1/* 2 * menu.c - the menu idle governor 3 * 4 * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com> 5 * Copyright (C) 2009 Intel Corporation 6 * Author: 7 * Arjan van de Ven <arjan@linux.intel.com> 8 * 9 * This code is licenced under the GPL version 2 as described 10 * in the COPYING file that acompanies the Linux Kernel. 11 */ 12 13#include <linux/kernel.h> 14#include <linux/cpuidle.h> 15#include <linux/pm_qos.h> 16#include <linux/time.h> 17#include <linux/ktime.h> 18#include <linux/hrtimer.h> 19#include <linux/tick.h> 20#include <linux/sched.h> 21#include <linux/math64.h> 22#include <linux/module.h> 23 24/* 25 * Please note when changing the tuning values: 26 * If (MAX_INTERESTING-1) * RESOLUTION > UINT_MAX, the result of 27 * a scaling operation multiplication may overflow on 32 bit platforms. 28 * In that case, #define RESOLUTION as ULL to get 64 bit result: 29 * #define RESOLUTION 1024ULL 30 * 31 * The default values do not overflow. 32 */ 33#define BUCKETS 12 34#define INTERVAL_SHIFT 3 35#define INTERVALS (1UL << INTERVAL_SHIFT) 36#define RESOLUTION 1024 37#define DECAY 8 38#define MAX_INTERESTING 50000 39 40 41/* 42 * Concepts and ideas behind the menu governor 43 * 44 * For the menu governor, there are 3 decision factors for picking a C 45 * state: 46 * 1) Energy break even point 47 * 2) Performance impact 48 * 3) Latency tolerance (from pmqos infrastructure) 49 * These these three factors are treated independently. 50 * 51 * Energy break even point 52 * ----------------------- 53 * C state entry and exit have an energy cost, and a certain amount of time in 54 * the C state is required to actually break even on this cost. CPUIDLE 55 * provides us this duration in the "target_residency" field. So all that we 56 * need is a good prediction of how long we'll be idle. Like the traditional 57 * menu governor, we start with the actual known "next timer event" time. 58 * 59 * Since there are other source of wakeups (interrupts for example) than 60 * the next timer event, this estimation is rather optimistic. To get a 61 * more realistic estimate, a correction factor is applied to the estimate, 62 * that is based on historic behavior. For example, if in the past the actual 63 * duration always was 50% of the next timer tick, the correction factor will 64 * be 0.5. 65 * 66 * menu uses a running average for this correction factor, however it uses a 67 * set of factors, not just a single factor. This stems from the realization 68 * that the ratio is dependent on the order of magnitude of the expected 69 * duration; if we expect 500 milliseconds of idle time the likelihood of 70 * getting an interrupt very early is much higher than if we expect 50 micro 71 * seconds of idle time. A second independent factor that has big impact on 72 * the actual factor is if there is (disk) IO outstanding or not. 73 * (as a special twist, we consider every sleep longer than 50 milliseconds 74 * as perfect; there are no power gains for sleeping longer than this) 75 * 76 * For these two reasons we keep an array of 12 independent factors, that gets 77 * indexed based on the magnitude of the expected duration as well as the 78 * "is IO outstanding" property. 79 * 80 * Repeatable-interval-detector 81 * ---------------------------- 82 * There are some cases where "next timer" is a completely unusable predictor: 83 * Those cases where the interval is fixed, for example due to hardware 84 * interrupt mitigation, but also due to fixed transfer rate devices such as 85 * mice. 86 * For this, we use a different predictor: We track the duration of the last 8 87 * intervals and if the stand deviation of these 8 intervals is below a 88 * threshold value, we use the average of these intervals as prediction. 