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.
35 #define RESOLUTION 1024
37 #define MAX_INTERESTING 50000
41 * Concepts and ideas behind the menu governor
43 * For the menu governor, there are 3 decision factors for picking a C
45 * 1) Energy break even point
46 * 2) Performance impact
47 * 3) Latency tolerance (from pmqos infrastructure)
48 * These these three factors are treated independently.
50 * Energy break even point
51 * -----------------------
52 * C state entry and exit have an energy cost, and a certain amount of time in
53 * the C state is required to actually break even on this cost. CPUIDLE
54 * provides us this duration in the "target_residency" field. So all that we
55 * need is a good prediction of how long we'll be idle. Like the traditional
56 * menu governor, we start with the actual known "next timer event" time.
58 * Since there are other source of wakeups (interrupts for example) than
59 * the next timer event, this estimation is rather optimistic. To get a
60 * more realistic estimate, a correction factor is applied to the estimate,
61 * that is based on historic behavior. For example, if in the past the actual
62 * duration always was 50% of the next timer tick, the correction factor will
65 * menu uses a running average for this correction factor, however it uses a
66 * set of factors, not just a single factor. This stems from the realization
67 * that the ratio is dependent on the order of magnitude of the expected
68 * duration; if we expect 500 milliseconds of idle time the likelihood of
69 * getting an interrupt very early is much higher than if we expect 50 micro
70 * seconds of idle time. A second independent factor that has big impact on
71 * the actual factor is if there is (disk) IO outstanding or not.
72 * (as a special twist, we consider every sleep longer than 50 milliseconds
73 * as perfect; there are no power gains for sleeping longer than this)
75 * For these two reasons we keep an array of 12 independent factors, that gets
76 * indexed based on the magnitude of the expected duration as well as the
77 * "is IO outstanding" property.
79 * Repeatable-interval-detector
80 * ----------------------------
81 * There are some cases where "next timer" is a completely unusable predictor:
82 * Those cases where the interval is fixed, for example due to hardware
83 * interrupt mitigation, but also due to fixed transfer rate devices such as
85 * For this, we use a different predictor: We track the duration of the last 8
86 * intervals and if the stand deviation of these 8 intervals is below a
87 * threshold value, we use the average of these intervals as prediction.
89 * Limiting Performance Impact
90 * ---------------------------
91 * C states, especially those with large exit latencies, can have a real
92 * noticeable impact on workloads, which is not acceptable for most sysadmins,
93 * and in addition, less performance has a power price of its own.
95 * As a general rule of thumb, menu assumes that the following heuristic
97 * The busier the system, the less impact of C states is acceptable
99 * This rule-of-thumb is implemented using a performance-multiplier:
100 * If the exit latency times the performance multiplier is longer than
101 * the predicted duration, the C state is not considered a candidate
102 * for selection due to a too high performance impact. So the higher
103 * this multiplier is, the longer we need to be idle to pick a deep C
104 * state, and thus the less likely a busy CPU will hit such a deep
107 * Two factors are used in determing this multiplier:
108 * a value of 10 is added for each point of "per cpu load average" we have.
109 * a value of 5 points is added for each process that is waiting for
111 * (these values are experimentally determined)
113 * The load average factor gives a longer term (few seconds) input to the
114 * decision, while the iowait value gives a cpu local instantanious input.
115 * The iowait factor may look low, but realize that this is also already
116 * represented in the system load average.
124 unsigned int next_timer_us;
125 unsigned int predicted_us;
127 unsigned int correction_factor[BUCKETS];
128 unsigned int intervals[INTERVALS];
133 #define LOAD_INT(x) ((x) >> FSHIFT)
134 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
136 static int get_loadavg(void)
138 unsigned long this = this_cpu_load();
141 return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10;
144 static inline int which_bucket(unsigned int duration)
149 * We keep two groups of stats; one with no
150 * IO pending, one without.
151 * This allows us to calculate
154 if (nr_iowait_cpu(smp_processor_id()))
163 if (duration < 10000)
165 if (duration < 100000)
171 * Return a multiplier for the exit latency that is intended
172 * to take performance requirements into account.
