• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 INTERVAL_SHIFT 3
35 #define INTERVALS (1UL << INTERVAL_SHIFT)
36 #define RESOLUTION 1024
37 #define DECAY 8
38 #define MAX_INTERESTING 50000
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 
get_loadavg(unsigned long load)137 static inline int get_loadavg(unsigned long load)
138 {
139 	return LOAD_INT(load) * 10 + LOAD_FRAC(load) / 10;
140 }
141 
which_bucket(unsigned int duration,unsigned long nr_iowaiters)142 static inline int which_bucket(unsigned int duration, unsigned long nr_iowaiters)
143 {
144 	int bucket = 0;
145 
146 	/*
147 	 * We keep two groups of stats; one with no
148 	 * IO pending, one without.
149 	 * This allows us to calculate
150 	 * E(duration)|iowait
151 	 */
152 	if (nr_iowaiters)
153 		bucket = BUCKETS/2;
154 
155 	if (duration < 10)
156 		return bucket;
157 	if (duration < 100)
158 		return bucket + 1;
159 	if (duration < 1000)
160 		return bucket + 2;
161 	if (duration < 10000)
162 		return bucket + 3;
163 	if (duration < 100000)
164 		return bucket + 4;
165 	return bucket + 5;
166 }
167 
168 /*
169  * Return a multiplier for the exit latency that is intended
170  * to take performance requirements into account.
171  * The more performance critical we estimate the system
172  * to be, the higher this multiplier, and thus the higher
173  * the barrier to go to an expensive C state.
174  */
performance_multiplier(unsigned long nr_iowaiters,unsigned long load)175 static inline int performance_multiplier(unsigned long nr_iowaiters, unsigned long load)
176 {
177 	int mult = 1;
178 
179 	/* for higher loadavg, we are more reluctant */
180 
181 	/*
182 	 * this doesn't work as intended - it is almost always 0, but can
183 	 * sometimes, depending on workload, spike very high into the hundreds
184 	 * even when the average cpu load is under 10%.
185 	 */
186 	/* mult += 2 * get_loadavg(); */
187 
188 	/* for IO wait tasks (per cpu!) we add 5x each */
189 	mult += 10 * nr_iowaiters;
190 
191 	return mult;
192 }
193 
194 static DEFINE_PER_CPU(struct menu_device, menu_devices);
195 
196 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
197 
198 /*
199  * Try detecting repeating patterns by keeping track of the last 8
200  * intervals, and checking if the standard deviation of that set
201  * of points is below a threshold. If it is... then use the
202  * average of these 8 points as the estimated value.
203  */
get_typical_interval(struct menu_device * data)204 static void get_typical_interval(struct menu_device *data)
205 {
206 	int i, divisor;
207 	unsigned int max, thresh;
208 	uint64_t avg, stddev;
209 
210 	thresh = UINT_MAX; /* Discard outliers above this value */
211 
212 again:
213 
214 	/* First calculate the average of past intervals */
215 	max = 0;
216 	avg = 0;
217 	divisor = 0;
218 	for (i = 0; i < INTERVALS; i++) {
219 		unsigned int value = data->intervals[i];
220 		if (value <= thresh) {
221 			avg += value;
222 			divisor++;
223 			if (value > max)
224 				max = value;
225 		}
226 	}
227 	if (divisor == INTERVALS)
228 		avg >>= INTERVAL_SHIFT;
229 	else
230 		do_div(avg, divisor);
231 
232 	/* Then try to determine standard deviation */
233 	stddev = 0;
234 	for (i = 0; i < INTERVALS; i++) {
235 		unsigned int value = data->intervals[i];
236 		if (value <= thresh) {
237 			int64_t diff = value - avg;
238 			stddev += diff * diff;
239 		}
240 	}
241 	if (divisor == INTERVALS)
242 		stddev >>= INTERVAL_SHIFT;
243 	else
244 		do_div(stddev, divisor);
245 
246 	/*
247 	 * The typical interval is obtained when standard deviation is small
248 	 * or standard deviation is small compared to the average interval.
249 	 *
250 	 * int_sqrt() formal parameter type is unsigned long. When the
251 	 * greatest difference to an outlier exceeds ~65 ms * sqrt(divisor)
252 	 * the resulting squared standard deviation exceeds the input domain
253 	 * of int_sqrt on platforms where unsigned long is 32 bits in size.
