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>
24 * Please note when changing the tuning values:
25 * If (MAX_INTERESTING-1) * RESOLUTION > UINT_MAX, the result of
26 * a scaling operation multiplication may overflow on 32 bit platforms.
27 * In that case, #define RESOLUTION as ULL to get 64 bit result:
28 * #define RESOLUTION 1024ULL
30 * The default values do not overflow.
33 #define INTERVAL_SHIFT 3
34 #define INTERVALS (1UL << INTERVAL_SHIFT)
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 inline int get_loadavg(unsigned long load
)
138 return LOAD_INT(load
) * 10 + LOAD_FRAC(load
) / 10;
141 static inline int which_bucket(unsigned int duration
, unsigned long nr_iowaiters
)
146 * We keep two groups of stats; one with no
147 * IO pending, one without.
148 * This allows us to calculate
160 if (duration
< 10000)
162 if (duration
< 100000)
168 * Return a multiplier for the exit latency that is intended
169 * to take performance requirements into account.
170 * The more performance critical we estimate the system
171 * to be, the higher this multiplier, and thus the higher
172 * the barrier to go to an expensive C state.
174 static inline int performance_multiplier(unsigned long nr_iowaiters
, unsigned long load
)
178 /* for higher loadavg, we are more reluctant */
180 mult
+= 2 * get_loadavg(load
);
182 /* for IO wait tasks (per cpu!) we add 5x each */
183 mult
+= 10 * nr_iowaiters
;
188 static DEFINE_PER_CPU(struct menu_device
, menu_devices
);
190 static void menu_update(struct cpuidle_driver
*drv
, struct cpuidle_device
*dev
);
193 * Try detecting repeating patterns by keeping track of the last 8
194 * intervals, and checking if the standard deviation of that set
195 * of points is below a threshold. If it is... then use the
196 * average of these 8 points as the estimated value.
198 static unsigned int get_typical_interval(struct menu_device
*data
)
201 unsigned int max
, thresh
, avg
;
202 uint64_t sum
, variance
;
204 thresh
= UINT_MAX
; /* Discard outliers above this value */
208 /* First calculate the average of past intervals */
212 for (i
= 0; i
< INTERVALS
; i
++) {
213 unsigned int value
= data
->intervals
[i
];
214 if (value
<= thresh
) {
221 if (divisor
== INTERVALS
)
222 avg
= sum
>> INTERVAL_SHIFT
;
224 avg
= div_u64(sum
, divisor
);
226 /* Then try to determine variance */
228 for (i
= 0; i
< INTERVALS
; i
++) {
229 unsigned int value
= data
->intervals
[i
];
230 if (value
<= thresh
) {
231 int64_t diff
= (int64_t)value
- avg
;
232 variance
+= diff
* diff
;
235 if (divisor
== INTERVALS
)
236 variance
>>= INTERVAL_SHIFT
;
238 do_div(variance
, divisor
);
241 * The typical interval is obtained when standard deviation is
242 * small (stddev <= 20 us, variance <= 400 us^2) or standard
243 * deviation is small compared to the average interval (avg >
244 * 6*stddev, avg^2 > 36*variance). The average is smaller than
245 * UINT_MAX aka U32_MAX, so computing its square does not
246 * overflow a u64. We simply reject this candidate average if
247 * the standard deviation is greater than 715 s (which is
250 * Use this result only if there is no timer to wake us up sooner.
252 if (likely(variance
<= U64_MAX
/36)) {
253 if ((((u64
)avg
*avg
> variance
*36) && (divisor
* 4 >= INTERVALS
* 3))
254 || variance
<= 400) {
260 * If we have outliers to the upside in our distribution, discard
261 * those by setting the threshold to exclude these outliers, then
262 * calculate the average and standard deviation again. Once we get
263 * down to the bottom 3/4 of our samples, stop excluding samples.
265 * This can deal with workloads that have long pauses interspersed
266 * with sporadic activity with a bunch of short pauses.
268 if ((divisor
* 4) <= INTERVALS
* 3)
276 * menu_select - selects the next idle state to enter
277 * @drv: cpuidle driver containing state data
280 static int menu_select(struct cpuidle_driver
*drv
, struct cpuidle_device
*dev
)
282 struct menu_device
*data
= this_cpu_ptr(&menu_devices
);
283 int latency_req
= pm_qos_request(PM_QOS_CPU_DMA_LATENCY
);
285 unsigned int interactivity_req
;
286 unsigned int expected_interval
;
287 unsigned long nr_iowaiters
, cpu_load
;
289 if (data
->needs_update
) {
290 menu_update(drv
, dev
);
291 data
->needs_update
= 0;
294 /* Special case when user has set very strict latency requirement */
295 if (unlikely(latency_req
== 0))
298 /* determine the expected residency time, round up */
299 data
->next_timer_us
= ktime_to_us(tick_nohz_get_sleep_length());
301 get_iowait_load(&nr_iowaiters
, &cpu_load
);
302 data
->bucket
= which_bucket(data
->next_timer_us
, nr_iowaiters
);
305 * Force the result of multiplication to be 64 bits even if both
306 * operands are 32 bits.
307 * Make sure to round up for half microseconds.
309 data
->predicted_us
= DIV_ROUND_CLOSEST_ULL((uint64_t)data
->next_timer_us
*
310 data
->correction_factor
[data
->bucket
],
313 expected_interval
= get_typical_interval(data
);
314 expected_interval
= min(expected_interval
, data
->next_timer_us
);
316 if (CPUIDLE_DRIVER_STATE_START
> 0) {
317 struct cpuidle_state
*s
= &drv
->states
[CPUIDLE_DRIVER_STATE_START
];
318 unsigned int polling_threshold
;
321 * We want to default to C1 (hlt), not to busy polling
322 * unless the timer is happening really really soon, or
323 * C1's exit latency exceeds the user configured limit.
