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
70struct 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
81static 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
98static DEFINE_PER_CPU(struct menu_device, menu_devices);
99
100static void menu_update_intervals(struct menu_device *data, unsigned int interval_us)
101{
102 /* Update the repeating-pattern data. */
103 data->intervals[data->interval_ptr++] = interval_us;
104 if (data->interval_ptr >= INTERVALS)
105 data->interval_ptr = 0;
106}
107
108static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
109
110/*
111 * Try detecting repeating patterns by keeping track of the last 8
112 * intervals, and checking if the standard deviation of that set
113 * of points is below a threshold. If it is... then use the
114 * average of these 8 points as the estimated value.
115 */
116static unsigned int get_typical_interval(struct menu_device *data)
117{
118 s64 value, min_thresh = -1, max_thresh = UINT_MAX;
119 unsigned int max, min, divisor;
120 u64 avg, variance, avg_sq;
121 int i;
122
123again:
124 /* Compute the average and variance of past intervals. */
125 max = 0;
126 min = UINT_MAX;
127 avg = 0;
128 variance = 0;
129 divisor = 0;
130 for (i = 0; i < INTERVALS; i++) {
131 value = data->intervals[i];
132 /*
133 * Discard the samples outside the interval between the min and
134 * max thresholds.
135 */
136 if (value <= min_thresh || value >= max_thresh)
137 continue;
138
139 divisor++;
140
141 avg += value;
142 variance += value * value;
143
144 if (value > max)
145 max = value;
146
147 if (value < min)
148 min = value;
149 }
150
151 if (!max)
152 return UINT_MAX;
153
154 if (divisor == INTERVALS) {
155 avg >>= INTERVAL_SHIFT;
156 variance >>= INTERVAL_SHIFT;
157 } else {
158 do_div(avg, divisor);
159 do_div(variance, divisor);
160 }
161
162 avg_sq = avg * avg;
163 variance -= avg_sq;
164
165 /*
166 * The typical interval is obtained when standard deviation is
167 * small (stddev <= 20 us, variance <= 400 us^2) or standard
168 * deviation is small compared to the average interval (avg >
169 * 6*stddev, avg^2 > 36*variance). The average is smaller than
170 * UINT_MAX aka U32_MAX, so computing its square does not
171 * overflow a u64. We simply reject this candidate average if
172 * the standard deviation is greater than 715 s (which is
173 * rather unlikely).
174 *
175 * Use this result only if there is no timer to wake us up sooner.
176 */
177 if (likely(variance <= U64_MAX/36)) {
178 if ((avg_sq > variance * 36 && divisor * 4 >= INTERVALS * 3) ||
179 variance <= 400)
180 return avg;
181 }
182
183 /*
184 * If there are outliers, discard them by setting thresholds to exclude
185 * data points at a large enough distance from the average, then
186 * calculate the average and standard deviation again. Once we get
187 * down to the last 3/4 of our samples, stop excluding samples.
188 *
189 * This can deal with workloads that have long pauses interspersed
190 * with sporadic activity with a bunch of short pauses.
191 */
192 if (divisor * 4 <= INTERVALS * 3) {
193 /*
194 * If there are sufficiently many data points still under
195 * consideration after the outliers have been eliminated,
196 * returning without a prediction would be a mistake because it
197 * is likely that the next interval will not exceed the current
198 * maximum, so return the latter in that case.
199 */
200 if (divisor >= INTERVALS / 2)
201 return max;
202
203 return UINT_MAX;
204 }
205
206 /* Update the thresholds for the next round. */
207 if (avg - min > max - avg)
208 min_thresh = min;
209 else
210 max_thresh = max;
211
212 goto again;
213}
214
215/**
216 * menu_select - selects the next idle state to enter
217 * @drv: cpuidle driver containing state data
218 * @dev: the CPU
219 * @stop_tick: indication on whether or not to stop the tick
220 */
221static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
222 bool *stop_tick)
223{
224 struct menu_device *data = this_cpu_ptr(&menu_devices);
225 s64 latency_req = cpuidle_governor_latency_req(cpu: dev->cpu);
226 u64 predicted_ns;
227 ktime_t delta, delta_tick;
228 int i, idx;
229
230 if (data->needs_update) {
231 menu_update(drv, dev);
232 data->needs_update = 0;
233 } else if (!dev->last_residency_ns) {
234 /*
235 * This happens when the driver rejects the previously selected
236 * idle state and returns an error, so update the recent
237 * intervals table to prevent invalid information from being
238 * used going forward.