254 	 * In such case reject the candidate average.
255 	 *
256 	 * Use this result only if there is no timer to wake us up sooner.
257 	 */
258 	if (likely(stddev <= ULONG_MAX)) {
259 		stddev = int_sqrt(stddev);
260 		if (((avg > stddev * 6) && (divisor * 4 >= INTERVALS * 3))
261 							|| stddev <= 20) {
262 			if (data->next_timer_us > avg)
263 				data->predicted_us = avg;
264 			return;
265 		}
266 	}
267 
268 	/*
269 	 * If we have outliers to the upside in our distribution, discard
270 	 * those by setting the threshold to exclude these outliers, then
271 	 * calculate the average and standard deviation again. Once we get
272 	 * down to the bottom 3/4 of our samples, stop excluding samples.
273 	 *
274 	 * This can deal with workloads that have long pauses interspersed
275 	 * with sporadic activity with a bunch of short pauses.
276 	 */
277 	if ((divisor * 4) <= INTERVALS * 3)
278 		return;
279 
280 	thresh = max - 1;
281 	goto again;
282 }
283 
284 /**
285  * menu_select - selects the next idle state to enter
286  * @drv: cpuidle driver containing state data
287  * @dev: the CPU
288  */
menu_select(struct cpuidle_driver * drv,struct cpuidle_device * dev)289 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
290 {
291 	struct menu_device *data = this_cpu_ptr(&menu_devices);
292 	int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY);
293 	int i;
294 	unsigned int interactivity_req;
295 	unsigned long nr_iowaiters, cpu_load;
296 
297 	if (data->needs_update) {
298 		menu_update(drv, dev);
299 		data->needs_update = 0;
300 	}
301 
302 	data->last_state_idx = CPUIDLE_DRIVER_STATE_START - 1;
303 
304 	/* Special case when user has set very strict latency requirement */
305 	if (unlikely(latency_req == 0))
306 		return 0;
307 
308 	/* determine the expected residency time, round up */
309 	data->next_timer_us = ktime_to_us(tick_nohz_get_sleep_length());
310 
311 	get_iowait_load(&nr_iowaiters, &cpu_load);
312 	data->bucket = which_bucket(data->next_timer_us, nr_iowaiters);
313 
314 	/*
315 	 * Force the result of multiplication to be 64 bits even if both
316 	 * operands are 32 bits.
317 	 * Make sure to round up for half microseconds.
318 	 */
319 	data->predicted_us = DIV_ROUND_CLOSEST_ULL((uint64_t)data->next_timer_us *
320 					 data->correction_factor[data->bucket],
321 					 RESOLUTION * DECAY);
322 
323 	get_typical_interval(data);
324 
325 	/*
326 	 * Performance multiplier defines a minimum predicted idle
327 	 * duration / latency ratio. Adjust the latency limit if
328 	 * necessary.
329 	 */
330 	interactivity_req = data->predicted_us / performance_multiplier(nr_iowaiters, cpu_load);
331 	if (latency_req > interactivity_req)
332 		latency_req = interactivity_req;
333 
334 	/*
335 	 * We want to default to C1 (hlt), not to busy polling
336 	 * unless the timer is happening really really soon.
337 	 */
338 	if (data->next_timer_us > 5 &&
339 	    !drv->states[CPUIDLE_DRIVER_STATE_START].disabled &&
340 		dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0)
341 		data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
342 
343 	/*
344 	 * Find the idle state with the lowest power while satisfying
345 	 * our constraints.
346 	 */
347 	for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
348 		struct cpuidle_state *s = &drv->states[i];
349 		struct cpuidle_state_usage *su = &dev->states_usage[i];
350 
351 		if (s->disabled || su->disable)
352 			continue;
353 		if (s->target_residency > data->predicted_us)
354 			continue;
355 		if (s->exit_latency > latency_req)
356 			continue;
357 
358 		data->last_state_idx = i;
359 	}
360 
361 	return data->last_state_idx;
362 }
363 
364 /**
365  * menu_reflect - records that data structures need update
366  * @dev: the CPU
367  * @index: the index of actual entered state
368  *
369  * NOTE: it's important to be fast here because this operation will add to
370  *       the overall exit latency.
