1 // SPDX-License-Identifier: GPL-2.0-only
3 * menu.c - the menu idle governor
5 * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com>
6 * Copyright (C) 2009 Intel Corporation
8 * Arjan van de Ven <arjan@linux.intel.com>
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>
23 * Please note when changing the tuning values:
24 * If (MAX_INTERESTING-1) * RESOLUTION > UINT_MAX, the result of
25 * a scaling operation multiplication may overflow on 32 bit platforms.
26 * In that case, #define RESOLUTION as ULL to get 64 bit result:
27 * #define RESOLUTION 1024ULL
29 * The default values do not overflow.
32 #define INTERVAL_SHIFT 3
33 #define INTERVALS (1UL << INTERVAL_SHIFT)
34 #define RESOLUTION 1024
36 #define MAX_INTERESTING 50000
40 * Concepts and ideas behind the menu governor
42 * For the menu governor, there are 3 decision factors for picking a C
44 * 1) Energy break even point
45 * 2) Performance impact
46 * 3) Latency tolerance (from pmqos infrastructure)
47 * These these three factors are treated independently.
49 * Energy break even point
50 * -----------------------
51 * C state entry and exit have an energy cost, and a certain amount of time in
52 * the C state is required to actually break even on this cost. CPUIDLE
53 * provides us this duration in the "target_residency" field. So all that we
54 * need is a good prediction of how long we'll be idle. Like the traditional
55 * menu governor, we start with the actual known "next timer event" time.
57 * Since there are other source of wakeups (interrupts for example) than
58 * the next timer event, this estimation is rather optimistic. To get a
59 * more realistic estimate, a correction factor is applied to the estimate,
60 * that is based on historic behavior. For example, if in the past the actual
61 * duration always was 50% of the next timer tick, the correction factor will
64 * menu uses a running average for this correction factor, however it uses a
65 * set of factors, not just a single factor. This stems from the realization
66 * that the ratio is dependent on the order of magnitude of the expected
67 * duration; if we expect 500 milliseconds of idle time the likelihood of
68 * getting an interrupt very early is much higher than if we expect 50 micro
69 * seconds of idle time. A second independent factor that has big impact on
70 * the actual factor is if there is (disk) IO outstanding or not.
71 * (as a special twist, we consider every sleep longer than 50 milliseconds
72 * as perfect; there are no power gains for sleeping longer than this)
74 * For these two reasons we keep an array of 12 independent factors, that gets
75 * indexed based on the magnitude of the expected duration as well as the
76 * "is IO outstanding" property.
78 * Repeatable-interval-detector
79 * ----------------------------
80 * There are some cases where "next timer" is a completely unusable predictor:
81 * Those cases where the interval is fixed, for example due to hardware
82 * interrupt mitigation, but also due to fixed transfer rate devices such as
84 * For this, we use a different predictor: We track the duration of the last 8
85 * intervals and if the stand deviation of these 8 intervals is below a
86 * threshold value, we use the average of these intervals as prediction.
88 * Limiting Performance Impact
89 * ---------------------------
90 * C states, especially those with large exit latencies, can have a real
91 * noticeable impact on workloads, which is not acceptable for most sysadmins,
92 * and in addition, less performance has a power price of its own.
94 * As a general rule of thumb, menu assumes that the following heuristic
96 * The busier the system, the less impact of C states is acceptable
98 * This rule-of-thumb is implemented using a performance-multiplier:
99 * If the exit latency times the performance multiplier is longer than
100 * the predicted duration, the C state is not considered a candidate
101 * for selection due to a too high performance impact. So the higher
102 * this multiplier is, the longer we need to be idle to pick a deep C
103 * state, and thus the less likely a busy CPU will hit such a deep
106 * Two factors are used in determing this multiplier:
107 * a value of 10 is added for each point of "per cpu load average" we have.
108 * a value of 5 points is added for each process that is waiting for
110 * (these values are experimentally determined)
112 * The load average factor gives a longer term (few seconds) input to the
113 * decision, while the iowait value gives a cpu local instantanious input.
