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/sched/loadavg.h>
22 #include <linux/math64.h>
23 #include <linux/cpu.h>
26 * Please note when changing the tuning values:
27 * If (MAX_INTERESTING-1) * RESOLUTION > UINT_MAX, the result of
28 * a scaling operation multiplication may overflow on 32 bit platforms.
29 * In that case, #define RESOLUTION as ULL to get 64 bit result:
30 * #define RESOLUTION 1024ULL
32 * The default values do not overflow.
35 #define INTERVAL_SHIFT 3
36 #define INTERVALS (1UL << INTERVAL_SHIFT)
37 #define RESOLUTION 1024
39 #define MAX_INTERESTING 50000
43 * Concepts and ideas behind the menu governor
45 * For the menu governor, there are 3 decision factors for picking a C
47 * 1) Energy break even point
48 * 2) Performance impact
49 * 3) Latency tolerance (from pmqos infrastructure)
50 * These these three factors are treated independently.
52 * Energy break even point
53 * -----------------------
54 * C state entry and exit have an energy cost, and a certain amount of time in
55 * the C state is required to actually break even on this cost. CPUIDLE
56 * provides us this duration in the "target_residency" field. So all that we
57 * need is a good prediction of how long we'll be idle. Like the traditional
58 * menu governor, we start with the actual known "next timer event" time.
60 * Since there are other source of wakeups (interrupts for example) than
61 * the next timer event, this estimation is rather optimistic. To get a
62 * more realistic estimate, a correction factor is applied to the estimate,
63 * that is based on historic behavior. For example, if in the past the actual
64 * duration always was 50% of the next timer tick, the correction factor will
67 * menu uses a running average for this correction factor, however it uses a
68 * set of factors, not just a single factor. This stems from the realization
69 * that the ratio is dependent on the order of magnitude of the expected
70 * duration; if we expect 500 milliseconds of idle time the likelihood of
71 * getting an interrupt very early is much higher than if we expect 50 micro
72 * seconds of idle time. A second independent factor that has big impact on
73 * the actual factor is if there is (disk) IO outstanding or not.
74 * (as a special twist, we consider every sleep longer than 50 milliseconds
75 * as perfect; there are no power gains for sleeping longer than this)
77 * For these two reasons we keep an array of 12 independent factors, that gets
78 * indexed based on the magnitude of the expected duration as well as the
79 * "is IO outstanding" property.
81 * Repeatable-interval-detector
82 * ----------------------------
83 * There are some cases where "next timer" is a completely unusable predictor:
84 * Those cases where the interval is fixed, for example due to hardware
85 * interrupt mitigation, but also due to fixed transfer rate devices such as
87 * For this, we use a different predictor: We track the duration of the last 8
88 * intervals and if the stand deviation of these 8 intervals is below a
89 * threshold value, we use the average of these intervals as prediction.
91 * Limiting Performance Impact
92 * ---------------------------
93 * C states, especially those with large exit latencies, can have a real
94 * noticeable impact on workloads, which is not acceptable for most sysadmins,
95 * and in addition, less performance has a power price of its own.
97 * As a general rule of thumb, menu assumes that the following heuristic
99 * The busier the system, the less impact of C states is acceptable
101 * This rule-of-thumb is implemented using a performance-multiplier:
102 * If the exit latency times the performance multiplier is longer than
103 * the predicted duration, the C state is not considered a candidate
104 * for selection due to a too high performance impact. So the higher
105 * this multiplier is, the longer we need to be idle to pick a deep C
106 * state, and thus the less likely a busy CPU will hit such a deep
109 * Two factors are used in determing this multiplier:
110 * a value of 10 is added for each point of "per cpu load average" we have.
111 * a value of 5 points is added for each process that is waiting for
113 * (these values are experimentally determined)
115 * The load average factor gives a longer term (few seconds) input to the
116 * decision, while the iowait value gives a cpu local instantanious input.
117 * The iowait factor may look low, but realize that this is also already
118 * represented in the system load average.
126 unsigned int next_timer_us;
127 unsigned int predicted_us;
129 unsigned int correction_factor[BUCKETS];
130 unsigned int intervals[INTERVALS];
135 #define LOAD_INT(x) ((x) >> FSHIFT)
136 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
138 static inline int get_loadavg(unsigned long load)
140 return LOAD_INT(load) * 10 + LOAD_FRAC(load) / 10;
143 static inline int which_bucket(unsigned int duration, unsigned long nr_iowaiters)
148 * We keep two groups of stats; one with no
149 * IO pending, one without.
150 * This allows us to calculate
162 if (duration < 10000)
164 if (duration < 100000)
170 * Return a multiplier for the exit latency that is intended
171 * to take performance requirements into account.
