4 * Kernel load calculations, forked from sched/core.c
7 #include <linux/export.h>
11 unsigned long this_cpu_load(void)
13 struct rq
*this = this_rq();
14 return this->cpu_load
[0];
19 * Global load-average calculations
21 * We take a distributed and async approach to calculating the global load-avg
22 * in order to minimize overhead.
24 * The global load average is an exponentially decaying average of nr_running +
27 * Once every LOAD_FREQ:
30 * for_each_possible_cpu(cpu)
31 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
33 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
35 * Due to a number of reasons the above turns in the mess below:
37 * - for_each_possible_cpu() is prohibitively expensive on machines with
38 * serious number of cpus, therefore we need to take a distributed approach
39 * to calculating nr_active.
41 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
42 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
44 * So assuming nr_active := 0 when we start out -- true per definition, we
45 * can simply take per-cpu deltas and fold those into a global accumulate
46 * to obtain the same result. See calc_load_fold_active().
48 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
49 * across the machine, we assume 10 ticks is sufficient time for every
50 * cpu to have completed this task.
52 * This places an upper-bound on the IRQ-off latency of the machine. Then
53 * again, being late doesn't loose the delta, just wrecks the sample.
55 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
56 * this would add another cross-cpu cacheline miss and atomic operation
57 * to the wakeup path. Instead we increment on whatever cpu the task ran
58 * when it went into uninterruptible state and decrement on whatever cpu
59 * did the wakeup. This means that only the sum of nr_uninterruptible over
60 * all cpus yields the correct result.
62 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
65 /* Variables and functions for calc_load */
66 atomic_long_t calc_load_tasks
;
67 unsigned long calc_load_update
;
68 unsigned long avenrun
[3];
69 EXPORT_SYMBOL(avenrun
); /* should be removed */
72 * get_avenrun - get the load average array
73 * @loads: pointer to dest load array
74 * @offset: offset to add
75 * @shift: shift count to shift the result left
77 * These values are estimates at best, so no need for locking.
79 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
81 loads
[0] = (avenrun
[0] + offset
) << shift
;
82 loads
[1] = (avenrun
[1] + offset
) << shift
;
83 loads
[2] = (avenrun
[2] + offset
) << shift
;
86 long calc_load_fold_active(struct rq
*this_rq
)
88 long nr_active
, delta
= 0;
90 nr_active
= this_rq
->nr_running
;
91 nr_active
+= (long) this_rq
->nr_uninterruptible
;
93 if (nr_active
!= this_rq
->calc_load_active
) {
94 delta
= nr_active
- this_rq
->calc_load_active
;
95 this_rq
->calc_load_active
= nr_active
;
102 * a1 = a0 * e + a * (1 - e)
105 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
108 load
+= active
* (FIXED_1
- exp
);
109 load
+= 1UL << (FSHIFT
- 1);
110 return load
>> FSHIFT
;
113 #ifdef CONFIG_NO_HZ_COMMON
115 * Handle NO_HZ for the global load-average.
117 * Since the above described distributed algorithm to compute the global
118 * load-average relies on per-cpu sampling from the tick, it is affected by
121 * The basic idea is to fold the nr_active delta into a global idle-delta upon
122 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
123 * when we read the global state.
125 * Obviously reality has to ruin such a delightfully simple scheme:
127 * - When we go NO_HZ idle during the window, we can negate our sample
128 * contribution, causing under-accounting.
130 * We avoid this by keeping two idle-delta counters and flipping them
131 * when the window starts, thus separating old and new NO_HZ load.
133 * The only trick is the slight shift in index flip for read vs write.
137 * |-|-----------|-|-----------|-|-----------|-|
138 * r:0 0 1 1 0 0 1 1 0
139 * w:0 1 1 0 0 1 1 0 0
141 * This ensures we'll fold the old idle contribution in this window while
142 * accumlating the new one.
144 * - When we wake up from NO_HZ idle during the window, we push up our
145 * contribution, since we effectively move our sample point to a known
148 * This is solved by pushing the window forward, and thus skipping the
149 * sample, for this cpu (effectively using the idle-delta for this cpu which
150 * was in effect at the time the window opened). This also solves the issue
151 * of having to deal with a cpu having been in NOHZ idle for multiple
152 * LOAD_FREQ intervals.
154 * When making the ILB scale, we should try to pull this in as well.
156 static atomic_long_t calc_load_idle
[2];
157 static int calc_load_idx
;
159 static inline int calc_load_write_idx(void)
161 int idx
= calc_load_idx
;
164 * See calc_global_nohz(), if we observe the new index, we also
165 * need to observe the new update time.
