Now it works.
[cbs-scheduler.git] / kernel / sched_fair.c
blob8a3bdf35d55897ce6b418ca45fbd2c2b592bfdac
1 /*
2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
26 * Targeted preemption latency for CPU-bound tasks:
27 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
29 * NOTE: this latency value is not the same as the concept of
30 * 'timeslice length' - timeslices in CFS are of variable length
31 * and have no persistent notion like in traditional, time-slice
32 * based scheduling concepts.
34 * (to see the precise effective timeslice length of your workload,
35 * run vmstat and monitor the context-switches (cs) field)
37 unsigned int sysctl_sched_latency = 20000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity = 4000000ULL;
46 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
48 static unsigned int sched_nr_latency = 5;
51 * After fork, child runs first. (default) If set to 0 then
52 * parent will (try to) run first.
54 const_debug unsigned int sysctl_sched_child_runs_first = 1;
57 * sys_sched_yield() compat mode
59 * This option switches the agressive yield implementation of the
60 * old scheduler back on.
62 unsigned int __read_mostly sysctl_sched_compat_yield;
65 * SCHED_OTHER wake-up granularity.
66 * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 * This option delays the preemption effects of decoupled workloads
69 * and reduces their over-scheduling. Synchronous workloads will still
70 * have immediate wakeup/sleep latencies.
72 unsigned int sysctl_sched_wakeup_granularity = 5000000UL;
74 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
76 static const struct sched_class fair_sched_class;
78 /**************************************************************
79 * CFS operations on generic schedulable entities:
82 static inline struct task_struct *task_of(struct sched_entity *se)
84 return container_of(se, struct task_struct, se);
87 #ifdef CONFIG_FAIR_GROUP_SCHED
89 /* cpu runqueue to which this cfs_rq is attached */
90 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
92 return cfs_rq->rq;
95 /* An entity is a task if it doesn't "own" a runqueue */
96 #define entity_is_task(se) (!se->my_q)
98 /* Walk up scheduling entities hierarchy */
99 #define for_each_sched_entity(se) \
100 for (; se; se = se->parent)
102 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
104 return p->se.cfs_rq;
107 /* runqueue on which this entity is (to be) queued */
108 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
110 return se->cfs_rq;
113 /* runqueue "owned" by this group */
114 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
116 return grp->my_q;
119 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
120 * another cpu ('this_cpu')
122 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
124 return cfs_rq->tg->cfs_rq[this_cpu];
127 /* Iterate thr' all leaf cfs_rq's on a runqueue */
128 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
129 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
131 /* Do the two (enqueued) entities belong to the same group ? */
132 static inline int
133 is_same_group(struct sched_entity *se, struct sched_entity *pse)
135 if (se->cfs_rq == pse->cfs_rq)
136 return 1;
138 return 0;
141 static inline struct sched_entity *parent_entity(struct sched_entity *se)
143 return se->parent;
146 /* return depth at which a sched entity is present in the hierarchy */
147 static inline int depth_se(struct sched_entity *se)
149 int depth = 0;
151 for_each_sched_entity(se)
152 depth++;
154 return depth;
157 static void
158 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
160 int se_depth, pse_depth;
163 * preemption test can be made between sibling entities who are in the
164 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
165 * both tasks until we find their ancestors who are siblings of common
166 * parent.
169 /* First walk up until both entities are at same depth */
170 se_depth = depth_se(*se);
171 pse_depth = depth_se(*pse);
173 while (se_depth > pse_depth) {
174 se_depth--;
175 *se = parent_entity(*se);
178 while (pse_depth > se_depth) {
179 pse_depth--;
180 *pse = parent_entity(*pse);
183 while (!is_same_group(*se, *pse)) {
184 *se = parent_entity(*se);
185 *pse = parent_entity(*pse);
189 #else /* CONFIG_FAIR_GROUP_SCHED */
191 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
193 return container_of(cfs_rq, struct rq, cfs);
196 #define entity_is_task(se) 1
198 #define for_each_sched_entity(se) \
199 for (; se; se = NULL)
201 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
203 return &task_rq(p)->cfs;
206 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
208 struct task_struct *p = task_of(se);
209 struct rq *rq = task_rq(p);
211 return &rq->cfs;
214 /* runqueue "owned" by this group */
215 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
217 return NULL;
220 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
222 return &cpu_rq(this_cpu)->cfs;
225 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
226 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
228 static inline int
229 is_same_group(struct sched_entity *se, struct sched_entity *pse)
231 return 1;
234 static inline struct sched_entity *parent_entity(struct sched_entity *se)
236 return NULL;
239 static inline void
240 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
244 #endif /* CONFIG_FAIR_GROUP_SCHED */
247 /**************************************************************
248 * Scheduling class tree data structure manipulation methods:
251 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
253 s64 delta = (s64)(vruntime - min_vruntime);
254 if (delta > 0)
255 min_vruntime = vruntime;
257 return min_vruntime;
260 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
262 s64 delta = (s64)(vruntime - min_vruntime);
263 if (delta < 0)
264 min_vruntime = vruntime;
266 return min_vruntime;
269 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
271 return se->vruntime - cfs_rq->min_vruntime;
274 static void update_min_vruntime(struct cfs_rq *cfs_rq)
276 u64 vruntime = cfs_rq->min_vruntime;
278 if (cfs_rq->curr)
279 vruntime = cfs_rq->curr->vruntime;
281 if (cfs_rq->rb_leftmost) {
282 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
283 struct sched_entity,
284 run_node);
286 if (!cfs_rq->curr)
287 vruntime = se->vruntime;
288 else
289 vruntime = min_vruntime(vruntime, se->vruntime);
292 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
296 * Enqueue an entity into the rb-tree:
298 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
300 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
301 struct rb_node *parent = NULL;
302 struct sched_entity *entry;
303 s64 key = entity_key(cfs_rq, se);
304 int leftmost = 1;
307 * Find the right place in the rbtree:
309 while (*link) {
310 parent = *link;
311 entry = rb_entry(parent, struct sched_entity, run_node);
313 * We dont care about collisions. Nodes with
314 * the same key stay together.
