xref: /linux/drivers/cpuidle/governors/teo.c (revision d59fec29b131f30b27343d54bdf1071ee98eda8e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Timer events oriented CPU idle governor
4  *
5  * TEO governor:
6  * Copyright (C) 2018 - 2021 Intel Corporation
7  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
8  *
9  * Util-awareness mechanism:
10  * Copyright (C) 2022 Arm Ltd.
11  * Author: Kajetan Puchalski <kajetan.puchalski@arm.com>
12  */
13 
14 /**
15  * DOC: teo-description
16  *
17  * The idea of this governor is based on the observation that on many systems
18  * timer events are two or more orders of magnitude more frequent than any
19  * other interrupts, so they are likely to be the most significant cause of CPU
20  * wakeups from idle states.  Moreover, information about what happened in the
21  * (relatively recent) past can be used to estimate whether or not the deepest
22  * idle state with target residency within the (known) time till the closest
23  * timer event, referred to as the sleep length, is likely to be suitable for
24  * the upcoming CPU idle period and, if not, then which of the shallower idle
25  * states to choose instead of it.
26  *
27  * Of course, non-timer wakeup sources are more important in some use cases
28  * which can be covered by taking a few most recent idle time intervals of the
29  * CPU into account.  However, even in that context it is not necessary to
30  * consider idle duration values greater than the sleep length, because the
31  * closest timer will ultimately wake up the CPU anyway unless it is woken up
32  * earlier.
33  *
34  * Thus this governor estimates whether or not the prospective idle duration of
35  * a CPU is likely to be significantly shorter than the sleep length and selects
36  * an idle state for it accordingly.
37  *
38  * The computations carried out by this governor are based on using bins whose
39  * boundaries are aligned with the target residency parameter values of the CPU
40  * idle states provided by the %CPUIdle driver in the ascending order.  That is,
41  * the first bin spans from 0 up to, but not including, the target residency of
42  * the second idle state (idle state 1), the second bin spans from the target
43  * residency of idle state 1 up to, but not including, the target residency of
44  * idle state 2, the third bin spans from the target residency of idle state 2
45  * up to, but not including, the target residency of idle state 3 and so on.
46  * The last bin spans from the target residency of the deepest idle state
47  * supplied by the driver to infinity.
48  *
49  * Two metrics called "hits" and "intercepts" are associated with each bin.
50  * They are updated every time before selecting an idle state for the given CPU
51  * in accordance with what happened last time.
52  *
53  * The "hits" metric reflects the relative frequency of situations in which the
54  * sleep length and the idle duration measured after CPU wakeup fall into the
55  * same bin (that is, the CPU appears to wake up "on time" relative to the sleep
56  * length).  In turn, the "intercepts" metric reflects the relative frequency of
57  * situations in which the measured idle duration is so much shorter than the
58  * sleep length that the bin it falls into corresponds to an idle state
59  * shallower than the one whose bin is fallen into by the sleep length (these
60  * situations are referred to as "intercepts" below).
61  *
62  * In addition to the metrics described above, the governor counts recent
63  * intercepts (that is, intercepts that have occurred during the last
64  * %NR_RECENT invocations of it for the given CPU) for each bin.
65  *
66  * In order to select an idle state for a CPU, the governor takes the following
67  * steps (modulo the possible latency constraint that must be taken into account
68  * too):
69  *
70  * 1. Find the deepest CPU idle state whose target residency does not exceed
71  *    the current sleep length (the candidate idle state) and compute 3 sums as
72  *    follows:
73  *
74  *    - The sum of the "hits" and "intercepts" metrics for the candidate state
75  *      and all of the deeper idle states (it represents the cases in which the
76  *      CPU was idle long enough to avoid being intercepted if the sleep length
77  *      had been equal to the current one).
78  *
79  *    - The sum of the "intercepts" metrics for all of the idle states shallower
80  *      than the candidate one (it represents the cases in which the CPU was not
81  *      idle long enough to avoid being intercepted if the sleep length had been
82  *      equal to the current one).
83  *
84  *    - The sum of the numbers of recent intercepts for all of the idle states
85  *      shallower than the candidate one.
