1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * menu.c - the menu idle governor
4 *
5 * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com>
6 * Copyright (C) 2009 Intel Corporation
7 * Author:
8 * Arjan van de Ven <arjan@linux.intel.com>
9 */
10
11 #include <linux/kernel.h>
12 #include <linux/cpuidle.h>
13 #include <linux/time.h>
14 #include <linux/ktime.h>
15 #include <linux/hrtimer.h>
16 #include <linux/tick.h>
17 #include <linux/sched/stat.h>
18 #include <linux/math64.h>
19
20 #include "gov.h"
21
22 #define BUCKETS 6
23 #define INTERVAL_SHIFT 3
24 #define INTERVALS (1UL << INTERVAL_SHIFT)
25 #define RESOLUTION 1024
26 #define DECAY 8
27 #define MAX_INTERESTING (50000 * NSEC_PER_USEC)
28
29 /*
30 * Concepts and ideas behind the menu governor
31 *
32 * For the menu governor, there are 2 decision factors for picking a C
33 * state:
34 * 1) Energy break even point
35 * 2) Latency tolerance (from pmqos infrastructure)
36 * These two factors are treated independently.
37 *
38 * Energy break even point
39 * -----------------------
40 * C state entry and exit have an energy cost, and a certain amount of time in
41 * the C state is required to actually break even on this cost. CPUIDLE
42 * provides us this duration in the "target_residency" field. So all that we
43 * need is a good prediction of how long we'll be idle. Like the traditional
44 * menu governor, we take the actual known "next timer event" time.
45 *
46 * Since there are other source of wakeups (interrupts for example) than
47 * the next timer event, this estimation is rather optimistic. To get a
48 * more realistic estimate, a correction factor is applied to the estimate,
49 * that is based on historic behavior. For example, if in the past the actual
50 * duration always was 50% of the next timer tick, the correction factor will
51 * be 0.5.
52 *
53 * menu uses a running average for this correction factor, but it uses a set of
54 * factors, not just a single factor. This stems from the realization that the
55 * ratio is dependent on the order of magnitude of the expected duration; if we
56 * expect 500 milliseconds of idle time the likelihood of getting an interrupt
57 * very early is much higher than if we expect 50 micro seconds of idle time.
58 * For this reason, menu keeps an array of 6 independent factors, that gets
59 * indexed based on the magnitude of the expected duration.
60 *
61 * Repeatable-interval-detector
62 * ----------------------------
63 * There are some cases where "next timer" is a completely unusable predictor:
64 * Those cases where the interval is fixed, for example due to hardware
65 * interrupt mitigation, but also due to fixed transfer rate devices like mice.
66 * For this, we use a different predictor: We track the duration of the last 8
67 * intervals and use them to estimate the duration of the next one.
68 */
69
70 struct menu_device {
71 int needs_update;
72 int tick_wakeup;
73
74 u64 next_timer_ns;
75 unsigned int bucket;
76 unsigned int correction_factor[BUCKETS];
77 unsigned int intervals[INTERVALS];
78 int interval_ptr;
79 };
80
which_bucket(u64 duration_ns)81 static inline int which_bucket(u64 duration_ns)
82 {
83 int bucket = 0;
84
85 if (duration_ns < 10ULL * NSEC_PER_USEC)
86 return bucket;
87 if (duration_ns < 100ULL * NSEC_PER_USEC)
88 return bucket + 1;
89 if (duration_ns < 1000ULL * NSEC_PER_USEC)
90 return bucket + 2;
91 if (duration_ns < 10000ULL * NSEC_PER_USEC)
92 return bucket + 3;
93 if (duration_ns < 100000ULL * NSEC_PER_USEC)
94 return bucket + 4;
95 return bucket + 5;
96 }
97
98 static DEFINE_PER_CPU(struct menu_device, menu_devices);
99
100 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
101
102 /*
103 * Try detecting repeating patterns by keeping track of the last 8
104 * intervals, and checking if the standard deviation of that set
105 * of points is below a threshold. If it is... then use the
106 * average of these 8 points as the estimated value.
107 */
get_typical_interval(struct menu_device * data)108 static unsigned int get_typical_interval(struct menu_device *data)
109 {
110 s64 value, min_thresh = -1, max_thresh = UINT_MAX;
111 unsigned int max, min, divisor;
112 u64 avg, variance, avg_sq;
113 int i;
114
115 again:
116 /* Compute the average and variance of past intervals. */
117 max = 0;
118 min = UINT_MAX;
119 avg = 0;
120 variance = 0;
121 divisor = 0;
122 for (i = 0; i < INTERVALS; i++) {
123 value = data->intervals[i];
124 /*
125 * Discard the samples outside the interval between the min and
126 * max thresholds.