114 * The iowait factor may look low, but realize that this is also already
115 * represented in the system load average.
123 unsigned int next_timer_us;
125 unsigned int correction_factor[BUCKETS];
126 unsigned int intervals[INTERVALS];
130 static inline int which_bucket(unsigned int duration, unsigned long nr_iowaiters)
135 * We keep two groups of stats; one with no
136 * IO pending, one without.
137 * This allows us to calculate
149 if (duration < 10000)
151 if (duration < 100000)
157 * Return a multiplier for the exit latency that is intended
158 * to take performance requirements into account.
159 * The more performance critical we estimate the system
160 * to be, the higher this multiplier, and thus the higher
161 * the barrier to go to an expensive C state.
163 static inline int performance_multiplier(unsigned long nr_iowaiters)
165 /* for IO wait tasks (per cpu!) we add 10x each */
166 return 1 + 10 * nr_iowaiters;
169 static DEFINE_PER_CPU(struct menu_device, menu_devices);
171 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
174 * Try detecting repeating patterns by keeping track of the last 8
175 * intervals, and checking if the standard deviation of that set
176 * of points is below a threshold. If it is... then use the
177 * average of these 8 points as the estimated value.
179 static unsigned int get_typical_interval(struct menu_device *data,
180 unsigned int predicted_us)
183 unsigned int min, max, thresh, avg;
184 uint64_t sum, variance;
186 thresh = INT_MAX; /* Discard outliers above this value */
190 /* First calculate the average of past intervals */
195 for (i = 0; i < INTERVALS; i++) {
196 unsigned int value = data->intervals[i];
197 if (value <= thresh) {
209 * If the result of the computation is going to be discarded anyway,
210 * avoid the computation altogether.
212 if (min >= predicted_us)
215 if (divisor == INTERVALS)
216 avg = sum >> INTERVAL_SHIFT;
218 avg = div_u64(sum, divisor);
220 /* Then try to determine variance */
222 for (i = 0; i < INTERVALS; i++) {
223 unsigned int value = data->intervals[i];
224 if (value <= thresh) {
225 int64_t diff = (int64_t)value - avg;
226 variance += diff * diff;
229 if (divisor == INTERVALS)
230 variance >>= INTERVAL_SHIFT;
232 do_div(variance, divisor);
235 * The typical interval is obtained when standard deviation is
236 * small (stddev <= 20 us, variance <= 400 us^2) or standard
237 * deviation is small compared to the average interval (avg >
238 * 6*stddev, avg^2 > 36*variance). The average is smaller than
239 * UINT_MAX aka U32_MAX, so computing its square does not
240 * overflow a u64. We simply reject this candidate average if
241 * the standard deviation is greater than 715 s (which is
244 * Use this result only if there is no timer to wake us up sooner.
246 if (likely(variance <= U64_MAX/36)) {
247 if ((((u64)avg*avg > variance*36) && (divisor * 4 >= INTERVALS * 3))
248 || variance <= 400) {
254 * If we have outliers to the upside in our distribution, discard
255 * those by setting the threshold to exclude these outliers, then
256 * calculate the average and standard deviation again. Once we get
257 * down to the bottom 3/4 of our samples, stop excluding samples.
259 * This can deal with workloads that have long pauses interspersed
260 * with sporadic activity with a bunch of short pauses.
262 if ((divisor * 4) <= INTERVALS * 3)
270 * menu_select - selects the next idle state to enter
271 * @drv: cpuidle driver containing state data
273 * @stop_tick: indication on whether or not to stop the tick
275 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
278 struct menu_device *data = this_cpu_ptr(&menu_devices);
279 int latency_req = cpuidle_governor_latency_req(dev->cpu);
282 unsigned int interactivity_req;
283 unsigned int predicted_us;
284 unsigned long nr_iowaiters;
287 if (data->needs_update) {
288 menu_update(drv, dev);
289 data->needs_update = 0;
292 /* determine the expected residency time, round up */
293 data->next_timer_us = ktime_to_us(tick_nohz_get_sleep_length(&delta_next));
295 nr_iowaiters = nr_iowait_cpu(dev->cpu);
296 data->bucket = which_bucket(data->next_timer_us, nr_iowaiters);
298 if (unlikely(drv->state_count <= 1 || latency_req == 0) ||
299 ((data->next_timer_us < drv->states[1].target_residency ||
300 latency_req < drv->states[1].exit_latency) &&
301 !drv->states[0].disabled && !dev->states_usage[0].disable)) {
303 * In this case state[0] will be used no matter what, so return
304 * it right away and keep the tick running if state[0] is a
307 *stop_tick = !(drv->states[0].flags & CPUIDLE_FLAG_POLLING);
312 * Force the result of multiplication to be 64 bits even if both
313 * operands are 32 bits.
