xref: /linux/drivers/cpuidle/governors/teo.c (revision a634dda26186cf9a51567020fcce52bcba5e1e59)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Timer events oriented CPU idle governor
4  *
5  * Copyright (C) 2018 - 2021 Intel Corporation
6  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
7  */
8 
9 /**
10  * DOC: teo-description
11  *
12  * The idea of this governor is based on the observation that on many systems
13  * timer interrupts are two or more orders of magnitude more frequent than any
14  * other interrupt types, so they are likely to dominate CPU wakeup patterns.
15  * Moreover, in principle, the time when the next timer event is going to occur
16  * can be determined at the idle state selection time, although doing that may
17  * be costly, so it can be regarded as the most reliable source of information
18  * for idle state selection.
19  *
20  * Of course, non-timer wakeup sources are more important in some use cases,
21  * but even then it is generally unnecessary to consider idle duration values
22  * greater than the time time till the next timer event, referred as the sleep
23  * length in what follows, because the closest timer will ultimately wake up the
24  * CPU anyway unless it is woken up earlier.
25  *
26  * However, since obtaining the sleep length may be costly, the governor first
27  * checks if it can select a shallow idle state using wakeup pattern information
28  * from recent times, in which case it can do without knowing the sleep length
29  * at all.  For this purpose, it counts CPU wakeup events and looks for an idle
30  * state whose target residency has not exceeded the idle duration (measured
31  * after wakeup) in the majority of relevant recent cases.  If the target
32  * residency of that state is small enough, it may be used right away and the
33  * sleep length need not be determined.
34  *
35  * The computations carried out by this governor are based on using bins whose
36  * boundaries are aligned with the target residency parameter values of the CPU
37  * idle states provided by the %CPUIdle driver in the ascending order.  That is,
38  * the first bin spans from 0 up to, but not including, the target residency of
39  * the second idle state (idle state 1), the second bin spans from the target
40  * residency of idle state 1 up to, but not including, the target residency of
41  * idle state 2, the third bin spans from the target residency of idle state 2
42  * up to, but not including, the target residency of idle state 3 and so on.
43  * The last bin spans from the target residency of the deepest idle state
44  * supplied by the driver to the scheduler tick period length or to infinity if
45  * the tick period length is less than the target residency of that state.  In
46  * the latter case, the governor also counts events with the measured idle
47  * duration between the tick period length and the target residency of the
48  * deepest idle state.
49  *
50  * Two metrics called "hits" and "intercepts" are associated with each bin.
51  * They are updated every time before selecting an idle state for the given CPU
52  * in accordance with what happened last time.
53  *
54  * The "hits" metric reflects the relative frequency of situations in which the
55  * sleep length and the idle duration measured after CPU wakeup fall into the
56  * same bin (that is, the CPU appears to wake up "on time" relative to the sleep
57  * length).  In turn, the "intercepts" metric reflects the relative frequency of
58  * non-timer wakeup events for which the measured idle duration falls into a bin
59  * that corresponds to an idle state shallower than the one whose bin is fallen
60  * into by the sleep length (these events are also referred to as "intercepts"
61  * below).
62  *
63  * In order to select an idle state for a CPU, the governor takes the following
64  * steps (modulo the possible latency constraint that must be taken into account
65  * too):
66  *
67  * 1. Find the deepest enabled CPU idle state (the candidate idle state) and
68  *    compute 2 sums as follows:
69  *
70  *    - The sum of the "hits" metric for all of the idle states shallower than
71  *      the candidate one (it represents the cases in which the CPU was likely
72  *      woken up by a timer).
73  *
74  *    - The sum of the "intercepts" metric for all of the idle states shallower
75  *      than the candidate one (it represents the cases in which the CPU was
76  *      likely woken up by a non-timer wakeup source).
77  *
78  * 2. If the second sum computed in step 1 is greater than a half of the sum of
79  *    both metrics for the candidate state bin and all subsequent bins(if any),
80  *    a shallower idle state is likely to be more suitable, so look for it.
81  *
82  *    - Traverse the enabled idle states shallower than the candidate one in the
83  *      descending order.
84  *
85  *    - For each of them compute the sum of the "intercepts" metrics over all
86  *      of the idle states between it and the candidate one (including the
87  *      former and excluding the latter).
