xref: /openbmc/linux/drivers/cpuidle/governors/menu.c (revision 908dbf02)
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.h>
18 #include <linux/sched/loadavg.h>
19 #include <linux/sched/stat.h>
20 #include <linux/math64.h>
21 
22 #define BUCKETS 12
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 3 decision factors for picking a C
33  * state:
34  * 1) Energy break even point
35  * 2) Performance impact
36  * 3) Latency tolerance (from pmqos infrastructure)
37  * These these three factors are treated independently.
38  *
39  * Energy break even point
40  * -----------------------
41  * C state entry and exit have an energy cost, and a certain amount of time in
42  * the  C state is required to actually break even on this cost. CPUIDLE
43  * provides us this duration in the "target_residency" field. So all that we
44  * need is a good prediction of how long we'll be idle. Like the traditional
45  * menu governor, we start with the actual known "next timer event" time.
46  *
47  * Since there are other source of wakeups (interrupts for example) than
48  * the next timer event, this estimation is rather optimistic. To get a
49  * more realistic estimate, a correction factor is applied to the estimate,
50  * that is based on historic behavior. For example, if in the past the actual
51  * duration always was 50% of the next timer tick, the correction factor will
52  * be 0.5.
53  *
54  * menu uses a running average for this correction factor, however it uses a
55  * set of factors, not just a single factor. This stems from the realization
56  * that the ratio is dependent on the order of magnitude of the expected
57  * duration; if we expect 500 milliseconds of idle time the likelihood of
58  * getting an interrupt very early is much higher than if we expect 50 micro
59  * seconds of idle time. A second independent factor that has big impact on
60  * the actual factor is if there is (disk) IO outstanding or not.
61  * (as a special twist, we consider every sleep longer than 50 milliseconds
62  * as perfect; there are no power gains for sleeping longer than this)
63  *
64  * For these two reasons we keep an array of 12 independent factors, that gets
65  * indexed based on the magnitude of the expected duration as well as the
66  * "is IO outstanding" property.
67  *
68  * Repeatable-interval-detector
69  * ----------------------------
70  * There are some cases where "next timer" is a completely unusable predictor:
71  * Those cases where the interval is fixed, for example due to hardware
72  * interrupt mitigation, but also due to fixed transfer rate devices such as
73  * mice.
74  * For this, we use a different predictor: We track the duration of the last 8
75  * intervals and if the stand deviation of these 8 intervals is below a
76  * threshold value, we use the average of these intervals as prediction.
77  *
78  * Limiting Performance Impact
79  * ---------------------------
80  * C states, especially those with large exit latencies, can have a real
81  * noticeable impact on workloads, which is not acceptable for most sysadmins,
82  * and in addition, less performance has a power price of its own.
83  *
84  * As a general rule of thumb, menu assumes that the following heuristic
85  * holds:
86  *     The busier the system, the less impact of C states is acceptable
87  *
88  * This rule-of-thumb is implemented using a performance-multiplier:
89  * If the exit latency times the performance multiplier is longer than
90  * the predicted duration, the C state is not considered a candidate
91  * for selection due to a too high performance impact. So the higher
92  * this multiplier is, the longer we need to be idle to pick a deep C
93  * state, and thus the less likely a busy CPU will hit such a deep
94  * C state.
95  *
96  * Two factors are used in determing this multiplier:
97  * a value of 10 is added for each point of "per cpu load average" we have.
98  * a value of 5 points is added for each process that is waiting for
99  * IO on this CPU.
100  * (these values are experimentally determined)
101  *
102  * The load average factor gives a longer term (few seconds) input to the
103  * decision, while the iowait value gives a cpu local instantanious input.
104  * The iowait factor may look low, but realize that this is also already
105  * represented in the system load average.
106  *
107  */
108 
109 struct menu_device {
110 	int             needs_update;
111 	int             tick_wakeup;
112 
113 	u64		next_timer_ns;
114 	unsigned int	bucket;
115 	unsigned int	correction_factor[BUCKETS];
116 	unsigned int	intervals[INTERVALS];
117 	int		interval_ptr;
118 };
119 
120 static inline int which_bucket(u64 duration_ns, unsigned int nr_iowaiters)
121 {
122 	int bucket = 0;
123 
124 	/*
125 	 * We keep two groups of stats; one with no
126 	 * IO pending, one without.
