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_params.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 23#define BUCKETS 12 24#define INTERVALS 8 25#define RESOLUTION 1024 26#define DECAY 8 27#define MAX_INTERESTING 50000 28#define STDDEV_THRESH 400 29 30 31/* 32 * Concepts and ideas behind the menu governor 33 * 34 * For the menu governor, there are 3 decision factors for picking a C 35 * state: 36 * 1) Energy break even point 37 * 2) Performance impact 38 * 3) Latency tolerance (from pmqos infrastructure) 39 * These these three factors are treated independently. 40 * 41 * Energy break even point 42 * ----------------------- 43 * C state entry and exit have an energy cost, and a certain amount of time in 44 * the C state is required to actually break even on this cost. CPUIDLE 45 * provides us this duration in the "target_residency" field. So all that we 46 * need is a good prediction of how long we'll be idle. Like the traditional 47 * menu governor, we start with the actual known "next timer event" time. 48 * 49 * Since there are other source of wakeups (interrupts for example) than 50 * the next timer event, this estimation is rather optimistic. To get a 51 * more realistic estimate, a correction factor is applied to the estimate, 52 * that is based on historic behavior. For example, if in the past the actual 53 * duration always was 50% of the next timer tick, the correction factor will 54 * be 0.5. 55 * 56 * menu uses a running average for this correction factor, however it uses a 57 * set of factors, not just a single factor. This stems from the realization 58 * that the ratio is dependent on the order of magnitude of the expected 59 * duration; if we expect 500 milliseconds of idle time the likelihood of 60 * getting an interrupt very early is much higher than if we expect 50 micro 61 * seconds of idle time. A second independent factor that has big impact on 62 * the actual factor is if there is (disk) IO outstanding or not. 63 * (as a special twist, we consider every sleep longer than 50 milliseconds 64 * as perfect; there are no power gains for sleeping longer than this) 65 * 66 * For these two reasons we keep an array of 12 independent factors, that gets 67 * indexed based on the magnitude of the expected duration as well as the 68 * "is IO outstanding" property. 69 * 70 * Repeatable-interval-detector 71 * ---------------------------- 72 * There are some cases where "next timer" is a completely unusable predictor: 73 * Those cases where the interval is fixed, for example due to hardware 74 * interrupt mitigation, but also due to fixed transfer rate devices such as 75 * mice. 76 * For this, we use a different predictor: We track the duration of the last 8 77 * intervals and if the stand deviation of these 8 intervals is below a 78 * threshold value, we use the average of these intervals as prediction. 79 * 80 * Limiting Performance Impact 81 * --------------------------- 82 * C states, especially those with large exit latencies, can have a real 83 * noticeable impact on workloads, which is not acceptable for most sysadmins, 84 * and in addition, less performance has a power price of its own. 85 * 86 * As a general rule of thumb, menu assumes that the following heuristic 87 * holds: 88 * The busier the system, the less impact of C states is acceptable 89 * 90 * This rule-of-thumb is implemented using a performance-multiplier: 91 * If the exit latency times the performance multiplier is longer than 92 * the predicted duration, the C state is not considered a candidate 93 * for selection due to a too high performance impact. So the higher 94 * this multiplier is, the longer we need to be idle to pick a deep C 95 * state, and thus the less likely a busy CPU will hit such a deep 96 * C state. 97 * 98 * Two factors are used in determing this multiplier: 99 * a value of 10 is added for each point of "per cpu load average" we have. 100 * a value of 5 points is added for each process that is waiting for 101 * IO on this CPU. 102 * (these values are experimentally determined) 103 * 104 * The load average factor gives a longer term (few seconds) input to the 105 * decision, while the iowait value gives a cpu local instantanious input. 