88  *
89  *    - If this sum is greater than a half of the second sum computed in step 1,
90  *      use the given idle state as the new candidate one.
91  *
92  * 3. If the current candidate state is state 0 or its target residency is short
93  *    enough, return it and prevent the scheduler tick from being stopped.
94  *
95  * 4. Obtain the sleep length value and check if it is below the target
96  *    residency of the current candidate state, in which case a new shallower
97  *    candidate state needs to be found, so look for it.
98  */
99 
100 #include <linux/cpuidle.h>
101 #include <linux/jiffies.h>
102 #include <linux/kernel.h>
103 #include <linux/sched/clock.h>
104 #include <linux/tick.h>
105 
106 #include "gov.h"
107 
108 /*
109  * The PULSE value is added to metrics when they grow and the DECAY_SHIFT value
110  * is used for decreasing metrics on a regular basis.
111  */
112 #define PULSE		1024
113 #define DECAY_SHIFT	3
114 
115 /**
116  * struct teo_bin - Metrics used by the TEO cpuidle governor.
117  * @intercepts: The "intercepts" metric.
118  * @hits: The "hits" metric.
119  */
120 struct teo_bin {
121 	unsigned int intercepts;
122 	unsigned int hits;
123 };
124 
125 /**
126  * struct teo_cpu - CPU data used by the TEO cpuidle governor.
127  * @time_span_ns: Time between idle state selection and post-wakeup update.
128  * @sleep_length_ns: Time till the closest timer event (at the selection time).
129  * @state_bins: Idle state data bins for this CPU.
130  * @total: Grand total of the "intercepts" and "hits" metrics for all bins.
131  * @tick_hits: Number of "hits" after TICK_NSEC.
132  */
133 struct teo_cpu {
134 	s64 time_span_ns;
135 	s64 sleep_length_ns;
136 	struct teo_bin state_bins[CPUIDLE_STATE_MAX];
137 	unsigned int total;
138 	unsigned int tick_hits;
139 };
140 
141 static DEFINE_PER_CPU(struct teo_cpu, teo_cpus);
142 
143 /**
144  * teo_update - Update CPU metrics after wakeup.
145  * @drv: cpuidle driver containing state data.
146  * @dev: Target CPU.
147  */
148 static void teo_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
149 {
150 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
151 	int i, idx_timer = 0, idx_duration = 0;
152 	s64 target_residency_ns;
153 	u64 measured_ns;
154 
155 	if (cpu_data->time_span_ns >= cpu_data->sleep_length_ns) {
156 		/*
157 		 * One of the safety nets has triggered or the wakeup was close
158 		 * enough to the closest timer event expected at the idle state
159 		 * selection time to be discarded.
160 		 */
161 		measured_ns = U64_MAX;
162 	} else {
163 		u64 lat_ns = drv->states[dev->last_state_idx].exit_latency_ns;
164 
165 		/*
166 		 * The computations below are to determine whether or not the
167 		 * (saved) time till the next timer event and the measured idle
168 		 * duration fall into the same "bin", so use last_residency_ns
169 		 * for that instead of time_span_ns which includes the cpuidle
170 		 * overhead.
171 		 */
172 		measured_ns = dev->last_residency_ns;
173 		/*
174 		 * The delay between the wakeup and the first instruction
175 		 * executed by the CPU is not likely to be worst-case every
176 		 * time, so take 1/2 of the exit latency as a very rough
177 		 * approximation of the average of it.
178 		 */
179 		if (measured_ns >= lat_ns)
180 			measured_ns -= lat_ns / 2;
181 		else
182 			measured_ns /= 2;
183 	}
184 
185 	cpu_data->total = 0;
186 
187 	/*
188 	 * Decay the "hits" and "intercepts" metrics for all of the bins and
189 	 * find the bins that the sleep length and the measured idle duration
190 	 * fall into.
