aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--include/linux/sched.h8
-rw-r--r--kernel/sched_fair.c1131
2 files changed, 1139 insertions, 0 deletions
diff --git a/include/linux/sched.h b/include/linux/sched.h
index d5084e7c48cf..90420321994f 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1243,6 +1243,14 @@ static inline void idle_task_exit(void) {}
1243 1243
1244extern void sched_idle_next(void); 1244extern void sched_idle_next(void);
1245 1245
1246extern unsigned int sysctl_sched_granularity;
1247extern unsigned int sysctl_sched_wakeup_granularity;
1248extern unsigned int sysctl_sched_batch_wakeup_granularity;
1249extern unsigned int sysctl_sched_stat_granularity;
1250extern unsigned int sysctl_sched_runtime_limit;
1251extern unsigned int sysctl_sched_child_runs_first;
1252extern unsigned int sysctl_sched_features;
1253
1246#ifdef CONFIG_RT_MUTEXES 1254#ifdef CONFIG_RT_MUTEXES
1247extern int rt_mutex_getprio(struct task_struct *p); 1255extern int rt_mutex_getprio(struct task_struct *p);
1248extern void rt_mutex_setprio(struct task_struct *p, int prio); 1256extern void rt_mutex_setprio(struct task_struct *p, int prio);
diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
new file mode 100644
index 000000000000..6971db0a7160
--- /dev/null
+++ b/kernel/sched_fair.c
@@ -0,0 +1,1131 @@
1/*
2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3 *
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5 *
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
8 *
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11 *
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15 *
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
18 */
19
20/*
21 * Preemption granularity:
22 * (default: 2 msec, units: nanoseconds)
23 *
24 * NOTE: this granularity value is not the same as the concept of
25 * 'timeslice length' - timeslices in CFS will typically be somewhat
26 * larger than this value. (to see the precise effective timeslice
27 * length of your workload, run vmstat and monitor the context-switches
28 * field)
29 *
30 * On SMP systems the value of this is multiplied by the log2 of the
31 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
32 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
33 */
34unsigned int sysctl_sched_granularity __read_mostly = 2000000000ULL/HZ;
35
36/*
37 * SCHED_BATCH wake-up granularity.
38 * (default: 10 msec, units: nanoseconds)
39 *
40 * This option delays the preemption effects of decoupled workloads
41 * and reduces their over-scheduling. Synchronous workloads will still
42 * have immediate wakeup/sleep latencies.
43 */
44unsigned int sysctl_sched_batch_wakeup_granularity __read_mostly =
45 10000000000ULL/HZ;
46
47/*
48 * SCHED_OTHER wake-up granularity.
49 * (default: 1 msec, units: nanoseconds)
50 *
51 * This option delays the preemption effects of decoupled workloads
52 * and reduces their over-scheduling. Synchronous workloads will still
53 * have immediate wakeup/sleep latencies.
54 */
55unsigned int sysctl_sched_wakeup_granularity __read_mostly = 1000000000ULL/HZ;
56
57unsigned int sysctl_sched_stat_granularity __read_mostly;
58
59/*
60 * Initialized in sched_init_granularity():
61 */
62unsigned int sysctl_sched_runtime_limit __read_mostly;
63
64/*
65 * Debugging: various feature bits
66 */
67enum {
68 SCHED_FEAT_FAIR_SLEEPERS = 1,
69 SCHED_FEAT_SLEEPER_AVG = 2,
70 SCHED_FEAT_SLEEPER_LOAD_AVG = 4,
71 SCHED_FEAT_PRECISE_CPU_LOAD = 8,
72 SCHED_FEAT_START_DEBIT = 16,
73 SCHED_FEAT_SKIP_INITIAL = 32,
74};
75
76unsigned int sysctl_sched_features __read_mostly =
77 SCHED_FEAT_FAIR_SLEEPERS *1 |
78 SCHED_FEAT_SLEEPER_AVG *1 |
79 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
80 SCHED_FEAT_PRECISE_CPU_LOAD *1 |
81 SCHED_FEAT_START_DEBIT *1 |
82 SCHED_FEAT_SKIP_INITIAL *0;
83
84extern struct sched_class fair_sched_class;
85
86/**************************************************************
87 * CFS operations on generic schedulable entities:
88 */
89
90#ifdef CONFIG_FAIR_GROUP_SCHED
91
92/* cpu runqueue to which this cfs_rq is attached */
93static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
94{
95 return cfs_rq->rq;
96}
97
98/* currently running entity (if any) on this cfs_rq */
99static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
100{
101 return cfs_rq->curr;
102}
103
104/* An entity is a task if it doesn't "own" a runqueue */
105#define entity_is_task(se) (!se->my_q)
106
107static inline void
108set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
