blob: 8ecd552fe4f2229eacabe8d1e9bb1941078fb577 [file] [log] [blame]
Paul Gortmaker45ceebf2013-04-19 15:10:49 -04001/*
2 * kernel/sched/proc.c
3 *
4 * Kernel load calculations, forked from sched/core.c
5 */
6
7#include <linux/export.h>
8
9#include "sched.h"
10
Paul Gortmaker45ceebf2013-04-19 15:10:49 -040011/*
12 * Global load-average calculations
13 *
14 * We take a distributed and async approach to calculating the global load-avg
15 * in order to minimize overhead.
16 *
17 * The global load average is an exponentially decaying average of nr_running +
18 * nr_uninterruptible.
19 *
20 * Once every LOAD_FREQ:
21 *
22 * nr_active = 0;
23 * for_each_possible_cpu(cpu)
24 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
25 *
26 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
27 *
28 * Due to a number of reasons the above turns in the mess below:
29 *
30 * - for_each_possible_cpu() is prohibitively expensive on machines with
31 * serious number of cpus, therefore we need to take a distributed approach
32 * to calculating nr_active.
33 *
34 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
35 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
36 *
37 * So assuming nr_active := 0 when we start out -- true per definition, we
38 * can simply take per-cpu deltas and fold those into a global accumulate
39 * to obtain the same result. See calc_load_fold_active().
40 *
41 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
42 * across the machine, we assume 10 ticks is sufficient time for every
43 * cpu to have completed this task.
44 *
45 * This places an upper-bound on the IRQ-off latency of the machine. Then
46 * again, being late doesn't loose the delta, just wrecks the sample.
47 *
48 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
49 * this would add another cross-cpu cacheline miss and atomic operation
50 * to the wakeup path. Instead we increment on whatever cpu the task ran
51 * when it went into uninterruptible state and decrement on whatever cpu
52 * did the wakeup. This means that only the sum of nr_uninterruptible over
53 * all cpus yields the correct result.
54 *
55 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
56 */
57
58/* Variables and functions for calc_load */
59atomic_long_t calc_load_tasks;
60unsigned long calc_load_update;
61unsigned long avenrun[3];
62EXPORT_SYMBOL(avenrun); /* should be removed */
63
64/**
65 * get_avenrun - get the load average array
66 * @loads: pointer to dest load array
67 * @offset: offset to add
68 * @shift: shift count to shift the result left
69 *
70 * These values are estimates at best, so no need for locking.
71 */
72void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
73{
74 loads[0] = (avenrun[0] + offset) << shift;
75 loads[1] = (avenrun[1] + offset) << shift;
76 loads[2] = (avenrun[2] + offset) << shift;
77}
78
79long calc_load_fold_active(struct rq *this_rq)
80{
81 long nr_active, delta = 0;
82
83 nr_active = this_rq->nr_running;
84 nr_active += (long) this_rq->nr_uninterruptible;
85
86 if (nr_active != this_rq->calc_load_active) {
87 delta = nr_active - this_rq->calc_load_active;
88 this_rq->calc_load_active = nr_active;
89 }
90
91 return delta;
92}
93
94/*
95 * a1 = a0 * e + a * (1 - e)
96 */
97static unsigned long
98calc_load(unsigned long load, unsigned long exp, unsigned long active)
99{
100 load *= exp;
101 load += active * (FIXED_1 - exp);
102 load += 1UL << (FSHIFT - 1);
103 return load >> FSHIFT;
104}
105
106#ifdef CONFIG_NO_HZ_COMMON
107/*
108 * Handle NO_HZ for the global load-average.
109 *
110 * Since the above described distributed algorithm to compute the global
111 * load-average relies on per-cpu sampling from the tick, it is affected by
112 * NO_HZ.
113 *
114 * The basic idea is to fold the nr_active delta into a global idle-delta upon
115 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
116 * when we read the global state.
117 *
118 * Obviously reality has to ruin such a delightfully simple scheme:
119 *
120 * - When we go NO_HZ idle during the window, we can negate our sample
121 * contribution, causing under-accounting.
122 *
123 * We avoid this by keeping two idle-delta counters and flipping them
124 * when the window starts, thus separating old and new NO_HZ load.
125 *
126 * The only trick is the slight shift in index flip for read vs write.
