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 32 /* 33 * Concepts and ideas behind the menu governor 34 * 35 * For the menu governor, there are 3 decision factors for picking a C 36 * state: 37 * 1) Energy break even point 38 * 2) Performance impact 39 * 3) Latency tolerance (from pmqos infrastructure) 40 * These these three factors are treated independently. 41 * 42 * Energy break even point 43 * ----------------------- 44 * C state entry and exit have an energy cost, and a certain amount of time in 45 * the C state is required to actually break even on this cost. CPUIDLE 46 * provides us this duration in the "target_residency" field. So all that we 47 * need is a good prediction of how long we'll be idle. Like the traditional 48 * menu governor, we start with the actual known "next timer event" time. 49 * 50 * Since there are other source of wakeups (interrupts for example) than 51 * the next timer event, this estimation is rather optimistic. To get a 52 * more realistic estimate, a correction factor is applied to the estimate, 53 * that is based on historic behavior. For example, if in the past the actual 54 * duration always was 50% of the next timer tick, the correction factor will 55 * be 0.5. 56 * 57 * menu uses a running average for this correction factor, however it uses a 58 * set of factors, not just a single factor. This stems from the realization 59 * that the ratio is dependent on the order of magnitude of the expected 60 * duration; if we expect 500 milliseconds of idle time the likelihood of 61 * getting an interrupt very early is much higher than if we expect 50 micro 62 * seconds of idle time. A second independent factor that has big impact on 63 * the actual factor is if there is (disk) IO outstanding or not. 64 * (as a special twist, we consider every sleep longer than 50 milliseconds 65 * as perfect; there are no power gains for sleeping longer than this) 66 * 67 * For these two reasons we keep an array of 12 independent factors, that gets 68 * indexed based on the magnitude of the expected duration as well as the 69 * "is IO outstanding" property. 70 * 71 * Repeatable-interval-detector 72 * ---------------------------- 73 * There are some cases where "next timer" is a completely unusable predictor: 74 * Those cases where the interval is fixed, for example due to hardware 75 * interrupt mitigation, but also due to fixed transfer rate devices such as 76 * mice. 77 * For this, we use a different predictor: We track the duration of the last 8 78 * intervals and if the stand deviation of these 8 intervals is below a 79 * threshold value, we use the average of these intervals as prediction. 80 * 81 * Limiting Performance Impact 82 * --------------------------- 83 * C states, especially those with large exit latencies, can have a real 84 * noticeable impact on workloads, which is not acceptable for most sysadmins, 85 * and in addition, less performance has a power price of its own. 86 * 87 * As a general rule of thumb, menu assumes that the following heuristic 88 * holds: 89 * The busier the system, the less impact of C states is acceptable 90 * 91 * This rule-of-thumb is implemented using a performance-multiplier: 92 * If the exit latency times the performance multiplier is longer than 93 * the predicted duration, the C state is not considered a candidate 94 * for selection due to a too high performance impact. So the higher 95 * this multiplier is, the longer we need to be idle to pick a deep C 96 * state, and thus the less likely a busy CPU will hit such a deep 97 * C state. 98 * 99 * Two factors are used in determing this multiplier: 100 * a value of 10 is added for each point of "per cpu load average" we have. 101 * a value of 5 points is added for each process that is waiting for 102 * IO on this CPU. 103 * (these values are experimentally determined) 104 * 105 * The load average factor gives a longer term (few seconds) input to the 106 * decision, while the iowait value gives a cpu local instantanious input. 107 * The iowait factor may look low, but realize that this is also already 108 * represented in the system load average. 