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 #define BUCKETS 12 25 #define INTERVALS 8 26 #define RESOLUTION 1024 27 #define DECAY 8 28 #define MAX_INTERESTING 50000 29 #define STDDEV_THRESH 400 30 31 /* 60 * 60 > STDDEV_THRESH * INTERVALS = 400 * 8 */ 32 #define MAX_DEVIATION 60 33 34 static DEFINE_PER_CPU(struct hrtimer, menu_hrtimer); 35 static DEFINE_PER_CPU(int, hrtimer_status); 36 /* menu hrtimer mode */ 37 enum {MENU_HRTIMER_STOP, MENU_HRTIMER_REPEAT, MENU_HRTIMER_GENERAL}; 38 39 /* 40 * Concepts and ideas behind the menu governor 41 * 42 * For the menu governor, there are 3 decision factors for picking a C 43 * state: 44 * 1) Energy break even point 45 * 2) Performance impact 46 * 3) Latency tolerance (from pmqos infrastructure) 47 * These these three factors are treated independently. 48 * 49 * Energy break even point 50 * ----------------------- 51 * C state entry and exit have an energy cost, and a certain amount of time in 52 * the C state is required to actually break even on this cost. CPUIDLE 53 * provides us this duration in the "target_residency" field. So all that we 54 * need is a good prediction of how long we'll be idle. Like the traditional 55 * menu governor, we start with the actual known "next timer event" time. 56 * 57 * Since there are other source of wakeups (interrupts for example) than 58 * the next timer event, this estimation is rather optimistic. To get a 59 * more realistic estimate, a correction factor is applied to the estimate, 60 * that is based on historic behavior. For example, if in the past the actual 61 * duration always was 50% of the next timer tick, the correction factor will 62 * be 0.5. 63 * 64 * menu uses a running average for this correction factor, however it uses a 65 * set of factors, not just a single factor. This stems from the realization 66 * that the ratio is dependent on the order of magnitude of the expected 67 * duration; if we expect 500 milliseconds of idle time the likelihood of 68 * getting an interrupt very early is much higher than if we expect 50 micro 69 * seconds of idle time. A second independent factor that has big impact on 70 * the actual factor is if there is (disk) IO outstanding or not. 71 * (as a special twist, we consider every sleep longer than 50 milliseconds 72 * as perfect; there are no power gains for sleeping longer than this) 73 * 74 * For these two reasons we keep an array of 12 independent factors, that gets 75 * indexed based on the magnitude of the expected duration as well as the 76 * "is IO outstanding" property. 77 * 78 * Repeatable-interval-detector 79 * ---------------------------- 80 * There are some cases where "next timer" is a completely unusable predictor: 81 * Those cases where the interval is fixed, for example due to hardware 82 * interrupt mitigation, but also due to fixed transfer rate devices such as 83 * mice. 84 * For this, we use a different predictor: We track the duration of the last 8 85 * intervals and if the stand deviation of these 8 intervals is below a 86 * threshold value, we use the average of these intervals as prediction. 87 * 88 * Limiting Performance Impact 89 * --------------------------- 90 * C states, especially those with large exit latencies, can have a real 91 * noticeable impact on workloads, which is not acceptable for most sysadmins, 92 * and in addition, less performance has a power price of its own. 93 * 94 * As a general rule of thumb, menu assumes that the following heuristic 95 * holds: 96 * The busier the system, the less impact of C states is acceptable 97 * 98 * This rule-of-thumb is implemented using a performance-multiplier: 99 * If the exit latency times the performance multiplier is longer than 100 * the predicted duration, the C state is not considered a candidate 101 * for selection due to a too high performance impact. So the higher 102 * this multiplier is, the longer we need to be idle to pick a deep C 103 * state, and thus the less likely a busy CPU will hit such a deep 104 * C state. 105 * 106 * Two factors are used in determing this multiplier: 107 * a value of 10 is added for each point of "per cpu load average" we have. 108 * a value of 5 points is added for each process that is waiting for 109 * IO on this CPU. 110 * (these values are experimentally determined) 111 * 112 * The load average factor gives a longer term (few seconds) input to the 113 * decision, while the iowait value gives a cpu local instantanious input. 