xref: /openbmc/linux/drivers/cpuidle/governors/menu.c (revision d2999e1b)
1 /*
2  * menu.c - the menu idle governor
3  *
4  * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com>
5  * Copyright (C) 2009 Intel Corporation
6  * Author:
7  *        Arjan van de Ven <arjan@linux.intel.com>
8  *
9  * This code is licenced under the GPL version 2 as described
10  * in the COPYING file that acompanies the Linux Kernel.
11  */
12 
13 #include <linux/kernel.h>
14 #include <linux/cpuidle.h>
15 #include <linux/pm_qos.h>
16 #include <linux/time.h>
17 #include <linux/ktime.h>
18 #include <linux/hrtimer.h>
19 #include <linux/tick.h>
20 #include <linux/sched.h>
21 #include <linux/math64.h>
22 #include <linux/module.h>
23 
24 /*
25  * Please note when changing the tuning values:
26  * If (MAX_INTERESTING-1) * RESOLUTION > UINT_MAX, the result of
27  * a scaling operation multiplication may overflow on 32 bit platforms.
28  * In that case, #define RESOLUTION as ULL to get 64 bit result:
29  * #define RESOLUTION 1024ULL
30  *
31  * The default values do not overflow.
32  */
33 #define BUCKETS 12
34 #define INTERVALS 8
35 #define RESOLUTION 1024
36 #define DECAY 8
37 #define MAX_INTERESTING 50000
38 #define STDDEV_THRESH 400
39 
40 
41 /*
42  * Concepts and ideas behind the menu governor
43  *
44  * For the menu governor, there are 3 decision factors for picking a C
45  * state:
46  * 1) Energy break even point
47  * 2) Performance impact
48  * 3) Latency tolerance (from pmqos infrastructure)
49  * These these three factors are treated independently.
50  *
51  * Energy break even point
52  * -----------------------
53  * C state entry and exit have an energy cost, and a certain amount of time in
54  * the  C state is required to actually break even on this cost. CPUIDLE
55  * provides us this duration in the "target_residency" field. So all that we
56  * need is a good prediction of how long we'll be idle. Like the traditional
57  * menu governor, we start with the actual known "next timer event" time.
58  *
59  * Since there are other source of wakeups (interrupts for example) than
60  * the next timer event, this estimation is rather optimistic. To get a
61  * more realistic estimate, a correction factor is applied to the estimate,
62  * that is based on historic behavior. For example, if in the past the actual
63  * duration always was 50% of the next timer tick, the correction factor will
64  * be 0.5.
65  *
66  * menu uses a running average for this correction factor, however it uses a
67  * set of factors, not just a single factor. This stems from the realization
68  * that the ratio is dependent on the order of magnitude of the expected
69  * duration; if we expect 500 milliseconds of idle time the likelihood of
70  * getting an interrupt very early is much higher than if we expect 50 micro
71  * seconds of idle time. A second independent factor that has big impact on
72  * the actual factor is if there is (disk) IO outstanding or not.
73  * (as a special twist, we consider every sleep longer than 50 milliseconds
74  * as perfect; there are no power gains for sleeping longer than this)
75  *
76  * For these two reasons we keep an array of 12 independent factors, that gets
77  * indexed based on the magnitude of the expected duration as well as the
78  * "is IO outstanding" property.
79  *
80  * Repeatable-interval-detector
81  * ----------------------------
82  * There are some cases where "next timer" is a completely unusable predictor:
83  * Those cases where the interval is fixed, for example due to hardware
84  * interrupt mitigation, but also due to fixed transfer rate devices such as
85  * mice.
86  * For this, we use a different predictor: We track the duration of the last 8
87  * intervals and if the stand deviation of these 8 intervals is below a
88  * threshold value, we use the average of these intervals as prediction.
89  *
90  * Limiting Performance Impact
91  * ---------------------------
92  * C states, especially those with large exit latencies, can have a real
93  * noticeable impact on workloads, which is not acceptable for most sysadmins,
94  * and in addition, less performance has a power price of its own.
95  *
96  * As a general rule of thumb, menu assumes that the following heuristic
97  * holds:
98  *     The busier the system, the less impact of C states is acceptable
99  *
100  * This rule-of-thumb is implemented using a performance-multiplier:
101  * If the exit latency times the performance multiplier is longer than
102  * the predicted duration, the C state is not considered a candidate
103  * for selection due to a too high performance impact. So the higher
104  * this multiplier is, the longer we need to be idle to pick a deep C
105  * state, and thus the less likely a busy CPU will hit such a deep
106  * C state.
