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 |
ae779300 MG |
34 | #define INTERVAL_SHIFT 3 |
35 | #define INTERVALS (1UL << INTERVAL_SHIFT) | |
69d25870 | 36 | #define RESOLUTION 1024 |
1f85f87d | 37 | #define DECAY 8 |
69d25870 | 38 | #define MAX_INTERESTING 50000 |
1f85f87d | 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 | ||
372ba8cb | 137 | static inline int get_loadavg(unsigned long load) |
69d25870 | 138 | { |
372ba8cb | 139 | return LOAD_INT(load) * 10 + LOAD_FRAC(load) / 10; |
69d25870 AV |
140 | } |
141 | ||
64b4ca5c | 142 | static inline int which_bucket(unsigned int duration, unsigned long nr_iowaiters) |
69d25870 AV |
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 | */ | |
64b4ca5c | 152 | if (nr_iowaiters) |
69d25870 AV |
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 | */ | |
372ba8cb | 175 | static inline int performance_multiplier(unsigned long nr_iowaiters, unsigned long load) |
69d25870 AV |
176 | { |
177 | int mult = 1; | |
178 | ||
179 | /* for higher loadavg, we are more reluctant */ | |
180 | ||
372ba8cb | 181 | mult += 2 * get_loadavg(load); |
69d25870 AV |
182 | |
183 | /* for IO wait tasks (per cpu!) we add 5x each */ | |
64b4ca5c | 184 | mult += 10 * nr_iowaiters; |
69d25870 AV |
185 | |
186 | return mult; | |
187 | } | |
188 | ||
4f86d3a8 LB |
189 | static DEFINE_PER_CPU(struct menu_device, menu_devices); |
190 | ||
46bcfad7 | 191 | static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev); |
672917dc | 192 | |
1f85f87d AV |
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 | */ | |
14851912 | 199 | static void get_typical_interval(struct menu_device *data) |
1f85f87d | 200 | { |
4cd46bca | 201 | int i, divisor; |
0e96d5ad | 202 | unsigned int max, thresh; |
7024b18c | 203 | uint64_t avg, variance; |
0e96d5ad TT |
204 | |
205 | thresh = UINT_MAX; /* Discard outliers above this value */ | |
1f85f87d | 206 | |
c96ca4fb | 207 | again: |
1f85f87d | 208 | |
0e96d5ad | 209 | /* First calculate the average of past intervals */ |
4cd46bca TT |
210 | max = 0; |
211 | avg = 0; | |
212 | divisor = 0; | |
c96ca4fb | 213 | for (i = 0; i < INTERVALS; i++) { |
0e96d5ad | 214 | unsigned int value = data->intervals[i]; |
c96ca4fb YS |
215 | if (value <= thresh) { |
216 | avg += value; | |
217 | divisor++; | |
218 | if (value > max) | |
219 | max = value; | |
220 | } | |
221 | } | |
ae779300 MG |
222 | if (divisor == INTERVALS) |
223 | avg >>= INTERVAL_SHIFT; | |
224 | else | |
225 | do_div(avg, divisor); | |
c96ca4fb | 226 | |
7024b18c RV |
227 | /* Then try to determine variance */ |
228 | variance = 0; | |
c96ca4fb | 229 | for (i = 0; i < INTERVALS; i++) { |
0e96d5ad | 230 | unsigned int value = data->intervals[i]; |
c96ca4fb YS |
231 | if (value <= thresh) { |
232 | int64_t diff = value - avg; | |
7024b18c | 233 | variance += diff * diff; |
c96ca4fb YS |
234 | } |
235 | } | |
ae779300 | 236 | if (divisor == INTERVALS) |
7024b18c | 237 | variance >>= INTERVAL_SHIFT; |
ae779300 | 238 | else |
7024b18c | 239 | do_div(variance, divisor); |
ae779300 | 240 | |
1f85f87d | 241 | /* |
7024b18c RV |
242 | * The typical interval is obtained when standard deviation is |
243 | * small (stddev <= 20 us, variance <= 400 us^2) or standard | |
244 | * deviation is small compared to the average interval (avg > | |
245 | * 6*stddev, avg^2 > 36*variance). The average is smaller than | |
246 | * UINT_MAX aka U32_MAX, so computing its square does not | |
