2 * menu.c - the menu idle governor
4 * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com>
5 * Copyright (C) 2009 Intel Corporation
7 * Arjan van de Ven <arjan@linux.intel.com>
9 * This code is licenced under the GPL version 2 as described
10 * in the COPYING file that acompanies the Linux Kernel.
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>
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
31 * The default values do not overflow.
34 #define INTERVAL_SHIFT 3
35 #define INTERVALS (1UL << INTERVAL_SHIFT)
36 #define RESOLUTION 1024
38 #define MAX_INTERESTING 50000
42 * Concepts and ideas behind the menu governor
44 * For the menu governor, there are 3 decision factors for picking a C
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.
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.
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
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)
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.
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
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.
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.
96 * As a general rule of thumb, menu assumes that the following heuristic
98 * The busier the system, the less impact of C states is acceptable
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
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
112 * (these values are experimentally determined)
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.
125 unsigned int next_timer_us;
126 unsigned int predicted_us;
128 unsigned int correction_factor[BUCKETS];
129 unsigned int intervals[INTERVALS];
134 #define LOAD_INT(x) ((x) >> FSHIFT)
135 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
137 static inline int get_loadavg(unsigned long load)
139 return LOAD_INT(load) * 10 + LOAD_FRAC(load) / 10;
142 static inline int which_bucket(unsigned int duration, unsigned long nr_iowaiters)
147 * We keep two groups of stats; one with no
148 * IO pending, one without.
149 * This allows us to calculate
161 if (duration < 10000)
163 if (duration < 100000)
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.
175 static inline int performance_multiplier(unsigned long nr_iowaiters, unsigned long load)
179 /* for higher loadavg, we are more reluctant */
181 mult += 2 * get_loadavg(load);
183 /* for IO wait tasks (per cpu!) we add 5x each */
184 mult += 10 * nr_iowaiters;
189 static DEFINE_PER_CPU(struct menu_device, menu_devices);
191 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
194 * Try detecting repeating patterns by keeping track of the last 8
195 * intervals, and checking if the standard deviation of that set
196 * of points is below a threshold. If it is... then use the
197 * average of these 8 points as the estimated value.
199 static void get_typical_interval(struct menu_device *data)
202 unsigned int max, thresh;
203 uint64_t avg, stddev;
205 thresh = UINT_MAX; /* Discard outliers above this value */
209 /* First calculate the average of past intervals */
213 for (i = 0; i < INTERVALS; i++) {
214 unsigned int value = data->intervals[i];
215 if (value <= thresh) {
222 if (divisor == INTERVALS)
223 avg >>= INTERVAL_SHIFT;
225 do_div(avg, divisor);
227 /* Then try to determine standard deviation */
229 for (i = 0; i < INTERVALS; i++) {
230 unsigned int value = data->intervals[i];
231 if (value <= thresh) {
232 int64_t diff = value - avg;
233 stddev += diff * diff;
236 if (divisor == INTERVALS)
237 stddev >>= INTERVAL_SHIFT;
239 do_div(stddev, divisor);
242 * The typical interval is obtained when standard deviation is small
243 * or standard deviation is small compared to the average interval.
245 * int_sqrt() formal parameter type is unsigned long. When the
246 * greatest difference to an outlier exceeds ~65 ms * sqrt(divisor)
247 * the resulting squared standard deviation exceeds the input domain
248 * of int_sqrt on platforms where unsigned long is 32 bits in size.
249 * In such case reject the candidate average.
251 * Use this result only if there is no timer to wake us up sooner.
253 if (likely(stddev <= ULONG_MAX)) {
254 stddev = int_sqrt(stddev);
255 if (((avg > stddev * 6) && (divisor * 4 >= INTERVALS * 3))
257 if (data->next_timer_us > avg)
258 data->predicted_us = avg;
264 * If we have outliers to the upside in our distribution, discard
265 * those by setting the threshold to exclude these outliers, then
266 * calculate the average and standard deviation again. Once we get
267 * down to the bottom 3/4 of our samples, stop excluding samples.
269 * This can deal with workloads that have long pauses interspersed
270 * with sporadic activity with a bunch of short pauses.
272 if ((divisor * 4) <= INTERVALS * 3)
280 * menu_select - selects the next idle state to enter
281 * @drv: cpuidle driver containing state data
284 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
286 struct menu_device *data = this_cpu_ptr(&menu_devices);
287 int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY);
289 unsigned int interactivity_req;
290 unsigned long nr_iowaiters, cpu_load;
292 if (data->needs_update) {
293 menu_update(drv, dev);
294 data->needs_update = 0;
297 data->last_state_idx = CPUIDLE_DRIVER_STATE_START - 1;
299 /* Special case when user has set very strict latency requirement */
300 if (unlikely(latency_req == 0))
303 /* determine the expected residency time, round up */
304 data->next_timer_us = ktime_to_us(tick_nohz_get_sleep_length());
306 get_iowait_load(&nr_iowaiters, &cpu_load);
307 data->bucket = which_bucket(data->next_timer_us, nr_iowaiters);
310 * Force the result of multiplication to be 64 bits even if both
311 * operands are 32 bits.