114 * The iowait factor may look low, but realize that this is also already 115 * represented in the system load average. 116 * 117 */ 118 119 /* 120 * The C-state residency is so long that is is worthwhile to exit 121 * from the shallow C-state and re-enter into a deeper C-state. 122 */ 123 static unsigned int perfect_cstate_ms __read_mostly = 30; 124 module_param(perfect_cstate_ms, uint, 0000); 125 126 struct menu_device { 127 int last_state_idx; 128 int needs_update; 129 130 unsigned int expected_us; 131 u64 predicted_us; 132 unsigned int exit_us; 133 unsigned int bucket; 134 u64 correction_factor[BUCKETS]; 135 u32 intervals[INTERVALS]; 136 int interval_ptr; 137 }; 138 139 140 #define LOAD_INT(x) ((x) >> FSHIFT) 141 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) 142 143 static int get_loadavg(void) 144 { 145 unsigned long this = this_cpu_load(); 146 147 148 return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10; 149 } 150 151 static inline int which_bucket(unsigned int duration) 152 { 153 int bucket = 0; 154 155 /* 156 * We keep two groups of stats; one with no 157 * IO pending, one without. 158 * This allows us to calculate 159 * E(duration)|iowait 160 */ 161 if (nr_iowait_cpu(smp_processor_id())) 162 bucket = BUCKETS/2; 163 164 if (duration < 10) 165 return bucket; 166 if (duration < 100) 167 return bucket + 1; 168 if (duration < 1000) 169 return bucket + 2; 170 if (duration < 10000) 171 return bucket + 3; 172 if (duration < 100000) 173 return bucket + 4; 174 return bucket + 5; 175 } 176 177 /* 178 * Return a multiplier for the exit latency that is intended 179 * to take performance requirements into account. 180 * The more performance critical we estimate the system 181 * to be, the higher this multiplier, and thus the higher 182 * the barrier to go to an expensive C state. 183 */ 184 static inline int performance_multiplier(void) 185 { 186 int mult = 1; 187 188 /* for higher loadavg, we are more reluctant */ 189 190 mult += 2 * get_loadavg(); 191 192 /* for IO wait tasks (per cpu!) we add 5x each */ 193 mult += 10 * nr_iowait_cpu(smp_processor_id()); 194 195 return mult; 196 } 197 198 static DEFINE_PER_CPU(struct menu_device, menu_devices); 199 200 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev); 201 202 /* This implements DIV_ROUND_CLOSEST but avoids 64 bit division */ 203 static u64 div_round64(u64 dividend, u32 divisor) 204 { 205 return div_u64(dividend + (divisor / 2), divisor); 206 } 207 208 /* Cancel the hrtimer if it is not triggered yet */ 209 void menu_hrtimer_cancel(void) 210 { 211 int cpu = smp_processor_id(); 212 struct hrtimer *hrtmr = &per_cpu(menu_hrtimer, cpu); 213 214 /* The timer is still not time out*/ 215 if (per_cpu(hrtimer_status, cpu)) { 216 hrtimer_cancel(hrtmr); 217 per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_STOP; 218 } 219 } 220 EXPORT_SYMBOL_GPL(menu_hrtimer_cancel); 221 222 /* Call back for hrtimer is triggered */ 223 static enum hrtimer_restart menu_hrtimer_notify(struct hrtimer *hrtimer) 224 { 225 int cpu = smp_processor_id(); 226 struct menu_device *data = &per_cpu(menu_devices, cpu); 227 228 /* In general case, the expected residency is much larger than 229 * deepest C-state target residency, but prediction logic still 230 * predicts a small predicted residency, so the prediction 231 * history is totally broken if the timer is triggered. 232 * So reset the correction factor. 233 */ 234 if (per_cpu(hrtimer_status, cpu) == MENU_HRTIMER_GENERAL) 235 data->correction_factor[data->bucket] = RESOLUTION * DECAY; 236 237 per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_STOP; 238 239 return HRTIMER_NORESTART; 240 } 241 242 /* 243 * Try detecting repeating patterns by keeping track of the last 8 244 * intervals, and checking if the standard deviation of that set 245 * of points is below a threshold. If it is... then use the 246 * average of these 8 points as the estimated value. 