314 * Make sure to round up for half microseconds.
316 predicted_us = DIV_ROUND_CLOSEST_ULL((uint64_t)data->next_timer_us *
317 data->correction_factor[data->bucket],
320 * Use the lowest expected idle interval to pick the idle state.
322 predicted_us = min(predicted_us, get_typical_interval(data, predicted_us));
324 if (tick_nohz_tick_stopped()) {
326 * If the tick is already stopped, the cost of possible short
327 * idle duration misprediction is much higher, because the CPU
328 * may be stuck in a shallow idle state for a long time as a
329 * result of it. In that case say we might mispredict and use
330 * the known time till the closest timer event for the idle
333 if (predicted_us < TICK_USEC)
334 predicted_us = ktime_to_us(delta_next);
337 * Use the performance multiplier and the user-configurable
338 * latency_req to determine the maximum exit latency.
340 interactivity_req = predicted_us / performance_multiplier(nr_iowaiters);
341 if (latency_req > interactivity_req)
342 latency_req = interactivity_req;
346 * Find the idle state with the lowest power while satisfying
350 for (i = 0; i < drv->state_count; i++) {
351 struct cpuidle_state *s = &drv->states[i];
352 struct cpuidle_state_usage *su = &dev->states_usage[i];
354 if (s->disabled || su->disable)
358 idx = i; /* first enabled state */
360 if (s->target_residency > predicted_us) {
362 * Use a physical idle state, not busy polling, unless
363 * a timer is going to trigger soon enough.
365 if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
366 s->exit_latency <= latency_req &&
367 s->target_residency <= data->next_timer_us) {
368 predicted_us = s->target_residency;
372 if (predicted_us < TICK_USEC)
375 if (!tick_nohz_tick_stopped()) {
377 * If the state selected so far is shallow,
378 * waking up early won't hurt, so retain the
379 * tick in that case and let the governor run
380 * again in the next iteration of the loop.
382 predicted_us = drv->states[idx].target_residency;
387 * If the state selected so far is shallow and this
388 * state's target residency matches the time till the
389 * closest timer event, select this one to avoid getting
390 * stuck in the shallow one for too long.
392 if (drv->states[idx].target_residency < TICK_USEC &&
393 s->target_residency <= ktime_to_us(delta_next))
398 if (s->exit_latency > latency_req)
405 idx = 0; /* No states enabled. Must use 0. */
408 * Don't stop the tick if the selected state is a polling one or if the
409 * expected idle duration is shorter than the tick period length.
411 if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
412 predicted_us < TICK_USEC) && !tick_nohz_tick_stopped()) {
413 unsigned int delta_next_us = ktime_to_us(delta_next);
417 if (idx > 0 && drv->states[idx].target_residency > delta_next_us) {
419 * The tick is not going to be stopped and the target
420 * residency of the state to be returned is not within
421 * the time until the next timer event including the
422 * tick, so try to correct that.
424 for (i = idx - 1; i >= 0; i--) {
425 if (drv->states[i].disabled ||
426 dev->states_usage[i].disable)
430 if (drv->states[i].target_residency <= delta_next_us)
440 * menu_reflect - records that data structures need update
442 * @index: the index of actual entered state
444 * NOTE: it's important to be fast here because this operation will add to
445 * the overall exit latency.
