Commit | Line | Data |
---|---|---|
4f86d3a8 LB |
1 | /* |
2 | * menu.c - the menu idle governor | |
3 | * | |
4 | * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com> | |
69d25870 AV |
5 | * Copyright (C) 2009 Intel Corporation |
6 | * Author: | |
7 | * Arjan van de Ven <arjan@linux.intel.com> | |
4f86d3a8 | 8 | * |
69d25870 AV |
9 | * This code is licenced under the GPL version 2 as described |
10 | * in the COPYING file that acompanies the Linux Kernel. | |
4f86d3a8 LB |
11 | */ |
12 | ||
13 | #include <linux/kernel.h> | |
14 | #include <linux/cpuidle.h> | |
e8db0be1 | 15 | #include <linux/pm_qos.h> |
4f86d3a8 LB |
16 | #include <linux/time.h> |
17 | #include <linux/ktime.h> | |
18 | #include <linux/hrtimer.h> | |
19 | #include <linux/tick.h> | |
69d25870 | 20 | #include <linux/sched.h> |
5787536e | 21 | #include <linux/math64.h> |
884b17e1 | 22 | #include <linux/module.h> |
4f86d3a8 | 23 | |
decd51bb TT |
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 | */ | |
69d25870 | 33 | #define BUCKETS 12 |
1f85f87d | 34 | #define INTERVALS 8 |
69d25870 | 35 | #define RESOLUTION 1024 |
1f85f87d | 36 | #define DECAY 8 |
69d25870 | 37 | #define MAX_INTERESTING 50000 |
1f85f87d AV |
38 | #define STDDEV_THRESH 400 |
39 | ||
69d25870 AV |
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 | * | |
1f85f87d AV |
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 | * | |
69d25870 AV |
90 | * Limiting Performance Impact |
91 | * --------------------------- | |
92 | * C states, especially those with large exit latencies, can have a real | |
20e3341b | 93 | * noticeable impact on workloads, which is not acceptable for most sysadmins, |
69d25870 AV |
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 | */ | |
4f86d3a8 LB |
120 | |
121 | struct menu_device { | |
122 | int last_state_idx; | |
672917dc | 123 | int needs_update; |
4f86d3a8 | 124 | |
5dc2f5a3 | 125 | unsigned int next_timer_us; |
51f245b8 | 126 | unsigned int predicted_us; |
69d25870 | 127 | unsigned int bucket; |
51f245b8 | 128 | unsigned int correction_factor[BUCKETS]; |
939e33b7 | 129 | unsigned int intervals[INTERVALS]; |
1f85f87d | 130 | int interval_ptr; |
4f86d3a8 LB |
131 | }; |
132 | ||
69d25870 AV |
133 | |
134 | #define LOAD_INT(x) ((x) >> FSHIFT) | |
135 | #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) | |
136 | ||
137 | static int get_loadavg(void) | |
138 | { | |
139 | unsigned long this = this_cpu_load(); | |
140 | ||
141 | ||
142 | return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10; | |
143 | } | |
144 | ||
145 | static inline int which_bucket(unsigned int duration) | |
146 | { | |
147 | int bucket = 0; | |
148 | ||
149 | /* | |
150 | * We keep two groups of stats; one with no | |
151 | * IO pending, one without. | |
152 | * This allows us to calculate | |
153 | * E(duration)|iowait | |
154 | */ | |
8c215bd3 | 155 | if (nr_iowait_cpu(smp_processor_id())) |
69d25870 AV |
156 | bucket = BUCKETS/2; |
157 | ||
158 | if (duration < 10) | |
159 | return bucket; | |
160 | if (duration < 100) | |
161 | return bucket + 1; | |
162 | if (duration < 1000) | |
163 | return bucket + 2; | |
164 | if (duration < 10000) | |
165 | return bucket + 3; | |
166 | if (duration < 100000) | |
167 | return bucket + 4; | |
168 | return bucket + 5; | |
169 | } | |
170 | ||
171 | /* | |
172 | * Return a multiplier for the exit latency that is intended | |
173 | * to take performance requirements into account. | |
174 | * The more performance critical we estimate the system | |
175 | * to be, the higher this multiplier, and thus the higher | |
176 | * the barrier to go to an expensive C state. | |
177 | */ | |
178 | static inline int performance_multiplier(void) | |
179 | { | |
180 | int mult = 1; | |
181 | ||
182 | /* for higher loadavg, we are more reluctant */ | |
183 | ||
184 | mult += 2 * get_loadavg(); | |
185 | ||
186 | /* for IO wait tasks (per cpu!) we add 5x each */ | |
8c215bd3 | 187 | mult += 10 * nr_iowait_cpu(smp_processor_id()); |