191 	 */
192 	for (i = 0; i < drv->state_count; i++) {
193 		struct teo_bin *bin = &cpu_data->state_bins[i];
194 
195 		bin->hits -= bin->hits >> DECAY_SHIFT;
196 		bin->intercepts -= bin->intercepts >> DECAY_SHIFT;
197 
198 		cpu_data->total += bin->hits + bin->intercepts;
199 
200 		target_residency_ns = drv->states[i].target_residency_ns;
201 
202 		if (target_residency_ns <= cpu_data->sleep_length_ns) {
203 			idx_timer = i;
204 			if (target_residency_ns <= measured_ns)
205 				idx_duration = i;
206 		}
207 	}
208 
209 	/*
210 	 * If the deepest state's target residency is below the tick length,
211 	 * make a record of it to help teo_select() decide whether or not
212 	 * to stop the tick.  This effectively adds an extra hits-only bin
213 	 * beyond the last state-related one.
214 	 */
215 	if (target_residency_ns < TICK_NSEC) {
216 		cpu_data->tick_hits -= cpu_data->tick_hits >> DECAY_SHIFT;
217 
218 		cpu_data->total += cpu_data->tick_hits;
219 
220 		if (TICK_NSEC <= cpu_data->sleep_length_ns) {
221 			idx_timer = drv->state_count;
222 			if (TICK_NSEC <= measured_ns) {
223 				cpu_data->tick_hits += PULSE;
224 				goto end;
225 			}
226 		}
227 	}
228 
229 	/*
230 	 * If the measured idle duration falls into the same bin as the sleep
231 	 * length, this is a "hit", so update the "hits" metric for that bin.
232 	 * Otherwise, update the "intercepts" metric for the bin fallen into by
233 	 * the measured idle duration.
234 	 */
235 	if (idx_timer == idx_duration)
236 		cpu_data->state_bins[idx_timer].hits += PULSE;
237 	else
238 		cpu_data->state_bins[idx_duration].intercepts += PULSE;
239 
240 end:
241 	cpu_data->total += PULSE;
242 }
243 
244 static bool teo_state_ok(int i, struct cpuidle_driver *drv)
245 {
246 	return !tick_nohz_tick_stopped() ||
247 		drv->states[i].target_residency_ns >= TICK_NSEC;
248 }
249 
250 /**
251  * teo_find_shallower_state - Find shallower idle state matching given duration.
252  * @drv: cpuidle driver containing state data.
253  * @dev: Target CPU.
254  * @state_idx: Index of the capping idle state.
255  * @duration_ns: Idle duration value to match.
256  * @no_poll: Don't consider polling states.
257  */
258 static int teo_find_shallower_state(struct cpuidle_driver *drv,
259 				    struct cpuidle_device *dev, int state_idx,
260 				    s64 duration_ns, bool no_poll)
261 {
262 	int i;
263 
264 	for (i = state_idx - 1; i >= 0; i--) {
265 		if (dev->states_usage[i].disable ||
266 				(no_poll && drv->states[i].flags & CPUIDLE_FLAG_POLLING))
267 			continue;
268 
269 		state_idx = i;
270 		if (drv->states[i].target_residency_ns <= duration_ns)
271 			break;
272 	}
273 	return state_idx;
274 }
275 
276 /**
277  * teo_select - Selects the next idle state to enter.
278  * @drv: cpuidle driver containing state data.
279  * @dev: Target CPU.
280  * @stop_tick: Indication on whether or not to stop the scheduler tick.
281  */
282 static int teo_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
283 		      bool *stop_tick)
284 {
285 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
286 	s64 latency_req = cpuidle_governor_latency_req(dev->cpu);
287 	ktime_t delta_tick = TICK_NSEC / 2;
288 	unsigned int tick_intercept_sum = 0;
289 	unsigned int idx_intercept_sum = 0;
290 	unsigned int intercept_sum = 0;
291 	unsigned int idx_hit_sum = 0;
292 	unsigned int hit_sum = 0;
293 	int constraint_idx = 0;
294 	int idx0 = 0, idx = -1;
295 	int prev_intercept_idx;
296 	s64 duration_ns;
297 	int i;
298 
299 	if (dev->last_state_idx >= 0) {
300 		teo_update(drv, dev);
301 		dev->last_state_idx = -1;
302 	}
303 
304 	cpu_data->time_span_ns = local_clock();
305 	/*
306 	 * Set the expected sleep length to infinity in case of an early
307 	 * return.