109{
110 cfs_rq->curr = se;
111}
112
113#else /* CONFIG_FAIR_GROUP_SCHED */
114
115static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
116{
117 return container_of(cfs_rq, struct rq, cfs);
118}
119
120static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
121{
122 struct rq *rq = rq_of(cfs_rq);
123
124 if (unlikely(rq->curr->sched_class != &fair_sched_class))
125 return NULL;
126
127 return &rq->curr->se;
128}
129
130#define entity_is_task(se) 1
131
132static inline void
133set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
134
135#endif /* CONFIG_FAIR_GROUP_SCHED */
136
137static inline struct task_struct *task_of(struct sched_entity *se)
138{
139 return container_of(se, struct task_struct, se);
140}
141
142
143/**************************************************************
144 * Scheduling class tree data structure manipulation methods:
145 */
146
147/*
148 * Enqueue an entity into the rb-tree:
149 */
150static inline void
151__enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
152{
153 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
154 struct rb_node *parent = NULL;
155 struct sched_entity *entry;
156 s64 key = se->fair_key;
157 int leftmost = 1;
158
159 /*
160 * Find the right place in the rbtree:
161 */
162 while (*link) {
163 parent = *link;
164 entry = rb_entry(parent, struct sched_entity, run_node);
165 /*
166 * We dont care about collisions. Nodes with
167 * the same key stay together.
168 */
169 if (key - entry->fair_key < 0) {
170 link = &parent->rb_left;
171 } else {
172 link = &parent->rb_right;
173 leftmost = 0;
174 }
175 }
176
177 /*
178 * Maintain a cache of leftmost tree entries (it is frequently
179 * used):
180 */
181 if (leftmost)
182 cfs_rq->rb_leftmost = &se->run_node;
183
184 rb_link_node(&se->run_node, parent, link);
185 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
186 update_load_add(&cfs_rq->load, se->load.weight);
187 cfs_rq->nr_running++;
188 se->on_rq = 1;
189}
190
191static inline void
192__dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
193{
194 if (cfs_rq->rb_leftmost == &se->run_node)
195 cfs_rq->rb_leftmost = rb_next(&se->run_node);
196 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
197 update_load_sub(&cfs_rq->load, se->load.weight);
198 cfs_rq->nr_running--;
199 se->on_rq = 0;
200}
201
202static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
203{
204 return cfs_rq->rb_leftmost;
205}
206
207static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
208{
209 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
210}
211
212/**************************************************************
213 * Scheduling class statistics methods:
214 */
215
216/*
217 * We rescale the rescheduling granularity of tasks according to their
218 * nice level, but only linearly, not exponentially:
219 */
220static long
221niced_granularity(struct sched_entity *curr, unsigned long granularity)
222{
223 u64 tmp;
224
225 /*
226 * Negative nice levels get the same granularity as nice-0:
227 */
228 if (likely(curr->load.weight >= NICE_0_LOAD))
229 return granularity;
230 /*
231 * Positive nice level tasks get linearly finer
232 * granularity:
233 */
234 tmp = curr->load.weight * (u64)granularity;
235
236 /*
237 * It will always fit into 'long':
238 */
239 return (long) (tmp >> NICE_0_SHIFT);
240}
241
242static inline void
243limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
244{
245 long limit = sysctl_sched_runtime_limit;
246
247 /*
248 * Niced tasks have the same history dynamic range as
249 * non-niced tasks:
250 */
251 if (unlikely(se->wait_runtime > limit)) {
252 se->wait_runtime = limit;
253 schedstat_inc(se, wait_runtime_overruns);
254 schedstat_inc(cfs_rq, wait_runtime_overruns);
255 }
256 if (unlikely(se->wait_runtime < -limit)) {
257 se->wait_runtime = -limit;
258 schedstat_inc(se, wait_runtime_underruns);
259 schedstat_inc(cfs_rq, wait_runtime_underruns);
260 }
261}
262
263static inline void
264__add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
265{
266 se->wait_runtime += delta;
267 schedstat_add(se, sum_wait_runtime, delta);
268 limit_wait_runtime(cfs_rq, se);
269}
270
271static void
272add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
273{
274 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
275 __add_wait_runtime(cfs_rq, se, delta);
276 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
277}
278
279/*
280 * Update the current task's runtime statistics. Skip current tasks that
281 * are not in our scheduling class.