69d25870 AV |
188 | |
189 | return mult; | |
190 | } | |
191 | ||
4f86d3a8 LB |
192 | static DEFINE_PER_CPU(struct menu_device, menu_devices); |
193 | ||
46bcfad7 | 194 | static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev); |
672917dc | 195 | |
5787536e SH |
196 | /* This implements DIV_ROUND_CLOSEST but avoids 64 bit division */ |
197 | static u64 div_round64(u64 dividend, u32 divisor) | |
198 | { | |
199 | return div_u64(dividend + (divisor / 2), divisor); | |
200 | } | |
201 | ||
1f85f87d AV |
202 | /* |
203 | * Try detecting repeating patterns by keeping track of the last 8 | |
204 | * intervals, and checking if the standard deviation of that set | |
205 | * of points is below a threshold. If it is... then use the | |
206 | * average of these 8 points as the estimated value. | |
207 | */ | |
14851912 | 208 | static void get_typical_interval(struct menu_device *data) |
1f85f87d | 209 | { |
4cd46bca | 210 | int i, divisor; |
0e96d5ad TT |
211 | unsigned int max, thresh; |
212 | uint64_t avg, stddev; | |
213 | ||
214 | thresh = UINT_MAX; /* Discard outliers above this value */ | |
1f85f87d | 215 | |
c96ca4fb | 216 | again: |
1f85f87d | 217 | |
0e96d5ad | 218 | /* First calculate the average of past intervals */ |
4cd46bca TT |
219 | max = 0; |
220 | avg = 0; | |
221 | divisor = 0; | |
c96ca4fb | 222 | for (i = 0; i < INTERVALS; i++) { |
0e96d5ad | 223 | unsigned int value = data->intervals[i]; |
c96ca4fb YS |
224 | if (value <= thresh) { |
225 | avg += value; | |
226 | divisor++; | |
227 | if (value > max) | |
228 | max = value; | |
229 | } | |
230 | } | |
231 | do_div(avg, divisor); | |
232 | ||
0e96d5ad TT |
233 | /* Then try to determine standard deviation */ |
234 | stddev = 0; | |
c96ca4fb | 235 | for (i = 0; i < INTERVALS; i++) { |
0e96d5ad | 236 | unsigned int value = data->intervals[i]; |
c96ca4fb YS |
237 | if (value <= thresh) { |
238 | int64_t diff = value - avg; | |
239 | stddev += diff * diff; | |
240 | } | |
241 | } | |
242 | do_div(stddev, divisor); | |
1f85f87d | 243 | /* |
c96ca4fb YS |
244 | * The typical interval is obtained when standard deviation is small |
245 | * or standard deviation is small compared to the average interval. | |
330647a9 | 246 | * |
0d6a7ffa TT |
247 | * int_sqrt() formal parameter type is unsigned long. When the |
248 | * greatest difference to an outlier exceeds ~65 ms * sqrt(divisor) | |
249 | * the resulting squared standard deviation exceeds the input domain | |
250 | * of int_sqrt on platforms where unsigned long is 32 bits in size. | |
251 | * In such case reject the candidate average. | |
252 | * | |
330647a9 | 253 | * Use this result only if there is no timer to wake us up sooner. |
1f85f87d | 254 | */ |
0d6a7ffa TT |
255 | if (likely(stddev <= ULONG_MAX)) { |
256 | stddev = int_sqrt(stddev); | |
257 | if (((avg > stddev * 6) && (divisor * 4 >= INTERVALS * 3)) | |
c96ca4fb | 258 | || stddev <= 20) { |
5dc2f5a3 | 259 | if (data->next_timer_us > avg) |
0d6a7ffa TT |
260 | data->predicted_us = avg; |
261 | return; | |
262 | } | |
69a37bea | 263 | } |
017099e2 TT |
264 | |
265 | /* | |
266 | * If we have outliers to the upside in our distribution, discard | |
267 | * those by setting the threshold to exclude these outliers, then | |
268 | * calculate the average and standard deviation again. Once we get | |
269 | * down to the bottom 3/4 of our samples, stop excluding samples. | |
270 | * | |
271 | * This can deal with workloads that have long pauses interspersed | |
272 | * with sporadic activity with a bunch of short pauses. | |
273 | */ | |
274 | if ((divisor * 4) <= INTERVALS * 3) | |
275 | return; | |
276 | ||
277 | thresh = max - 1; | |
278 | goto again; | |
1f85f87d AV |
279 | } |
280 | ||
4f86d3a8 LB |
281 | /** |
282 | * menu_select - selects the next idle state to enter | |
46bcfad7 | 283 | * @drv: cpuidle driver containing state data |
4f86d3a8 LB |
284 | * @dev: the CPU |
285 | */ | |
46bcfad7 | 286 | static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev) |