247 | * overflow a u64. We simply reject this candidate average if | |
248 | * the standard deviation is greater than 715 s (which is | |
249 | * rather unlikely). | |
0d6a7ffa | 250 | * |
330647a9 | 251 | * Use this result only if there is no timer to wake us up sooner. |
1f85f87d | 252 | */ |
7024b18c RV |
253 | if (likely(variance <= U64_MAX/36)) { |
254 | if (((avg*avg > variance*36) && (divisor * 4 >= INTERVALS * 3)) | |
255 | || variance <= 400) { | |
5dc2f5a3 | 256 | if (data->next_timer_us > avg) |
0d6a7ffa TT |
257 | data->predicted_us = avg; |
258 | return; | |
259 | } | |
69a37bea | 260 | } |
017099e2 TT |
261 | |
262 | /* | |
263 | * If we have outliers to the upside in our distribution, discard | |
264 | * those by setting the threshold to exclude these outliers, then | |
265 | * calculate the average and standard deviation again. Once we get | |
266 | * down to the bottom 3/4 of our samples, stop excluding samples. | |
267 | * | |
268 | * This can deal with workloads that have long pauses interspersed | |
269 | * with sporadic activity with a bunch of short pauses. | |
270 | */ | |
271 | if ((divisor * 4) <= INTERVALS * 3) | |
272 | return; | |
273 | ||
274 | thresh = max - 1; | |
275 | goto again; | |
1f85f87d AV |
276 | } |
277 | ||
4f86d3a8 LB |
278 | /** |
279 | * menu_select - selects the next idle state to enter | |
46bcfad7 | 280 | * @drv: cpuidle driver containing state data |
4f86d3a8 LB |
281 | * @dev: the CPU |
282 | */ | |
46bcfad7 | 283 | static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev) |
4f86d3a8 | 284 | { |
229b6863 | 285 | struct menu_device *data = this_cpu_ptr(&menu_devices); |
ed77134b | 286 | int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY); |
4f86d3a8 | 287 | int i; |
96e95182 | 288 | unsigned int interactivity_req; |
372ba8cb | 289 | unsigned long nr_iowaiters, cpu_load; |
69d25870 | 290 | |
672917dc | 291 | if (data->needs_update) { |
46bcfad7 | 292 | menu_update(drv, dev); |
672917dc CZ |
293 | data->needs_update = 0; |
294 | } | |
295 | ||
a2bd9202 | 296 | /* Special case when user has set very strict latency requirement */ |
69d25870 | 297 | if (unlikely(latency_req == 0)) |
a2bd9202 | 298 | return 0; |
a2bd9202 | 299 | |
69d25870 | 300 | /* determine the expected residency time, round up */ |
107d4f46 | 301 | data->next_timer_us = ktime_to_us(tick_nohz_get_sleep_length()); |
69d25870 | 302 | |
372ba8cb | 303 | get_iowait_load(&nr_iowaiters, &cpu_load); |
64b4ca5c | 304 | data->bucket = which_bucket(data->next_timer_us, nr_iowaiters); |
69d25870 | 305 | |
51f245b8 TT |
306 | /* |
307 | * Force the result of multiplication to be 64 bits even if both | |
308 | * operands are 32 bits. | |
309 | * Make sure to round up for half microseconds. | |
310 | */ | |
ee3c86f3 | 311 | data->predicted_us = DIV_ROUND_CLOSEST_ULL((uint64_t)data->next_timer_us * |
51f245b8 | 312 | data->correction_factor[data->bucket], |
5787536e | 313 | RESOLUTION * DECAY); |
69d25870 | 314 | |
14851912 | 315 | get_typical_interval(data); |
1f85f87d | 316 | |
96e95182 | 317 | /* |
318 | * Performance multiplier defines a minimum predicted idle | |
319 | * duration / latency ratio. Adjust the latency limit if | |
320 | * necessary. | |
321 | */ | |
372ba8cb | 322 | interactivity_req = data->predicted_us / performance_multiplier(nr_iowaiters, cpu_load); |
96e95182 | 323 | if (latency_req > interactivity_req) |