282 */
283static inline void
284__update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, u64 now)
285{
286 unsigned long delta, delta_exec, delta_fair;
287 long delta_mine;
288 struct load_weight *lw = &cfs_rq->load;
289 unsigned long load = lw->weight;
290
291 if (unlikely(!load))
292 return;
293
294 delta_exec = curr->delta_exec;
295#ifdef CONFIG_SCHEDSTATS
296 if (unlikely(delta_exec > curr->exec_max))
297 curr->exec_max = delta_exec;
298#endif
299
300 curr->sum_exec_runtime += delta_exec;
301 cfs_rq->exec_clock += delta_exec;
302
303 delta_fair = calc_delta_fair(delta_exec, lw);
304 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
305
306 if (cfs_rq->sleeper_bonus > sysctl_sched_stat_granularity) {
307 delta = calc_delta_mine(cfs_rq->sleeper_bonus,
308 curr->load.weight, lw);
309 if (unlikely(delta > cfs_rq->sleeper_bonus))
310 delta = cfs_rq->sleeper_bonus;
311
312 cfs_rq->sleeper_bonus -= delta;
313 delta_mine -= delta;
314 }
315
316 cfs_rq->fair_clock += delta_fair;
317 /*
318 * We executed delta_exec amount of time on the CPU,
319 * but we were only entitled to delta_mine amount of
320 * time during that period (if nr_running == 1 then
321 * the two values are equal)
322 * [Note: delta_mine - delta_exec is negative]:
323 */
324 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
325}
326
327static void update_curr(struct cfs_rq *cfs_rq, u64 now)
328{
329 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
330 unsigned long delta_exec;
331
332 if (unlikely(!curr))
333 return;
334
335 /*
336 * Get the amount of time the current task was running
337 * since the last time we changed load (this cannot
338 * overflow on 32 bits):
339 */
340 delta_exec = (unsigned long)(now - curr->exec_start);
341
342 curr->delta_exec += delta_exec;
343
344 if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
345 __update_curr(cfs_rq, curr, now);
346 curr->delta_exec = 0;
347 }
348 curr->exec_start = now;
349}
350
351static inline void
352update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
353{
354 se->wait_start_fair = cfs_rq->fair_clock;
355 se->wait_start = now;
356}
357
358/*
359 * We calculate fair deltas here, so protect against the random effects
360 * of a multiplication overflow by capping it to the runtime limit:
361 */
362#if BITS_PER_LONG == 32
363static inline unsigned long
364calc_weighted(unsigned long delta, unsigned long weight, int shift)
365{
366 u64 tmp = (u64)delta * weight >> shift;
367
368 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
369 return sysctl_sched_runtime_limit*2;
370 return tmp;
371}
372#else
373static inline unsigned long
374calc_weighted(unsigned long delta, unsigned long weight, int shift)
375{
376 return delta * weight >> shift;
377}
378#endif
379
380/*
381 * Task is being enqueued - update stats:
382 */
383static void
384update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
385{
386 s64 key;
387
388 /*
389 * Are we enqueueing a waiting task? (for current tasks
390 * a dequeue/enqueue event is a NOP)
391 */
392 if (se != cfs_rq_curr(cfs_rq))
393 update_stats_wait_start(cfs_rq, se, now);
394 /*
395 * Update the key:
396 */
397 key = cfs_rq->fair_clock;
398
399 /*
400 * Optimize the common nice 0 case:
401 */
402 if (likely(se->load.weight == NICE_0_LOAD)) {
403 key -= se->wait_runtime;
404 } else {
405 u64 tmp;
406
407 if (se->wait_runtime < 0) {
408 tmp = -se->wait_runtime;
409 key += (tmp * se->load.inv_weight) >>
410 (WMULT_SHIFT - NICE_0_SHIFT);
411 } else {
412 tmp = se->wait_runtime;
413 key -= (tmp * se->load.weight) >> NICE_0_SHIFT;
414 }
415 }
416
417 se->fair_key = key;
418}
419
420/*
421 * Note: must be called with a freshly updated rq->fair_clock.