89 * 90 * Limiting Performance Impact 91 * --------------------------- 92 * C states, especially those with large exit latencies, can have a real 93 * noticeable impact on workloads, which is not acceptable for most sysadmins, 94 * and in addition, less performance has a power price of its own. 95 * 96 * As a general rule of thumb, menu assumes that the following heuristic 97 * holds: 98 * The busier the system, the less impact of C states is acceptable 99 * 100 * This rule-of-thumb is implemented using a performance-multiplier: 101 * If the exit latency times the performance multiplier is longer than 102 * the predicted duration, the C state is not considered a candidate 103 * for selection due to a too high performance impact. So the higher 104 * this multiplier is, the longer we need to be idle to pick a deep C 105 * state, and thus the less likely a busy CPU will hit such a deep 106 * C state. 107 * 108 * Two factors are used in determing this multiplier: 109 * a value of 10 is added for each point of "per cpu load average" we have. 110 * a value of 5 points is added for each process that is waiting for 111 * IO on this CPU. 112 * (these values are experimentally determined) 113 * 114 * The load average factor gives a longer term (few seconds) input to the 115 * decision, while the iowait value gives a cpu local instantanious input. 116 * The iowait factor may look low, but realize that this is also already 117 * represented in the system load average. 118 * 119 */ 120 121struct menu_device { 122 int last_state_idx; 123 int needs_update; 124 125 unsigned int next_timer_us; 126 unsigned int predicted_us; 127 unsigned int bucket; 128 unsigned int correction_factor[BUCKETS]; 129 unsigned int intervals[INTERVALS]; 130 int interval_ptr; 131}; 132 133 134#define LOAD_INT(x) ((x) >> FSHIFT) 135#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) 136 137static inline int get_loadavg(unsigned long load) 138{ 139 return LOAD_INT(load) * 10 + LOAD_FRAC(load) / 10; 140} 141 142static inline int which_bucket(unsigned int duration, unsigned long nr_iowaiters) 143{ 144 int bucket = 0; 145 146 /* 147 * We keep two groups of stats; one with no 148 * IO pending, one without. 149 * This allows us to calculate 150 * E(duration)|iowait 151 */ 152 if (nr_iowaiters) 153 bucket = BUCKETS/2; 154 155 if (duration < 10) 156 return bucket; 157 if (duration < 100) 158 return bucket + 1; 159 if (duration < 1000) 160 return bucket + 2; 161 if (duration < 10000) 162 return bucket + 3; 163 if (duration < 100000) 164 return bucket + 4; 165 return bucket + 5; 166} 167 168/* 169 * Return a multiplier for the exit latency that is intended 170 * to take performance requirements into account. 171 * The more performance critical we estimate the system 172 * to be, the higher this multiplier, and thus the higher 173 * the barrier to go to an expensive C state. 174 */ 175static inline int performance_multiplier(unsigned long nr_iowaiters, unsigned long load) 176{ 177 int mult = 1; 178 179 /* for higher loadavg, we are more reluctant */ 180 181 mult += 2 * get_loadavg(load); 182 183 /* for IO wait tasks (per cpu!) we add 5x each */ 184 mult += 10 * nr_iowaiters; 185 186 return mult; 187} 188 189static DEFINE_PER_CPU(struct menu_device, menu_devices); 190 191static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev); 192 193/* 194 * Try detecting repeating patterns by keeping track of the last 8 195 * intervals, and checking if the standard deviation of that set 196 * of points is below a threshold. If it is... then use the 197 * average of these 8 points as the estimated value. 198 */ 199static void get_typical_interval(struct menu_device *data) 200{ 201 int i, divisor; 202 unsigned int max, thresh; 203 uint64_t avg, stddev; 204 205 thresh = UINT_MAX; /* Discard outliers above this value */ 206 207again: 208 209 /* First calculate the average of past intervals */ 210 max = 0; 211 avg = 0; 212 divisor = 0; 213 for (i = 0; i < INTERVALS; i++) { 214 unsigned int value = data->intervals[i]; 215 if (value <= thresh) { 216 avg += value; 217 divisor++; 218 if (value > max) 219 max = value; 220 } 221 } 222 if (divisor == INTERVALS) 223 avg >>= INTERVAL_SHIFT; 224 else 225 do_div(avg, divisor); 226 227 /* Then try to determine standard deviation */ 228 stddev = 0; 229 for (i = 0; i < INTERVALS; i++) { 230 unsigned int value = data->intervals[i]; 231 if (value <= thresh) { 232 int64_t diff = value - avg; 233 stddev += diff * diff; 234 } 235 } 236 if (divisor == INTERVALS) 237 stddev >>= INTERVAL_SHIFT; 238 else 239 do_div(stddev, divisor); 240 241 /* 242 * The typical interval is obtained when standard deviation is small 243 * or standard deviation is small compared to the average interval. 