316 if (key < entity_key(cfs_rq, entry)) {
317 link = &parent->rb_left;
318 } else {
319 link = &parent->rb_right;
320 leftmost = 0;
325 * Maintain a cache of leftmost tree entries (it is frequently
326 * used):
328 if (leftmost)
329 cfs_rq->rb_leftmost = &se->run_node;
331 rb_link_node(&se->run_node, parent, link);
332 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
335 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
337 if (cfs_rq->rb_leftmost == &se->run_node) {
338 struct rb_node *next_node;
340 next_node = rb_next(&se->run_node);
341 cfs_rq->rb_leftmost = next_node;
344 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
347 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
349 struct rb_node *left = cfs_rq->rb_leftmost;
351 if (!left)
352 return NULL;
354 return rb_entry(left, struct sched_entity, run_node);
357 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
359 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
361 if (!last)
362 return NULL;
364 return rb_entry(last, struct sched_entity, run_node);
367 /**************************************************************
368 * Scheduling class statistics methods:
371 #ifdef CONFIG_SCHED_DEBUG
372 int sched_nr_latency_handler(struct ctl_table *table, int write,
373 struct file *filp, void __user *buffer, size_t *lenp,
374 loff_t *ppos)
376 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
378 if (ret || !write)
379 return ret;
381 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
382 sysctl_sched_min_granularity);
384 return 0;
386 #endif
389 * delta /= w
391 static inline unsigned long
392 calc_delta_fair(unsigned long delta, struct sched_entity *se)
394 if (unlikely(se->load.weight != NICE_0_LOAD))
395 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
397 return delta;
401 * The idea is to set a period in which each task runs once.
403 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
404 * this period because otherwise the slices get too small.
406 * p = (nr <= nl) ? l : l*nr/nl
408 static u64 __sched_period(unsigned long nr_running)
410 u64 period = sysctl_sched_latency;
411 unsigned long nr_latency = sched_nr_latency;
413 if (unlikely(nr_running > nr_latency)) {
414 period = sysctl_sched_min_granularity;
415 period *= nr_running;
418 return period;
422 * We calculate the wall-time slice from the period by taking a part
423 * proportional to the weight.
425 * s = p*P[w/rw]
427 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
429 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
431 for_each_sched_entity(se) {
432 struct load_weight *load;
434 cfs_rq = cfs_rq_of(se);
435 load = &cfs_rq->load;
437 if (unlikely(!se->on_rq)) {
438 struct load_weight lw = cfs_rq->load;
440 update_load_add(&lw, se->load.weight);
441 load = &lw;
443 slice = calc_delta_mine(slice, se->load.weight, load);
445 return slice;
449 * We calculate the vruntime slice of a to be inserted task
451 * vs = s/w
453 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
455 return calc_delta_fair(sched_slice(cfs_rq, se), se);
459 * Update the current task's runtime statistics. Skip current tasks that
460 * are not in our scheduling class.