422 */
423static inline void
424__update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
425{
426 unsigned long delta_fair = se->delta_fair_run;
427
428#ifdef CONFIG_SCHEDSTATS
429 {
430 s64 delta_wait = now - se->wait_start;
431 if (unlikely(delta_wait > se->wait_max))
432 se->wait_max = delta_wait;
433 }
434#endif
435
436 if (unlikely(se->load.weight != NICE_0_LOAD))
437 delta_fair = calc_weighted(delta_fair, se->load.weight,
438 NICE_0_SHIFT);
439
440 add_wait_runtime(cfs_rq, se, delta_fair);
441}
442
443static void
444update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
445{
446 unsigned long delta_fair;
447
448 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
449 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
450
451 se->delta_fair_run += delta_fair;
452 if (unlikely(abs(se->delta_fair_run) >=
453 sysctl_sched_stat_granularity)) {
454 __update_stats_wait_end(cfs_rq, se, now);
455 se->delta_fair_run = 0;
456 }
457
458 se->wait_start_fair = 0;
459 se->wait_start = 0;
460}
461
462static inline void
463update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
464{
465 update_curr(cfs_rq, now);
466 /*
467 * Mark the end of the wait period if dequeueing a
468 * waiting task:
469 */
470 if (se != cfs_rq_curr(cfs_rq))
471 update_stats_wait_end(cfs_rq, se, now);
472}
473
474/*
475 * We are picking a new current task - update its stats:
476 */
477static inline void
478update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
479{
480 /*
481 * We are starting a new run period:
482 */
483 se->exec_start = now;
484}
485
486/*
487 * We are descheduling a task - update its stats:
488 */
489static inline void
490update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
491{
492 se->exec_start = 0;
493}
494
495/**************************************************
496 * Scheduling class queueing methods:
497 */
498
499static void
500__enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
501{
502 unsigned long load = cfs_rq->load.weight, delta_fair;
503 long prev_runtime;
504
505 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
506 load = rq_of(cfs_rq)->cpu_load[2];
507
508 delta_fair = se->delta_fair_sleep;
509
510 /*
511 * Fix up delta_fair with the effect of us running
512 * during the whole sleep period:
513 */
514 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
515 delta_fair = div64_likely32((u64)delta_fair * load,
516 load + se->load.weight);
517
518 if (unlikely(se->load.weight != NICE_0_LOAD))
519 delta_fair = calc_weighted(delta_fair, se->load.weight,
520 NICE_0_SHIFT);
521
522 prev_runtime = se->wait_runtime;
523 __add_wait_runtime(cfs_rq, se, delta_fair);
524 delta_fair = se->wait_runtime - prev_runtime;
525
526 /*
527 * Track the amount of bonus we've given to sleepers:
528 */
529 cfs_rq->sleeper_bonus += delta_fair;
530
531 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
532}
533
534static void
535enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
536{
537 struct task_struct *tsk = task_of(se);
538 unsigned long delta_fair;
539
540 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
541 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
542 return;
543
544 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
545 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
546
547 se->delta_fair_sleep += delta_fair;
548 if (unlikely(abs(se->delta_fair_sleep) >=
549 sysctl_sched_stat_granularity)) {
550 __enqueue_sleeper(cfs_rq, se, now);
551 se->delta_fair_sleep = 0;
552 }
553
554 se->sleep_start_fair = 0;
555
556#ifdef CONFIG_SCHEDSTATS
557 if (se->sleep_start) {
558 u64 delta = now - se->sleep_start;
559
560 if ((s64)delta < 0)
561 delta = 0;
562
563 if (unlikely(delta > se->sleep_max))
564 se->sleep_max = delta;
565
566 se->sleep_start = 0;
567 se->sum_sleep_runtime += delta;
568 }
569 if (se->block_start) {
570 u64 delta = now - se->block_start;
571
572 if ((s64)delta < 0)
573 delta = 0;
574
575 if (unlikely(delta > se->block_max))
576 se->block_max = delta;
577
578 se->block_start = 0;
579 se->sum_sleep_runtime += delta;
580 }
581#endif
582}
583
584static void
585enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
586 int wakeup, u64 now)
587{
588 /*
589 * Update the fair clock.