86  *
87  * 2. If the second sum is greater than the first one or the third sum is
88  *    greater than %NR_RECENT / 2, the CPU is likely to wake up early, so look
89  *    for an alternative idle state to select.
90  *
91  *    - Traverse the idle states shallower than the candidate one in the
92  *      descending order.
93  *
94  *    - For each of them compute the sum of the "intercepts" metrics and the sum
95  *      of the numbers of recent intercepts over all of the idle states between
96  *      it and the candidate one (including the former and excluding the
97  *      latter).
98  *
99  *    - If each of these sums that needs to be taken into account (because the
100  *      check related to it has indicated that the CPU is likely to wake up
101  *      early) is greater than a half of the corresponding sum computed in step
102  *      1 (which means that the target residency of the state in question had
103  *      not exceeded the idle duration in over a half of the relevant cases),
104  *      select the given idle state instead of the candidate one.
105  *
106  * 3. By default, select the candidate state.
107  *
108  * Util-awareness mechanism:
109  *
110  * The idea behind the util-awareness extension is that there are two distinct
111  * scenarios for the CPU which should result in two different approaches to idle
112  * state selection - utilized and not utilized.
113  *
114  * In this case, 'utilized' means that the average runqueue util of the CPU is
115  * above a certain threshold.
116  *
117  * When the CPU is utilized while going into idle, more likely than not it will
118  * be woken up to do more work soon and so a shallower idle state should be
119  * selected to minimise latency and maximise performance. When the CPU is not
120  * being utilized, the usual metrics-based approach to selecting the deepest
121  * available idle state should be preferred to take advantage of the power
122  * saving.
123  *
124  * In order to achieve this, the governor uses a utilization threshold.
125  * The threshold is computed per-CPU as a percentage of the CPU's capacity
126  * by bit shifting the capacity value. Based on testing, the shift of 6 (~1.56%)
127  * seems to be getting the best results.
128  *
129  * Before selecting the next idle state, the governor compares the current CPU
130  * util to the precomputed util threshold. If it's below, it defaults to the
131  * TEO metrics mechanism. If it's above, the closest shallower idle state will
132  * be selected instead, as long as is not a polling state.
133  */
134 
135 #include <linux/cpuidle.h>
136 #include <linux/jiffies.h>
137 #include <linux/kernel.h>
138 #include <linux/sched.h>
139 #include <linux/sched/clock.h>
140 #include <linux/sched/topology.h>
141 #include <linux/tick.h>
142 
143 /*
144  * The number of bits to shift the CPU's capacity by in order to determine
145  * the utilized threshold.
146  *
147  * 6 was chosen based on testing as the number that achieved the best balance
148  * of power and performance on average.
149  *
150  * The resulting threshold is high enough to not be triggered by background
151  * noise and low enough to react quickly when activity starts to ramp up.
152  */
153 #define UTIL_THRESHOLD_SHIFT 6
154 
155 
156 /*
157  * The PULSE value is added to metrics when they grow and the DECAY_SHIFT value
158  * is used for decreasing metrics on a regular basis.
159  */
160 #define PULSE		1024
161 #define DECAY_SHIFT	3
162 
163 /*
164  * Number of the most recent idle duration values to take into consideration for
165  * the detection of recent early wakeup patterns.
166  */
167 #define NR_RECENT	9
168 
169 /**
170  * struct teo_bin - Metrics used by the TEO cpuidle governor.
171  * @intercepts: The "intercepts" metric.
172  * @hits: The "hits" metric.
173  * @recent: The number of recent "intercepts".
174  */
175 struct teo_bin {
176 	unsigned int intercepts;
177 	unsigned int hits;
178 	unsigned int recent;
179 };
180 
181 /**
182  * struct teo_cpu - CPU data used by the TEO cpuidle governor.
183  * @time_span_ns: Time between idle state selection and post-wakeup update.
184  * @sleep_length_ns: Time till the closest timer event (at the selection time).
185  * @state_bins: Idle state data bins for this CPU.
186  * @total: Grand total of the "intercepts" and "hits" metrics for all bins.