127 	 * This allows us to calculate
128 	 * E(duration)|iowait
129 	 */
130 	if (nr_iowaiters)
131 		bucket = BUCKETS/2;
132 
133 	if (duration_ns < 10ULL * NSEC_PER_USEC)
134 		return bucket;
135 	if (duration_ns < 100ULL * NSEC_PER_USEC)
136 		return bucket + 1;
137 	if (duration_ns < 1000ULL * NSEC_PER_USEC)
138 		return bucket + 2;
139 	if (duration_ns < 10000ULL * NSEC_PER_USEC)
140 		return bucket + 3;
141 	if (duration_ns < 100000ULL * NSEC_PER_USEC)
142 		return bucket + 4;
143 	return bucket + 5;
144 }
145 
146 /*
147  * Return a multiplier for the exit latency that is intended
148  * to take performance requirements into account.
149  * The more performance critical we estimate the system
150  * to be, the higher this multiplier, and thus the higher
151  * the barrier to go to an expensive C state.
152  */
153 static inline int performance_multiplier(unsigned int nr_iowaiters)
154 {
155 	/* for IO wait tasks (per cpu!) we add 10x each */
156 	return 1 + 10 * nr_iowaiters;
157 }
158 
159 static DEFINE_PER_CPU(struct menu_device, menu_devices);
160 
161 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
162 
163 /*
164  * Try detecting repeating patterns by keeping track of the last 8
165  * intervals, and checking if the standard deviation of that set
166  * of points is below a threshold. If it is... then use the
167  * average of these 8 points as the estimated value.
168  */
169 static unsigned int get_typical_interval(struct menu_device *data,
170 					 unsigned int predicted_us)
171 {
172 	int i, divisor;
173 	unsigned int min, max, thresh, avg;
174 	uint64_t sum, variance;
175 
176 	thresh = INT_MAX; /* Discard outliers above this value */
177 
178 again:
179 
180 	/* First calculate the average of past intervals */
181 	min = UINT_MAX;
182 	max = 0;
183 	sum = 0;
184 	divisor = 0;
185 	for (i = 0; i < INTERVALS; i++) {
186 		unsigned int value = data->intervals[i];
187 		if (value <= thresh) {
188 			sum += value;
189 			divisor++;
190 			if (value > max)
191 				max = value;
192 
193 			if (value < min)
194 				min = value;
195 		}
196 	}
197 
198 	/*
199 	 * If the result of the computation is going to be discarded anyway,
200 	 * avoid the computation altogether.
201 	 */
202 	if (min >= predicted_us)
203 		return UINT_MAX;
204 
205 	if (divisor == INTERVALS)
206 		avg = sum >> INTERVAL_SHIFT;
207 	else
208 		avg = div_u64(sum, divisor);
209 
210 	/* Then try to determine variance */
211 	variance = 0;
212 	for (i = 0; i < INTERVALS; i++) {
213 		unsigned int value = data->intervals[i];
214 		if (value <= thresh) {
215 			int64_t diff = (int64_t)value - avg;
216 			variance += diff * diff;
217 		}
218 	}
219 	if (divisor == INTERVALS)
220 		variance >>= INTERVAL_SHIFT;
221 	else
222 		do_div(variance, divisor);
223 
224 	/*
225 	 * The typical interval is obtained when standard deviation is
226 	 * small (stddev <= 20 us, variance <= 400 us^2) or standard
227 	 * deviation is small compared to the average interval (avg >
228 	 * 6*stddev, avg^2 > 36*variance). The average is smaller than
229 	 * UINT_MAX aka U32_MAX, so computing its square does not
230 	 * overflow a u64. We simply reject this candidate average if
231 	 * the standard deviation is greater than 715 s (which is
232 	 * rather unlikely).
233 	 *
234 	 * Use this result only if there is no timer to wake us up sooner.