127 */
128 if (value <= min_thresh || value >= max_thresh)
129 continue;
130
131 divisor++;
132
133 avg += value;
134 variance += value * value;
135
136 if (value > max)
137 max = value;
138
139 if (value < min)
140 min = value;
141 }
142
143 if (!max)
144 return UINT_MAX;
145
146 if (divisor == INTERVALS) {
147 avg >>= INTERVAL_SHIFT;
148 variance >>= INTERVAL_SHIFT;
149 } else {
150 do_div(avg, divisor);
151 do_div(variance, divisor);
152 }
153
154 avg_sq = avg * avg;
155 variance -= avg_sq;
156
157 /*
158 * The typical interval is obtained when standard deviation is
159 * small (stddev <= 20 us, variance <= 400 us^2) or standard
160 * deviation is small compared to the average interval (avg >
161 * 6*stddev, avg^2 > 36*variance). The average is smaller than
162 * UINT_MAX aka U32_MAX, so computing its square does not
163 * overflow a u64. We simply reject this candidate average if
164 * the standard deviation is greater than 715 s (which is
165 * rather unlikely).
166 *
167 * Use this result only if there is no timer to wake us up sooner.
168 */
169 if (likely(variance <= U64_MAX/36)) {
170 if ((avg_sq > variance * 36 && divisor * 4 >= INTERVALS * 3) ||
171 variance <= 400)
172 return avg;
173 }
174
175 /*
176 * If there are outliers, discard them by setting thresholds to exclude
177 * data points at a large enough distance from the average, then
178 * calculate the average and standard deviation again. Once we get
179 * down to the last 3/4 of our samples, stop excluding samples.
180 *
181 * This can deal with workloads that have long pauses interspersed
182 * with sporadic activity with a bunch of short pauses.
183 */
184 if (divisor * 4 <= INTERVALS * 3) {
185 /*
186 * If there are sufficiently many data points still under
187 * consideration after the outliers have been eliminated,
188 * returning without a prediction would be a mistake because it
189 * is likely that the next interval will not exceed the current
190 * maximum, so return the latter in that case.
191 */
192 if (divisor >= INTERVALS / 2)
193 return max;
194
195 return UINT_MAX;
196 }
197
198 /* Update the thresholds for the next round. */
199 if (avg - min > max - avg)
200 min_thresh = min;
201 else
202 max_thresh = max;
203
204 goto again;
205 }
206
207 /**
208 * menu_select - selects the next idle state to enter
209 * @drv: cpuidle driver containing state data
210 * @dev: the CPU
211 * @stop_tick: indication on whether or not to stop the tick
212 */
menu_select(struct cpuidle_driver * drv,struct cpuidle_device * dev,bool * stop_tick)213 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
214 bool *stop_tick)
215 {
216 struct menu_device *data = this_cpu_ptr(&menu_devices);
217 s64 latency_req = cpuidle_governor_latency_req(dev->cpu);
218 u64 predicted_ns;
219 ktime_t delta, delta_tick;
220 int i, idx;
221
222 if (data->needs_update) {
223 menu_update(drv, dev);
224 data->needs_update = 0;
225 }
226
227 /* Find the shortest expected idle interval. */
228 predicted_ns = get_typical_interval(data) * NSEC_PER_USEC;
229 if (predicted_ns > RESIDENCY_THRESHOLD_NS) {
230 unsigned int timer_us;
231
232 /* Determine the time till the closest timer. */
233 delta = tick_nohz_get_sleep_length(&delta_tick);
234 if (unlikely(delta < 0)) {
235 delta = 0;
236 delta_tick = 0;
237 }
238
239 data->next_timer_ns = delta;
240 data->bucket = which_bucket(data->next_timer_ns);
241
242 /* Round up the result for half microseconds. */
243 timer_us = div_u64((RESOLUTION * DECAY * NSEC_PER_USEC) / 2 +
244 data->next_timer_ns *
245 data->correction_factor[data->bucket],
246 RESOLUTION * DECAY * NSEC_PER_USEC);
247 /* Use the lowest expected idle interval to pick the idle state. */
248 predicted_ns = min((u64)timer_us * NSEC_PER_USEC, predicted_ns);
249 } else {
250 /*
251 * Because the next timer event is not going to be determined
252 * in this case, assume that without the tick the closest timer
253 * will be in distant future and that the closest tick will occur
254 * after 1/2 of the tick period.