173 * The more performance critical we estimate the system
174 * to be, the higher this multiplier, and thus the higher
175 * the barrier to go to an expensive C state.
177 static inline int performance_multiplier(void)
181 /* for higher loadavg, we are more reluctant */
183 mult += 2 * get_loadavg();
185 /* for IO wait tasks (per cpu!) we add 5x each */
186 mult += 10 * nr_iowait_cpu(smp_processor_id());
191 static DEFINE_PER_CPU(struct menu_device, menu_devices);
193 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
195 /* This implements DIV_ROUND_CLOSEST but avoids 64 bit division */
196 static u64 div_round64(u64 dividend, u32 divisor)
198 return div_u64(dividend + (divisor / 2), divisor);
202 * Try detecting repeating patterns by keeping track of the last 8
203 * intervals, and checking if the standard deviation of that set
204 * of points is below a threshold. If it is... then use the
205 * average of these 8 points as the estimated value.
207 static void get_typical_interval(struct menu_device *data)
210 unsigned int max, thresh;
211 uint64_t avg, stddev;
213 thresh = UINT_MAX; /* Discard outliers above this value */
217 /* First calculate the average of past intervals */
221 for (i = 0; i < INTERVALS; i++) {
222 unsigned int value = data->intervals[i];
223 if (value <= thresh) {
230 do_div(avg, divisor);
232 /* Then try to determine standard deviation */
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;
241 do_div(stddev, divisor);
243 * The typical interval is obtained when standard deviation is small
244 * or standard deviation is small compared to the average interval.
246 * int_sqrt() formal parameter type is unsigned long. When the
247 * greatest difference to an outlier exceeds ~65 ms * sqrt(divisor)
248 * the resulting squared standard deviation exceeds the input domain
249 * of int_sqrt on platforms where unsigned long is 32 bits in size.
250 * In such case reject the candidate average.
252 * Use this result only if there is no timer to wake us up sooner.
254 if (likely(stddev <= ULONG_MAX)) {
255 stddev = int_sqrt(stddev);
256 if (((avg > stddev * 6) && (divisor * 4 >= INTERVALS * 3))
258 if (data->next_timer_us > avg)
259 data->predicted_us = avg;
265 * If we have outliers to the upside in our distribution, discard
266 * those by setting the threshold to exclude these outliers, then
267 * calculate the average and standard deviation again. Once we get
268 * down to the bottom 3/4 of our samples, stop excluding samples.
270 * This can deal with workloads that have long pauses interspersed
271 * with sporadic activity with a bunch of short pauses.
273 if ((divisor * 4) <= INTERVALS * 3)
281 * menu_select - selects the next idle state to enter
282 * @drv: cpuidle driver containing state data
285 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
287 struct menu_device *data = &__get_cpu_var(menu_devices);
288 int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY);
290 unsigned int interactivity_req;
293 if (data->needs_update) {
294 menu_update(drv, dev);
295 data->needs_update = 0;
298 data->last_state_idx = CPUIDLE_DRIVER_STATE_START - 1;
300 /* Special case when user has set very strict latency requirement */
301 if (unlikely(latency_req == 0))
304 /* determine the expected residency time, round up */
305 t = ktime_to_timespec(tick_nohz_get_sleep_length());
306 data->next_timer_us =
307 t.tv_sec * USEC_PER_SEC + t.tv_nsec / NSEC_PER_USEC;
310 data->bucket = which_bucket(data->next_timer_us);
313 * Force the result of multiplication to be 64 bits even if both
314 * operands are 32 bits.
315 * Make sure to round up for half microseconds.
317 data->predicted_us = div_round64((uint64_t)data->next_timer_us *
318 data->correction_factor[data->bucket],
321 get_typical_interval(data);
324 * Performance multiplier defines a minimum predicted idle
325 * duration / latency ratio. Adjust the latency limit if
328 interactivity_req = data->predicted_us / performance_multiplier();
329 if (latency_req > interactivity_req)
330 latency_req = interactivity_req;
333 * We want to default to C1 (hlt), not to busy polling
334 * unless the timer is happening really really soon.