590 */
591 update_curr(cfs_rq, now);
592
593 if (wakeup)
594 enqueue_sleeper(cfs_rq, se, now);
595
596 update_stats_enqueue(cfs_rq, se, now);
597 __enqueue_entity(cfs_rq, se);
598}
599
600static void
601dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
602 int sleep, u64 now)
603{
604 update_stats_dequeue(cfs_rq, se, now);
605 if (sleep) {
606 se->sleep_start_fair = cfs_rq->fair_clock;
607#ifdef CONFIG_SCHEDSTATS
608 if (entity_is_task(se)) {
609 struct task_struct *tsk = task_of(se);
610
611 if (tsk->state & TASK_INTERRUPTIBLE)
612 se->sleep_start = now;
613 if (tsk->state & TASK_UNINTERRUPTIBLE)
614 se->block_start = now;
615 }
616 cfs_rq->wait_runtime -= se->wait_runtime;
617#endif
618 }
619 __dequeue_entity(cfs_rq, se);
620}
621
622/*
623 * Preempt the current task with a newly woken task if needed:
624 */
625static void
626__check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
627 struct sched_entity *curr, unsigned long granularity)
628{
629 s64 __delta = curr->fair_key - se->fair_key;
630
631 /*
632 * Take scheduling granularity into account - do not
633 * preempt the current task unless the best task has
634 * a larger than sched_granularity fairness advantage:
635 */
636 if (__delta > niced_granularity(curr, granularity))
637 resched_task(rq_of(cfs_rq)->curr);
638}
639
640static inline void
641set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
642{
643 /*
644 * Any task has to be enqueued before it get to execute on
645 * a CPU. So account for the time it spent waiting on the
646 * runqueue. (note, here we rely on pick_next_task() having
647 * done a put_prev_task_fair() shortly before this, which
648 * updated rq->fair_clock - used by update_stats_wait_end())
649 */
650 update_stats_wait_end(cfs_rq, se, now);
651 update_stats_curr_start(cfs_rq, se, now);
652 set_cfs_rq_curr(cfs_rq, se);
653}
654
655static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq, u64 now)
656{
657 struct sched_entity *se = __pick_next_entity(cfs_rq);
658
659 set_next_entity(cfs_rq, se, now);
660
661 return se;
662}
663
664static void
665put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev, u64 now)
666{
667 /*
668 * If still on the runqueue then deactivate_task()
669 * was not called and update_curr() has to be done:
670 */
671 if (prev->on_rq)
672 update_curr(cfs_rq, now);
673
674 update_stats_curr_end(cfs_rq, prev, now);
675
676 if (prev->on_rq)
677 update_stats_wait_start(cfs_rq, prev, now);
678 set_cfs_rq_curr(cfs_rq, NULL);
679}
680
681static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
682{
683 struct rq *rq = rq_of(cfs_rq);
684 struct sched_entity *next;
685 u64 now = __rq_clock(rq);
686
687 /*
688 * Dequeue and enqueue the task to update its
689 * position within the tree:
690 */
691 dequeue_entity(cfs_rq, curr, 0, now);
692 enqueue_entity(cfs_rq, curr, 0, now);
693
694 /*
695 * Reschedule if another task tops the current one.