239 */
240 menu_update_intervals(data, UINT_MAX);
241 }
242
243 /* Find the shortest expected idle interval. */
244 predicted_ns = get_typical_interval(data) * NSEC_PER_USEC;
245 if (predicted_ns > RESIDENCY_THRESHOLD_NS) {
246 unsigned int timer_us;
247
248 /* Determine the time till the closest timer. */
249 delta = tick_nohz_get_sleep_length(delta_next: &delta_tick);
250 if (unlikely(delta < 0)) {
251 delta = 0;
252 delta_tick = 0;
253 }
254
255 data->next_timer_ns = delta;
256 data->bucket = which_bucket(duration_ns: data->next_timer_ns);
257
258 /* Round up the result for half microseconds. */
259 timer_us = div_u64(dividend: (RESOLUTION * DECAY * NSEC_PER_USEC) / 2 +
260 data->next_timer_ns *
261 data->correction_factor[data->bucket],
262 RESOLUTION * DECAY * NSEC_PER_USEC);
263 /* Use the lowest expected idle interval to pick the idle state. */
264 predicted_ns = min((u64)timer_us * NSEC_PER_USEC, predicted_ns);
265 } else {
266 /*
267 * Because the next timer event is not going to be determined
268 * in this case, assume that without the tick the closest timer
269 * will be in distant future and that the closest tick will occur
270 * after 1/2 of the tick period.
271 */
272 data->next_timer_ns = KTIME_MAX;
273 delta_tick = TICK_NSEC / 2;
274 data->bucket = BUCKETS - 1;
275 }
276
277 if (unlikely(drv->state_count <= 1 || latency_req == 0) ||
278 ((data->next_timer_ns < drv->states[1].target_residency_ns ||
279 latency_req < drv->states[1].exit_latency_ns) &&
280 !dev->states_usage[0].disable)) {
281 /*
282 * In this case state[0] will be used no matter what, so return
283 * it right away and keep the tick running if state[0] is a
284 * polling one.
285 */
286 *stop_tick = !(drv->states[0].flags & CPUIDLE_FLAG_POLLING);
287 return 0;
288 }
289
290 /*
291 * If the tick is already stopped, the cost of possible short idle
292 * duration misprediction is much higher, because the CPU may be stuck
293 * in a shallow idle state for a long time as a result of it. In that
294 * case, say we might mispredict and use the known time till the closest
295 * timer event for the idle state selection.
296 */
297 if (tick_nohz_tick_stopped() && predicted_ns < TICK_NSEC)
298 predicted_ns = data->next_timer_ns;
299
300 /*
301 * Find the idle state with the lowest power while satisfying
302 * our constraints.
303 */
304 idx = -1;
305 for (i = 0; i < drv->state_count; i++) {
306 struct cpuidle_state *s = &drv->states[i];
307
308 if (dev->states_usage[i].disable)
309 continue;
310
311 if (idx == -1)
312 idx = i; /* first enabled state */
313
314 if (s->exit_latency_ns > latency_req)
315 break;
316
317 if (s->target_residency_ns <= predicted_ns) {
318 idx = i;
319 continue;
320 }
321
322 /*
323 * Use a physical idle state, not busy polling, unless a timer
324 * is going to trigger soon enough.
325 */
326 if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
327 s->target_residency_ns <= data->next_timer_ns) {
328 predicted_ns = s->target_residency_ns;
329 idx = i;
330 break;
331 }
332
333 if (predicted_ns < TICK_NSEC)
334 break;
335
336 if (!tick_nohz_tick_stopped()) {
337 /*
338 * If the state selected so far is shallow, waking up
339 * early won't hurt, so retain the tick in that case and
340 * let the governor run again in the next iteration of
341 * the idle loop.
342 */
343 predicted_ns = drv->states[idx].target_residency_ns;
344 break;
345 }
346
347 /*
348 * If the state selected so far is shallow and this state's
349 * target residency matches the time till the closest timer
350 * event, select this one to avoid getting stuck in the shallow
351 * one for too long.
352 */
353 if (drv->states[idx].target_residency_ns < TICK_NSEC &&
354 s->target_residency_ns <= delta_tick)
355 idx = i;
356
357 return idx;
358 }
359
360 if (idx == -1)
361 idx = 0; /* No states enabled. Must use 0. */
362
363 /*
364 * Don't stop the tick if the selected state is a polling one or if the
365 * expected idle duration is shorter than the tick period length.
366 */
367 if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
368 predicted_ns < TICK_NSEC) && !tick_nohz_tick_stopped()) {
369 *stop_tick = false;
370
371 if (idx > 0 && drv->states[idx].target_residency_ns > delta_tick) {
372 /*
373 * The tick is not going to be stopped and the target
374 * residency of the state to be returned is not within
375 * the time until the next timer event including the
376 * tick, so try to correct that.
377 */
378 for (i = idx - 1; i >= 0; i--) {
379 if (dev->states_usage[i].disable)
380 continue;
381
382 idx = i;
383 if (drv->states[i].target_residency_ns <= delta_tick)
384 break;
385 }
386 }
387 }
388
389 return idx;
390}
391
392/**
393 * menu_reflect - records that data structures need update
394 * @dev: the CPU
395 * @index: the index of actual entered state
396 *
397 * NOTE: it's important to be fast here because this operation will add to
398 * the overall exit latency.