107  *
108  * Two factors are used in determing this multiplier:
109  * a value of 10 is added for each point of "per cpu load average" we have.
110  * a value of 5 points is added for each process that is waiting for
111  * IO on this CPU.
112  * (these values are experimentally determined)
113  *
114  * The load average factor gives a longer term (few seconds) input to the
115  * decision, while the iowait value gives a cpu local instantanious input.
116  * The iowait factor may look low, but realize that this is also already
117  * represented in the system load average.
118  *
119  */
120 
121 struct menu_device {
122 	int		last_state_idx;
123 	int             needs_update;
124 
125 	unsigned int	next_timer_us;
126 	unsigned int	predicted_us;
127 	unsigned int	bucket;
128 	unsigned int	correction_factor[BUCKETS];
129 	unsigned int	intervals[INTERVALS];
130 	int		interval_ptr;
131 };
132 
133 
134 #define LOAD_INT(x) ((x) >> FSHIFT)
135 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
136 
137 static int get_loadavg(void)
138 {
139 	unsigned long this = this_cpu_load();
140 
141 
142 	return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10;
143 }
144 
145 static inline int which_bucket(unsigned int duration)
146 {
147 	int bucket = 0;
148 
149 	/*
150 	 * We keep two groups of stats; one with no
151 	 * IO pending, one without.
152 	 * This allows us to calculate
153 	 * E(duration)|iowait
154 	 */
155 	if (nr_iowait_cpu(smp_processor_id()))
156 		bucket = BUCKETS/2;
157 
158 	if (duration < 10)
159 		return bucket;
160 	if (duration < 100)
161 		return bucket + 1;
162 	if (duration < 1000)
163 		return bucket + 2;
164 	if (duration < 10000)
165 		return bucket + 3;
166 	if (duration < 100000)
167 		return bucket + 4;
168 	return bucket + 5;
169 }
170 
171 /*
172  * Return a multiplier for the exit latency that is intended
173  * to take performance requirements into account.
174  * The more performance critical we estimate the system
175  * to be, the higher this multiplier, and thus the higher
176  * the barrier to go to an expensive C state.
177  */
178 static inline int performance_multiplier(void)
179 {
180 	int mult = 1;
181 
182 	/* for higher loadavg, we are more reluctant */
183 
184 	mult += 2 * get_loadavg();
185 
186 	/* for IO wait tasks (per cpu!) we add 5x each */
187 	mult += 10 * nr_iowait_cpu(smp_processor_id());
188 
189 	return mult;
190 }
191 
192 static DEFINE_PER_CPU(struct menu_device, menu_devices);
193 
194 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
195 
196 /* This implements DIV_ROUND_CLOSEST but avoids 64 bit division */
197 static u64 div_round64(u64 dividend, u32 divisor)
198 {
199 	return div_u64(dividend + (divisor / 2), divisor);
200 }
201 
202 /*
203  * Try detecting repeating patterns by keeping track of the last 8
204  * intervals, and checking if the standard deviation of that set
205  * of points is below a threshold. If it is... then use the
206  * average of these 8 points as the estimated value.
207  */
208 static void get_typical_interval(struct menu_device *data)
209 {
210 	int i, divisor;
211 	unsigned int max, thresh;
212 	uint64_t avg, stddev;
213 
214 	thresh = UINT_MAX; /* Discard outliers above this value */
215 
216 again:
217 
218 	/* First calculate the average of past intervals */
219 	max = 0;
220 	avg = 0;
221 	divisor = 0;
222 	for (i = 0; i < INTERVALS; i++) {
223 		unsigned int value = data->intervals[i];
224 		if (value <= thresh) {
225 			avg += value;
226 			divisor++;
227 			if (value > max)
228 				max = value;
229 		}
230 	}
231 	do_div(avg, divisor);
232 
233 	/* Then try to determine standard deviation */
234 	stddev = 0;
235 	for (i = 0; i < INTERVALS; i++) {
236 		unsigned int value = data->intervals[i];
237 		if (value <= thresh) {
238 			int64_t diff = value - avg;
239 			stddev += diff * diff;
240 		}
241 	}
242 	do_div(stddev, divisor);
243 	/*
244 	 * The typical interval is obtained when standard deviation is small
245 	 * or standard deviation is small compared to the average interval.