247 */ 248 static u32 get_typical_interval(struct menu_device *data) 249 { 250 int i = 0, divisor = 0; 251 uint64_t max = 0, avg = 0, stddev = 0; 252 int64_t thresh = LLONG_MAX; /* Discard outliers above this value. */ 253 unsigned int ret = 0; 254 255 again: 256 257 /* first calculate average and standard deviation of the past */ 258 max = avg = divisor = stddev = 0; 259 for (i = 0; i < INTERVALS; i++) { 260 int64_t value = data->intervals[i]; 261 if (value <= thresh) { 262 avg += value; 263 divisor++; 264 if (value > max) 265 max = value; 266 } 267 } 268 do_div(avg, divisor); 269 270 for (i = 0; i < INTERVALS; i++) { 271 int64_t value = data->intervals[i]; 272 if (value <= thresh) { 273 int64_t diff = value - avg; 274 stddev += diff * diff; 275 } 276 } 277 do_div(stddev, divisor); 278 stddev = int_sqrt(stddev); 279 /* 280 * If we have outliers to the upside in our distribution, discard 281 * those by setting the threshold to exclude these outliers, then 282 * calculate the average and standard deviation again. Once we get 283 * down to the bottom 3/4 of our samples, stop excluding samples. 284 * 285 * This can deal with workloads that have long pauses interspersed 286 * with sporadic activity with a bunch of short pauses. 287 * 288 * The typical interval is obtained when standard deviation is small 289 * or standard deviation is small compared to the average interval. 290 */ 291 if (((avg > stddev * 6) && (divisor * 4 >= INTERVALS * 3)) 292 || stddev <= 20) { 293 data->predicted_us = avg; 294 ret = 1; 295 return ret; 296 297 } else if ((divisor * 4) > INTERVALS * 3) { 298 /* Exclude the max interval */ 299 thresh = max - 1; 300 goto again; 301 } 302 303 return ret; 304 } 305 306 /** 307 * menu_select - selects the next idle state to enter 308 * @drv: cpuidle driver containing state data 309 * @dev: the CPU 310 */ 311 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev) 312 { 313 struct menu_device *data = &__get_cpu_var(menu_devices); 314 int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); 315 int i; 316 int multiplier; 317 struct timespec t; 318 int repeat = 0, low_predicted = 0; 319 int cpu = smp_processor_id(); 320 struct hrtimer *hrtmr = &per_cpu(menu_hrtimer, cpu); 321 322 if (data->needs_update) { 323 menu_update(drv, dev); 324 data->needs_update = 0; 325 } 326 327 data->last_state_idx = 0; 328 data->exit_us = 0; 329 330 /* Special case when user has set very strict latency requirement */ 331 if (unlikely(latency_req == 0)) 332 return 0; 333 334 /* determine the expected residency time, round up */ 335 t = ktime_to_timespec(tick_nohz_get_sleep_length()); 336 data->expected_us = 337 t.tv_sec * USEC_PER_SEC + t.tv_nsec / NSEC_PER_USEC; 338 339 340 data->bucket = which_bucket(data->expected_us); 341 342 multiplier = performance_multiplier(); 343 344 /* 345 * if the correction factor is 0 (eg first time init or cpu hotplug 346 * etc), we actually want to start out with a unity factor. 347 */ 348 if (data->correction_factor[data->bucket] == 0) 349 data->correction_factor[data->bucket] = RESOLUTION * DECAY; 350 351 /* Make sure to round up for half microseconds */ 352 data->predicted_us = div_round64(data->expected_us * data->correction_factor[data->bucket], 353 RESOLUTION * DECAY); 354 355 repeat = get_typical_interval(data); 356 357 /* 358 * We want to default to C1 (hlt), not to busy polling 359 * unless the timer is happening really really soon. 360 */ 361 if (data->expected_us > 5 && 362 !drv->states[CPUIDLE_DRIVER_STATE_START].disabled && 363 dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0) 364 data->last_state_idx = CPUIDLE_DRIVER_STATE_START; 365 366 /* 367 * Find the idle state with the lowest power while satisfying 368 * our constraints. 369 */ 370 for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) { 371 struct cpuidle_state *s = &drv->states[i]; 372 struct cpuidle_state_usage *su = &dev->states_usage[i]; 373 374 if (s->disabled || su->disable) 375 continue; 376 if (s->target_residency > data->predicted_us) { 377 low_predicted = 1; 378 continue; 379 } 380 if (s->exit_latency > latency_req) 381 continue; 382 if (s->exit_latency * multiplier > data->predicted_us) 383 continue; 384 385 data->last_state_idx = i; 386 data->exit_us = s->exit_latency; 387 } 388 389 /* not deepest C-state chosen for low predicted residency */ 390 if (low_predicted) { 391 unsigned int timer_us = 0; 392 unsigned int perfect_us = 0; 393 394 /* 395 * Set a timer to detect whether this sleep is much 396 * longer than repeat mode predicted. If the timer 397 * triggers, the code will evaluate whether to put 398 * the CPU into a deeper C-state. 399 * The timer is cancelled on CPU wakeup. 