4f86d3a8 LB |
287 | { |
288 | struct menu_device *data = &__get_cpu_var(menu_devices); | |
ed77134b | 289 | int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); |
4f86d3a8 | 290 | int i; |
69d25870 | 291 | int multiplier; |
7467571f | 292 | struct timespec t; |
69d25870 | 293 | |
672917dc | 294 | if (data->needs_update) { |
46bcfad7 | 295 | menu_update(drv, dev); |
672917dc CZ |
296 | data->needs_update = 0; |
297 | } | |
298 | ||
1c6fe036 | 299 | data->last_state_idx = 0; |
1c6fe036 | 300 | |
a2bd9202 | 301 | /* Special case when user has set very strict latency requirement */ |
69d25870 | 302 | if (unlikely(latency_req == 0)) |
a2bd9202 | 303 | return 0; |
a2bd9202 | 304 | |
69d25870 | 305 | /* determine the expected residency time, round up */ |
7467571f | 306 | t = ktime_to_timespec(tick_nohz_get_sleep_length()); |
5dc2f5a3 | 307 | data->next_timer_us = |
7467571f | 308 | t.tv_sec * USEC_PER_SEC + t.tv_nsec / NSEC_PER_USEC; |
69d25870 AV |
309 | |
310 | ||
5dc2f5a3 | 311 | data->bucket = which_bucket(data->next_timer_us); |
69d25870 AV |
312 | |
313 | multiplier = performance_multiplier(); | |
314 | ||
315 | /* | |
316 | * if the correction factor is 0 (eg first time init or cpu hotplug | |
317 | * etc), we actually want to start out with a unity factor. | |
318 | */ | |
319 | if (data->correction_factor[data->bucket] == 0) | |
320 | data->correction_factor[data->bucket] = RESOLUTION * DECAY; | |
321 | ||
51f245b8 TT |
322 | /* |
323 | * Force the result of multiplication to be 64 bits even if both | |
324 | * operands are 32 bits. | |
325 | * Make sure to round up for half microseconds. | |
326 | */ | |
5dc2f5a3 | 327 | data->predicted_us = div_round64((uint64_t)data->next_timer_us * |
51f245b8 | 328 | data->correction_factor[data->bucket], |
5787536e | 329 | RESOLUTION * DECAY); |
69d25870 | 330 | |
14851912 | 331 | get_typical_interval(data); |
1f85f87d | 332 | |
69d25870 AV |
333 | /* |
334 | * We want to default to C1 (hlt), not to busy polling | |
335 | * unless the timer is happening really really soon. | |
336 | */ | |
5dc2f5a3 | 337 | if (data->next_timer_us > 5 && |
cbc9ef02 | 338 | !drv->states[CPUIDLE_DRIVER_STATE_START].disabled && |
dc7fd275 | 339 | dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0) |
69d25870 | 340 | data->last_state_idx = CPUIDLE_DRIVER_STATE_START; |
4f86d3a8 | 341 | |
71abbbf8 AL |
342 | /* |
343 | * Find the idle state with the lowest power while satisfying | |
344 | * our constraints. | |
345 | */ | |
46bcfad7 DD |
346 | for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) { |
347 | struct cpuidle_state *s = &drv->states[i]; | |
dc7fd275 | 348 | struct cpuidle_state_usage *su = &dev->states_usage[i]; |
4f86d3a8 | 349 | |
cbc9ef02 | 350 | if (s->disabled || su->disable) |
3a53396b | 351 | continue; |
14851912 | 352 | if (s->target_residency > data->predicted_us) |
71abbbf8 | 353 | continue; |
a2bd9202 | 354 | if (s->exit_latency > latency_req) |
71abbbf8 | 355 | continue; |
69d25870 | 356 | if (s->exit_latency * multiplier > data->predicted_us) |
71abbbf8 AL |
357 | continue; |
358 | ||
8aef33a7 | 359 | data->last_state_idx = i; |
4f86d3a8 LB |
360 | } |
361 | ||
69d25870 | 362 | return data->last_state_idx; |
4f86d3a8 LB |
363 | } |
364 | ||
365 | /** | |
672917dc | 366 | * menu_reflect - records that data structures need update |
4f86d3a8 | 367 | * @dev: the CPU |
e978aa7d | 368 | * @index: the index of actual entered state |
4f86d3a8 LB |
369 | * |
370 | * NOTE: it's important to be fast here because this operation will add to | |
371 | * the overall exit latency. | |
372 | */ | |
e978aa7d | 373 | static void menu_reflect(struct cpuidle_device *dev, int index) |
672917dc CZ |
374 | { |
375 | struct menu_device *data = &__get_cpu_var(menu_devices); | |
e978aa7d DD |
376 | data->last_state_idx = index; |
377 | if (index >= 0) | |
378 | data->needs_update = 1; | |
672917dc CZ |
379 | } |
380 | ||
381 | /** | |