187  * @next_recent_idx: Index of the next @recent_idx entry to update.
188  * @recent_idx: Indices of bins corresponding to recent "intercepts".
189  * @util_threshold: Threshold above which the CPU is considered utilized
190  * @utilized: Whether the last sleep on the CPU happened while utilized
191  */
192 struct teo_cpu {
193 	s64 time_span_ns;
194 	s64 sleep_length_ns;
195 	struct teo_bin state_bins[CPUIDLE_STATE_MAX];
196 	unsigned int total;
197 	int next_recent_idx;
198 	int recent_idx[NR_RECENT];
199 	unsigned long util_threshold;
200 	bool utilized;
201 };
202 
203 static DEFINE_PER_CPU(struct teo_cpu, teo_cpus);
204 
205 /**
206  * teo_cpu_is_utilized - Check if the CPU's util is above the threshold
207  * @cpu: Target CPU
208  * @cpu_data: Governor CPU data for the target CPU
209  */
210 #ifdef CONFIG_SMP
211 static bool teo_cpu_is_utilized(int cpu, struct teo_cpu *cpu_data)
212 {
213 	return sched_cpu_util(cpu) > cpu_data->util_threshold;
214 }
215 #else
216 static bool teo_cpu_is_utilized(int cpu, struct teo_cpu *cpu_data)
217 {
218 	return false;
219 }
220 #endif
221 
222 /**
223  * teo_update - Update CPU metrics after wakeup.
224  * @drv: cpuidle driver containing state data.
225  * @dev: Target CPU.
226  */
227 static void teo_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
228 {
229 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
230 	int i, idx_timer = 0, idx_duration = 0;
231 	u64 measured_ns;
232 
233 	if (cpu_data->time_span_ns >= cpu_data->sleep_length_ns) {
234 		/*
235 		 * One of the safety nets has triggered or the wakeup was close
236 		 * enough to the closest timer event expected at the idle state
237 		 * selection time to be discarded.
238 		 */
239 		measured_ns = U64_MAX;
240 	} else {
241 		u64 lat_ns = drv->states[dev->last_state_idx].exit_latency_ns;
242 
243 		/*
244 		 * The computations below are to determine whether or not the
245 		 * (saved) time till the next timer event and the measured idle
246 		 * duration fall into the same "bin", so use last_residency_ns
247 		 * for that instead of time_span_ns which includes the cpuidle
248 		 * overhead.
249 		 */
250 		measured_ns = dev->last_residency_ns;
251 		/*
252 		 * The delay between the wakeup and the first instruction
253 		 * executed by the CPU is not likely to be worst-case every
254 		 * time, so take 1/2 of the exit latency as a very rough
255 		 * approximation of the average of it.
256 		 */
257 		if (measured_ns >= lat_ns)
258 			measured_ns -= lat_ns / 2;
259 		else
260 			measured_ns /= 2;
261 	}
262 
263 	cpu_data->total = 0;
264 
265 	/*
266 	 * Decay the "hits" and "intercepts" metrics for all of the bins and
267 	 * find the bins that the sleep length and the measured idle duration
268 	 * fall into.
269 	 */
270 	for (i = 0; i < drv->state_count; i++) {
271 		s64 target_residency_ns = drv->states[i].target_residency_ns;
272 		struct teo_bin *bin = &cpu_data->state_bins[i];
273 
274 		bin->hits -= bin->hits >> DECAY_SHIFT;
275 		bin->intercepts -= bin->intercepts >> DECAY_SHIFT;
276 
277 		cpu_data->total += bin->hits + bin->intercepts;
278 
279 		if (target_residency_ns <= cpu_data->sleep_length_ns) {
280 			idx_timer = i;
281 			if (target_residency_ns <= measured_ns)
282 				idx_duration = i;
283 		}
284 	}
285 
286 	i = cpu_data->next_recent_idx++;
287 	if (cpu_data->next_recent_idx >= NR_RECENT)
288 		cpu_data->next_recent_idx = 0;
289 
290 	if (cpu_data->recent_idx[i] >= 0)
291 		cpu_data->state_bins[cpu_data->recent_idx[i]].recent--;
292 
293 	/*
294 	 * If the measured idle duration falls into the same bin as the sleep
295 	 * length, this is a "hit", so update the "hits" metric for that bin.