235 	 */
236 	if (likely(variance <= U64_MAX/36)) {
237 		if ((((u64)avg*avg > variance*36) && (divisor * 4 >= INTERVALS * 3))
238 							|| variance <= 400) {
239 			return avg;
240 		}
241 	}
242 
243 	/*
244 	 * If we have outliers to the upside in our distribution, discard
245 	 * those by setting the threshold to exclude these outliers, then
246 	 * calculate the average and standard deviation again. Once we get
247 	 * down to the bottom 3/4 of our samples, stop excluding samples.
248 	 *
249 	 * This can deal with workloads that have long pauses interspersed
250 	 * with sporadic activity with a bunch of short pauses.
251 	 */
252 	if ((divisor * 4) <= INTERVALS * 3)
253 		return UINT_MAX;
254 
255 	thresh = max - 1;
256 	goto again;
257 }
258 
259 /**
260  * menu_select - selects the next idle state to enter
261  * @drv: cpuidle driver containing state data
262  * @dev: the CPU
263  * @stop_tick: indication on whether or not to stop the tick
264  */
265 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
266 		       bool *stop_tick)
267 {
268 	struct menu_device *data = this_cpu_ptr(&menu_devices);
269 	s64 latency_req = cpuidle_governor_latency_req(dev->cpu);
270 	unsigned int predicted_us;
271 	u64 predicted_ns;
272 	u64 interactivity_req;
273 	unsigned int nr_iowaiters;
274 	ktime_t delta, delta_tick;
275 	int i, idx;
276 
277 	if (data->needs_update) {
278 		menu_update(drv, dev);
279 		data->needs_update = 0;
280 	}
281 
282 	/* determine the expected residency time, round up */
283 	delta = tick_nohz_get_sleep_length(&delta_tick);
284 	if (unlikely(delta < 0)) {
285 		delta = 0;
286 		delta_tick = 0;
287 	}
288 	data->next_timer_ns = delta;
289 
290 	nr_iowaiters = nr_iowait_cpu(dev->cpu);
291 	data->bucket = which_bucket(data->next_timer_ns, nr_iowaiters);
292 
293 	if (unlikely(drv->state_count <= 1 || latency_req == 0) ||
294 	    ((data->next_timer_ns < drv->states[1].target_residency_ns ||
295 	      latency_req < drv->states[1].exit_latency_ns) &&
296 	     !dev->states_usage[0].disable)) {
297 		/*
298 		 * In this case state[0] will be used no matter what, so return
299 		 * it right away and keep the tick running if state[0] is a
300 		 * polling one.
301 		 */
302 		*stop_tick = !(drv->states[0].flags & CPUIDLE_FLAG_POLLING);
303 		return 0;
304 	}
305 
306 	/* Round up the result for half microseconds. */
307 	predicted_us = div_u64(data->next_timer_ns *
308 			       data->correction_factor[data->bucket] +
309 			       (RESOLUTION * DECAY * NSEC_PER_USEC) / 2,
310 			       RESOLUTION * DECAY * NSEC_PER_USEC);
311 	/* Use the lowest expected idle interval to pick the idle state. */
312 	predicted_ns = (u64)min(predicted_us,
313 				get_typical_interval(data, predicted_us)) *
314 				NSEC_PER_USEC;
315 
316 	if (tick_nohz_tick_stopped()) {
317 		/*
318 		 * If the tick is already stopped, the cost of possible short
319 		 * idle duration misprediction is much higher, because the CPU
320 		 * may be stuck in a shallow idle state for a long time as a
321 		 * result of it.  In that case say we might mispredict and use
322 		 * the known time till the closest timer event for the idle
323 		 * state selection.
324 		 */
325 		if (predicted_ns < TICK_NSEC)
326 			predicted_ns = data->next_timer_ns;
327 	} else {
328 		/*
329 		 * Use the performance multiplier and the user-configurable
330 		 * latency_req to determine the maximum exit latency.
331 		 */
332 		interactivity_req = div64_u64(predicted_ns,
333 					      performance_multiplier(nr_iowaiters));
334 		if (latency_req > interactivity_req)
335 			latency_req = interactivity_req;
336 	}
337 
338 	/*
339 	 * Find the idle state with the lowest power while satisfying
340 	 * our constraints.