462 static inline void
463 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
464 unsigned long delta_exec)
466 unsigned long delta_exec_weighted;
468 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
470 curr->sum_exec_runtime += delta_exec;
471 schedstat_add(cfs_rq, exec_clock, delta_exec);
472 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
473 curr->vruntime += delta_exec_weighted;
474 update_min_vruntime(cfs_rq);
477 static void update_curr(struct cfs_rq *cfs_rq)
479 struct sched_entity *curr = cfs_rq->curr;
480 u64 now = rq_of(cfs_rq)->clock;
481 unsigned long delta_exec;
483 if (unlikely(!curr))
484 return;
487 * Get the amount of time the current task was running
488 * since the last time we changed load (this cannot
489 * overflow on 32 bits):
491 delta_exec = (unsigned long)(now - curr->exec_start);
492 if (!delta_exec)
493 return;
495 __update_curr(cfs_rq, curr, delta_exec);
496 curr->exec_start = now;
498 if (entity_is_task(curr)) {
499 struct task_struct *curtask = task_of(curr);
501 cpuacct_charge(curtask, delta_exec);
502 account_group_exec_runtime(curtask, delta_exec);
506 static inline void
507 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
509 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
513 * Task is being enqueued - update stats:
515 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
518 * Are we enqueueing a waiting task? (for current tasks
519 * a dequeue/enqueue event is a NOP)
521 if (se != cfs_rq->curr)
522 update_stats_wait_start(cfs_rq, se);
525 static void
526 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
528 schedstat_set(se->wait_max, max(se->wait_max,
529 rq_of(cfs_rq)->clock - se->wait_start));
530 schedstat_set(se->wait_count, se->wait_count + 1);
531 schedstat_set(se->wait_sum, se->wait_sum +
532 rq_of(cfs_rq)->clock - se->wait_start);
533 schedstat_set(se->wait_start, 0);
536 static inline void
537 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
540 * Mark the end of the wait period if dequeueing a
541 * waiting task:
543 if (se != cfs_rq->curr)
544 update_stats_wait_end(cfs_rq, se);
548 * We are picking a new current task - update its stats:
550 static inline void
551 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
554 * We are starting a new run period:
556 se->exec_start = rq_of(cfs_rq)->clock;
559 /**************************************************
560 * Scheduling class queueing methods:
563 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
564 static void
565 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
567 cfs_rq->task_weight += weight;
569 #else
570 static inline void
571 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
574 #endif
576 static void
577 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
579 update_load_add(&cfs_rq->load, se->load.weight);
580 if (!parent_entity(se))
581 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
582 if (entity_is_task(se)) {
583 add_cfs_task_weight(cfs_rq, se->load.weight);
584 list_add(&se->group_node, &cfs_rq->tasks);
586 cfs_rq->nr_running++;
587 se->on_rq = 1;
590 static void
591 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
593 update_load_sub(&cfs_rq->load, se->load.weight);
594 if (!parent_entity(se))
595 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
596 if (entity_is_task(se)) {
597 add_cfs_task_weight(cfs_rq, -se->load.weight);
598 list_del_init(&se->group_node);
600 cfs_rq->nr_running--;
601 se->on_rq = 0;
604 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
606 #ifdef CONFIG_SCHEDSTATS
607 if (se->sleep_start) {
608 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
609 struct task_struct *tsk = task_of(se);
611 if ((s64)delta < 0)
612 delta = 0;
614 if (unlikely(delta > se->sleep_max))
615 se->sleep_max = delta;
617 se->sleep_start = 0;
618 se->sum_sleep_runtime += delta;
620 account_scheduler_latency(tsk, delta >> 10, 1);
622 if (se->block_start) {
623 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
624 struct task_struct *tsk = task_of(se);
626 if ((s64)delta < 0)
627 delta = 0;
629 if (unlikely(delta > se->block_max))
630 se->block_max = delta;
632 se->block_start = 0;
633 se->sum_sleep_runtime += delta;
636 * Blocking time is in units of nanosecs, so shift by 20 to
637 * get a milliseconds-range estimation of the amount of
638 * time that the task spent sleeping:
640 if (unlikely(prof_on == SLEEP_PROFILING)) {
642 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
643 delta >> 20);
645 account_scheduler_latency(tsk, delta >> 10, 0);
647 #endif
650 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
652 #ifdef CONFIG_SCHED_DEBUG
653 s64 d = se->vruntime - cfs_rq->min_vruntime;
655 if (d < 0)
656 d = -d;
658 if (d > 3*sysctl_sched_latency)
659 schedstat_inc(cfs_rq, nr_spread_over);
660 #endif
663 static void
664 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
666 u64 vruntime = cfs_rq->min_vruntime;
669 * The 'current' period is already promised to the current tasks,
670 * however the extra weight of the new task will slow them down a
671 * little, place the new task so that it fits in the slot that
672 * stays open at the end.
674 if (initial && sched_feat(START_DEBIT))
675 vruntime += sched_vslice(cfs_rq, se);
677 if (!initial) {
678 /* sleeps upto a single latency don't count. */
679 if (sched_feat(NEW_FAIR_SLEEPERS)) {
680 unsigned long thresh = sysctl_sched_latency;
683 * Convert the sleeper threshold into virtual time.