371  */
menu_reflect(struct cpuidle_device * dev,int index)372 static void menu_reflect(struct cpuidle_device *dev, int index)
373 {
374 	struct menu_device *data = this_cpu_ptr(&menu_devices);
375 
376 	data->last_state_idx = index;
377 	data->needs_update = 1;
378 }
379 
380 /**
381  * menu_update - attempts to guess what happened after entry
382  * @drv: cpuidle driver containing state data
383  * @dev: the CPU
384  */
menu_update(struct cpuidle_driver * drv,struct cpuidle_device * dev)385 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
386 {
387 	struct menu_device *data = this_cpu_ptr(&menu_devices);
388 	int last_idx = data->last_state_idx;
389 	struct cpuidle_state *target = &drv->states[last_idx];
390 	unsigned int measured_us;
391 	unsigned int new_factor;
392 
393 	/*
394 	 * Try to figure out how much time passed between entry to low
395 	 * power state and occurrence of the wakeup event.
396 	 *
397 	 * If the entered idle state didn't support residency measurements,
398 	 * we use them anyway if they are short, and if long,
399 	 * truncate to the whole expected time.
400 	 *
401 	 * Any measured amount of time will include the exit latency.
402 	 * Since we are interested in when the wakeup begun, not when it
403 	 * was completed, we must subtract the exit latency. However, if
404 	 * the measured amount of time is less than the exit latency,
405 	 * assume the state was never reached and the exit latency is 0.
406 	 */
407 
408 	/* measured value */
409 	measured_us = cpuidle_get_last_residency(dev);
410 
411 	/* Deduct exit latency */
412 	if (measured_us > target->exit_latency)
413 		measured_us -= target->exit_latency;
414 
415 	/* Make sure our coefficients do not exceed unity */
416 	if (measured_us > data->next_timer_us)
417 		measured_us = data->next_timer_us;
418 
419 	/* Update our correction ratio */
420 	new_factor = data->correction_factor[data->bucket];
421 	new_factor -= new_factor / DECAY;
422 
423 	if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING)
424 		new_factor += RESOLUTION * measured_us / data->next_timer_us;
425 	else
426 		/*
427 		 * we were idle so long that we count it as a perfect
428 		 * prediction
429 		 */
430 		new_factor += RESOLUTION;
431 
432 	/*
433 	 * We don't want 0 as factor; we always want at least
434 	 * a tiny bit of estimated time. Fortunately, due to rounding,
435 	 * new_factor will stay nonzero regardless of measured_us values
436 	 * and the compiler can eliminate this test as long as DECAY > 1.
437 	 */
438 	if (DECAY == 1 && unlikely(new_factor == 0))
439 		new_factor = 1;
440 
441 	data->correction_factor[data->bucket] = new_factor;
442 
443 	/* update the repeating-pattern data */
444 	data->intervals[data->interval_ptr++] = measured_us;
445 	if (data->interval_ptr >= INTERVALS)
446 		data->interval_ptr = 0;
447 }
448 
449 /**
450  * menu_enable_device - scans a CPU's states and does setup
451  * @drv: cpuidle driver
452  * @dev: the CPU
453  */
menu_enable_device(struct cpuidle_driver * drv,struct cpuidle_device * dev)454 static int menu_enable_device(struct cpuidle_driver *drv,
455 				struct cpuidle_device *dev)
456 {
457 	struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
458 	int i;
459 
460 	memset(data, 0, sizeof(struct menu_device));
461 
462 	/*
463 	 * if the correction factor is 0 (eg first time init or cpu hotplug
464 	 * etc), we actually want to start out with a unity factor.
465 	 */
466 	for(i = 0; i < BUCKETS; i++)
467 		data->correction_factor[i] = RESOLUTION * DECAY;
468 
469 	return 0;
470 }
471 
472 static struct cpuidle_governor menu_governor = {
473 	.name =		"menu",
474 	.rating =	20,
475 	.enable =	menu_enable_device,
476 	.select =	menu_select,
477 	.reflect =	menu_reflect,
478 	.owner =	THIS_MODULE,
479 };
480 
481 /**
482  * init_menu - initializes the governor
483  */
init_menu(void)484 static int __init init_menu(void)
485 {
486 	return cpuidle_register_governor(&menu_governor);
487 }
488 
489 postcore_initcall(init_menu);
490