246 	 *
247 	 * int_sqrt() formal parameter type is unsigned long. When the
248 	 * greatest difference to an outlier exceeds ~65 ms * sqrt(divisor)
249 	 * the resulting squared standard deviation exceeds the input domain
250 	 * of int_sqrt on platforms where unsigned long is 32 bits in size.
251 	 * In such case reject the candidate average.
252 	 *
253 	 * Use this result only if there is no timer to wake us up sooner.
254 	 */
255 	if (likely(stddev <= ULONG_MAX)) {
256 		stddev = int_sqrt(stddev);
257 		if (((avg > stddev * 6) && (divisor * 4 >= INTERVALS * 3))
258 							|| stddev <= 20) {
259 			if (data->next_timer_us > avg)
260 				data->predicted_us = avg;
261 			return;
262 		}
263 	}
264 
265 	/*
266 	 * If we have outliers to the upside in our distribution, discard
267 	 * those by setting the threshold to exclude these outliers, then
268 	 * calculate the average and standard deviation again. Once we get
269 	 * down to the bottom 3/4 of our samples, stop excluding samples.
270 	 *
271 	 * This can deal with workloads that have long pauses interspersed
272 	 * with sporadic activity with a bunch of short pauses.
273 	 */
274 	if ((divisor * 4) <= INTERVALS * 3)
275 		return;
276 
277 	thresh = max - 1;
278 	goto again;
279 }
280 
281 /**
282  * menu_select - selects the next idle state to enter
283  * @drv: cpuidle driver containing state data
284  * @dev: the CPU
285  */
286 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
287 {
288 	struct menu_device *data = &__get_cpu_var(menu_devices);
289 	int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY);
290 	int i;
291 	unsigned int interactivity_req;
292 	struct timespec t;
293 
294 	if (data->needs_update) {
295 		menu_update(drv, dev);
296 		data->needs_update = 0;
297 	}
298 
299 	data->last_state_idx = CPUIDLE_DRIVER_STATE_START - 1;
300 
301 	/* Special case when user has set very strict latency requirement */
302 	if (unlikely(latency_req == 0))
303 		return 0;
304 
305 	/* determine the expected residency time, round up */
306 	t = ktime_to_timespec(tick_nohz_get_sleep_length());
307 	data->next_timer_us =
308 		t.tv_sec * USEC_PER_SEC + t.tv_nsec / NSEC_PER_USEC;
309 
310 
311 	data->bucket = which_bucket(data->next_timer_us);
312 
313 	/*
314 	 * Force the result of multiplication to be 64 bits even if both
315 	 * operands are 32 bits.
316 	 * Make sure to round up for half microseconds.
317 	 */
318 	data->predicted_us = div_round64((uint64_t)data->next_timer_us *
319 					 data->correction_factor[data->bucket],
320 					 RESOLUTION * DECAY);
321 
322 	get_typical_interval(data);
323 
324 	/*
325 	 * Performance multiplier defines a minimum predicted idle
326 	 * duration / latency ratio. Adjust the latency limit if
327 	 * necessary.
328 	 */
329 	interactivity_req = data->predicted_us / performance_multiplier();
330 	if (latency_req > interactivity_req)
331 		latency_req = interactivity_req;
332 
333 	/*
334 	 * We want to default to C1 (hlt), not to busy polling
335 	 * unless the timer is happening really really soon.
336 	 */
337 	if (data->next_timer_us > 5 &&
338 	    !drv->states[CPUIDLE_DRIVER_STATE_START].disabled &&
339 		dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0)
340 		data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
341 
342 	/*
343 	 * Find the idle state with the lowest power while satisfying
344 	 * our constraints.
345 	 */
346 	for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
347 		struct cpuidle_state *s = &drv->states[i];
348 		struct cpuidle_state_usage *su = &dev->states_usage[i];
349 
350 		if (s->disabled || su->disable)
351 			continue;
352 		if (s->target_residency > data->predicted_us)
353 			continue;
354 		if (s->exit_latency > latency_req)
355 			continue;
356 
357 		data->last_state_idx = i;
358 	}
359 
360 	return data->last_state_idx;
361 }
362 
363 /**
364  * menu_reflect - records that data structures need update
365  * @dev: the CPU
366  * @index: the index of actual entered state
367  *
368  * NOTE: it's important to be fast here because this operation will add to
369  *       the overall exit latency.