106 * The iowait factor may look low, but realize that this is also already 107 * represented in the system load average. 108 * 109 */ 110 111struct menu_device { 112 int last_state_idx; 113 int needs_update; 114 115 unsigned int expected_us; 116 u64 predicted_us; 117 unsigned int exit_us; 118 unsigned int bucket; 119 u64 correction_factor[BUCKETS]; 120 u32 intervals[INTERVALS]; 121 int interval_ptr; 122}; 123 124 125#define LOAD_INT(x) ((x) >> FSHIFT) 126#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) 127 128static int get_loadavg(void) 129{ 130 unsigned long this = this_cpu_load(); 131 132 133 return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10; 134} 135 136static inline int which_bucket(unsigned int duration) 137{ 138 int bucket = 0; 139 140 /* 141 * We keep two groups of stats; one with no 142 * IO pending, one without. 143 * This allows us to calculate 144 * E(duration)|iowait 145 */ 146 if (nr_iowait_cpu(smp_processor_id())) 147 bucket = BUCKETS/2; 148 149 if (duration < 10) 150 return bucket; 151 if (duration < 100) 152 return bucket + 1; 153 if (duration < 1000) 154 return bucket + 2; 155 if (duration < 10000) 156 return bucket + 3; 157 if (duration < 100000) 158 return bucket + 4; 159 return bucket + 5; 160} 161 162/* 163 * Return a multiplier for the exit latency that is intended 164 * to take performance requirements into account. 165 * The more performance critical we estimate the system 166 * to be, the higher this multiplier, and thus the higher 167 * the barrier to go to an expensive C state. 168 */ 169static inline int performance_multiplier(void) 170{ 171 int mult = 1; 172 173 /* for higher loadavg, we are more reluctant */ 174 175 mult += 2 * get_loadavg(); 176 177 /* for IO wait tasks (per cpu!) we add 5x each */ 178 mult += 10 * nr_iowait_cpu(smp_processor_id()); 179 180 return mult; 181} 182 183static DEFINE_PER_CPU(struct menu_device, menu_devices); 184 185static void menu_update(struct cpuidle_device *dev); 186 187/* This implements DIV_ROUND_CLOSEST but avoids 64 bit division */ 188static u64 div_round64(u64 dividend, u32 divisor) 189{ 190 return div_u64(dividend + (divisor / 2), divisor); 191} 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 detect_repeating_patterns(struct menu_device *data) 200{ 201 int i; 202 uint64_t avg = 0; 203 uint64_t stddev = 0; /* contains the square of the std deviation */ 204 205 /* first calculate average and standard deviation of the past */ 206 for (i = 0; i < INTERVALS; i++) 207 avg += data->intervals[i]; 208 avg = avg / INTERVALS; 209 210 /* if the avg is beyond the known next tick, it's worthless */ 211 if (avg > data->expected_us) 212 return; 213 214 for (i = 0; i < INTERVALS; i++) 215 stddev += (data->intervals[i] - avg) * 216 (data->intervals[i] - avg); 217 218 stddev = stddev / INTERVALS; 219 220 /* 221 * now.. if stddev is small.. then assume we have a 222 * repeating pattern and predict we keep doing this. 223 */ 224 225 if (avg && stddev < STDDEV_THRESH) 226 data->predicted_us = avg; 227} 228 229/** 230 * menu_select - selects the next idle state to enter 231 * @dev: the CPU 232 */ 233static int menu_select(struct cpuidle_device *dev) 234{ 235 struct menu_device *data = &__get_cpu_var(menu_devices); 236 int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); 237 unsigned int power_usage = -1; 238 int i; 239 int multiplier; 240 241 if (data->needs_update) { 242 menu_update(dev); 243 data->needs_update = 0; 244 } 245 246 data->last_state_idx = 0; 247 data->exit_us = 0; 248 249 /* Special case when user has set very strict latency requirement */ 250 if (unlikely(latency_req == 0)) 251 return 0; 252 253 /* determine the expected residency time, round up */ 254 data->expected_us = 255 DIV_ROUND_UP((u32)ktime_to_ns(tick_nohz_get_sleep_length()), 1000); 256 257 258 data->bucket = which_bucket(data->expected_us); 259 260 multiplier = performance_multiplier(); 261 262 /* 263 * if the correction factor is 0 (eg first time init or cpu hotplug 264 * etc), we actually want to start out with a unity factor. 