170 * If the folding window started, make sure we start writing in the
173 if (!time_before(jiffies
, calc_load_update
))
179 static inline int calc_load_read_idx(void)
181 return calc_load_idx
& 1;
184 void calc_load_enter_idle(void)
186 struct rq
*this_rq
= this_rq();
190 * We're going into NOHZ mode, if there's any pending delta, fold it
191 * into the pending idle delta.
193 delta
= calc_load_fold_active(this_rq
);
195 int idx
= calc_load_write_idx();
196 atomic_long_add(delta
, &calc_load_idle
[idx
]);
200 void calc_load_exit_idle(void)
202 struct rq
*this_rq
= this_rq();
205 * If we're still before the sample window, we're done.
207 if (time_before(jiffies
, this_rq
->calc_load_update
))
211 * We woke inside or after the sample window, this means we're already
212 * accounted through the nohz accounting, so skip the entire deal and
213 * sync up for the next window.
215 this_rq
->calc_load_update
= calc_load_update
;
216 if (time_before(jiffies
, this_rq
->calc_load_update
+ 10))
217 this_rq
->calc_load_update
+= LOAD_FREQ
;
220 static long calc_load_fold_idle(void)
222 int idx
= calc_load_read_idx();
225 if (atomic_long_read(&calc_load_idle
[idx
]))
226 delta
= atomic_long_xchg(&calc_load_idle
[idx
], 0);
232 * fixed_power_int - compute: x^n, in O(log n) time
234 * @x: base of the power
235 * @frac_bits: fractional bits of @x
236 * @n: power to raise @x to.
238 * By exploiting the relation between the definition of the natural power
239 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
240 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
241 * (where: n_i \elem {0, 1}, the binary vector representing n),
242 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
243 * of course trivially computable in O(log_2 n), the length of our binary
247 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
249 unsigned long result
= 1UL << frac_bits
;
254 result
+= 1UL << (frac_bits
- 1);
255 result
>>= frac_bits
;
261 x
+= 1UL << (frac_bits
- 1);
269 * a1 = a0 * e + a * (1 - e)
271 * a2 = a1 * e + a * (1 - e)
272 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
273 * = a0 * e^2 + a * (1 - e) * (1 + e)
275 * a3 = a2 * e + a * (1 - e)
276 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
277 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
281 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
282 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
283 * = a0 * e^n + a * (1 - e^n)
285 * [1] application of the geometric series:
288 * S_n := \Sum x^i = -------------
292 calc_load_n(unsigned long load
, unsigned long exp
,
293 unsigned long active
, unsigned int n
)
296 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
300 * NO_HZ can leave us missing all per-cpu ticks calling
301 * calc_load_account_active(), but since an idle CPU folds its delta into
302 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
303 * in the pending idle delta if our idle period crossed a load cycle boundary.
305 * Once we've updated the global active value, we need to apply the exponential
306 * weights adjusted to the number of cycles missed.
308 static void calc_global_nohz(void)
310 long delta
, active
, n
;
312 if (!time_before(jiffies
, calc_load_update
+ 10)) {
314 * Catch-up, fold however many we are behind still
316 delta
= jiffies
- calc_load_update
- 10;
317 n
= 1 + (delta
/ LOAD_FREQ
);
319 active
= atomic_long_read(&calc_load_tasks
);
320 active
= active
> 0 ? active
* FIXED_1
: 0;
322 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
323 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
324 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
326 calc_load_update
+= n
* LOAD_FREQ
;
330 * Flip the idle index...
332 * Make sure we first write the new time then flip the index, so that
333 * calc_load_write_idx() will see the new time when it reads the new
334 * index, this avoids a double flip messing things up.
339 #else /* !CONFIG_NO_HZ_COMMON */
341 static inline long calc_load_fold_idle(void) { return 0; }
342 static inline void calc_global_nohz(void) { }
344 #endif /* CONFIG_NO_HZ_COMMON */
347 * calc_load - update the avenrun load estimates 10 ticks after the
348 * CPUs have updated calc_load_tasks.
350 void calc_global_load(unsigned long ticks
)
354 if (time_before(jiffies
, calc_load_update
+ 10))
358 * Fold the 'old' idle-delta to include all NO_HZ cpus.
360 delta
= calc_load_fold_idle();
362 atomic_long_add(delta
, &calc_load_tasks
);
364 active
= atomic_long_read(&calc_load_tasks
);
365 active
= active
> 0 ? active
* FIXED_1
: 0;
367 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
368 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
369 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
371 calc_load_update
+= LOAD_FREQ
;
374 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
380 * Called from update_cpu_load() to periodically update this CPU's
383 static void calc_load_account_active(struct rq
*this_rq
)
387 if (time_before(jiffies
, this_rq
->calc_load_update
))
390 delta
= calc_load_fold_active(this_rq
);
392 atomic_long_add(delta
, &calc_load_tasks
);
394 this_rq
->calc_load_update
+= LOAD_FREQ
;
398 * End of global load-average stuff
402 * The exact cpuload at various idx values, calculated at every tick would be
403 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
405 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
406 * on nth tick when cpu may be busy, then we have:
407 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
408 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
410 * decay_load_missed() below does efficient calculation of
411 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
412 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
414 * The calculation is approximated on a 128 point scale.