324 | latency_req = interactivity_req; | |
325 | ||
9c4b2867 RW |
326 | if (CPUIDLE_DRIVER_STATE_START > 0) { |
327 | data->last_state_idx = CPUIDLE_DRIVER_STATE_START - 1; | |
328 | /* | |
329 | * We want to default to C1 (hlt), not to busy polling | |
330 | * unless the timer is happening really really soon. | |
331 | */ | |
332 | if (interactivity_req > 20 && | |
333 | !drv->states[CPUIDLE_DRIVER_STATE_START].disabled && | |
334 | dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0) | |
335 | data->last_state_idx = CPUIDLE_DRIVER_STATE_START; | |
336 | } else { | |
69d25870 | 337 | data->last_state_idx = CPUIDLE_DRIVER_STATE_START; |
9c4b2867 | 338 | } |
4f86d3a8 | 339 | |
71abbbf8 AL |
340 | /* |
341 | * Find the idle state with the lowest power while satisfying | |
342 | * our constraints. | |
343 | */ | |
5bb1729c | 344 | for (i = data->last_state_idx + 1; i < drv->state_count; i++) { |
46bcfad7 | 345 | struct cpuidle_state *s = &drv->states[i]; |
dc7fd275 | 346 | struct cpuidle_state_usage *su = &dev->states_usage[i]; |
4f86d3a8 | 347 | |
cbc9ef02 | 348 | if (s->disabled || su->disable) |
3a53396b | 349 | continue; |
14851912 | 350 | if (s->target_residency > data->predicted_us) |
71abbbf8 | 351 | continue; |
a2bd9202 | 352 | if (s->exit_latency > latency_req) |
71abbbf8 | 353 | continue; |
71abbbf8 | 354 | |
8aef33a7 | 355 | data->last_state_idx = i; |
4f86d3a8 LB |
356 | } |
357 | ||
69d25870 | 358 | return data->last_state_idx; |
4f86d3a8 LB |
359 | } |
360 | ||
361 | /** | |
672917dc | 362 | * menu_reflect - records that data structures need update |
4f86d3a8 | 363 | * @dev: the CPU |
e978aa7d | 364 | * @index: the index of actual entered state |
4f86d3a8 LB |
365 | * |
366 | * NOTE: it's important to be fast here because this operation will add to | |
367 | * the overall exit latency. | |
368 | */ | |
e978aa7d | 369 | static void menu_reflect(struct cpuidle_device *dev, int index) |
672917dc | 370 | { |
229b6863 | 371 | struct menu_device *data = this_cpu_ptr(&menu_devices); |
a802ea96 | 372 | |
e978aa7d | 373 | data->last_state_idx = index; |
a802ea96 | 374 | data->needs_update = 1; |
672917dc CZ |
375 | } |
376 | ||
377 | /** | |
378 | * menu_update - attempts to guess what happened after entry | |
46bcfad7 | 379 | * @drv: cpuidle driver containing state data |
672917dc CZ |
380 | * @dev: the CPU |
381 | */ | |
46bcfad7 | 382 | static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev) |
4f86d3a8 | 383 | { |
229b6863 | 384 | struct menu_device *data = this_cpu_ptr(&menu_devices); |
4f86d3a8 | 385 | int last_idx = data->last_state_idx; |
46bcfad7 | 386 | struct cpuidle_state *target = &drv->states[last_idx]; |
320eee77 | 387 | unsigned int measured_us; |
51f245b8 | 388 | unsigned int new_factor; |
4f86d3a8 LB |
389 | |
390 | /* | |
61c66d6e | 391 | * Try to figure out how much time passed between entry to low |
392 | * power state and occurrence of the wakeup event. | |
393 | * | |
394 | * If the entered idle state didn't support residency measurements, | |
4108b3d9 LB |
395 | * we use them anyway if they are short, and if long, |
396 | * truncate to the whole expected time. | |
61c66d6e | 397 | * |
398 | * Any measured amount of time will include the exit latency. | |
399 | * Since we are interested in when the wakeup begun, not when it | |
2fba5376 | 400 | * was completed, we must subtract the exit latency. However, if |