447 static void menu_reflect(struct cpuidle_device *dev, int index)
449 struct menu_device *data = this_cpu_ptr(&menu_devices);
451 dev->last_state_idx = index;
452 data->needs_update = 1;
453 data->tick_wakeup = tick_nohz_idle_got_tick();
457 * menu_update - attempts to guess what happened after entry
458 * @drv: cpuidle driver containing state data
461 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
463 struct menu_device *data = this_cpu_ptr(&menu_devices);
464 int last_idx = dev->last_state_idx;
465 struct cpuidle_state *target = &drv->states[last_idx];
466 unsigned int measured_us;
467 unsigned int new_factor;
470 * Try to figure out how much time passed between entry to low
471 * power state and occurrence of the wakeup event.
473 * If the entered idle state didn't support residency measurements,
474 * we use them anyway if they are short, and if long,
475 * truncate to the whole expected time.
477 * Any measured amount of time will include the exit latency.
478 * Since we are interested in when the wakeup begun, not when it
479 * was completed, we must subtract the exit latency. However, if
480 * the measured amount of time is less than the exit latency,
481 * assume the state was never reached and the exit latency is 0.
484 if (data->tick_wakeup && data->next_timer_us > TICK_USEC) {
486 * The nohz code said that there wouldn't be any events within
487 * the tick boundary (if the tick was stopped), but the idle
488 * duration predictor had a differing opinion. Since the CPU
489 * was woken up by a tick (that wasn't stopped after all), the
490 * predictor was not quite right, so assume that the CPU could
491 * have been idle long (but not forever) to help the idle
492 * duration predictor do a better job next time.
494 measured_us = 9 * MAX_INTERESTING / 10;
495 } else if ((drv->states[last_idx].flags & CPUIDLE_FLAG_POLLING) &&
496 dev->poll_time_limit) {
498 * The CPU exited the "polling" state due to a time limit, so
499 * the idle duration prediction leading to the selection of that
500 * state was inaccurate. If a better prediction had been made,
501 * the CPU might have been woken up from idle by the next timer.
502 * Assume that to be the case.
504 measured_us = data->next_timer_us;
507 measured_us = dev->last_residency;
509 /* Deduct exit latency */
510 if (measured_us > 2 * target->exit_latency)
511 measured_us -= target->exit_latency;
516 /* Make sure our coefficients do not exceed unity */
517 if (measured_us > data->next_timer_us)
518 measured_us = data->next_timer_us;
520 /* Update our correction ratio */
521 new_factor = data->correction_factor[data->bucket];
522 new_factor -= new_factor / DECAY;
524 if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING)
525 new_factor += RESOLUTION * measured_us / data->next_timer_us;
528 * we were idle so long that we count it as a perfect
531 new_factor += RESOLUTION;
534 * We don't want 0 as factor; we always want at least
535 * a tiny bit of estimated time. Fortunately, due to rounding,
536 * new_factor will stay nonzero regardless of measured_us values
537 * and the compiler can eliminate this test as long as DECAY > 1.
539 if (DECAY == 1 && unlikely(new_factor == 0))
542 data->correction_factor[data->bucket] = new_factor;
544 /* update the repeating-pattern data */
545 data->intervals[data->interval_ptr++] = measured_us;
546 if (data->interval_ptr >= INTERVALS)
547 data->interval_ptr = 0;
551 * menu_enable_device - scans a CPU's states and does setup
552 * @drv: cpuidle driver
555 static int menu_enable_device(struct cpuidle_driver *drv,
556 struct cpuidle_device *dev)
558 struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
561 memset(data, 0, sizeof(struct menu_device));
564 * if the correction factor is 0 (eg first time init or cpu hotplug
565 * etc), we actually want to start out with a unity factor.
567 for(i = 0; i < BUCKETS; i++)
568 data->correction_factor[i] = RESOLUTION * DECAY;
573 static struct cpuidle_governor menu_governor = {
576 .enable = menu_enable_device,
577 .select = menu_select,
578 .reflect = menu_reflect,
582 * init_menu - initializes the governor
584 static int __init init_menu(void)
586 return cpuidle_register_governor(&menu_governor);
589 postcore_initcall(init_menu);