244 * 245 * int_sqrt() formal parameter type is unsigned long. When the 246 * greatest difference to an outlier exceeds ~65 ms * sqrt(divisor) 247 * the resulting squared standard deviation exceeds the input domain 248 * of int_sqrt on platforms where unsigned long is 32 bits in size. 249 * In such case reject the candidate average. 250 * 251 * Use this result only if there is no timer to wake us up sooner. 252 */ 253 if (likely(stddev <= ULONG_MAX)) { 254 stddev = int_sqrt(stddev); 255 if (((avg > stddev * 6) && (divisor * 4 >= INTERVALS * 3)) 256 || stddev <= 20) { 257 if (data->next_timer_us > avg) 258 data->predicted_us = avg; 259 return; 260 } 261 } 262 263 /* 264 * If we have outliers to the upside in our distribution, discard 265 * those by setting the threshold to exclude these outliers, then 266 * calculate the average and standard deviation again. Once we get 267 * down to the bottom 3/4 of our samples, stop excluding samples. 268 * 269 * This can deal with workloads that have long pauses interspersed 270 * with sporadic activity with a bunch of short pauses. 271 */ 272 if ((divisor * 4) <= INTERVALS * 3) 273 return; 274 275 thresh = max - 1; 276 goto again; 277} 278 279/** 280 * menu_select - selects the next idle state to enter 281 * @drv: cpuidle driver containing state data 282 * @dev: the CPU 283 */ 284static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev) 285{ 286 struct menu_device *data = this_cpu_ptr(&menu_devices); 287 int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); 288 int i; 289 unsigned int interactivity_req; 290 unsigned long nr_iowaiters, cpu_load; 291 292 if (data->needs_update) { 293 menu_update(drv, dev); 294 data->needs_update = 0; 295 } 296 297 data->last_state_idx = CPUIDLE_DRIVER_STATE_START - 1; 298 299 /* Special case when user has set very strict latency requirement */ 300 if (unlikely(latency_req == 0)) 301 return 0; 302 303 /* determine the expected residency time, round up */ 304 data->next_timer_us = ktime_to_us(tick_nohz_get_sleep_length()); 305 306 get_iowait_load(&nr_iowaiters, &cpu_load); 307 data->bucket = which_bucket(data->next_timer_us, nr_iowaiters); 308 309 /* 310 * Force the result of multiplication to be 64 bits even if both 311 * operands are 32 bits. 312 * Make sure to round up for half microseconds. 313 */ 314 data->predicted_us = DIV_ROUND_CLOSEST_ULL((uint64_t)data->next_timer_us * 315 data->correction_factor[data->bucket], 316 RESOLUTION * DECAY); 317 318 get_typical_interval(data); 319 320 /* 321 * Performance multiplier defines a minimum predicted idle 322 * duration / latency ratio. Adjust the latency limit if 323 * necessary. 324 */ 325 interactivity_req = data->predicted_us / performance_multiplier(nr_iowaiters, cpu_load); 326 if (latency_req > interactivity_req) 327 latency_req = interactivity_req; 328 329 /* 330 * We want to default to C1 (hlt), not to busy polling 331 * unless the timer is happening really really soon. 332 */ 333 if (data->next_timer_us > 5 && 334 !drv->states[CPUIDLE_DRIVER_STATE_START].disabled && 335 dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0) 336 data->last_state_idx = CPUIDLE_DRIVER_STATE_START; 337 338 /* 339 * Find the idle state with the lowest power while satisfying 340 * our constraints. 