61c66d6e | 401 | * the measured amount of time is less than the exit latency, |
402 | * assume the state was never reached and the exit latency is 0. | |
4f86d3a8 | 403 | */ |
69d25870 | 404 | |
4108b3d9 LB |
405 | /* measured value */ |
406 | measured_us = cpuidle_get_last_residency(dev); | |
4f86d3a8 | 407 | |
4108b3d9 | 408 | /* Deduct exit latency */ |
efddfd90 | 409 | if (measured_us > 2 * target->exit_latency) |
4108b3d9 | 410 | measured_us -= target->exit_latency; |
efddfd90 RR |
411 | else |
412 | measured_us /= 2; | |
69d25870 | 413 | |
4108b3d9 LB |
414 | /* Make sure our coefficients do not exceed unity */ |
415 | if (measured_us > data->next_timer_us) | |
416 | measured_us = data->next_timer_us; | |
69d25870 | 417 | |
51f245b8 TT |
418 | /* Update our correction ratio */ |
419 | new_factor = data->correction_factor[data->bucket]; | |
420 | new_factor -= new_factor / DECAY; | |
69d25870 | 421 | |
5dc2f5a3 | 422 | if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING) |
423 | new_factor += RESOLUTION * measured_us / data->next_timer_us; | |
320eee77 | 424 | else |
69d25870 AV |
425 | /* |
426 | * we were idle so long that we count it as a perfect | |
427 | * prediction | |
428 | */ | |
429 | new_factor += RESOLUTION; | |
320eee77 | 430 | |
69d25870 AV |
431 | /* |
432 | * We don't want 0 as factor; we always want at least | |
51f245b8 TT |
433 | * a tiny bit of estimated time. Fortunately, due to rounding, |
434 | * new_factor will stay nonzero regardless of measured_us values | |
435 | * and the compiler can eliminate this test as long as DECAY > 1. | |
69d25870 | 436 | */ |
51f245b8 | 437 | if (DECAY == 1 && unlikely(new_factor == 0)) |
69d25870 | 438 | new_factor = 1; |
320eee77 | 439 | |
69d25870 | 440 | data->correction_factor[data->bucket] = new_factor; |
1f85f87d AV |
441 | |
442 | /* update the repeating-pattern data */ | |
61c66d6e | 443 | data->intervals[data->interval_ptr++] = measured_us; |
1f85f87d AV |
444 | if (data->interval_ptr >= INTERVALS) |
445 | data->interval_ptr = 0; | |
4f86d3a8 LB |
446 | } |
447 | ||
448 | /** | |
449 | * menu_enable_device - scans a CPU's states and does setup | |
46bcfad7 | 450 | * @drv: cpuidle driver |
4f86d3a8 LB |
451 | * @dev: the CPU |
452 | */ | |
46bcfad7 DD |
453 | static int menu_enable_device(struct cpuidle_driver *drv, |
454 | struct cpuidle_device *dev) | |
4f86d3a8 LB |
455 | { |
456 | struct menu_device *data = &per_cpu(menu_devices, dev->cpu); | |
bed4d597 | 457 | int i; |
4f86d3a8 LB |
458 | |
459 | memset(data, 0, sizeof(struct menu_device)); | |
460 | ||
bed4d597 CK |
461 | /* |
462 | * if the correction factor is 0 (eg first time init or cpu hotplug | |
463 | * etc), we actually want to start out with a unity factor. | |
464 | */ | |
465 | for(i = 0; i < BUCKETS; i++) | |
466 | data->correction_factor[i] = RESOLUTION * DECAY; | |
467 | ||
4f86d3a8 LB |
468 | return 0; |
469 | } | |
470 | ||
471 | static struct cpuidle_governor menu_governor = { | |
472 | .name = "menu", | |
473 | .rating = 20, | |
474 | .enable = menu_enable_device, | |
475 | .select = menu_select, | |
476 | .reflect = menu_reflect, | |
477 | .owner = THIS_MODULE, | |
478 | }; | |
479 | ||
480 | /** | |
481 | * init_menu - initializes the governor | |
482 | */ | |
483 | static int __init init_menu(void) | |
484 | { | |
485 | return cpuidle_register_governor(&menu_governor); | |
486 | } | |
487 | ||
137b944e | 488 | postcore_initcall(init_menu); |