255 */
256 data->next_timer_ns = KTIME_MAX;
257 delta_tick = TICK_NSEC / 2;
258 data->bucket = BUCKETS - 1;
259 }
260
261 if (unlikely(drv->state_count <= 1 || latency_req == 0) ||
262 ((data->next_timer_ns < drv->states[1].target_residency_ns ||
263 latency_req < drv->states[1].exit_latency_ns) &&
264 !dev->states_usage[0].disable)) {
265 /*
266 * In this case state[0] will be used no matter what, so return
267 * it right away and keep the tick running if state[0] is a
268 * polling one.
269 */
270 *stop_tick = !(drv->states[0].flags & CPUIDLE_FLAG_POLLING);
271 return 0;
272 }
273
274 if (tick_nohz_tick_stopped()) {
275 /*
276 * If the tick is already stopped, the cost of possible short
277 * idle duration misprediction is much higher, because the CPU
278 * may be stuck in a shallow idle state for a long time as a
279 * result of it. In that case say we might mispredict and use
280 * the known time till the closest timer event for the idle
281 * state selection.
282 */
283 if (predicted_ns < TICK_NSEC)
284 predicted_ns = data->next_timer_ns;
285 } else if (latency_req > predicted_ns) {
286 latency_req = predicted_ns;
287 }
288
289 /*
290 * Find the idle state with the lowest power while satisfying
291 * our constraints.
292 */
293 idx = -1;
294 for (i = 0; i < drv->state_count; i++) {
295 struct cpuidle_state *s = &drv->states[i];
296
297 if (dev->states_usage[i].disable)
298 continue;
299
300 if (idx == -1)
301 idx = i; /* first enabled state */
302
303 if (s->target_residency_ns > predicted_ns) {
304 /*
305 * Use a physical idle state, not busy polling, unless
306 * a timer is going to trigger soon enough.
307 */
308 if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
309 s->exit_latency_ns <= latency_req &&
310 s->target_residency_ns <= data->next_timer_ns) {
311 predicted_ns = s->target_residency_ns;
312 idx = i;
313 break;
314 }
315 if (predicted_ns < TICK_NSEC)
316 break;
317
318 if (!tick_nohz_tick_stopped()) {
319 /*
320 * If the state selected so far is shallow,
321 * waking up early won't hurt, so retain the
322 * tick in that case and let the governor run
323 * again in the next iteration of the loop.
324 */
325 predicted_ns = drv->states[idx].target_residency_ns;
326 break;
327 }
328
329 /*
330 * If the state selected so far is shallow and this
331 * state's target residency matches the time till the
332 * closest timer event, select this one to avoid getting
333 * stuck in the shallow one for too long.
334 */
335 if (drv->states[idx].target_residency_ns < TICK_NSEC &&
336 s->target_residency_ns <= delta_tick)
337 idx = i;
338
339 return idx;
340 }
341 if (s->exit_latency_ns > latency_req)
342 break;
343
344 idx = i;
345 }
346
347 if (idx == -1)
348 idx = 0; /* No states enabled. Must use 0. */
349
350 /*
351 * Don't stop the tick if the selected state is a polling one or if the
352 * expected idle duration is shorter than the tick period length.
353 */
354 if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
355 predicted_ns < TICK_NSEC) && !tick_nohz_tick_stopped()) {
356 *stop_tick = false;
357
358 if (idx > 0 && drv->states[idx].target_residency_ns > delta_tick) {
359 /*
360 * The tick is not going to be stopped and the target
361 * residency of the state to be returned is not within
362 * the time until the next timer event including the
363 * tick, so try to correct that.
364 */
365 for (i = idx - 1; i >= 0; i--) {
366 if (dev->states_usage[i].disable)
367 continue;
368
369 idx = i;
370 if (drv->states[i].target_residency_ns <= delta_tick)
371 break;
372 }
373 }
374 }
375
376 return idx;
377 }
378
379 /**
380 * menu_reflect - records that data structures need update
381 * @dev: the CPU
382 * @index: the index of actual entered state
383 *
384 * NOTE: it's important to be fast here because this operation will add to
385 * the overall exit latency.
386 */
menu_reflect(struct cpuidle_device * dev,int index)387 static void menu_reflect(struct cpuidle_device *dev, int index)
388 {
389 struct menu_device *data = this_cpu_ptr(&menu_devices);
390
391 dev->last_state_idx = index;
392 data->needs_update = 1;
393 data->tick_wakeup = tick_nohz_idle_got_tick();
394 }
395
396 /**
397 * menu_update - attempts to guess what happened after entry
398 * @drv: cpuidle driver containing state data
399 * @dev: the CPU
400 */
menu_update(struct cpuidle_driver * drv,struct cpuidle_device * dev)401 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
402 {
403 struct menu_device *data = this_cpu_ptr(&menu_devices);
404 int last_idx = dev->last_state_idx;
405 struct cpuidle_state *target = &drv->states[last_idx];
406 u64 measured_ns;
407 unsigned int new_factor;
408
409 /*
410 * Try to figure out how much time passed between entry to low
411 * power state and occurrence of the wakeup event.
