xref: /linux/drivers/cpuidle/governors/teo.c (revision 1d1997db870f4058676439ef7014390ba9e24eb2)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Timer events oriented CPU idle governor
4  *
5  * Copyright (C) 2018 Intel Corporation
6  * Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
7  *
8  * The idea of this governor is based on the observation that on many systems
9  * timer events are two or more orders of magnitude more frequent than any
10  * other interrupts, so they are likely to be the most significant source of CPU
11  * wakeups from idle states.  Moreover, information about what happened in the
12  * (relatively recent) past can be used to estimate whether or not the deepest
13  * idle state with target residency within the time to the closest timer is
14  * likely to be suitable for the upcoming idle time of the CPU and, if not, then
15  * which of the shallower idle states to choose.
16  *
17  * Of course, non-timer wakeup sources are more important in some use cases and
18  * they can be covered by taking a few most recent idle time intervals of the
19  * CPU into account.  However, even in that case it is not necessary to consider
20  * idle duration values greater than the time till the closest timer, as the
21  * patterns that they may belong to produce average values close enough to
22  * the time till the closest timer (sleep length) anyway.
23  *
24  * Thus this governor estimates whether or not the upcoming idle time of the CPU
25  * is likely to be significantly shorter than the sleep length and selects an
26  * idle state for it in accordance with that, as follows:
27  *
28  * - Find an idle state on the basis of the sleep length and state statistics
29  *   collected over time:
30  *
31  *   o Find the deepest idle state whose target residency is less than or equal
32  *     to the sleep length.
33  *
34  *   o Select it if it matched both the sleep length and the observed idle
35  *     duration in the past more often than it matched the sleep length alone
36  *     (i.e. the observed idle duration was significantly shorter than the sleep
37  *     length matched by it).
38  *
39  *   o Otherwise, select the shallower state with the greatest matched "early"
40  *     wakeups metric.
41  *
42  * - If the majority of the most recent idle duration values are below the
43  *   target residency of the idle state selected so far, use those values to
44  *   compute the new expected idle duration and find an idle state matching it
45  *   (which has to be shallower than the one selected so far).
46  */
47 
48 #include <linux/cpuidle.h>
49 #include <linux/jiffies.h>
50 #include <linux/kernel.h>
51 #include <linux/sched/clock.h>
52 #include <linux/tick.h>
53 
54 /*
55  * The PULSE value is added to metrics when they grow and the DECAY_SHIFT value
56  * is used for decreasing metrics on a regular basis.
57  */
58 #define PULSE		1024
59 #define DECAY_SHIFT	3
60 
61 /*
62  * Number of the most recent idle duration values to take into consideration for
63  * the detection of wakeup patterns.
64  */
65 #define INTERVALS	8
66 
67 /**
68  * struct teo_idle_state - Idle state data used by the TEO cpuidle governor.
69  * @early_hits: "Early" CPU wakeups "matching" this state.
70  * @hits: "On time" CPU wakeups "matching" this state.
71  * @misses: CPU wakeups "missing" this state.
72  *
73  * A CPU wakeup is "matched" by a given idle state if the idle duration measured
74  * after the wakeup is between the target residency of that state and the target
75  * residency of the next one (or if this is the deepest available idle state, it
76  * "matches" a CPU wakeup when the measured idle duration is at least equal to
77  * its target residency).
78  *
79  * Also, from the TEO governor perspective, a CPU wakeup from idle is "early" if
80  * it occurs significantly earlier than the closest expected timer event (that
81  * is, early enough to match an idle state shallower than the one matching the
82  * time till the closest timer event).  Otherwise, the wakeup is "on time", or
83  * it is a "hit".
84  *
85  * A "miss" occurs when the given state doesn't match the wakeup, but it matches
86  * the time till the closest timer event used for idle state selection.
87  */
88 struct teo_idle_state {
89 	unsigned int early_hits;
90 	unsigned int hits;
91 	unsigned int misses;
92 };
93 
94 /**
95  * struct teo_cpu - CPU data used by the TEO cpuidle governor.
96  * @time_span_ns: Time between idle state selection and post-wakeup update.
97  * @sleep_length_ns: Time till the closest timer event (at the selection time).
98  * @states: Idle states data corresponding to this CPU.
99  * @interval_idx: Index of the most recent saved idle interval.
100  * @intervals: Saved idle duration values.
101  */
102 struct teo_cpu {
103 	u64 time_span_ns;
104 	u64 sleep_length_ns;
105 	struct teo_idle_state states[CPUIDLE_STATE_MAX];
106 	int interval_idx;
107 	u64 intervals[INTERVALS];
108 };
109 
110 static DEFINE_PER_CPU(struct teo_cpu, teo_cpus);
111 
112 /**
113  * teo_update - Update CPU data after wakeup.