296 	 * Otherwise, update the "intercepts" metric for the bin fallen into by
297 	 * the measured idle duration.
298 	 */
299 	if (idx_timer == idx_duration) {
300 		cpu_data->state_bins[idx_timer].hits += PULSE;
301 		cpu_data->recent_idx[i] = -1;
302 	} else {
303 		cpu_data->state_bins[idx_duration].intercepts += PULSE;
304 		cpu_data->state_bins[idx_duration].recent++;
305 		cpu_data->recent_idx[i] = idx_duration;
306 	}
307 
308 	cpu_data->total += PULSE;
309 }
310 
311 static bool teo_time_ok(u64 interval_ns)
312 {
313 	return !tick_nohz_tick_stopped() || interval_ns >= TICK_NSEC;
314 }
315 
316 static s64 teo_middle_of_bin(int idx, struct cpuidle_driver *drv)
317 {
318 	return (drv->states[idx].target_residency_ns +
319 		drv->states[idx+1].target_residency_ns) / 2;
320 }
321 
322 /**
323  * teo_find_shallower_state - Find shallower idle state matching given duration.
324  * @drv: cpuidle driver containing state data.
325  * @dev: Target CPU.
326  * @state_idx: Index of the capping idle state.
327  * @duration_ns: Idle duration value to match.
328  * @no_poll: Don't consider polling states.
329  */
330 static int teo_find_shallower_state(struct cpuidle_driver *drv,
331 				    struct cpuidle_device *dev, int state_idx,
332 				    s64 duration_ns, bool no_poll)
333 {
334 	int i;
335 
336 	for (i = state_idx - 1; i >= 0; i--) {
337 		if (dev->states_usage[i].disable ||
338 				(no_poll && drv->states[i].flags & CPUIDLE_FLAG_POLLING))
339 			continue;
340 
341 		state_idx = i;
342 		if (drv->states[i].target_residency_ns <= duration_ns)
343 			break;
344 	}
345 	return state_idx;
346 }
347 
348 /**
349  * teo_select - Selects the next idle state to enter.
350  * @drv: cpuidle driver containing state data.
351  * @dev: Target CPU.
352  * @stop_tick: Indication on whether or not to stop the scheduler tick.
353  */
354 static int teo_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
355 		      bool *stop_tick)
356 {
357 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
358 	s64 latency_req = cpuidle_governor_latency_req(dev->cpu);
359 	unsigned int idx_intercept_sum = 0;
360 	unsigned int intercept_sum = 0;
361 	unsigned int idx_recent_sum = 0;
362 	unsigned int recent_sum = 0;
363 	unsigned int idx_hit_sum = 0;
364 	unsigned int hit_sum = 0;
365 	int constraint_idx = 0;
366 	int idx0 = 0, idx = -1;
367 	bool alt_intercepts, alt_recent;
368 	ktime_t delta_tick;
369 	s64 duration_ns;
370 	int i;
371 
372 	if (dev->last_state_idx >= 0) {
373 		teo_update(drv, dev);
374 		dev->last_state_idx = -1;
375 	}
376 
377 	cpu_data->time_span_ns = local_clock();
378 
379 	duration_ns = tick_nohz_get_sleep_length(&delta_tick);
380 	cpu_data->sleep_length_ns = duration_ns;
381 
382 	/* Check if there is any choice in the first place. */
383 	if (drv->state_count < 2) {
384 		idx = 0;
385 		goto end;
386 	}
387 	if (!dev->states_usage[0].disable) {
388 		idx = 0;
389 		if (drv->states[1].target_residency_ns > duration_ns)
390 			goto end;
391 	}
392 
393 	cpu_data->utilized = teo_cpu_is_utilized(dev->cpu, cpu_data);
394 	/*
395 	 * If the CPU is being utilized over the threshold and there are only 2
396 	 * states to choose from, the metrics need not be considered, so choose
397 	 * the shallowest non-polling state and exit.