336 if (data->next_timer_us > 5 &&
337 !drv->states[CPUIDLE_DRIVER_STATE_START].disabled &&
338 dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable == 0)
339 data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
342 * Find the idle state with the lowest power while satisfying
345 for (i = CPUIDLE_DRIVER_STATE_START; i < drv->state_count; i++) {
346 struct cpuidle_state *s = &drv->states[i];
347 struct cpuidle_state_usage *su = &dev->states_usage[i];
349 if (s->disabled || su->disable)
351 if (s->target_residency > data->predicted_us)
353 if (s->exit_latency > latency_req)
356 data->last_state_idx = i;
359 return data->last_state_idx;
363 * menu_reflect - records that data structures need update
365 * @index: the index of actual entered state
367 * NOTE: it's important to be fast here because this operation will add to
368 * the overall exit latency.
370 static void menu_reflect(struct cpuidle_device *dev, int index)
372 struct menu_device *data = &__get_cpu_var(menu_devices);
373 data->last_state_idx = index;
375 data->needs_update = 1;
379 * menu_update - attempts to guess what happened after entry
380 * @drv: cpuidle driver containing state data
383 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
385 struct menu_device *data = &__get_cpu_var(menu_devices);
386 int last_idx = data->last_state_idx;
387 struct cpuidle_state *target = &drv->states[last_idx];
388 unsigned int measured_us;
389 unsigned int new_factor;
392 * Try to figure out how much time passed between entry to low
393 * power state and occurrence of the wakeup event.
395 * If the entered idle state didn't support residency measurements,
396 * we are basically lost in the dark how much time passed.
397 * As a compromise, assume we slept for the whole expected time.
399 * Any measured amount of time will include the exit latency.
400 * Since we are interested in when the wakeup begun, not when it
401 * was completed, we must substract the exit latency. However, if
402 * the measured amount of time is less than the exit latency,
403 * assume the state was never reached and the exit latency is 0.
405 if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID))) {
406 /* Use timer value as is */
407 measured_us = data->next_timer_us;
410 /* Use measured value */
411 measured_us = cpuidle_get_last_residency(dev);
413 /* Deduct exit latency */
414 if (measured_us > target->exit_latency)
415 measured_us -= target->exit_latency;
417 /* Make sure our coefficients do not exceed unity */
418 if (measured_us > data->next_timer_us)
419 measured_us = data->next_timer_us;
422 /* Update our correction ratio */
423 new_factor = data->correction_factor[data->bucket];
424 new_factor -= new_factor / DECAY;
426 if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING)
427 new_factor += RESOLUTION * measured_us / data->next_timer_us;
430 * we were idle so long that we count it as a perfect
433 new_factor += RESOLUTION;
436 * We don't want 0 as factor; we always want at least
437 * a tiny bit of estimated time. Fortunately, due to rounding,
438 * new_factor will stay nonzero regardless of measured_us values
439 * and the compiler can eliminate this test as long as DECAY > 1.
441 if (DECAY == 1 && unlikely(new_factor == 0))
444 data->correction_factor[data->bucket] = new_factor;
446 /* update the repeating-pattern data */
447 data->intervals[data->interval_ptr++] = measured_us;
448 if (data->interval_ptr >= INTERVALS)
449 data->interval_ptr = 0;
453 * menu_enable_device - scans a CPU's states and does setup
454 * @drv: cpuidle driver
457 static int menu_enable_device(struct cpuidle_driver *drv,
458 struct cpuidle_device *dev)
460 struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
463 memset(data, 0, sizeof(struct menu_device));
466 * if the correction factor is 0 (eg first time init or cpu hotplug
467 * etc), we actually want to start out with a unity factor.
469 for(i = 0; i < BUCKETS; i++)
470 data->correction_factor[i] = RESOLUTION * DECAY;
475 static struct cpuidle_governor menu_governor = {
478 .enable = menu_enable_device,
479 .select = menu_select,
480 .reflect = menu_reflect,
481 .owner = THIS_MODULE,
485 * init_menu - initializes the governor
487 static int __init init_menu(void)
489 return cpuidle_register_governor(&menu_governor);
492 postcore_initcall(init_menu);