308 	 */
309 	cpu_data->sleep_length_ns = KTIME_MAX;
310 
311 	/* Check if there is any choice in the first place. */
312 	if (drv->state_count < 2) {
313 		idx = 0;
314 		goto out_tick;
315 	}
316 
317 	if (!dev->states_usage[0].disable)
318 		idx = 0;
319 
320 	/* Compute the sums of metrics for early wakeup pattern detection. */
321 	for (i = 1; i < drv->state_count; i++) {
322 		struct teo_bin *prev_bin = &cpu_data->state_bins[i-1];
323 		struct cpuidle_state *s = &drv->states[i];
324 
325 		/*
326 		 * Update the sums of idle state mertics for all of the states
327 		 * shallower than the current one.
328 		 */
329 		intercept_sum += prev_bin->intercepts;
330 		hit_sum += prev_bin->hits;
331 
332 		if (dev->states_usage[i].disable)
333 			continue;
334 
335 		if (idx < 0)
336 			idx0 = i; /* first enabled state */
337 
338 		idx = i;
339 
340 		if (s->exit_latency_ns <= latency_req)
341 			constraint_idx = i;
342 
343 		/* Save the sums for the current state. */
344 		idx_intercept_sum = intercept_sum;
345 		idx_hit_sum = hit_sum;
346 	}
347 
348 	/* Avoid unnecessary overhead. */
349 	if (idx < 0) {
350 		idx = 0; /* No states enabled, must use 0. */
351 		goto out_tick;
352 	}
353 
354 	if (idx == idx0) {
355 		/*
356 		 * Only one idle state is enabled, so use it, but do not
357 		 * allow the tick to be stopped it is shallow enough.
358 		 */
359 		duration_ns = drv->states[idx].target_residency_ns;
360 		goto end;
361 	}
362 
363 	tick_intercept_sum = intercept_sum +
364 			cpu_data->state_bins[drv->state_count-1].intercepts;
365 
366 	/*
367 	 * If the sum of the intercepts metric for all of the idle states
368 	 * shallower than the current candidate one (idx) is greater than the
369 	 * sum of the intercepts and hits metrics for the candidate state and
370 	 * all of the deeper states a shallower idle state is likely to be a
371 	 * better choice.
372 	 */
373 	prev_intercept_idx = idx;
374 	if (2 * idx_intercept_sum > cpu_data->total - idx_hit_sum) {
375 		int first_suitable_idx = idx;
376 
377 		/*
378 		 * Look for the deepest idle state whose target residency had
379 		 * not exceeded the idle duration in over a half of the relevant
380 		 * cases in the past.
381 		 *
382 		 * Take the possible duration limitation present if the tick
383 		 * has been stopped already into account.
384 		 */
385 		intercept_sum = 0;
386 
387 		for (i = idx - 1; i >= 0; i--) {
388 			struct teo_bin *bin = &cpu_data->state_bins[i];
389 
390 			intercept_sum += bin->intercepts;
391 
392 			if (2 * intercept_sum > idx_intercept_sum) {
393 				/*
394 				 * Use the current state unless it is too
395 				 * shallow or disabled, in which case take the
396 				 * first enabled state that is deep enough.
397 				 */
398 				if (teo_state_ok(i, drv) &&
399 				    !dev->states_usage[i].disable)
400 					idx = i;
401 				else
402 					idx = first_suitable_idx;
403 
404 				break;
405 			}
406 
407 			if (dev->states_usage[i].disable)
408 				continue;
409 
410 			if (!teo_state_ok(i, drv)) {
411 				/*
412 				 * The current state is too shallow, but if an
413 				 * alternative candidate state has been found,
414 				 * it may still turn out to be a better choice.
415 				 */
416 				if (first_suitable_idx != idx)
417 					continue;
418 
419 				break;
420 			}
421 
422 			first_suitable_idx = i;
423 		}
424 	}
425 	if (!idx && prev_intercept_idx) {
426 		/*
427 		 * We have to query the sleep length here otherwise we don't
428 		 * know after wakeup if our guess was correct.
429 		 */
430 		duration_ns = tick_nohz_get_sleep_length(&delta_tick);
431 		cpu_data->sleep_length_ns = duration_ns;
432 		goto out_tick;
433 	}
434 
435 	/*
436 	 * If there is a latency constraint, it may be necessary to select an
437 	 * idle state shallower than the current candidate one.