415 * degrade_zero_ticks is the number of ticks after which load at any
416 * particular idx is approximated to be zero.
417 * degrade_factor is a precomputed table, a row for each load idx.
418 * Each column corresponds to degradation factor for a power of two ticks,
419 * based on 128 point scale.
421 * row 2, col 3 (=12) says that the degradation at load idx 2 after
422 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
424 * With this power of 2 load factors, we can degrade the load n times
425 * by looking at 1 bits in n and doing as many mult/shift instead of
426 * n mult/shifts needed by the exact degradation.
428 #define DEGRADE_SHIFT 7
429 static const unsigned char
430 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
431 static const unsigned char
432 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
433 {0, 0, 0, 0, 0, 0, 0, 0},
434 {64, 32, 8, 0, 0, 0, 0, 0},
435 {96, 72, 40, 12, 1, 0, 0},
436 {112, 98, 75, 43, 15, 1, 0},
437 {120, 112, 98, 76, 45, 16, 2} };
440 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
441 * would be when CPU is idle and so we just decay the old load without
442 * adding any new load.
445 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
452 if (missed_updates
>= degrade_zero_ticks
[idx
])
456 return load
>> missed_updates
;
458 while (missed_updates
) {
459 if (missed_updates
% 2)
460 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
462 missed_updates
>>= 1;
469 * Update rq->cpu_load[] statistics. This function is usually called every
470 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
471 * every tick. We fix it up based on jiffies.
473 static void __update_cpu_load(struct rq
*this_rq
, unsigned long this_load
,
474 unsigned long pending_updates
)
478 this_rq
->nr_load_updates
++;
480 /* Update our load: */
481 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
482 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
483 unsigned long old_load
, new_load
;
485 /* scale is effectively 1 << i now, and >> i divides by scale */
487 old_load
= this_rq
->cpu_load
[i
];
488 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
489 new_load
= this_load
;
491 * Round up the averaging division if load is increasing. This
492 * prevents us from getting stuck on 9 if the load is 10, for
495 if (new_load
> old_load
)
496 new_load
+= scale
- 1;
498 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
501 sched_avg_update(this_rq
);
505 static inline unsigned long get_rq_runnable_load(struct rq
*rq
)
507 return rq
->cfs
.runnable_load_avg
;
510 static inline unsigned long get_rq_runnable_load(struct rq
*rq
)
512 return rq
->load
.weight
;
516 #ifdef CONFIG_NO_HZ_COMMON
518 * There is no sane way to deal with nohz on smp when using jiffies because the
519 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
520 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
522 * Therefore we cannot use the delta approach from the regular tick since that
523 * would seriously skew the load calculation. However we'll make do for those
524 * updates happening while idle (nohz_idle_balance) or coming out of idle
525 * (tick_nohz_idle_exit).
527 * This means we might still be one tick off for nohz periods.
531 * Called from nohz_idle_balance() to update the load ratings before doing the
534 void update_idle_cpu_load(struct rq
*this_rq
)
536 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
537 unsigned long load
= get_rq_runnable_load(this_rq
);
538 unsigned long pending_updates
;
541 * bail if there's load or we're actually up-to-date.
543 if (load
|| curr_jiffies
== this_rq
->last_load_update_tick
)
546 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
547 this_rq
->last_load_update_tick
= curr_jiffies
;
549 __update_cpu_load(this_rq
, load
, pending_updates
);
553 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
555 void update_cpu_load_nohz(void)
557 struct rq
*this_rq
= this_rq();
558 unsigned long curr_jiffies
= ACCESS_ONCE(jiffies
);
559 unsigned long pending_updates
;
561 if (curr_jiffies
== this_rq
->last_load_update_tick
)
564 raw_spin_lock(&this_rq
->lock
);
565 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
566 if (pending_updates
) {
567 this_rq
->last_load_update_tick
= curr_jiffies
;
569 * We were idle, this means load 0, the current load might be
570 * !0 due to remote wakeups and the sort.
572 __update_cpu_load(this_rq
, 0, pending_updates
);
574 raw_spin_unlock(&this_rq
->lock
);
576 #endif /* CONFIG_NO_HZ */
579 * Called from scheduler_tick()
581 void update_cpu_load_active(struct rq
*this_rq
)
583 unsigned long load
= get_rq_runnable_load(this_rq
);
585 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
587 this_rq
->last_load_update_tick
= jiffies
;
588 __update_cpu_load(this_rq
, load
, 1);
590 calc_load_account_active(this_rq
);