398 	 */
399 	if (drv->state_count < 3 && cpu_data->utilized) {
400 		for (i = 0; i < drv->state_count; ++i) {
401 			if (!dev->states_usage[i].disable &&
402 			    !(drv->states[i].flags & CPUIDLE_FLAG_POLLING)) {
403 				idx = i;
404 				goto end;
405 			}
406 		}
407 	}
408 
409 	/*
410 	 * Find the deepest idle state whose target residency does not exceed
411 	 * the current sleep length and the deepest idle state not deeper than
412 	 * the former whose exit latency does not exceed the current latency
413 	 * constraint.  Compute the sums of metrics for early wakeup pattern
414 	 * detection.
415 	 */
416 	for (i = 1; i < drv->state_count; i++) {
417 		struct teo_bin *prev_bin = &cpu_data->state_bins[i-1];
418 		struct cpuidle_state *s = &drv->states[i];
419 
420 		/*
421 		 * Update the sums of idle state mertics for all of the states
422 		 * shallower than the current one.
423 		 */
424 		intercept_sum += prev_bin->intercepts;
425 		hit_sum += prev_bin->hits;
426 		recent_sum += prev_bin->recent;
427 
428 		if (dev->states_usage[i].disable)
429 			continue;
430 
431 		if (idx < 0) {
432 			idx = i; /* first enabled state */
433 			idx0 = i;
434 		}
435 
436 		if (s->target_residency_ns > duration_ns)
437 			break;
438 
439 		idx = i;
440 
441 		if (s->exit_latency_ns <= latency_req)
442 			constraint_idx = i;
443 
444 		idx_intercept_sum = intercept_sum;
445 		idx_hit_sum = hit_sum;
446 		idx_recent_sum = recent_sum;
447 	}
448 
449 	/* Avoid unnecessary overhead. */
450 	if (idx < 0) {
451 		idx = 0; /* No states enabled, must use 0. */
452 		goto end;
453 	} else if (idx == idx0) {
454 		goto end;
455 	}
456 
457 	/*
458 	 * If the sum of the intercepts metric for all of the idle states
459 	 * shallower than the current candidate one (idx) is greater than the
460 	 * sum of the intercepts and hits metrics for the candidate state and
461 	 * all of the deeper states, or the sum of the numbers of recent
462 	 * intercepts over all of the states shallower than the candidate one
463 	 * is greater than a half of the number of recent events taken into
464 	 * account, the CPU is likely to wake up early, so find an alternative
465 	 * idle state to select.
466 	 */
467 	alt_intercepts = 2 * idx_intercept_sum > cpu_data->total - idx_hit_sum;
468 	alt_recent = idx_recent_sum > NR_RECENT / 2;
469 	if (alt_recent || alt_intercepts) {
470 		s64 first_suitable_span_ns = duration_ns;
471 		int first_suitable_idx = idx;
472 
473 		/*
474 		 * Look for the deepest idle state whose target residency had
475 		 * not exceeded the idle duration in over a half of the relevant
476 		 * cases (both with respect to intercepts overall and with
477 		 * respect to the recent intercepts only) in the past.
478 		 *
479 		 * Take the possible latency constraint and duration limitation
480 		 * present if the tick has been stopped already into account.
481 		 */
482 		intercept_sum = 0;
483 		recent_sum = 0;
484 
485 		for (i = idx - 1; i >= 0; i--) {
486 			struct teo_bin *bin = &cpu_data->state_bins[i];
487 			s64 span_ns;
488 
489 			intercept_sum += bin->intercepts;
490 			recent_sum += bin->recent;
491 
492 			span_ns = teo_middle_of_bin(i, drv);
493 
494 			if ((!alt_recent || 2 * recent_sum > idx_recent_sum) &&
495 			    (!alt_intercepts ||
496 			     2 * intercept_sum > idx_intercept_sum)) {
497 				if (teo_time_ok(span_ns) &&
498 				    !dev->states_usage[i].disable) {
499 					idx = i;
500 					duration_ns = span_ns;
501 				} else {
502 					/*
503 					 * The current state is too shallow or
504 					 * disabled, so take the first enabled
505 					 * deeper state with suitable time span.