325 polling_threshold
= max_t(unsigned int, 20, s
->target_residency
);
326 if (data
->next_timer_us
> polling_threshold
&&
327 latency_req
> s
->exit_latency
&& !s
->disabled
&&
328 !dev
->states_usage
[CPUIDLE_DRIVER_STATE_START
].disable
)
329 data
->last_state_idx
= CPUIDLE_DRIVER_STATE_START
;
331 data
->last_state_idx
= CPUIDLE_DRIVER_STATE_START
- 1;
333 data
->last_state_idx
= CPUIDLE_DRIVER_STATE_START
;
337 * Use the lowest expected idle interval to pick the idle state.
339 data
->predicted_us
= min(data
->predicted_us
, expected_interval
);
342 * Use the performance multiplier and the user-configurable
343 * latency_req to determine the maximum exit latency.
345 interactivity_req
= data
->predicted_us
/ performance_multiplier(nr_iowaiters
, cpu_load
);
346 if (latency_req
> interactivity_req
)
347 latency_req
= interactivity_req
;
350 * Find the idle state with the lowest power while satisfying
353 for (i
= data
->last_state_idx
+ 1; i
< drv
->state_count
; i
++) {
354 struct cpuidle_state
*s
= &drv
->states
[i
];
355 struct cpuidle_state_usage
*su
= &dev
->states_usage
[i
];
357 if (s
->disabled
|| su
->disable
)
359 if (s
->target_residency
> data
->predicted_us
)
361 if (s
->exit_latency
> latency_req
)
364 data
->last_state_idx
= i
;
367 return data
->last_state_idx
;
371 * menu_reflect - records that data structures need update
373 * @index: the index of actual entered state
375 * NOTE: it's important to be fast here because this operation will add to
376 * the overall exit latency.
378 static void menu_reflect(struct cpuidle_device
*dev
, int index
)
380 struct menu_device
*data
= this_cpu_ptr(&menu_devices
);
382 data
->last_state_idx
= index
;
383 data
->needs_update
= 1;
387 * menu_update - attempts to guess what happened after entry
388 * @drv: cpuidle driver containing state data
391 static void menu_update(struct cpuidle_driver
*drv
, struct cpuidle_device
*dev
)
393 struct menu_device
*data
= this_cpu_ptr(&menu_devices
);
394 int last_idx
= data
->last_state_idx
;
395 struct cpuidle_state
*target
= &drv
->states
[last_idx
];
396 unsigned int measured_us
;
397 unsigned int new_factor
;
400 * Try to figure out how much time passed between entry to low
401 * power state and occurrence of the wakeup event.
403 * If the entered idle state didn't support residency measurements,
404 * we use them anyway if they are short, and if long,
405 * truncate to the whole expected time.
407 * Any measured amount of time will include the exit latency.
408 * Since we are interested in when the wakeup begun, not when it
409 * was completed, we must subtract the exit latency. However, if
410 * the measured amount of time is less than the exit latency,
411 * assume the state was never reached and the exit latency is 0.
415 measured_us
= cpuidle_get_last_residency(dev
);
417 /* Deduct exit latency */
418 if (measured_us
> 2 * target
->exit_latency
)
419 measured_us
-= target
->exit_latency
;
423 /* Make sure our coefficients do not exceed unity */
424 if (measured_us
> data
->next_timer_us
)
425 measured_us
= data
->next_timer_us
;
427 /* Update our correction ratio */
428 new_factor
= data
->correction_factor
[data
->bucket
];
429 new_factor
-= new_factor
/ DECAY
;
431 if (data
->next_timer_us
> 0 && measured_us
< MAX_INTERESTING
)
432 new_factor
+= RESOLUTION
* measured_us
/ data
->next_timer_us
;
435 * we were idle so long that we count it as a perfect
438 new_factor
+= RESOLUTION
;
441 * We don't want 0 as factor; we always want at least
442 * a tiny bit of estimated time. Fortunately, due to rounding,
443 * new_factor will stay nonzero regardless of measured_us values
444 * and the compiler can eliminate this test as long as DECAY > 1.
446 if (DECAY
== 1 && unlikely(new_factor
== 0))
449 data
->correction_factor
[data
->bucket
] = new_factor
;
451 /* update the repeating-pattern data */
452 data
->intervals
[data
->interval_ptr
++] = measured_us
;
453 if (data
->interval_ptr
>= INTERVALS
)
454 data
->interval_ptr
= 0;
458 * menu_enable_device - scans a CPU's states and does setup
459 * @drv: cpuidle driver
462 static int menu_enable_device(struct cpuidle_driver
*drv
,
463 struct cpuidle_device
*dev
)
465 struct menu_device
*data
= &per_cpu(menu_devices
, dev
->cpu
);
468 memset(data
, 0, sizeof(struct menu_device
));
471 * if the correction factor is 0 (eg first time init or cpu hotplug
472 * etc), we actually want to start out with a unity factor.
474 for(i
= 0; i
< BUCKETS
; i
++)
475 data
->correction_factor
[i
] = RESOLUTION
* DECAY
;
480 static struct cpuidle_governor menu_governor
= {
483 .enable
= menu_enable_device
,
484 .select
= menu_select
,
485 .reflect
= menu_reflect
,
489 * init_menu - initializes the governor
491 static int __init
init_menu(void)
493 return cpuidle_register_governor(&menu_governor
);
496 postcore_initcall(init_menu
);