265 */ 266 if (data->correction_factor[data->bucket] == 0) 267 data->correction_factor[data->bucket] = RESOLUTION * DECAY; 268 269 /* Make sure to round up for half microseconds */ 270 data->predicted_us = div_round64(data->expected_us * data->correction_factor[data->bucket], 271 RESOLUTION * DECAY); 272 273 detect_repeating_patterns(data); 274 275 /* 276 * We want to default to C1 (hlt), not to busy polling 277 * unless the timer is happening really really soon. 278 */ 279 if (data->expected_us > 5) 280 data->last_state_idx = CPUIDLE_DRIVER_STATE_START; 281 282 /* 283 * Find the idle state with the lowest power while satisfying 284 * our constraints. 285 */ 286 for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++) { 287 struct cpuidle_state *s = &dev->states[i]; 288 289 if (s->flags & CPUIDLE_FLAG_IGNORE) 290 continue; 291 if (s->target_residency > data->predicted_us) 292 continue; 293 if (s->exit_latency > latency_req) 294 continue; 295 if (s->exit_latency * multiplier > data->predicted_us) 296 continue; 297 298 if (s->power_usage < power_usage) { 299 power_usage = s->power_usage; 300 data->last_state_idx = i; 301 data->exit_us = s->exit_latency; 302 } 303 } 304 305 return data->last_state_idx; 306} 307 308/** 309 * menu_reflect - records that data structures need update 310 * @dev: the CPU 311 * 312 * NOTE: it's important to be fast here because this operation will add to 313 * the overall exit latency. 314 */ 315static void menu_reflect(struct cpuidle_device *dev) 316{ 317 struct menu_device *data = &__get_cpu_var(menu_devices); 318 data->needs_update = 1; 319} 320 321/** 322 * menu_update - attempts to guess what happened after entry 323 * @dev: the CPU 324 */ 325static void menu_update(struct cpuidle_device *dev) 326{ 327 struct menu_device *data = &__get_cpu_var(menu_devices); 328 int last_idx = data->last_state_idx; 329 unsigned int last_idle_us = cpuidle_get_last_residency(dev); 330 struct cpuidle_state *target = &dev->states[last_idx]; 331 unsigned int measured_us; 332 u64 new_factor; 333 334 /* 335 * Ugh, this idle state doesn't support residency measurements, so we 336 * are basically lost in the dark. As a compromise, assume we slept 337 * for the whole expected time. 338 */ 339 if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID))) 340 last_idle_us = data->expected_us; 341 342 343 measured_us = last_idle_us; 344 345 /* 346 * We correct for the exit latency; we are assuming here that the 347 * exit latency happens after the event that we're interested in. 348 */ 349 if (measured_us > data->exit_us) 350 measured_us -= data->exit_us; 351 352 353 /* update our correction ratio */ 354 355 new_factor = data->correction_factor[data->bucket] 356 * (DECAY - 1) / DECAY; 357 358 if (data->expected_us > 0 && measured_us < MAX_INTERESTING) 359 new_factor += RESOLUTION * measured_us / data->expected_us; 360 else 361 /* 362 * we were idle so long that we count it as a perfect 363 * prediction 364 */ 365 new_factor += RESOLUTION; 366 367 /* 368 * We don't want 0 as factor; we always want at least 369 * a tiny bit of estimated time. 370 */ 371 if (new_factor == 0) 372 new_factor = 1; 373 374 data->correction_factor[data->bucket] = new_factor; 375 376 /* update the repeating-pattern data */ 377 data->intervals[data->interval_ptr++] = last_idle_us; 378 if (data->interval_ptr >= INTERVALS) 379 data->interval_ptr = 0; 380} 381 382/** 383 * menu_enable_device - scans a CPU's states and does setup 384 * @dev: the CPU 385 */ 386static int menu_enable_device(struct cpuidle_device *dev) 387{ 388 struct menu_device *data = &per_cpu(menu_devices, dev->cpu); 389 390 memset(data, 0, sizeof(struct menu_device)); 391 392 return 0; 393} 394 395static struct cpuidle_governor menu_governor = { 396 .name = "menu", 397 .rating = 20, 398 .enable = menu_enable_device, 399 .select = menu_select, 400 .reflect = menu_reflect, 401 .owner = THIS_MODULE, 402}; 403 404/** 405 * init_menu - initializes the governor 406 */ 407static int __init init_menu(void) 408{ 409 return cpuidle_register_governor(&menu_governor); 410} 411 412/** 413 * exit_menu - exits the governor 414 */ 415static void __exit exit_menu(void) 416{ 417 cpuidle_unregister_governor(&menu_governor); 418} 419 420MODULE_LICENSE("GPL"); 421module_init(init_menu); 422module_exit(exit_menu); 423