127 *
128 * 0s 5s 10s 15s
129 * +10 +10 +10 +10
130 * |-|-----------|-|-----------|-|-----------|-|
131 * r:0 0 1 1 0 0 1 1 0
132 * w:0 1 1 0 0 1 1 0 0
133 *
134 * This ensures we'll fold the old idle contribution in this window while
135 * accumlating the new one.
136 *
137 * - When we wake up from NO_HZ idle during the window, we push up our
138 * contribution, since we effectively move our sample point to a known
139 * busy state.
140 *
141 * This is solved by pushing the window forward, and thus skipping the
142 * sample, for this cpu (effectively using the idle-delta for this cpu which
143 * was in effect at the time the window opened). This also solves the issue
144 * of having to deal with a cpu having been in NOHZ idle for multiple
145 * LOAD_FREQ intervals.
146 *
147 * When making the ILB scale, we should try to pull this in as well.
148 */
149static atomic_long_t calc_load_idle[2];
150static int calc_load_idx;
151
152static inline int calc_load_write_idx(void)
153{
154 int idx = calc_load_idx;
155
156 /*
157 * See calc_global_nohz(), if we observe the new index, we also
158 * need to observe the new update time.
159 */
160 smp_rmb();
161
162 /*
163 * If the folding window started, make sure we start writing in the
164 * next idle-delta.
165 */
166 if (!time_before(jiffies, calc_load_update))
167 idx++;
168
169 return idx & 1;
170}
171
172static inline int calc_load_read_idx(void)
173{
174 return calc_load_idx & 1;
175}
176
177void calc_load_enter_idle(void)
178{
179 struct rq *this_rq = this_rq();
180 long delta;
181
182 /*
183 * We're going into NOHZ mode, if there's any pending delta, fold it
184 * into the pending idle delta.
185 */
186 delta = calc_load_fold_active(this_rq);
187 if (delta) {
188 int idx = calc_load_write_idx();
189 atomic_long_add(delta, &calc_load_idle[idx]);
190 }
191}
192
193void calc_load_exit_idle(void)
194{
195 struct rq *this_rq = this_rq();
196
197 /*
198 * If we're still before the sample window, we're done.
199 */
200 if (time_before(jiffies, this_rq->calc_load_update))
201 return;
202
203 /*
204 * We woke inside or after the sample window, this means we're already
205 * accounted through the nohz accounting, so skip the entire deal and
206 * sync up for the next window.
207 */
208 this_rq->calc_load_update = calc_load_update;
209 if (time_before(jiffies, this_rq->calc_load_update + 10))
210 this_rq->calc_load_update += LOAD_FREQ;
211}
212
213static long calc_load_fold_idle(void)
214{
215 int idx = calc_load_read_idx();
216 long delta = 0;
217
218 if (atomic_long_read(&calc_load_idle[idx]))
219 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
220
221 return delta;
222}
223
224/**
225 * fixed_power_int - compute: x^n, in O(log n) time
226 *
227 * @x: base of the power
228 * @frac_bits: fractional bits of @x
229 * @n: power to raise @x to.
230 *
231 * By exploiting the relation between the definition of the natural power
232 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
233 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
234 * (where: n_i \elem {0, 1}, the binary vector representing n),
235 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
236 * of course trivially computable in O(log_2 n), the length of our binary
237 * vector.
238 */
239static unsigned long
240fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
241{
242 unsigned long result = 1UL << frac_bits;
243
244 if (n) for (;;) {
245 if (n & 1) {
246 result *= x;
247 result += 1UL << (frac_bits - 1);
248 result >>= frac_bits;
249 }
250 n >>= 1;
251 if (!n)
252 break;
253 x *= x;
254 x += 1UL << (frac_bits - 1);
255 x >>= frac_bits;
256 }
257
258 return result;
259}
260
261/*
262 * a1 = a0 * e + a * (1 - e)
263 *
264 * a2 = a1 * e + a * (1 - e)
265 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
266 * = a0 * e^2 + a * (1 - e) * (1 + e)
267 *
268 * a3 = a2 * e + a * (1 - e)
269 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
270 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
271 *
272 * ...
273 *
274 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
275 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
276 * = a0 * e^n + a * (1 - e^n)
277 *
278 * [1] application of the geometric series:
279 *
280 * n 1 - x^(n+1)
281 * S_n := \Sum x^i = -------------
282 * i=0 1 - x
283 */
284static unsigned long
285calc_load_n(unsigned long load, unsigned long exp,
286 unsigned long active, unsigned int n)
287{
288
289 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
290}
291
292/*
293 * NO_HZ can leave us missing all per-cpu ticks calling
294 * calc_load_account_active(), but since an idle CPU folds its delta into
295 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
296 * in the pending idle delta if our idle period crossed a load cycle boundary.
