; 526 527 data->correction_factor[data->bucket] = new_factor; 528 529 /* update the repeating-pattern data */ 530 data->intervals[data->interval_ptr++] = ktime_to_us(measured_ns); 531 if (data->interval_ptr >= INTERVALS) 532 data->interval_ptr = 0; 533} 534 535/** 536 * menu_enable_device - scans a CPU's states and does setup 537 * @drv: cpuidle driver 538 * @dev: the CPU 539 */ 540static int menu_enable_device(struct cpuidle_driver *drv, 541 struct cpuidle_device *dev) 542{ 543 struct menu_device *data = &per_cpu(menu_devices, dev->cpu); 544 int i; 545 546 memset(data, 0, sizeof(struct menu_device)); 547 548 /* 549 * if the correction factor is 0 (eg first time init or cpu hotplug 550 * etc), we actually want to start out with a unity factor. 551 */ 552 for(i = 0; i < BUCKETS; i++) 553 data->correction_factor[i] = RESOLUTION * DECAY; 554 555 return 0; 556} 557 558static struct cpuidle_governor menu_governor = { 559 .name = "menu", 560 .rating = 20, 561 .enable = menu_enable_device, 562 .select = menu_select, 563 .reflect = menu_reflect, 564}; 565 566/** 567 * init_menu - initializes the governor 568 */ 569static int __init init_menu(void) 570{ 571 return cpuidle_register_governor(&menu_governor); 572} 573 574postcore_initcall(init_menu); 575