370  */
371 static void menu_reflect(struct cpuidle_device *dev, int index)
372 {
373 	struct menu_device *data = &__get_cpu_var(menu_devices);
374 	data->last_state_idx = index;
375 	if (index >= 0)
376 		data->needs_update = 1;
377 }
378 
379 /**
380  * menu_update - attempts to guess what happened after entry
381  * @drv: cpuidle driver containing state data
382  * @dev: the CPU
383  */
384 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
385 {
386 	struct menu_device *data = &__get_cpu_var(menu_devices);
387 	int last_idx = data->last_state_idx;
388 	struct cpuidle_state *target = &drv->states[last_idx];
389 	unsigned int measured_us;
390 	unsigned int new_factor;
391 
392 	/*
393 	 * Try to figure out how much time passed between entry to low
394 	 * power state and occurrence of the wakeup event.
395 	 *
396 	 * If the entered idle state didn't support residency measurements,
397 	 * we are basically lost in the dark how much time passed.
398 	 * As a compromise, assume we slept for the whole expected time.
399 	 *
400 	 * Any measured amount of time will include the exit latency.
401 	 * Since we are interested in when the wakeup begun, not when it
402 	 * was completed, we must substract the exit latency. However, if
403 	 * the measured amount of time is less than the exit latency,
404 	 * assume the state was never reached and the exit latency is 0.
405 	 */
406 	if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID))) {
407 		/* Use timer value as is */
408 		measured_us = data->next_timer_us;
409 
410 	} else {
411 		/* Use measured value */
412 		measured_us = cpuidle_get_last_residency(dev);
413 
414 		/* Deduct exit latency */
415 		if (measured_us > target->exit_latency)
416 			measured_us -= target->exit_latency;
417 
418 		/* Make sure our coefficients do not exceed unity */
419 		if (measured_us > data->next_timer_us)
420 			measured_us = data->next_timer_us;
421 	}
422 
423 	/* Update our correction ratio */
424 	new_factor = data->correction_factor[data->bucket];
425 	new_factor -= new_factor / DECAY;
426 
427 	if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING)
428 		new_factor += RESOLUTION * measured_us / data->next_timer_us;
429 	else
430 		/*
431 		 * we were idle so long that we count it as a perfect
432 		 * prediction
433 		 */
434 		new_factor += RESOLUTION;
435 
436 	/*
437 	 * We don't want 0 as factor; we always want at least
438 	 * a tiny bit of estimated time. Fortunately, due to rounding,
439 	 * new_factor will stay nonzero regardless of measured_us values
440 	 * and the compiler can eliminate this test as long as DECAY > 1.
441 	 */
442 	if (DECAY == 1 && unlikely(new_factor == 0))
443 		new_factor = 1;
444 
445 	data->correction_factor[data->bucket] = new_factor;
446 
447 	/* update the repeating-pattern data */
448 	data->intervals[data->interval_ptr++] = measured_us;
449 	if (data->interval_ptr >= INTERVALS)
450 		data->interval_ptr = 0;
451 }
452 
453 /**
454  * menu_enable_device - scans a CPU's states and does setup
455  * @drv: cpuidle driver
456  * @dev: the CPU
457  */
458 static int menu_enable_device(struct cpuidle_driver *drv,
459 				struct cpuidle_device *dev)
460 {
461 	struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
462 	int i;
463 
464 	memset(data, 0, sizeof(struct menu_device));
465 
466 	/*
467 	 * if the correction factor is 0 (eg first time init or cpu hotplug
468 	 * etc), we actually want to start out with a unity factor.
469 	 */
470 	for(i = 0; i < BUCKETS; i++)
471 		data->correction_factor[i] = RESOLUTION * DECAY;
472 
473 	return 0;
474 }
475 
476 static struct cpuidle_governor menu_governor = {
477 	.name =		"menu",
478 	.rating =	20,
479 	.enable =	menu_enable_device,
480 	.select =	menu_select,
481 	.reflect =	menu_reflect,
482 	.owner =	THIS_MODULE,
483 };
484 
485 /**
486  * init_menu - initializes the governor
487  */
488 static int __init init_menu(void)
489 {
490 	return cpuidle_register_governor(&menu_governor);
491 }
492 
493 postcore_initcall(init_menu);
494