109 * 110 */ 111 112 struct menu_device { 113 int last_state_idx; 114 int needs_update; 115 116 unsigned int expected_us; 117 u64 predicted_us; 118 unsigned int exit_us; 119 unsigned int bucket; 120 u64 correction_factor[BUCKETS]; 121 u32 intervals[INTERVALS]; 122 int interval_ptr; 123 }; 124 125 126 #define LOAD_INT(x) ((x) >> FSHIFT) 127 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) 128 129 static int get_loadavg(void) 130 { 131 unsigned long this = this_cpu_load(); 132 133 134 return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10; 135 } 136 137 static inline int which_bucket(unsigned int duration) 138 { 139 int bucket = 0; 140 141 /* 142 * We keep two groups of stats; one with no 143 * IO pending, one without. 144 * This allows us to calculate 145 * E(duration)|iowait 146 */ 147 if (nr_iowait_cpu(smp_processor_id())) 148 bucket = BUCKETS/2; 149 150 if (duration < 10) 151 return bucket; 152 if (duration < 100) 153 return bucket + 1; 154 if (duration < 1000) 155 return bucket + 2; 156 if (duration < 10000) 157 return bucket + 3; 158 if (duration < 100000) 159 return bucket + 4; 160 return bucket + 5; 161 } 162 163 /* 164 * Return a multiplier for the exit latency that is intended 165 * to take performance requirements into account. 166 * The more performance critical we estimate the system 167 * to be, the higher this multiplier, and thus the higher 168 * the barrier to go to an expensive C state. 169 */ 170 static inline int performance_multiplier(void) 171 { 172 int mult = 1; 173 174 /* for higher loadavg, we are more reluctant */ 175 176 mult += 2 * get_loadavg(); 177 178 /* for IO wait tasks (per cpu!) we add 5x each */ 179 mult += 10 * nr_iowait_cpu(smp_processor_id()); 180 181 return mult; 182 } 183 184 static DEFINE_PER_CPU(struct menu_device, menu_devices); 185 186 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev); 187 188 /* This implements DIV_ROUND_CLOSEST but avoids 64 bit division */ 189 static u64 div_round64(u64 dividend, u32 divisor) 190 { 191 return div_u64(dividend + (divisor / 2), divisor); 192 } 193 194 /* 195 * Try detecting repeating patterns by keeping track of the last 8 196 * intervals, and checking if the standard deviation of that set 197 * of points is below a threshold. If it is... then use the 198 * average of these 8 points as the estimated value. 199 */ 200 static void detect_repeating_patterns(struct menu_device *data) 201 { 202 int i; 203 uint64_t avg = 0; 204 uint64_t stddev = 0; /* contains the square of the std deviation */ 205 206 /* first calculate average and standard deviation of the past */ 207 for (i = 0; i < INTERVALS; i++) 208 avg += data->intervals[i]; 209 avg = avg / INTERVALS; 210 211 /* if the avg is beyond the known next tick, it's worthless */ 212 if (avg > data->expected_us) 213 return; 214 215 for (i = 0; i < INTERVALS; i++) 216 stddev += (data->intervals[i] - avg) * 217 (data->intervals[i] - avg); 218 219 stddev = stddev / INTERVALS; 220 221 /* 222 * now.. if stddev is small.. then assume we have a 223 * repeating pattern and predict we keep doing this. 224 */ 225 226 if (avg && stddev < STDDEV_THRESH) 227 data->predicted_us = avg; 228 } 229 230 /** 231 * menu_select - selects the next idle state to enter 232 * @drv: cpuidle driver containing state data 233 * @dev: the CPU 234 */ 235 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev) 236 { 237 struct menu_device *data = &__get_cpu_var(menu_devices); 238 int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); 239 int power_usage = -1; 240 int i; 241 int multiplier; 242 struct timespec t; 243 244 if (data->needs_update) { 245 menu_update(drv, dev); 246 data->needs_update = 0; 247 } 248 249 data->last_state_idx = 0; 250 data->exit_us = 0; 251 252 /* Special case when user has set very strict latency requirement */ 253 if (unlikely(latency_req == 0)) 254 return 0; 255 256 /* determine the expected residency time, round up */ 257 t = ktime_to_timespec(tick_nohz_get_sleep_length()); 258 data->expected_us = 259 t.tv_sec * USEC_PER_SEC + t.tv_nsec / NSEC_PER_USEC; 260 261 262 data->bucket = which_bucket(data->expected_us); 263 264 multiplier = performance_multiplier(); 265 266 /* 267 * if the correction factor is 0 (eg first time init or cpu hotplug 268 * etc), we actually want to start out with a unity factor. 