506 					 */
507 					idx = first_suitable_idx;
508 					duration_ns = first_suitable_span_ns;
509 				}
510 				break;
511 			}
512 
513 			if (dev->states_usage[i].disable)
514 				continue;
515 
516 			if (!teo_time_ok(span_ns)) {
517 				/*
518 				 * The current state is too shallow, but if an
519 				 * alternative candidate state has been found,
520 				 * it may still turn out to be a better choice.
521 				 */
522 				if (first_suitable_idx != idx)
523 					continue;
524 
525 				break;
526 			}
527 
528 			first_suitable_span_ns = span_ns;
529 			first_suitable_idx = i;
530 		}
531 	}
532 
533 	/*
534 	 * If there is a latency constraint, it may be necessary to select an
535 	 * idle state shallower than the current candidate one.
536 	 */
537 	if (idx > constraint_idx)
538 		idx = constraint_idx;
539 
540 	/*
541 	 * If the CPU is being utilized over the threshold, choose a shallower
542 	 * non-polling state to improve latency
543 	 */
544 	if (cpu_data->utilized)
545 		idx = teo_find_shallower_state(drv, dev, idx, duration_ns, true);
546 
547 end:
548 	/*
549 	 * Don't stop the tick if the selected state is a polling one or if the
550 	 * expected idle duration is shorter than the tick period length.
551 	 */
552 	if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
553 	    duration_ns < TICK_NSEC) && !tick_nohz_tick_stopped()) {
554 		*stop_tick = false;
555 
556 		/*
557 		 * The tick is not going to be stopped, so if the target
558 		 * residency of the state to be returned is not within the time
559 		 * till the closest timer including the tick, try to correct
560 		 * that.
561 		 */
562 		if (idx > idx0 &&
563 		    drv->states[idx].target_residency_ns > delta_tick)
564 			idx = teo_find_shallower_state(drv, dev, idx, delta_tick, false);
565 	}
566 
567 	return idx;
568 }
569 
570 /**
571  * teo_reflect - Note that governor data for the CPU need to be updated.
572  * @dev: Target CPU.
573  * @state: Entered state.
574  */
575 static void teo_reflect(struct cpuidle_device *dev, int state)
576 {
577 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
578 
579 	dev->last_state_idx = state;
580 	/*
581 	 * If the wakeup was not "natural", but triggered by one of the safety
582 	 * nets, assume that the CPU might have been idle for the entire sleep
583 	 * length time.
584 	 */
585 	if (dev->poll_time_limit ||
586 	    (tick_nohz_idle_got_tick() && cpu_data->sleep_length_ns > TICK_NSEC)) {
587 		dev->poll_time_limit = false;
588 		cpu_data->time_span_ns = cpu_data->sleep_length_ns;
589 	} else {
590 		cpu_data->time_span_ns = local_clock() - cpu_data->time_span_ns;
591 	}
592 }
593 
594 /**
595  * teo_enable_device - Initialize the governor's data for the target CPU.
596  * @drv: cpuidle driver (not used).
597  * @dev: Target CPU.
598  */
599 static int teo_enable_device(struct cpuidle_driver *drv,
600 			     struct cpuidle_device *dev)
601 {
602 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
603 	unsigned long max_capacity = arch_scale_cpu_capacity(dev->cpu);
604 	int i;
605 
606 	memset(cpu_data, 0, sizeof(*cpu_data));
607 	cpu_data->util_threshold = max_capacity >> UTIL_THRESHOLD_SHIFT;
608 
609 	for (i = 0; i < NR_RECENT; i++)
610 		cpu_data->recent_idx[i] = -1;
611 
612 	return 0;
613 }
614 
615 static struct cpuidle_governor teo_governor = {
616 	.name =		"teo",
617 	.rating =	19,
618 	.enable =	teo_enable_device,
619 	.select =	teo_select,
620 	.reflect =	teo_reflect,
621 };
622 
623 static int __init teo_governor_init(void)
624 {
625 	return cpuidle_register_governor(&teo_governor);
626 }
627 
628 postcore_initcall(teo_governor_init);
629