172 * The more performance critical we estimate the system
173 * to be, the higher this multiplier, and thus the higher
174 * the barrier to go to an expensive C state.
176 static inline int performance_multiplier(unsigned long nr_iowaiters, unsigned long load)
180 /* for higher loadavg, we are more reluctant */
182 mult += 2 * get_loadavg(load);
184 /* for IO wait tasks (per cpu!) we add 5x each */
185 mult += 10 * nr_iowaiters;
190 static DEFINE_PER_CPU(struct menu_device, menu_devices);
192 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev);
195 * Try detecting repeating patterns by keeping track of the last 8
196 * intervals, and checking if the standard deviation of that set
197 * of points is below a threshold. If it is... then use the
198 * average of these 8 points as the estimated value.
200 static unsigned int get_typical_interval(struct menu_device *data)
203 unsigned int max, thresh, avg;
204 uint64_t sum, variance;
206 thresh = UINT_MAX; /* Discard outliers above this value */
210 /* First calculate the average of past intervals */
214 for (i = 0; i < INTERVALS; i++) {
215 unsigned int value = data->intervals[i];
216 if (value <= thresh) {
223 if (divisor == INTERVALS)
224 avg = sum >> INTERVAL_SHIFT;
226 avg = div_u64(sum, divisor);
228 /* Then try to determine variance */
230 for (i = 0; i < INTERVALS; i++) {
231 unsigned int value = data->intervals[i];
232 if (value <= thresh) {
233 int64_t diff = (int64_t)value - avg;
234 variance += diff * diff;
237 if (divisor == INTERVALS)
238 variance >>= INTERVAL_SHIFT;
240 do_div(variance, divisor);
243 * The typical interval is obtained when standard deviation is
244 * small (stddev <= 20 us, variance <= 400 us^2) or standard
245 * deviation is small compared to the average interval (avg >
246 * 6*stddev, avg^2 > 36*variance). The average is smaller than
247 * UINT_MAX aka U32_MAX, so computing its square does not
248 * overflow a u64. We simply reject this candidate average if
249 * the standard deviation is greater than 715 s (which is
252 * Use this result only if there is no timer to wake us up sooner.
254 if (likely(variance <= U64_MAX/36)) {
255 if ((((u64)avg*avg > variance*36) && (divisor * 4 >= INTERVALS * 3))
256 || variance <= 400) {
262 * If we have outliers to the upside in our distribution, discard
263 * those by setting the threshold to exclude these outliers, then
264 * calculate the average and standard deviation again. Once we get
265 * down to the bottom 3/4 of our samples, stop excluding samples.
267 * This can deal with workloads that have long pauses interspersed
268 * with sporadic activity with a bunch of short pauses.
270 if ((divisor * 4) <= INTERVALS * 3)
278 * menu_select - selects the next idle state to enter
279 * @drv: cpuidle driver containing state data
282 static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev)
284 struct menu_device *data = this_cpu_ptr(&menu_devices);
285 struct device *device = get_cpu_device(dev->cpu);
286 int latency_req = pm_qos_request(PM_QOS_CPU_DMA_LATENCY);
288 unsigned int interactivity_req;
289 unsigned int expected_interval;
290 unsigned long nr_iowaiters, cpu_load;
291 int resume_latency = dev_pm_qos_read_value(device);
293 if (data->needs_update) {
294 menu_update(drv, dev);
295 data->needs_update = 0;
298 /* resume_latency is 0 means no restriction */
299 if (resume_latency && resume_latency < latency_req)
300 latency_req = resume_latency;
302 /* Special case when user has set very strict latency requirement */
303 if (unlikely(latency_req == 0))
306 /* determine the expected residency time, round up */
307 data->next_timer_us = ktime_to_us(tick_nohz_get_sleep_length());
309 get_iowait_load(&nr_iowaiters, &cpu_load);
310 data->bucket = which_bucket(data->next_timer_us, nr_iowaiters);
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_ROUND_CLOSEST_ULL((uint64_t)data->next_timer_us *
318 data->correction_factor[data->bucket],
321 expected_interval = get_typical_interval(data);
322 expected_interval = min(expected_interval, data->next_timer_us);
324 if (CPUIDLE_DRIVER_STATE_START > 0) {
325 struct cpuidle_state *s = &drv->states[CPUIDLE_DRIVER_STATE_START];
326 unsigned int polling_threshold;
329 * We want to default to C1 (hlt), not to busy polling
330 * unless the timer is happening really really soon, or
331 * C1's exit latency exceeds the user configured limit.