438 	 */
439 	if (idx > constraint_idx)
440 		idx = constraint_idx;
441 
442 	/*
443 	 * Skip the timers check if state 0 is the current candidate one,
444 	 * because an immediate non-timer wakeup is expected in that case.
445 	 */
446 	if (!idx)
447 		goto out_tick;
448 
449 	/*
450 	 * If state 0 is a polling one, check if the target residency of
451 	 * the current candidate state is low enough and skip the timers
452 	 * check in that case too.
453 	 */
454 	if ((drv->states[0].flags & CPUIDLE_FLAG_POLLING) &&
455 	    drv->states[idx].target_residency_ns < RESIDENCY_THRESHOLD_NS)
456 		goto out_tick;
457 
458 	duration_ns = tick_nohz_get_sleep_length(&delta_tick);
459 	cpu_data->sleep_length_ns = duration_ns;
460 
461 	/*
462 	 * If the closest expected timer is before the target residency of the
463 	 * candidate state, a shallower one needs to be found.
464 	 */
465 	if (drv->states[idx].target_residency_ns > duration_ns) {
466 		i = teo_find_shallower_state(drv, dev, idx, duration_ns, false);
467 		if (teo_state_ok(i, drv))
468 			idx = i;
469 	}
470 
471 	/*
472 	 * If the selected state's target residency is below the tick length
473 	 * and intercepts occurring before the tick length are the majority of
474 	 * total wakeup events, do not stop the tick.
475 	 */
476 	if (drv->states[idx].target_residency_ns < TICK_NSEC &&
477 	    tick_intercept_sum > cpu_data->total / 2 + cpu_data->total / 8)
478 		duration_ns = TICK_NSEC / 2;
479 
480 end:
481 	/*
482 	 * Allow the tick to be stopped unless the selected state is a polling
483 	 * one or the expected idle duration is shorter than the tick period
484 	 * length.
485 	 */
486 	if ((!(drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
487 	    duration_ns >= TICK_NSEC) || tick_nohz_tick_stopped())
488 		return idx;
489 
490 	/*
491 	 * The tick is not going to be stopped, so if the target residency of
492 	 * the state to be returned is not within the time till the closest
493 	 * timer including the tick, try to correct that.
494 	 */
495 	if (idx > idx0 &&
496 	    drv->states[idx].target_residency_ns > delta_tick)
497 		idx = teo_find_shallower_state(drv, dev, idx, delta_tick, false);
498 
499 out_tick:
500 	*stop_tick = false;
501 	return idx;
502 }
503 
504 /**
505  * teo_reflect - Note that governor data for the CPU need to be updated.
506  * @dev: Target CPU.
507  * @state: Entered state.
508  */
509 static void teo_reflect(struct cpuidle_device *dev, int state)
510 {
511 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
512 
513 	dev->last_state_idx = state;
514 	/*
515 	 * If the wakeup was not "natural", but triggered by one of the safety
516 	 * nets, assume that the CPU might have been idle for the entire sleep
517 	 * length time.
518 	 */
519 	if (dev->poll_time_limit ||
520 	    (tick_nohz_idle_got_tick() && cpu_data->sleep_length_ns > TICK_NSEC)) {
521 		dev->poll_time_limit = false;
522 		cpu_data->time_span_ns = cpu_data->sleep_length_ns;
523 	} else {
524 		cpu_data->time_span_ns = local_clock() - cpu_data->time_span_ns;
525 	}
526 }
527 
528 /**
529  * teo_enable_device - Initialize the governor's data for the target CPU.
530  * @drv: cpuidle driver (not used).
531  * @dev: Target CPU.
532  */
533 static int teo_enable_device(struct cpuidle_driver *drv,
534 			     struct cpuidle_device *dev)
535 {
536 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
537 
538 	memset(cpu_data, 0, sizeof(*cpu_data));
539 
540 	return 0;
541 }
542 
543 static struct cpuidle_governor teo_governor = {
544 	.name =		"teo",
545 	.rating =	19,
546 	.enable =	teo_enable_device,
547 	.select =	teo_select,
548 	.reflect =	teo_reflect,
549 };
550 
551 static int __init teo_governor_init(void)
552 {
553 	return cpuidle_register_governor(&teo_governor);
554 }
555 
556 postcore_initcall(teo_governor_init);
557