399 */
400static void menu_reflect(struct cpuidle_device *dev, int index)
401{
402 struct menu_device *data = this_cpu_ptr(&menu_devices);
403
404 dev->last_state_idx = index;
405 data->needs_update = 1;
406 data->tick_wakeup = tick_nohz_idle_got_tick();
407}
408
409/**
410 * menu_update - attempts to guess what happened after entry
411 * @drv: cpuidle driver containing state data
412 * @dev: the CPU
413 */
414static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
415{
416 struct menu_device *data = this_cpu_ptr(&menu_devices);
417 int last_idx = dev->last_state_idx;
418 struct cpuidle_state *target = &drv->states[last_idx];
419 u64 measured_ns;
420 unsigned int new_factor;
421
422 /*
423 * Try to figure out how much time passed between entry to low
424 * power state and occurrence of the wakeup event.
425 *
426 * If the entered idle state didn't support residency measurements,
427 * we use them anyway if they are short, and if long,
428 * truncate to the whole expected time.
429 *
430 * Any measured amount of time will include the exit latency.
431 * Since we are interested in when the wakeup begun, not when it
432 * was completed, we must subtract the exit latency. However, if
433 * the measured amount of time is less than the exit latency,
434 * assume the state was never reached and the exit latency is 0.
435 */
436
437 if (data->tick_wakeup && data->next_timer_ns > TICK_NSEC) {
438 /*
439 * The nohz code said that there wouldn't be any events within
440 * the tick boundary (if the tick was stopped), but the idle
441 * duration predictor had a differing opinion. Since the CPU
442 * was woken up by a tick (that wasn't stopped after all), the
443 * predictor was not quite right, so assume that the CPU could
444 * have been idle long (but not forever) to help the idle
445 * duration predictor do a better job next time.
446 */
447 measured_ns = 9 * MAX_INTERESTING / 10;
448 } else if ((drv->states[last_idx].flags & CPUIDLE_FLAG_POLLING) &&
449 dev->poll_time_limit) {
450 /*
451 * The CPU exited the "polling" state due to a time limit, so
452 * the idle duration prediction leading to the selection of that
453 * state was inaccurate. If a better prediction had been made,
454 * the CPU might have been woken up from idle by the next timer.
455 * Assume that to be the case.
456 */
457 measured_ns = data->next_timer_ns;
458 } else {
459 /* measured value */
460 measured_ns = dev->last_residency_ns;
461
462 /* Deduct exit latency */
463 if (measured_ns > 2 * target->exit_latency_ns)
464 measured_ns -= target->exit_latency_ns;
465 else
466 measured_ns /= 2;
467 }
468
469 /* Make sure our coefficients do not exceed unity */
470 if (measured_ns > data->next_timer_ns)
471 measured_ns = data->next_timer_ns;
472
473 /* Update our correction ratio */
474 new_factor = data->correction_factor[data->bucket];
475 new_factor -= new_factor / DECAY;
476
477 if (data->next_timer_ns > 0 && measured_ns < MAX_INTERESTING)
478 new_factor += div64_u64(RESOLUTION * measured_ns,
479 divisor: data->next_timer_ns);
480 else
481 /*
482 * we were idle so long that we count it as a perfect
483 * prediction
484 */
485 new_factor += RESOLUTION;
486
487 /*
488 * We don't want 0 as factor; we always want at least
489 * a tiny bit of estimated time. Fortunately, due to rounding,
490 * new_factor will stay nonzero regardless of measured_us values
491 * and the compiler can eliminate this test as long as DECAY > 1.
492 */
493 if (DECAY == 1 && unlikely(new_factor == 0))
494 new_factor = 1;
495
496 data->correction_factor[data->bucket] = new_factor;
497
498 menu_update_intervals(data, interval_us: ktime_to_us(kt: measured_ns));
499}
500
501/**
502 * menu_enable_device - scans a CPU's states and does setup
503 * @drv: cpuidle driver
504 * @dev: the CPU
505 */
506static int menu_enable_device(struct cpuidle_driver *drv,
507 struct cpuidle_device *dev)
508{
509 struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
510 int i;
511
512 memset(s: data, c: 0, n: sizeof(struct menu_device));
513
514 /*
515 * if the correction factor is 0 (eg first time init or cpu hotplug
516 * etc), we actually want to start out with a unity factor.
517 */
518 for(i = 0; i < BUCKETS; i++)
519 data->correction_factor[i] = RESOLUTION * DECAY;
520
521 return 0;
522}
523
524static struct cpuidle_governor menu_governor = {
525 .name = "menu",
526 .rating = 20,
527 .enable = menu_enable_device,
528 .select = menu_select,
529 .reflect = menu_reflect,
530};
531
532/**
533 * init_menu - initializes the governor
534 */
535static int __init init_menu(void)
536{
537 return cpuidle_register_governor(gov: &menu_governor);
538}
539
540postcore_initcall(init_menu);
541