333 polling_threshold = max_t(unsigned int, 20, s->target_residency);
334 if (data->next_timer_us > polling_threshold &&
335 latency_req > s->exit_latency && !s->disabled &&
336 !dev->states_usage[CPUIDLE_DRIVER_STATE_START].disable)
337 data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
339 data->last_state_idx = CPUIDLE_DRIVER_STATE_START - 1;
341 data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
345 * Use the lowest expected idle interval to pick the idle state.
347 data->predicted_us = min(data->predicted_us, expected_interval);
350 * Use the performance multiplier and the user-configurable
351 * latency_req to determine the maximum exit latency.
353 interactivity_req = data->predicted_us / performance_multiplier(nr_iowaiters, cpu_load);
354 if (latency_req > interactivity_req)
355 latency_req = interactivity_req;
358 * Find the idle state with the lowest power while satisfying
361 for (i = data->last_state_idx + 1; i < drv->state_count; i++) {
362 struct cpuidle_state *s = &drv->states[i];
363 struct cpuidle_state_usage *su = &dev->states_usage[i];
365 if (s->disabled || su->disable)
367 if (s->target_residency > data->predicted_us)
369 if (s->exit_latency > latency_req)
372 data->last_state_idx = i;
375 return data->last_state_idx;
379 * menu_reflect - records that data structures need update
381 * @index: the index of actual entered state
383 * NOTE: it's important to be fast here because this operation will add to
384 * the overall exit latency.
386 static void menu_reflect(struct cpuidle_device *dev, int index)
388 struct menu_device *data = this_cpu_ptr(&menu_devices);
390 data->last_state_idx = index;
391 data->needs_update = 1;
395 * menu_update - attempts to guess what happened after entry
396 * @drv: cpuidle driver containing state data
399 static void menu_update(struct cpuidle_driver *drv, struct cpuidle_device *dev)
401 struct menu_device *data = this_cpu_ptr(&menu_devices);
402 int last_idx = data->last_state_idx;
403 struct cpuidle_state *target = &drv->states[last_idx];
404 unsigned int measured_us;
405 unsigned int new_factor;
408 * Try to figure out how much time passed between entry to low
409 * power state and occurrence of the wakeup event.
411 * If the entered idle state didn't support residency measurements,
412 * we use them anyway if they are short, and if long,
413 * truncate to the whole expected time.
415 * Any measured amount of time will include the exit latency.
416 * Since we are interested in when the wakeup begun, not when it
417 * was completed, we must subtract the exit latency. However, if
418 * the measured amount of time is less than the exit latency,
419 * assume the state was never reached and the exit latency is 0.
423 measured_us = cpuidle_get_last_residency(dev);
425 /* Deduct exit latency */
426 if (measured_us > 2 * target->exit_latency)
427 measured_us -= target->exit_latency;
431 /* Make sure our coefficients do not exceed unity */
432 if (measured_us > data->next_timer_us)
433 measured_us = data->next_timer_us;
435 /* Update our correction ratio */
436 new_factor = data->correction_factor[data->bucket];
437 new_factor -= new_factor / DECAY;
439 if (data->next_timer_us > 0 && measured_us < MAX_INTERESTING)
440 new_factor += RESOLUTION * measured_us / data->next_timer_us;
443 * we were idle so long that we count it as a perfect
446 new_factor += RESOLUTION;
449 * We don't want 0 as factor; we always want at least
450 * a tiny bit of estimated time. Fortunately, due to rounding,
451 * new_factor will stay nonzero regardless of measured_us values
452 * and the compiler can eliminate this test as long as DECAY > 1.
454 if (DECAY == 1 && unlikely(new_factor == 0))
457 data->correction_factor[data->bucket] = new_factor;
459 /* update the repeating-pattern data */
460 data->intervals[data->interval_ptr++] = measured_us;
461 if (data->interval_ptr >= INTERVALS)
462 data->interval_ptr = 0;
466 * menu_enable_device - scans a CPU's states and does setup
467 * @drv: cpuidle driver
470 static int menu_enable_device(struct cpuidle_driver *drv,
471 struct cpuidle_device *dev)
473 struct menu_device *data = &per_cpu(menu_devices, dev->cpu);
476 memset(data, 0, sizeof(struct menu_device));
479 * if the correction factor is 0 (eg first time init or cpu hotplug
480 * etc), we actually want to start out with a unity factor.
482 for(i = 0; i < BUCKETS; i++)
483 data->correction_factor[i] = RESOLUTION * DECAY;
488 static struct cpuidle_governor menu_governor = {
491 .enable = menu_enable_device,
492 .select = menu_select,
493 .reflect = menu_reflect,
497 * init_menu - initializes the governor
499 static int __init init_menu(void)
501 return cpuidle_register_governor(&menu_governor);
504 postcore_initcall(init_menu);