696 */
697 next = __pick_next_entity(cfs_rq);
698 if (next == curr)
699 return;
700
701 __check_preempt_curr_fair(cfs_rq, next, curr, sysctl_sched_granularity);
702}
703
704/**************************************************
705 * CFS operations on tasks:
706 */
707
708#ifdef CONFIG_FAIR_GROUP_SCHED
709
710/* Walk up scheduling entities hierarchy */
711#define for_each_sched_entity(se) \
712 for (; se; se = se->parent)
713
714static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
715{
716 return p->se.cfs_rq;
717}
718
719/* runqueue on which this entity is (to be) queued */
720static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
721{
722 return se->cfs_rq;
723}
724
725/* runqueue "owned" by this group */
726static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
727{
728 return grp->my_q;
729}
730
731/* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
732 * another cpu ('this_cpu')
733 */
734static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
735{
736 /* A later patch will take group into account */
737 return &cpu_rq(this_cpu)->cfs;
738}
739
740/* Iterate thr' all leaf cfs_rq's on a runqueue */
741#define for_each_leaf_cfs_rq(rq, cfs_rq) \
742 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
743
744/* Do the two (enqueued) tasks belong to the same group ? */
745static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
746{
747 if (curr->se.cfs_rq == p->se.cfs_rq)
748 return 1;
749
750 return 0;
751}
752
753#else /* CONFIG_FAIR_GROUP_SCHED */
754
755#define for_each_sched_entity(se) \
756 for (; se; se = NULL)
757
758static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
759{
760 return &task_rq(p)->cfs;
761}
762
763static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
764{
765 struct task_struct *p = task_of(se);
766 struct rq *rq = task_rq(p);
767
768 return &rq->cfs;
769}
770
771/* runqueue "owned" by this group */
772static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
773{
774 return NULL;
775}
776
777static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
778{
779 return &cpu_rq(this_cpu)->cfs;
780}
781
782#define for_each_leaf_cfs_rq(rq, cfs_rq) \
783 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
784
785static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
786{
787 return 1;
788}
789
790#endif /* CONFIG_FAIR_GROUP_SCHED */
791
792/*
793 * The enqueue_task method is called before nr_running is
794 * increased. Here we update the fair scheduling stats and
795 * then put the task into the rbtree:
796 */
797static void
798enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup, u64 now)
799{
800 struct cfs_rq *cfs_rq;
801 struct sched_entity *se = &p->se;
802
803 for_each_sched_entity(se) {
804 if (se->on_rq)
805 break;
806 cfs_rq = cfs_rq_of(se);
807 enqueue_entity(cfs_rq, se, wakeup, now);
808 }
809}
810
811/*
812 * The dequeue_task method is called before nr_running is
813 * decreased. We remove the task from the rbtree and
814 * update the fair scheduling stats:
815 */
816static void
817dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep, u64 now)
818{
819 struct cfs_rq *cfs_rq;
820 struct sched_entity *se = &p->se;
821
822 for_each_sched_entity(se) {
823 cfs_rq = cfs_rq_of(se);
824 dequeue_entity(cfs_rq, se, sleep, now);
825 /* Don't dequeue parent if it has other entities besides us */
826 if (cfs_rq->load.weight)
827 break;
828 }
829}
830
831/*
832 * sched_yield() support is very simple - we dequeue and enqueue
833 */
834static void yield_task_fair(struct rq *rq, struct task_struct *p)
835{
836 struct cfs_rq *cfs_rq = task_cfs_rq(p);
837 u64 now = __rq_clock(rq);
838
839 /*
840 * Dequeue and enqueue the task to update its
841 * position within the tree:
842 */
843 dequeue_entity(cfs_rq, &p->se, 0, now);
844 enqueue_entity(cfs_rq, &p->se, 0, now);
845}
846
847/*
848 * Preempt the current task with a newly woken task if needed:
849 */
850static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
851{
852 struct task_struct *curr = rq->curr;