114  * @drv: cpuidle driver containing state data.
115  * @dev: Target CPU.
116  */
117 static void teo_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
118 {
119 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
120 	int i, idx_hit = -1, idx_timer = -1;
121 	u64 measured_ns;
122 
123 	if (cpu_data->time_span_ns >= cpu_data->sleep_length_ns) {
124 		/*
125 		 * One of the safety nets has triggered or the wakeup was close
126 		 * enough to the closest timer event expected at the idle state
127 		 * selection time to be discarded.
128 		 */
129 		measured_ns = U64_MAX;
130 	} else {
131 		u64 lat_ns = drv->states[dev->last_state_idx].exit_latency_ns;
132 
133 		/*
134 		 * The computations below are to determine whether or not the
135 		 * (saved) time till the next timer event and the measured idle
136 		 * duration fall into the same "bin", so use last_residency_ns
137 		 * for that instead of time_span_ns which includes the cpuidle
138 		 * overhead.
139 		 */
140 		measured_ns = dev->last_residency_ns;
141 		/*
142 		 * The delay between the wakeup and the first instruction
143 		 * executed by the CPU is not likely to be worst-case every
144 		 * time, so take 1/2 of the exit latency as a very rough
145 		 * approximation of the average of it.
146 		 */
147 		if (measured_ns >= lat_ns)
148 			measured_ns -= lat_ns / 2;
149 		else
150 			measured_ns /= 2;
151 	}
152 
153 	/*
154 	 * Decay the "early hits" metric for all of the states and find the
155 	 * states matching the sleep length and the measured idle duration.
156 	 */
157 	for (i = 0; i < drv->state_count; i++) {
158 		unsigned int early_hits = cpu_data->states[i].early_hits;
159 
160 		cpu_data->states[i].early_hits -= early_hits >> DECAY_SHIFT;
161 
162 		if (drv->states[i].target_residency_ns <= cpu_data->sleep_length_ns) {
163 			idx_timer = i;
164 			if (drv->states[i].target_residency_ns <= measured_ns)
165 				idx_hit = i;
166 		}
167 	}
168 
169 	/*
170 	 * Update the "hits" and "misses" data for the state matching the sleep
171 	 * length.  If it matches the measured idle duration too, this is a hit,
172 	 * so increase the "hits" metric for it then.  Otherwise, this is a
173 	 * miss, so increase the "misses" metric for it.  In the latter case
174 	 * also increase the "early hits" metric for the state that actually
175 	 * matches the measured idle duration.
176 	 */
177 	if (idx_timer >= 0) {
178 		unsigned int hits = cpu_data->states[idx_timer].hits;
179 		unsigned int misses = cpu_data->states[idx_timer].misses;
180 
181 		hits -= hits >> DECAY_SHIFT;
182 		misses -= misses >> DECAY_SHIFT;
183 
184 		if (idx_timer > idx_hit) {
185 			misses += PULSE;
186 			if (idx_hit >= 0)
187 				cpu_data->states[idx_hit].early_hits += PULSE;
188 		} else {
189 			hits += PULSE;
190 		}
191 
192 		cpu_data->states[idx_timer].misses = misses;
193 		cpu_data->states[idx_timer].hits = hits;
194 	}
195 
196 	/*
197 	 * Save idle duration values corresponding to non-timer wakeups for
198 	 * pattern detection.
199 	 */
200 	cpu_data->intervals[cpu_data->interval_idx++] = measured_ns;
201 	if (cpu_data->interval_idx > INTERVALS)
202 		cpu_data->interval_idx = 0;
203 }
204 
205 static bool teo_time_ok(u64 interval_ns)
206 {
207 	return !tick_nohz_tick_stopped() || interval_ns >= TICK_NSEC;
208 }
209 
210 /**
211  * teo_find_shallower_state - Find shallower idle state matching given duration.
212  * @drv: cpuidle driver containing state data.
213  * @dev: Target CPU.
214  * @state_idx: Index of the capping idle state.
215  * @duration_ns: Idle duration value to match.
216  */
217 static int teo_find_shallower_state(struct cpuidle_driver *drv,
218 				    struct cpuidle_device *dev, int state_idx,
219 				    u64 duration_ns)
220 {
221 	int i;
222 
223 	for (i = state_idx - 1; i >= 0; i--) {
224 		if (dev->states_usage[i].disable)
225 			continue;
226 
227 		state_idx = i;
228 		if (drv->states[i].target_residency_ns <= duration_ns)
229 			break;
230 	}
231 	return state_idx;
232 }
233 
234 /**
235  * teo_select - Selects the next idle state to enter.