297 *
298 * Once we've updated the global active value, we need to apply the exponential
299 * weights adjusted to the number of cycles missed.
300 */
301static void calc_global_nohz(void)
302{
303 long delta, active, n;
304
305 if (!time_before(jiffies, calc_load_update + 10)) {
306 /*
307 * Catch-up, fold however many we are behind still
308 */
309 delta = jiffies - calc_load_update - 10;
310 n = 1 + (delta / LOAD_FREQ);
311
312 active = atomic_long_read(&calc_load_tasks);
313 active = active > 0 ? active * FIXED_1 : 0;
314
315 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
316 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
317 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
318
319 calc_load_update += n * LOAD_FREQ;
320 }
321
322 /*
323 * Flip the idle index...
324 *
325 * Make sure we first write the new time then flip the index, so that
326 * calc_load_write_idx() will see the new time when it reads the new
327 * index, this avoids a double flip messing things up.
328 */
329 smp_wmb();
330 calc_load_idx++;
331}
332#else /* !CONFIG_NO_HZ_COMMON */
333
334static inline long calc_load_fold_idle(void) { return 0; }
335static inline void calc_global_nohz(void) { }
336
337#endif /* CONFIG_NO_HZ_COMMON */
338
339/*
340 * calc_load - update the avenrun load estimates 10 ticks after the
341 * CPUs have updated calc_load_tasks.
342 */
343void calc_global_load(unsigned long ticks)
344{
345 long active, delta;
346
347 if (time_before(jiffies, calc_load_update + 10))
348 return;
349
350 /*
351 * Fold the 'old' idle-delta to include all NO_HZ cpus.
352 */
353 delta = calc_load_fold_idle();
354 if (delta)
355 atomic_long_add(delta, &calc_load_tasks);
356
357 active = atomic_long_read(&calc_load_tasks);
358 active = active > 0 ? active * FIXED_1 : 0;
359
360 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
361 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
362 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
363
364 calc_load_update += LOAD_FREQ;
365
366 /*
367 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
368 */
369 calc_global_nohz();
370}
371
372/*
373 * Called from update_cpu_load() to periodically update this CPU's
374 * active count.
375 */
376static void calc_load_account_active(struct rq *this_rq)
377{
378 long delta;
379
380 if (time_before(jiffies, this_rq->calc_load_update))
381 return;
382
383 delta = calc_load_fold_active(this_rq);
384 if (delta)
385 atomic_long_add(delta, &calc_load_tasks);
386
387 this_rq->calc_load_update += LOAD_FREQ;
388}
389
390/*
391 * End of global load-average stuff
392 */
393
394/*
395 * The exact cpuload at various idx values, calculated at every tick would be
396 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
397 *
398 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
399 * on nth tick when cpu may be busy, then we have:
400 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
401 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
402 *
403 * decay_load_missed() below does efficient calculation of
404 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
405 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
406 *
407 * The calculation is approximated on a 128 point scale.
408 * degrade_zero_ticks is the number of ticks after which load at any
409 * particular idx is approximated to be zero.
410 * degrade_factor is a precomputed table, a row for each load idx.
411 * Each column corresponds to degradation factor for a power of two ticks,
412 * based on 128 point scale.
413 * Example:
414 * row 2, col 3 (=12) says that the degradation at load idx 2 after
415 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
416 *
417 * With this power of 2 load factors, we can degrade the load n times
418 * by looking at 1 bits in n and doing as many mult/shift instead of
419 * n mult/shifts needed by the exact degradation.
420 */
421#define DEGRADE_SHIFT 7
422static const unsigned char
423 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
424static const unsigned char
425 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
426 {0, 0, 0, 0, 0, 0, 0, 0},
427 {64, 32, 8, 0, 0, 0, 0, 0},
428 {96, 72, 40, 12, 1, 0, 0},
429 {112, 98, 75, 43, 15, 1, 0},
430 {120, 112, 98, 76, 45, 16, 2} };
431
432/*
433 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
434 * would be when CPU is idle and so we just decay the old load without
435 * adding any new load.