853 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
854 unsigned long gran;
855
856 if (unlikely(rt_prio(p->prio))) {
857 update_curr(cfs_rq, rq_clock(rq));
858 resched_task(curr);
859 return;
860 }
861
862 gran = sysctl_sched_wakeup_granularity;
863 /*
864 * Batch tasks prefer throughput over latency:
865 */
866 if (unlikely(p->policy == SCHED_BATCH))
867 gran = sysctl_sched_batch_wakeup_granularity;
868
869 if (is_same_group(curr, p))
870 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
871}
872
873static struct task_struct *pick_next_task_fair(struct rq *rq, u64 now)
874{
875 struct cfs_rq *cfs_rq = &rq->cfs;
876 struct sched_entity *se;
877
878 if (unlikely(!cfs_rq->nr_running))
879 return NULL;
880
881 do {
882 se = pick_next_entity(cfs_rq, now);
883 cfs_rq = group_cfs_rq(se);
884 } while (cfs_rq);
885
886 return task_of(se);
887}
888
889/*
890 * Account for a descheduled task:
891 */
892static void put_prev_task_fair(struct rq *rq, struct task_struct *prev, u64 now)
893{
894 struct sched_entity *se = &prev->se;
895 struct cfs_rq *cfs_rq;
896
897 for_each_sched_entity(se) {
898 cfs_rq = cfs_rq_of(se);
899 put_prev_entity(cfs_rq, se, now);
900 }
901}
902
903/**************************************************
904 * Fair scheduling class load-balancing methods:
905 */
906
907/*
908 * Load-balancing iterator. Note: while the runqueue stays locked
909 * during the whole iteration, the current task might be
910 * dequeued so the iterator has to be dequeue-safe. Here we
911 * achieve that by always pre-iterating before returning
912 * the current task:
913 */
914static inline struct task_struct *
915__load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
916{
917 struct task_struct *p;
918
919 if (!curr)
920 return NULL;
921
922 p = rb_entry(curr, struct task_struct, se.run_node);
923 cfs_rq->rb_load_balance_curr = rb_next(curr);
924
925 return p;
926}
927
928static struct task_struct *load_balance_start_fair(void *arg)
929{
930 struct cfs_rq *cfs_rq = arg;
931
932 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
933}
934
935static struct task_struct *load_balance_next_fair(void *arg)
936{
937 struct cfs_rq *cfs_rq = arg;
938
939 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
940}
941
942static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
943{
944 struct sched_entity *curr;
945 struct task_struct *p;
946
947 if (!cfs_rq->nr_running)
948 return MAX_PRIO;
949
950 curr = __pick_next_entity(cfs_rq);
951 p = task_of(curr);
952
953 return p->prio;
954}
955
956static int
957load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
958 unsigned long max_nr_move, unsigned long max_load_move,
959 struct sched_domain *sd, enum cpu_idle_type idle,
960 int *all_pinned, unsigned long *total_load_moved)
961{
962 struct cfs_rq *busy_cfs_rq;
963 unsigned long load_moved, total_nr_moved = 0, nr_moved;
964 long rem_load_move = max_load_move;
965 struct rq_iterator cfs_rq_iterator;
966
967 cfs_rq_iterator.start = load_balance_start_fair;
968 cfs_rq_iterator.next = load_balance_next_fair;
969
970 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
971 struct cfs_rq *this_cfs_rq;
972 long imbalance;
973 unsigned long maxload;
974 int this_best_prio, best_prio, best_prio_seen = 0;
975
976 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
977
978 imbalance = busy_cfs_rq->load.weight -
979 this_cfs_rq->load.weight;
980 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
981 if (imbalance <= 0)
982 continue;
983
984 /* Don't pull more than imbalance/2 */
985 imbalance /= 2;
986 maxload = min(rem_load_move, imbalance);
987
988 this_best_prio = cfs_rq_best_prio(this_cfs_rq);
989 best_prio = cfs_rq_best_prio(busy_cfs_rq);
990
991 /*
992 * Enable handling of the case where there is more than one task
993 * with the best priority. If the current running task is one
994 * of those with prio==best_prio we know it won't be moved
995 * and therefore it's safe to override the skip (based on load)
996 * of any task we find with that prio.