236  * @drv: cpuidle driver containing state data.
237  * @dev: Target CPU.
238  * @stop_tick: Indication on whether or not to stop the scheduler tick.
239  */
240 static int teo_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,
241 		      bool *stop_tick)
242 {
243 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
244 	s64 latency_req = cpuidle_governor_latency_req(dev->cpu);
245 	u64 duration_ns;
246 	unsigned int hits, misses, early_hits;
247 	int max_early_idx, prev_max_early_idx, constraint_idx, idx, i;
248 	ktime_t delta_tick;
249 
250 	if (dev->last_state_idx >= 0) {
251 		teo_update(drv, dev);
252 		dev->last_state_idx = -1;
253 	}
254 
255 	cpu_data->time_span_ns = local_clock();
256 
257 	duration_ns = tick_nohz_get_sleep_length(&delta_tick);
258 	cpu_data->sleep_length_ns = duration_ns;
259 
260 	hits = 0;
261 	misses = 0;
262 	early_hits = 0;
263 	max_early_idx = -1;
264 	prev_max_early_idx = -1;
265 	constraint_idx = drv->state_count;
266 	idx = -1;
267 
268 	for (i = 0; i < drv->state_count; i++) {
269 		struct cpuidle_state *s = &drv->states[i];
270 
271 		if (dev->states_usage[i].disable) {
272 			/*
273 			 * Ignore disabled states with target residencies beyond
274 			 * the anticipated idle duration.
275 			 */
276 			if (s->target_residency_ns > duration_ns)
277 				continue;
278 
279 			/*
280 			 * This state is disabled, so the range of idle duration
281 			 * values corresponding to it is covered by the current
282 			 * candidate state, but still the "hits" and "misses"
283 			 * metrics of the disabled state need to be used to
284 			 * decide whether or not the state covering the range in
285 			 * question is good enough.
286 			 */
287 			hits = cpu_data->states[i].hits;
288 			misses = cpu_data->states[i].misses;
289 
290 			if (early_hits >= cpu_data->states[i].early_hits ||
291 			    idx < 0)
292 				continue;
293 
294 			/*
295 			 * If the current candidate state has been the one with
296 			 * the maximum "early hits" metric so far, the "early
297 			 * hits" metric of the disabled state replaces the
298 			 * current "early hits" count to avoid selecting a
299 			 * deeper state with lower "early hits" metric.
300 			 */
301 			if (max_early_idx == idx) {
302 				early_hits = cpu_data->states[i].early_hits;
303 				continue;
304 			}
305 
306 			/*
307 			 * The current candidate state is closer to the disabled
308 			 * one than the current maximum "early hits" state, so
309 			 * replace the latter with it, but in case the maximum
310 			 * "early hits" state index has not been set so far,
311 			 * check if the current candidate state is not too
312 			 * shallow for that role.
313 			 */
314 			if (teo_time_ok(drv->states[idx].target_residency_ns)) {
315 				prev_max_early_idx = max_early_idx;
316 				early_hits = cpu_data->states[i].early_hits;
317 				max_early_idx = idx;
318 			}
319 
320 			continue;
321 		}
322 
323 		if (idx < 0) {
324 			idx = i; /* first enabled state */
325 			hits = cpu_data->states[i].hits;
326 			misses = cpu_data->states[i].misses;
327 		}
328 
329 		if (s->target_residency_ns > duration_ns)
330 			break;
331 
332 		if (s->exit_latency_ns > latency_req && constraint_idx > i)
333 			constraint_idx = i;
334 
335 		idx = i;
336 		hits = cpu_data->states[i].hits;
337 		misses = cpu_data->states[i].misses;
338 
339 		if (early_hits < cpu_data->states[i].early_hits &&
340 		    teo_time_ok(drv->states[i].target_residency_ns)) {
341 			prev_max_early_idx = max_early_idx;
342 			early_hits = cpu_data->states[i].early_hits;
343 			max_early_idx = i;
344 		}
345 	}
346 
347 	/*
348 	 * If the "hits" metric of the idle state matching the sleep length is
349 	 * greater than its "misses" metric, that is the one to use.  Otherwise,
350 	 * it is more likely that one of the shallower states will match the
351 	 * idle duration observed after wakeup, so take the one with the maximum
352 	 * "early hits" metric, but if that cannot be determined, just use the
353 	 * state selected so far.
354 	 */
355 	if (hits <= misses) {
356 		/*
357 		 * The current candidate state is not suitable, so take the one
358 		 * whose "early hits" metric is the maximum for the range of
359 		 * shallower states.