341 */ 342 for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) { 343 struct cpuidle_state *s = &drv->states[i]; 344 struct cpuidle_state_usage *su = &dev->states_usage[i]; 345 346 if (s->disabled || su->disable) 347 continue; 348 if (s->target_residency > data->predicted_us) 349 continue; 350 if (s->exit_latency > latency_req) 351 continue; 352 353 data->last_state_idx = i; 354 } 355 356 return data->last_state_idx; 357} 358 359/** 360 * menu_reflect - records that data structures need update 361 * @dev: the CPU 362 * @index: the index of actual entered state 363 * 364 * NOTE: it's important to be fast here because this operation will add to 365 * the overall exit latency. 366 */ 367static void menu_reflect(struct cpuidle_device *dev, int index) 368{ 369 struct menu_device *data = this_cpu_ptr(&menu_devices); 370 data->last_state_idx = index; 371 if (index >= 0) 372 data->needs_update = 1; 373} 374 375/** 376 * menu_update - attempts to guess what happened after entry 377 * @drv: cpuidle driver containing state data 378 * @dev: the CPU 379 */ 380static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev) 381{ 382 struct menu_device *data = this_cpu_ptr(&menu_devices); 383 int last_idx = data->last_state_idx; 384 struct cpuidle_state *target = &drv->states[last_idx]; 385 unsigned int measured_us; 386 unsigned int new_factor; 387 388 /* 389 * Try to figure out how much time passed between entry to low 390 * power state and occurrence of the wakeup event. 391 * 392 * If the entered idle state didn't support residency measurements, 393 * we use them anyway if they are short, and if long, 394 * truncate to the whole expected time. 395 * 396 * Any measured amount of time will include the exit latency. 397 * Since we are interested in when the wakeup begun, not when it 398 * was completed, we must subtract the exit latency. However, if 399 * the measured amount of time is less than the exit latency, 400 * assume the state was never reached and the exit latency is 0. 401 */ 402 403 /* measured value */ 404 measured_us = cpuidle_get_last_residency(dev); 405 406 /* Deduct exit latency */ 407 if (measured_us > target->exit_latency) 408 measured_us -= target->exit_latency; 409 410 /* Make sure our coefficients do not exceed unity */ 411 if (measured_us > data->next_timer_us) 412 measured_us = data->next_timer_us; 413 414 /* Update our correction ratio */ 415 new_factor = data->correction_factor[data->bucket]; 416 new_factor -= new_factor / DECAY; 417 418 if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING) 419 new_factor += RESOLUTION * measured_us / data->next_timer_us; 420 else 421 /* 422 * we were idle so long that we count it as a perfect 423 * prediction 424 */ 425 new_factor += RESOLUTION; 426 427 /* 428 * We don't want 0 as factor; we always want at least 429 * a tiny bit of estimated time. Fortunately, due to rounding, 430 * new_factor will stay nonzero regardless of measured_us values 431 * and the compiler can eliminate this test as long as DECAY > 1. 432 */ 433 if (DECAY == 1 && unlikely(new_factor == 0)) 434 new_factor = 1; 435 436 data->correction_factor[data->bucket] = new_factor; 437 438 /* update the repeating-pattern data */ 439 data->intervals[data->interval_ptr++] = measured_us; 440 if (data->interval_ptr >= INTERVALS) 441 data->interval_ptr = 0; 442} 443 444/** 445 * menu_enable_device - scans a CPU's states and does setup 446 * @drv: cpuidle driver 447 * @dev: the CPU 448 */ 449static int menu_enable_device(struct cpuidle_driver *drv, 450 struct cpuidle_device *dev) 451{ 452 struct menu_device *data = &per_cpu(menu_devices, dev->cpu); 453 int i; 454 455 memset(data, 0, sizeof(struct menu_device)); 456 457 /* 458 * if the correction factor is 0 (eg first time init or cpu hotplug 459 * etc), we actually want to start out with a unity factor. 460 */ 461 for(i = 0; i < BUCKETS; i++) 462 data->correction_factor[i] = RESOLUTION * DECAY; 463 464 return 0; 465} 466 467static struct cpuidle_governor menu_governor = { 468 .name = "menu", 469 .rating = 20, 470 .enable = menu_enable_device, 471 .select = menu_select, 472 .reflect = menu_reflect, 473 .owner = THIS_MODULE, 474}; 475 476/** 477 * init_menu - initializes the governor 478 */ 479static int __init init_menu(void) 480{ 481 return cpuidle_register_governor(&menu_governor); 482} 483 484postcore_initcall(init_menu); 485