997 */
998 if (cfs_rq_curr(busy_cfs_rq) == &busiest->curr->se)
999 best_prio_seen = 1;
1000
1001 /* pass busy_cfs_rq argument into
1002 * load_balance_[start|next]_fair iterators
1003 */
1004 cfs_rq_iterator.arg = busy_cfs_rq;
1005 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1006 max_nr_move, maxload, sd, idle, all_pinned,
1007 &load_moved, this_best_prio, best_prio,
1008 best_prio_seen, &cfs_rq_iterator);
1009
1010 total_nr_moved += nr_moved;
1011 max_nr_move -= nr_moved;
1012 rem_load_move -= load_moved;
1013
1014 if (max_nr_move <= 0 || rem_load_move <= 0)
1015 break;
1016 }
1017
1018 *total_load_moved = max_load_move - rem_load_move;
1019
1020 return total_nr_moved;
1021}
1022
1023/*
1024 * scheduler tick hitting a task of our scheduling class:
1025 */
1026static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1027{
1028 struct cfs_rq *cfs_rq;
1029 struct sched_entity *se = &curr->se;
1030
1031 for_each_sched_entity(se) {
1032 cfs_rq = cfs_rq_of(se);
1033 entity_tick(cfs_rq, se);
1034 }
1035}
1036
1037/*
1038 * Share the fairness runtime between parent and child, thus the
1039 * total amount of pressure for CPU stays equal - new tasks
1040 * get a chance to run but frequent forkers are not allowed to
1041 * monopolize the CPU. Note: the parent runqueue is locked,
1042 * the child is not running yet.
1043 */
1044static void task_new_fair(struct rq *rq, struct task_struct *p)
1045{
1046 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1047 struct sched_entity *se = &p->se;
1048 u64 now = rq_clock(rq);
1049
1050 sched_info_queued(p);
1051
1052 update_stats_enqueue(cfs_rq, se, now);
1053 /*
1054 * Child runs first: we let it run before the parent
1055 * until it reschedules once. We set up the key so that
1056 * it will preempt the parent:
1057 */
1058 p->se.fair_key = current->se.fair_key -
1059 niced_granularity(&rq->curr->se, sysctl_sched_granularity) - 1;
1060 /*
1061 * The first wait is dominated by the child-runs-first logic,
1062 * so do not credit it with that waiting time yet:
1063 */
1064 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1065 p->se.wait_start_fair = 0;
1066
1067 /*
1068 * The statistical average of wait_runtime is about
1069 * -granularity/2, so initialize the task with that:
1070 */
1071 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1072 p->se.wait_runtime = -(sysctl_sched_granularity / 2);
1073
1074 __enqueue_entity(cfs_rq, se);
1075 inc_nr_running(p, rq, now);
1076}
1077
1078#ifdef CONFIG_FAIR_GROUP_SCHED
1079/* Account for a task changing its policy or group.
1080 *
1081 * This routine is mostly called to set cfs_rq->curr field when a task
1082 * migrates between groups/classes.
1083 */
1084static void set_curr_task_fair(struct rq *rq)
1085{
1086 struct task_struct *curr = rq->curr;
1087 struct sched_entity *se = &curr->se;
1088 u64 now = rq_clock(rq);
1089 struct cfs_rq *cfs_rq;
1090
1091 for_each_sched_entity(se) {
1092 cfs_rq = cfs_rq_of(se);
1093 set_next_entity(cfs_rq, se, now);
1094 }
1095}
1096#else
1097static void set_curr_task_fair(struct rq *rq)
1098{
1099}
1100#endif
1101
1102/*
1103 * All the scheduling class methods:
1104 */
1105struct sched_class fair_sched_class __read_mostly = {
1106 .enqueue_task = enqueue_task_fair,
1107 .dequeue_task = dequeue_task_fair,
1108 .yield_task = yield_task_fair,
1109
1110 .check_preempt_curr = check_preempt_curr_fair,
1111
1112 .pick_next_task = pick_next_task_fair,
1113 .put_prev_task = put_prev_task_fair,
1114
1115 .load_balance = load_balance_fair,
1116
1117 .set_curr_task = set_curr_task_fair,
1118 .task_tick = task_tick_fair,
1119 .task_new = task_new_fair,
1120};
1121
1122#ifdef CONFIG_SCHED_DEBUG
1123void print_cfs_stats(struct seq_file *m, int cpu, u64 now)
1124{
1125 struct rq *rq = cpu_rq(cpu);
1126 struct cfs_rq *cfs_rq;
1127
1128 for_each_leaf_cfs_rq(rq, cfs_rq)
1129 print_cfs_rq(m, cpu, cfs_rq, now);
1130}
1131#endif