360 		 */
361 		if (idx == max_early_idx)
362 			max_early_idx = prev_max_early_idx;
363 
364 		if (max_early_idx >= 0) {
365 			idx = max_early_idx;
366 			duration_ns = drv->states[idx].target_residency_ns;
367 		}
368 	}
369 
370 	/*
371 	 * If there is a latency constraint, it may be necessary to use a
372 	 * shallower idle state than the one selected so far.
373 	 */
374 	if (constraint_idx < idx)
375 		idx = constraint_idx;
376 
377 	if (idx < 0) {
378 		idx = 0; /* No states enabled. Must use 0. */
379 	} else if (idx > 0) {
380 		unsigned int count = 0;
381 		u64 sum = 0;
382 
383 		/*
384 		 * Count and sum the most recent idle duration values less than
385 		 * the current expected idle duration value.
386 		 */
387 		for (i = 0; i < INTERVALS; i++) {
388 			u64 val = cpu_data->intervals[i];
389 
390 			if (val >= duration_ns)
391 				continue;
392 
393 			count++;
394 			sum += val;
395 		}
396 
397 		/*
398 		 * Give up unless the majority of the most recent idle duration
399 		 * values are in the interesting range.
400 		 */
401 		if (count > INTERVALS / 2) {
402 			u64 avg_ns = div64_u64(sum, count);
403 
404 			/*
405 			 * Avoid spending too much time in an idle state that
406 			 * would be too shallow.
407 			 */
408 			if (teo_time_ok(avg_ns)) {
409 				duration_ns = avg_ns;
410 				if (drv->states[idx].target_residency_ns > avg_ns)
411 					idx = teo_find_shallower_state(drv, dev,
412 								       idx, avg_ns);
413 			}
414 		}
415 	}
416 
417 	/*
418 	 * Don't stop the tick if the selected state is a polling one or if the
419 	 * expected idle duration is shorter than the tick period length.
420 	 */
421 	if (((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) ||
422 	    duration_ns < TICK_NSEC) && !tick_nohz_tick_stopped()) {
423 		*stop_tick = false;
424 
425 		/*
426 		 * The tick is not going to be stopped, so if the target
427 		 * residency of the state to be returned is not within the time
428 		 * till the closest timer including the tick, try to correct
429 		 * that.
430 		 */
431 		if (idx > 0 && drv->states[idx].target_residency_ns > delta_tick)
432 			idx = teo_find_shallower_state(drv, dev, idx, delta_tick);
433 	}
434 
435 	return idx;
436 }
437 
438 /**
439  * teo_reflect - Note that governor data for the CPU need to be updated.
440  * @dev: Target CPU.
441  * @state: Entered state.
442  */
443 static void teo_reflect(struct cpuidle_device *dev, int state)
444 {
445 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
446 
447 	dev->last_state_idx = state;
448 	/*
449 	 * If the wakeup was not "natural", but triggered by one of the safety
450 	 * nets, assume that the CPU might have been idle for the entire sleep
451 	 * length time.
452 	 */
453 	if (dev->poll_time_limit ||
454 	    (tick_nohz_idle_got_tick() && cpu_data->sleep_length_ns > TICK_NSEC)) {
455 		dev->poll_time_limit = false;
456 		cpu_data->time_span_ns = cpu_data->sleep_length_ns;
457 	} else {
458 		cpu_data->time_span_ns = local_clock() - cpu_data->time_span_ns;
459 	}
460 }
461 
462 /**
463  * teo_enable_device - Initialize the governor's data for the target CPU.
464  * @drv: cpuidle driver (not used).
465  * @dev: Target CPU.
466  */
467 static int teo_enable_device(struct cpuidle_driver *drv,
468 			     struct cpuidle_device *dev)
469 {
470 	struct teo_cpu *cpu_data = per_cpu_ptr(&teo_cpus, dev->cpu);
471 	int i;
472 
473 	memset(cpu_data, 0, sizeof(*cpu_data));
474 
475 	for (i = 0; i < INTERVALS; i++)
476 		cpu_data->intervals[i] = U64_MAX;
477 
478 	return 0;
479 }
480 
481 static struct cpuidle_governor teo_governor = {
482 	.name =		"teo",
483 	.rating =	19,
484 	.enable =	teo_enable_device,
485 	.select =	teo_select,
486 	.reflect =	teo_reflect,
487 };
488 
489 static int __init teo_governor_init(void)
490 {
491 	return cpuidle_register_governor(&teo_governor);
492 }
493 
494 postcore_initcall(teo_governor_init);
495