400 */ 401 timer_us = 2 * (data->predicted_us + MAX_DEVIATION); 402 403 perfect_us = perfect_cstate_ms * 1000; 404 405 if (repeat && (4 * timer_us < data->expected_us)) { 406 RCU_NONIDLE(hrtimer_start(hrtmr, 407 ns_to_ktime(1000 * timer_us), 408 HRTIMER_MODE_REL_PINNED)); 409 /* In repeat case, menu hrtimer is started */ 410 per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_REPEAT; 411 } else if (perfect_us < data->expected_us) { 412 /* 413 * The next timer is long. This could be because 414 * we did not make a useful prediction. 415 * In that case, it makes sense to re-enter 416 * into a deeper C-state after some time. 417 */ 418 RCU_NONIDLE(hrtimer_start(hrtmr, 419 ns_to_ktime(1000 * timer_us), 420 HRTIMER_MODE_REL_PINNED)); 421 /* In general case, menu hrtimer is started */ 422 per_cpu(hrtimer_status, cpu) = MENU_HRTIMER_GENERAL; 423 } 424 425 } 426 427 return data->last_state_idx; 428 } 429 430 /** 431 * menu_reflect - records that data structures need update 432 * @dev: the CPU 433 * @index: the index of actual entered state 434 * 435 * NOTE: it's important to be fast here because this operation will add to 436 * the overall exit latency. 437 */ 438 static void menu_reflect(struct cpuidle_device *dev, int index) 439 { 440 struct menu_device *data = &__get_cpu_var(menu_devices); 441 data->last_state_idx = index; 442 if (index >= 0) 443 data->needs_update = 1; 444 } 445 446 /** 447 * menu_update - attempts to guess what happened after entry 448 * @drv: cpuidle driver containing state data 449 * @dev: the CPU 450 */ 451 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev) 452 { 453 struct menu_device *data = &__get_cpu_var(menu_devices); 454 int last_idx = data->last_state_idx; 455 unsigned int last_idle_us = cpuidle_get_last_residency(dev); 456 struct cpuidle_state *target = &drv->states[last_idx]; 457 unsigned int measured_us; 458 u64 new_factor; 459 460 /* 461 * Ugh, this idle state doesn't support residency measurements, so we 462 * are basically lost in the dark. As a compromise, assume we slept 463 * for the whole expected time. 464 */ 465 if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID))) 466 last_idle_us = data->expected_us; 467 468 469 measured_us = last_idle_us; 470 471 /* 472 * We correct for the exit latency; we are assuming here that the 473 * exit latency happens after the event that we're interested in. 474 */ 475 if (measured_us > data->exit_us) 476 measured_us -= data->exit_us; 477 478 479 /* update our correction ratio */ 480 481 new_factor = data->correction_factor[data->bucket] 482 * (DECAY - 1) / DECAY; 483 484 if (data->expected_us > 0 && measured_us < MAX_INTERESTING) 485 new_factor += RESOLUTION * measured_us / data->expected_us; 486 else 487 /* 488 * we were idle so long that we count it as a perfect 489 * prediction 490 */ 491 new_factor += RESOLUTION; 492 493 /* 494 * We don't want 0 as factor; we always want at least 495 * a tiny bit of estimated time. 496 */ 497 if (new_factor == 0) 498 new_factor = 1; 499 500 data->correction_factor[data->bucket] = new_factor; 501 502 /* update the repeating-pattern data */ 503 data->intervals[data->interval_ptr++] = last_idle_us; 504 if (data->interval_ptr >= INTERVALS) 505 data->interval_ptr = 0; 506 } 507 508 /** 509 * menu_enable_device - scans a CPU's states and does setup 510 * @drv: cpuidle driver 511 * @dev: the CPU 512 */ 513 static int menu_enable_device(struct cpuidle_driver *drv, 514 struct cpuidle_device *dev) 515 { 516 struct menu_device *data = &per_cpu(menu_devices, dev->cpu); 517 struct hrtimer *t = &per_cpu(menu_hrtimer, dev->cpu); 518 hrtimer_init(t, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 519 t->function = menu_hrtimer_notify; 520 521 memset(data, 0, sizeof(struct menu_device)); 522 523 return 0; 524 } 525 526 static struct cpuidle_governor menu_governor = { 527 .name = "menu", 528 .rating = 20, 529 .enable = menu_enable_device, 530 .select = menu_select, 531 .reflect = menu_reflect, 532 .owner = THIS_MODULE, 533 }; 534 535 /** 536 * init_menu - initializes the governor 537 */ 538 static int __init init_menu(void) 539 { 540 return cpuidle_register_governor(&menu_governor); 541 } 542 543 /** 544 * exit_menu - exits the governor 545 */ 546 static void __exit exit_menu(void) 547 { 548 cpuidle_unregister_governor(&menu_governor); 549 } 550 551 MODULE_LICENSE("GPL"); 552 module_init(init_menu); 553 module_exit(exit_menu); 554