341 	 */
342 	idx = -1;
343 	for (i = 0; i < drv->state_count; i++) {
344 		struct cpuidle_state *s = &drv->states[i];
345 
346 		if (dev->states_usage[i].disable)
347 			continue;
348 
349 		if (idx == -1)
350 			idx = i; /* first enabled state */
351 
352 		if (s->target_residency_ns > predicted_ns) {
353 			/*
354 			 * Use a physical idle state, not busy polling, unless
355 			 * a timer is going to trigger soon enough.
356 			 */
357 			if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
358 			    s->exit_latency_ns <= latency_req &&
359 			    s->target_residency_ns <= data->next_timer_ns) {
360 				predicted_ns = s->target_residency_ns;
361 				idx = i;
362 				break;
363 			}
364 			if (predicted_ns < TICK_NSEC)
365 				break;
366 
367 			if (!tick_nohz_tick_stopped()) {
368 				/*
369 				 * If the state selected so far is shallow,
370 				 * waking up early won't hurt, so retain the
371 				 * tick in that case and let the governor run
372 				 * again in the next iteration of the loop.
373 				 */
374 				predicted_ns = drv->states[idx].target_residency_ns;
375 				break;
376 			}
377 
378 			/*
379 			 * If the state selected so far is shallow and this
380 			 * state's target residency matches the time till the
381 			 * closest timer event, select this one to avoid getting
382 			 * stuck in the shallow one for too long.
383 			 */
384 			if (drv->states[idx].target_residency_ns < TICK_NSEC &&
385 			    s->target_residency_ns <= delta_tick)
386 				idx = i;
387 
388 			return idx;
389 		}
390 		if (s->exit_latency_ns > latency_req)
391 			break;
392 
393 		idx = i;
394 	}
395 
396 	if (idx == -1)
397 		idx = 0; /* No states enabled. Must use 0. */
398 
399 	/*
400 	 * Don't stop the tick if the selected state is a polling one or if the
401 	 * expected idle duration is shorter than the tick period length.
402 	 */
403 	if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
404 	     predicted_ns < TICK_NSEC) && !tick_nohz_tick_stopped()) {
405 		*stop_tick = false;
406 
407 		if (idx > 0 && drv->states[idx].target_residency_ns > delta_tick) {
408 			/*
409 			 * The tick is not going to be stopped and the target
410 			 * residency of the state to be returned is not within
411 			 * the time until the next timer event including the
412 			 * tick, so try to correct that.
413 			 */
414 			for (i = idx - 1; i >= 0; i--) {
415 				if (dev->states_usage[i].disable)
416 					continue;
417 
418 				idx = i;
419 				if (drv->states[i].target_residency_ns <= delta_tick)
420 					break;
421 			}
422 		}
423 	}
424 
425 	return idx;
426 }
427 
428 /**
429  * menu_reflect - records that data structures need update
430  * @dev: the CPU
431  * @index: the index of actual entered state
432  *
433  * NOTE: it's important to be fast here because this operation will add to
434  *       the overall exit latency.
435  */
436 static void menu_reflect(struct cpuidle_device *dev, int index)
437 {
438 	struct menu_device *data = this_cpu_ptr(&menu_devices);
439 
440 	dev->last_state_idx = index;
441 	data->needs_update = 1;
442 	data->tick_wakeup = tick_nohz_idle_got_tick();
443 }
444 
445 /**
446  * menu_update - attempts to guess what happened after entry
447  * @drv: cpuidle driver containing state data
448  * @dev: the CPU
449  */
450 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
451 {
452 	struct menu_device *data = this_cpu_ptr(&menu_devices);
453 	int last_idx = dev->last_state_idx;
454 	struct cpuidle_state *target = &drv->states[last_idx];
455 	u64 measured_ns;
456 	unsigned int new_factor;
457 
458 	/*
459 	 * Try to figure out how much time passed between entry to low
460 	 * power state and occurrence of the wakeup event.
461 	 *
462 	 * If the entered idle state didn't support residency measurements,
463 	 * we use them anyway if they are short, and if long,
464 	 * truncate to the whole expected time.
465 	 *
466 	 * Any measured amount of time will include the exit latency.
467 	 * Since we are interested in when the wakeup begun, not when it
468 	 * was completed, we must subtract the exit latency. However, if
469 	 * the measured amount of time is less than the exit latency,
470 	 * assume the state was never reached and the exit latency is 0.