412 *
413 * If the entered idle state didn't support residency measurements,
414 * we use them anyway if they are short, and if long,
415 * truncate to the whole expected time.
416 *
417 * Any measured amount of time will include the exit latency.
418 * Since we are interested in when the wakeup begun, not when it
419 * was completed, we must subtract the exit latency. However, if
420 * the measured amount of time is less than the exit latency,
421 * assume the state was never reached and the exit latency is 0.
422 */
423
424 if (data->tick_wakeup && data->next_timer_ns > TICK_NSEC) {
425 /*
426 * The nohz code said that there wouldn't be any events within
427 * the tick boundary (if the tick was stopped), but the idle
428 * duration predictor had a differing opinion. Since the CPU
429 * was woken up by a tick (that wasn't stopped after all), the
430 * predictor was not quite right, so assume that the CPU could
431 * have been idle long (but not forever) to help the idle
432 * duration predictor do a better job next time.
433 */
434 measured_ns = 9 * MAX_INTERESTING / 10;
435 } else if ((drv->states[last_idx].flags & CPUIDLE_FLAG_POLLING) &&
436 dev->poll_time_limit) {
437 /*
438 * The CPU exited the "polling" state due to a time limit, so
439 * the idle duration prediction leading to the selection of that
440 * state was inaccurate. If a better prediction had been made,
441 * the CPU might have been woken up from idle by the next timer.
442 * Assume that to be the case.
443 */
444 measured_ns = data->next_timer_ns;
445 } else {
446 /* measured value */
447 measured_ns = dev->last_residency_ns;
448
449 /* Deduct exit latency */
450 if (measured_ns > 2 * target->exit_latency_ns)
451 measured_ns -= target->exit_latency_ns;
452 else
453 measured_ns /= 2;
454 }
455
456 /* Make sure our coefficients do not exceed unity */
457 if (measured_ns > data->next_timer_ns)
458 measured_ns = data->next_timer_ns;
459
460 /* Update our correction ratio */
461 new_factor = data->correction_factor[data->bucket];
462 new_factor -= new_factor / DECAY;
463
464 if (data->next_timer_ns > 0 && measured_ns < MAX_INTERESTING)
465 new_factor += div64_u64(RESOLUTION * measured_ns,
466 data->next_timer_ns);
467 else
468 /*
469 * we were idle so long that we count it as a perfect
470 * prediction
471 */
472 new_factor += RESOLUTION;
473
474 /*
475 * We don't want 0 as factor; we always want at least
476 * a tiny bit of estimated time. Fortunately, due to rounding,
477 * new_factor will stay nonzero regardless of measured_us values
478 * and the compiler can eliminate this test as long as DECAY > 1.
479 */
480 if (DECAY == 1 && unlikely(new_factor == 0))
481 new_factor = 1;
482
483 data->correction_factor[data->bucket] = new_factor;
484
485 /* update the repeating-pattern data */
486 data->intervals[data->interval_ptr++] = ktime_to_us(measured_ns);
487 if (data->interval_ptr >= INTERVALS)
488 data->interval_ptr = 0;
489 }
490
491 /**
492 * menu_enable_device - scans a CPU's states and does setup
493 * @drv: cpuidle driver
494 * @dev: the CPU
495 */
menu_enable_device(struct cpuidle_driver * drv,struct cpuidle_device * dev)496 static int menu_enable_device(struct cpuidle_driver *drv,
497 struct cpuidle_device *dev)
498 {
499 struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
500 int i;
501
502 memset(data, 0, sizeof(struct menu_device));
503
504 /*
505 * if the correction factor is 0 (eg first time init or cpu hotplug
506 * etc), we actually want to start out with a unity factor.
507 */
508 for(i = 0; i < BUCKETS; i++)
509 data->correction_factor[i] = RESOLUTION * DECAY;
510
511 return 0;
512 }
513
514 static struct cpuidle_governor menu_governor = {
515 .name = "menu",
516 .rating = 20,
517 .enable = menu_enable_device,
518 .select = menu_select,
519 .reflect = menu_reflect,
520 };
521
522 /**
523 * init_menu - initializes the governor
524 */
init_menu(void)525 static int __init init_menu(void)
526 {
527 return cpuidle_register_governor(&menu_governor);
528 }
529
530 postcore_initcall(init_menu);
531