269 */ 270 if (data->correction_factor[data->bucket] == 0) 271 data->correction_factor[data->bucket] = RESOLUTION * DECAY; 272 273 /* Make sure to round up for half microseconds */ 274 data->predicted_us = div_round64(data->expected_us * data->correction_factor[data->bucket], 275 RESOLUTION * DECAY); 276 277 detect_repeating_patterns(data); 278 279 /* 280 * We want to default to C1 (hlt), not to busy polling 281 * unless the timer is happening really really soon. 282 */ 283 if (data->expected_us > 5 && 284 !drv->states[CPUIDLE_DRIVER_STATE_START].disabled && 285 dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0) 286 data->last_state_idx = CPUIDLE_DRIVER_STATE_START; 287 288 /* 289 * Find the idle state with the lowest power while satisfying 290 * our constraints. 291 */ 292 for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) { 293 struct cpuidle_state *s = &drv->states[i]; 294 struct cpuidle_state_usage *su = &dev->states_usage[i]; 295 296 if (s->disabled || su->disable) 297 continue; 298 if (s->target_residency > data->predicted_us) 299 continue; 300 if (s->exit_latency > latency_req) 301 continue; 302 if (s->exit_latency * multiplier > data->predicted_us) 303 continue; 304 305 if (s->power_usage < power_usage) { 306 power_usage = s->power_usage; 307 data->last_state_idx = i; 308 data->exit_us = s->exit_latency; 309 } 310 } 311 312 return data->last_state_idx; 313 } 314 315 /** 316 * menu_reflect - records that data structures need update 317 * @dev: the CPU 318 * @index: the index of actual entered state 319 * 320 * NOTE: it's important to be fast here because this operation will add to 321 * the overall exit latency. 322 */ 323 static void menu_reflect(struct cpuidle_device *dev, int index) 324 { 325 struct menu_device *data = &__get_cpu_var(menu_devices); 326 data->last_state_idx = index; 327 if (index >= 0) 328 data->needs_update = 1; 329 } 330 331 /** 332 * menu_update - attempts to guess what happened after entry 333 * @drv: cpuidle driver containing state data 334 * @dev: the CPU 335 */ 336 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev) 337 { 338 struct menu_device *data = &__get_cpu_var(menu_devices); 339 int last_idx = data->last_state_idx; 340 unsigned int last_idle_us = cpuidle_get_last_residency(dev); 341 struct cpuidle_state *target = &drv->states[last_idx]; 342 unsigned int measured_us; 343 u64 new_factor; 344 345 /* 346 * Ugh, this idle state doesn't support residency measurements, so we 347 * are basically lost in the dark. As a compromise, assume we slept 348 * for the whole expected time. 349 */ 350 if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID))) 351 last_idle_us = data->expected_us; 352 353 354 measured_us = last_idle_us; 355 356 /* 357 * We correct for the exit latency; we are assuming here that the 358 * exit latency happens after the event that we're interested in. 359 */ 360 if (measured_us > data->exit_us) 361 measured_us -= data->exit_us; 362 363 364 /* update our correction ratio */ 365 366 new_factor = data->correction_factor[data->bucket] 367 * (DECAY - 1) / DECAY; 368 369 if (data->expected_us > 0 && measured_us < MAX_INTERESTING) 370 new_factor += RESOLUTION * measured_us / data->expected_us; 371 else 372 /* 373 * we were idle so long that we count it as a perfect 374 * prediction 375 */ 376 new_factor += RESOLUTION; 377 378 /* 379 * We don't want 0 as factor; we always want at least 380 * a tiny bit of estimated time. 381 */ 382 if (new_factor == 0) 383 new_factor = 1; 384 385 data->correction_factor[data->bucket] = new_factor; 386 387 /* update the repeating-pattern data */ 388 data->intervals[data->interval_ptr++] = last_idle_us; 389 if (data->interval_ptr >= INTERVALS) 390 data->interval_ptr = 0; 391 } 392 393 /** 394 * menu_enable_device - scans a CPU's states and does setup 395 * @drv: cpuidle driver 396 * @dev: the CPU 397 */ 398 static int menu_enable_device(struct cpuidle_driver *drv, 399 struct cpuidle_device *dev) 400 { 401 struct menu_device *data = &per_cpu(menu_devices, dev->cpu); 402 403 memset(data, 0, sizeof(struct menu_device)); 404 405 return 0; 406 } 407 408 static struct cpuidle_governor menu_governor = { 409 .name = "menu", 410 .rating = 20, 411 .enable = menu_enable_device, 412 .select = menu_select, 413 .reflect = menu_reflect, 414 .owner = THIS_MODULE, 415 }; 416 417 /** 418 * init_menu - initializes the governor 419 */ 420 static int __init init_menu(void) 421 { 422 return cpuidle_register_governor(&menu_governor); 423 } 424 425 /** 426 * exit_menu - exits the governor 427 */ 428 static void __exit exit_menu(void) 429 { 430 cpuidle_unregister_governor(&menu_governor); 431 } 432 433 MODULE_LICENSE("GPL"); 434 module_init(init_menu); 435 module_exit(exit_menu); 436