312 * Make sure to round up for half microseconds.
314 data->predicted_us = DIV_ROUND_CLOSEST_ULL((uint64_t)data->next_timer_us *
315 data->correction_factor[data->bucket],
318 get_typical_interval(data);
321 * Performance multiplier defines a minimum predicted idle
322 * duration / latency ratio. Adjust the latency limit if
325 interactivity_req = data->predicted_us / performance_multiplier(nr_iowaiters, cpu_load);
326 if (latency_req > interactivity_req)
327 latency_req = interactivity_req;
330 * We want to default to C1 (hlt), not to busy polling
331 * unless the timer is happening really really soon.
333 if (data->next_timer_us > 5 &&
334 !drv->states[CPUIDLE_DRIVER_STATE_START].disabled &&
335 dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0)
336 data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
339 * Find the idle state with the lowest power while satisfying
342 for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
343 struct cpuidle_state *s = &drv->states[i];
344 struct cpuidle_state_usage *su = &dev->states_usage[i];
346 if (s->disabled || su->disable)
348 if (s->target_residency > data->predicted_us)
350 if (s->exit_latency > latency_req)
353 data->last_state_idx = i;
356 return data->last_state_idx;
360 * menu_reflect - records that data structures need update
362 * @index: the index of actual entered state
364 * NOTE: it's important to be fast here because this operation will add to
365 * the overall exit latency.
367 static void menu_reflect(struct cpuidle_device *dev, int index)
369 struct menu_device *data = this_cpu_ptr(&menu_devices);
371 data->last_state_idx = index;
372 data->needs_update = 1;
376 * menu_update - attempts to guess what happened after entry
377 * @drv: cpuidle driver containing state data
380 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
382 struct menu_device *data = this_cpu_ptr(&menu_devices);
383 int last_idx = data->last_state_idx;
384 struct cpuidle_state *target = &drv->states[last_idx];
385 unsigned int measured_us;
386 unsigned int new_factor;
389 * Try to figure out how much time passed between entry to low
390 * power state and occurrence of the wakeup event.
392 * If the entered idle state didn't support residency measurements,
393 * we use them anyway if they are short, and if long,
394 * truncate to the whole expected time.
396 * Any measured amount of time will include the exit latency.
397 * Since we are interested in when the wakeup begun, not when it
398 * was completed, we must subtract the exit latency. However, if
399 * the measured amount of time is less than the exit latency,
400 * assume the state was never reached and the exit latency is 0.
404 measured_us = cpuidle_get_last_residency(dev);
406 /* Deduct exit latency */
407 if (measured_us > target->exit_latency)
408 measured_us -= target->exit_latency;
410 /* Make sure our coefficients do not exceed unity */
411 if (measured_us > data->next_timer_us)
412 measured_us = data->next_timer_us;
414 /* Update our correction ratio */
415 new_factor = data->correction_factor[data->bucket];
416 new_factor -= new_factor / DECAY;
418 if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING)
419 new_factor += RESOLUTION * measured_us / data->next_timer_us;
422 * we were idle so long that we count it as a perfect
425 new_factor += RESOLUTION;
428 * We don't want 0 as factor; we always want at least
429 * a tiny bit of estimated time. Fortunately, due to rounding,
430 * new_factor will stay nonzero regardless of measured_us values
431 * and the compiler can eliminate this test as long as DECAY > 1.
433 if (DECAY == 1 && unlikely(new_factor == 0))
436 data->correction_factor[data->bucket] = new_factor;
438 /* update the repeating-pattern data */
439 data->intervals[data->interval_ptr++] = measured_us;
440 if (data->interval_ptr >= INTERVALS)
441 data->interval_ptr = 0;
445 * menu_enable_device - scans a CPU's states and does setup
446 * @drv: cpuidle driver
449 static int menu_enable_device(struct cpuidle_driver *drv,
450 struct cpuidle_device *dev)
452 struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
455 memset(data, 0, sizeof(struct menu_device));
458 * if the correction factor is 0 (eg first time init or cpu hotplug
459 * etc), we actually want to start out with a unity factor.
461 for(i = 0; i < BUCKETS; i++)
462 data->correction_factor[i] = RESOLUTION * DECAY;
467 static struct cpuidle_governor menu_governor = {
470 .enable = menu_enable_device,
471 .select = menu_select,
472 .reflect = menu_reflect,
473 .owner = THIS_MODULE,
477 * init_menu - initializes the governor
479 static int __init init_menu(void)
481 return cpuidle_register_governor(&menu_governor);
484 postcore_initcall(init_menu);