382 | * menu_update - attempts to guess what happened after entry | |
46bcfad7 | 383 | * @drv: cpuidle driver containing state data |
672917dc CZ |
384 | * @dev: the CPU |
385 | */ | |
46bcfad7 | 386 | static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev) |
4f86d3a8 LB |
387 | { |
388 | struct menu_device *data = &__get_cpu_var(menu_devices); | |
389 | int last_idx = data->last_state_idx; | |
320eee77 | 390 | unsigned int last_idle_us = cpuidle_get_last_residency(dev); |
46bcfad7 | 391 | struct cpuidle_state *target = &drv->states[last_idx]; |
320eee77 | 392 | unsigned int measured_us; |
51f245b8 | 393 | unsigned int new_factor; |
4f86d3a8 LB |
394 | |
395 | /* | |
396 | * Ugh, this idle state doesn't support residency measurements, so we | |
397 | * are basically lost in the dark. As a compromise, assume we slept | |
69d25870 | 398 | * for the whole expected time. |
4f86d3a8 | 399 | */ |
320eee77 | 400 | if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID))) |
5dc2f5a3 | 401 | last_idle_us = data->next_timer_us; |
69d25870 AV |
402 | |
403 | ||
404 | measured_us = last_idle_us; | |
4f86d3a8 | 405 | |
320eee77 | 406 | /* |
69d25870 AV |
407 | * We correct for the exit latency; we are assuming here that the |
408 | * exit latency happens after the event that we're interested in. | |
320eee77 | 409 | */ |
22695ab6 | 410 | if (measured_us > target->exit_latency) |
411 | measured_us -= target->exit_latency; | |
69d25870 | 412 | |
7ac26436 | 413 | /* Make sure our coefficients do not exceed unity */ |
414 | if (measured_us > data->next_timer_us) | |
415 | measured_us = data->next_timer_us; | |
69d25870 | 416 | |
51f245b8 TT |
417 | /* Update our correction ratio */ |
418 | new_factor = data->correction_factor[data->bucket]; | |
419 | new_factor -= new_factor / DECAY; | |
69d25870 | 420 | |
5dc2f5a3 | 421 | if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING) |
422 | new_factor += RESOLUTION * measured_us / data->next_timer_us; | |
320eee77 | 423 | else |
69d25870 AV |
424 | /* |
425 | * we were idle so long that we count it as a perfect | |
426 | * prediction | |
427 | */ | |
428 | new_factor += RESOLUTION; | |
320eee77 | 429 | |
69d25870 AV |
430 | /* |
431 | * We don't want 0 as factor; we always want at least | |
51f245b8 TT |
432 | * a tiny bit of estimated time. Fortunately, due to rounding, |
433 | * new_factor will stay nonzero regardless of measured_us values | |
434 | * and the compiler can eliminate this test as long as DECAY > 1. | |
69d25870 | 435 | */ |
51f245b8 | 436 | if (DECAY == 1 && unlikely(new_factor == 0)) |
69d25870 | 437 | new_factor = 1; |
320eee77 | 438 | |
69d25870 | 439 | data->correction_factor[data->bucket] = new_factor; |
1f85f87d AV |
440 | |
441 | /* update the repeating-pattern data */ | |
442 | data->intervals[data->interval_ptr++] = last_idle_us; | |
443 | if (data->interval_ptr >= INTERVALS) | |
444 | data->interval_ptr = 0; | |
4f86d3a8 LB |
445 | } |
446 | ||
447 | /** | |
448 | * menu_enable_device - scans a CPU's states and does setup | |
46bcfad7 | 449 | * @drv: cpuidle driver |
4f86d3a8 LB |
450 | * @dev: the CPU |
451 | */ | |
46bcfad7 DD |
452 | static int menu_enable_device(struct cpuidle_driver *drv, |
453 | struct cpuidle_device *dev) | |
4f86d3a8 LB |
454 | { |
455 | struct menu_device *data = &per_cpu(menu_devices, dev->cpu); | |
456 | ||
457 | memset(data, 0, sizeof(struct menu_device)); | |
458 | ||
459 | return 0; | |
460 | } | |
461 | ||
462 | static struct cpuidle_governor menu_governor = { | |
463 | .name = "menu", | |
464 | .rating = 20, | |
465 | .enable = menu_enable_device, | |
466 | .select = menu_select, | |
467 | .reflect = menu_reflect, | |
468 | .owner = THIS_MODULE, | |
469 | }; | |
470 | ||
471 | /** | |
472 | * init_menu - initializes the governor | |
473 | */ | |
474 | static int __init init_menu(void) | |
475 | { | |
476 | return cpuidle_register_governor(&menu_governor); | |
477 | } | |
478 | ||
137b944e | 479 | postcore_initcall(init_menu); |