471 	 */
472 
473 	if (data->tick_wakeup && data->next_timer_ns > TICK_NSEC) {
474 		/*
475 		 * The nohz code said that there wouldn't be any events within
476 		 * the tick boundary (if the tick was stopped), but the idle
477 		 * duration predictor had a differing opinion.  Since the CPU
478 		 * was woken up by a tick (that wasn't stopped after all), the
479 		 * predictor was not quite right, so assume that the CPU could
480 		 * have been idle long (but not forever) to help the idle
481 		 * duration predictor do a better job next time.
482 		 */
483 		measured_ns = 9 * MAX_INTERESTING / 10;
484 	} else if ((drv->states[last_idx].flags & CPUIDLE_FLAG_POLLING) &&
485 		   dev->poll_time_limit) {
486 		/*
487 		 * The CPU exited the "polling" state due to a time limit, so
488 		 * the idle duration prediction leading to the selection of that
489 		 * state was inaccurate.  If a better prediction had been made,
490 		 * the CPU might have been woken up from idle by the next timer.
491 		 * Assume that to be the case.
492 		 */
493 		measured_ns = data->next_timer_ns;
494 	} else {
495 		/* measured value */
496 		measured_ns = dev->last_residency_ns;
497 
498 		/* Deduct exit latency */
499 		if (measured_ns > 2 * target->exit_latency_ns)
500 			measured_ns -= target->exit_latency_ns;
501 		else
502 			measured_ns /= 2;
503 	}
504 
505 	/* Make sure our coefficients do not exceed unity */
506 	if (measured_ns > data->next_timer_ns)
507 		measured_ns = data->next_timer_ns;
508 
509 	/* Update our correction ratio */
510 	new_factor = data->correction_factor[data->bucket];
511 	new_factor -= new_factor / DECAY;
512 
513 	if (data->next_timer_ns > 0 && measured_ns < MAX_INTERESTING)
514 		new_factor += div64_u64(RESOLUTION * measured_ns,
515 					data->next_timer_ns);
516 	else
517 		/*
518 		 * we were idle so long that we count it as a perfect
519 		 * prediction
520 		 */
521 		new_factor += RESOLUTION;
522 
523 	/*
524 	 * We don't want 0 as factor; we always want at least
525 	 * a tiny bit of estimated time. Fortunately, due to rounding,
526 	 * new_factor will stay nonzero regardless of measured_us values
527 	 * and the compiler can eliminate this test as long as DECAY > 1.
528 	 */
529 	if (DECAY == 1 && unlikely(new_factor == 0))
530 		new_factor = 1;
531 
532 	data->correction_factor[data->bucket] = new_factor;
533 
534 	/* update the repeating-pattern data */
535 	data->intervals[data->interval_ptr++] = ktime_to_us(measured_ns);
536 	if (data->interval_ptr >= INTERVALS)
537 		data->interval_ptr = 0;
538 }
539 
540 /**
541  * menu_enable_device - scans a CPU's states and does setup
542  * @drv: cpuidle driver
543  * @dev: the CPU
544  */
545 static int menu_enable_device(struct cpuidle_driver *drv,
546 				struct cpuidle_device *dev)
547 {
548 	struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
549 	int i;
550 
551 	memset(data, 0, sizeof(struct menu_device));
552 
553 	/*
554 	 * if the correction factor is 0 (eg first time init or cpu hotplug
555 	 * etc), we actually want to start out with a unity factor.
556 	 */
557 	for(i = 0; i < BUCKETS; i++)
558 		data->correction_factor[i] = RESOLUTION * DECAY;
559 
560 	return 0;
561 }
562 
563 static struct cpuidle_governor menu_governor = {
564 	.name =		"menu",
565 	.rating =	20,
566 	.enable =	menu_enable_device,
567 	.select =	menu_select,
568 	.reflect =	menu_reflect,
569 };
570 
571 /**
572  * init_menu - initializes the governor
573  */
574 static int __init init_menu(void)
575 {
576 	return cpuidle_register_governor(&menu_governor);
577 }
578 
579 postcore_initcall(init_menu);
580