436 */
437static unsigned long
438decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
439{
440 int j = 0;
441
442 if (!missed_updates)
443 return load;
444
445 if (missed_updates >= degrade_zero_ticks[idx])
446 return 0;
447
448 if (idx == 1)
449 return load >> missed_updates;
450
451 while (missed_updates) {
452 if (missed_updates % 2)
453 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
454
455 missed_updates >>= 1;
456 j++;
457 }
458 return load;
459}
460
461/*
462 * Update rq->cpu_load[] statistics. This function is usually called every
463 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
464 * every tick. We fix it up based on jiffies.
465 */
466static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
467 unsigned long pending_updates)
468{
469 int i, scale;
470
471 this_rq->nr_load_updates++;
472
473 /* Update our load: */
474 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
475 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
476 unsigned long old_load, new_load;
477
478 /* scale is effectively 1 << i now, and >> i divides by scale */
479
480 old_load = this_rq->cpu_load[i];
481 old_load = decay_load_missed(old_load, pending_updates - 1, i);
482 new_load = this_load;
483 /*
484 * Round up the averaging division if load is increasing. This
485 * prevents us from getting stuck on 9 if the load is 10, for
486 * example.
487 */
488 if (new_load > old_load)
489 new_load += scale - 1;
490
491 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
492 }
493
494 sched_avg_update(this_rq);
495}
496
Alex Shib92486c2013-06-20 10:18:50 +0800497#ifdef CONFIG_SMP
Alex Shia9dc5d02013-06-20 10:18:57 +0800498static inline unsigned long get_rq_runnable_load(struct rq *rq)
Alex Shib92486c2013-06-20 10:18:50 +0800499{
500 return rq->cfs.runnable_load_avg;
501}
502#else
Alex Shia9dc5d02013-06-20 10:18:57 +0800503static inline unsigned long get_rq_runnable_load(struct rq *rq)
Alex Shib92486c2013-06-20 10:18:50 +0800504{
505 return rq->load.weight;
506}
507#endif
508
Paul Gortmaker45ceebf2013-04-19 15:10:49 -0400509#ifdef CONFIG_NO_HZ_COMMON
510/*
511 * There is no sane way to deal with nohz on smp when using jiffies because the
512 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
513 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
514 *
515 * Therefore we cannot use the delta approach from the regular tick since that
516 * would seriously skew the load calculation. However we'll make do for those
517 * updates happening while idle (nohz_idle_balance) or coming out of idle
518 * (tick_nohz_idle_exit).
519 *
520 * This means we might still be one tick off for nohz periods.
521 */
522
523/*
524 * Called from nohz_idle_balance() to update the load ratings before doing the
525 * idle balance.
526 */
527void update_idle_cpu_load(struct rq *this_rq)
528{
529 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
Alex Shib92486c2013-06-20 10:18:50 +0800530 unsigned long load = get_rq_runnable_load(this_rq);
Paul Gortmaker45ceebf2013-04-19 15:10:49 -0400531 unsigned long pending_updates;
532
533 /*
534 * bail if there's load or we're actually up-to-date.
535 */
536 if (load || curr_jiffies == this_rq->last_load_update_tick)
537 return;
538
539 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
540 this_rq->last_load_update_tick = curr_jiffies;
541
542 __update_cpu_load(this_rq, load, pending_updates);
543}
544
545/*
546 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
547 */
548void update_cpu_load_nohz(void)
549{
550 struct rq *this_rq = this_rq();
551 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
552 unsigned long pending_updates;
553
554 if (curr_jiffies == this_rq->last_load_update_tick)
555 return;
556
557 raw_spin_lock(&this_rq->lock);
558 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
559 if (pending_updates) {
560 this_rq->last_load_update_tick = curr_jiffies;
561 /*
562 * We were idle, this means load 0, the current load might be
563 * !0 due to remote wakeups and the sort.
564 */
565 __update_cpu_load(this_rq, 0, pending_updates);
566 }
567 raw_spin_unlock(&this_rq->lock);
568}
569#endif /* CONFIG_NO_HZ */
570
571/*
572 * Called from scheduler_tick()
573 */
574void update_cpu_load_active(struct rq *this_rq)
575{
Alex Shib92486c2013-06-20 10:18:50 +0800576 unsigned long load = get_rq_runnable_load(this_rq);
Paul Gortmaker45ceebf2013-04-19 15:10:49 -0400577 /*
578 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
579 */
580 this_rq->last_load_update_tick = jiffies;
Alex Shib92486c2013-06-20 10:18:50 +0800581 __update_cpu_load(this_rq, load, 1);
Paul Gortmaker45ceebf2013-04-19 15:10:49 -0400582
583 calc_load_account_active(this_rq);
584}