684 * SCHED_IDLE is a special sub-class. We care about
685 * fairness only relative to other SCHED_IDLE tasks,
686 * all of which have the same weight.
688 if (sched_feat(NORMALIZED_SLEEPER) &&
689 task_of(se)->policy != SCHED_IDLE)
690 thresh = calc_delta_fair(thresh, se);
692 vruntime -= thresh;
695 /* ensure we never gain time by being placed backwards. */
696 vruntime = max_vruntime(se->vruntime, vruntime);
699 se->vruntime = vruntime;
702 static void
703 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
706 * Update run-time statistics of the 'current'.
708 update_curr(cfs_rq);
709 account_entity_enqueue(cfs_rq, se);
711 if (wakeup) {
712 place_entity(cfs_rq, se, 0);
713 enqueue_sleeper(cfs_rq, se);
716 update_stats_enqueue(cfs_rq, se);
717 check_spread(cfs_rq, se);
718 if (se != cfs_rq->curr)
719 __enqueue_entity(cfs_rq, se);
722 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
724 if (cfs_rq->last == se)
725 cfs_rq->last = NULL;
727 if (cfs_rq->next == se)
728 cfs_rq->next = NULL;
731 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
733 for_each_sched_entity(se)
734 __clear_buddies(cfs_rq_of(se), se);
737 static void
738 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
741 * Update run-time statistics of the 'current'.
743 update_curr(cfs_rq);
745 update_stats_dequeue(cfs_rq, se);
746 if (sleep) {
747 #ifdef CONFIG_SCHEDSTATS
748 if (entity_is_task(se)) {
749 struct task_struct *tsk = task_of(se);
751 if (tsk->state & TASK_INTERRUPTIBLE)
752 se->sleep_start = rq_of(cfs_rq)->clock;
753 if (tsk->state & TASK_UNINTERRUPTIBLE)
754 se->block_start = rq_of(cfs_rq)->clock;
756 #endif
759 clear_buddies(cfs_rq, se);
761 if (se != cfs_rq->curr)
762 __dequeue_entity(cfs_rq, se);
763 account_entity_dequeue(cfs_rq, se);
764 update_min_vruntime(cfs_rq);
768 * Preempt the current task with a newly woken task if needed:
770 static void
771 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
773 unsigned long ideal_runtime, delta_exec;
775 ideal_runtime = sched_slice(cfs_rq, curr);
776 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
777 if (delta_exec > ideal_runtime) {
778 resched_task(rq_of(cfs_rq)->curr);
780 * The current task ran long enough, ensure it doesn't get
781 * re-elected due to buddy favours.
783 clear_buddies(cfs_rq, curr);
787 static void
788 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
790 /* 'current' is not kept within the tree. */
791 if (se->on_rq) {
793 * Any task has to be enqueued before it get to execute on
794 * a CPU. So account for the time it spent waiting on the
795 * runqueue.
797 update_stats_wait_end(cfs_rq, se);
798 __dequeue_entity(cfs_rq, se);
801 update_stats_curr_start(cfs_rq, se);
802 cfs_rq->curr = se;
803 #ifdef CONFIG_SCHEDSTATS
805 * Track our maximum slice length, if the CPU's load is at
806 * least twice that of our own weight (i.e. dont track it
807 * when there are only lesser-weight tasks around):
809 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
810 se->slice_max = max(se->slice_max,
811 se->sum_exec_runtime - se->prev_sum_exec_runtime);
813 #endif
814 se->prev_sum_exec_runtime = se->sum_exec_runtime;
817 static int
818 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
820 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
822 struct sched_entity *se = __pick_next_entity(cfs_rq);
824 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
825 return cfs_rq->next;
827 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
828 return cfs_rq->last;
830 return se;
833 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
836 * If still on the runqueue then deactivate_task()
837 * was not called and update_curr() has to be done:
839 if (prev->on_rq)
840 update_curr(cfs_rq);
842 check_spread(cfs_rq, prev);
843 if (prev->on_rq) {
844 update_stats_wait_start(cfs_rq, prev);
845 /* Put 'current' back into the tree. */
846 __enqueue_entity(cfs_rq, prev);
848 cfs_rq->curr = NULL;
851 static void
852 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
855 * Update run-time statistics of the 'current'.
857 update_curr(cfs_rq);
859 #ifdef CONFIG_SCHED_HRTICK
861 * queued ticks are scheduled to match the slice, so don't bother
862 * validating it and just reschedule.
864 if (queued) {
865 resched_task(rq_of(cfs_rq)->curr);
866 return;
869 * don't let the period tick interfere with the hrtick preemption
871 if (!sched_feat(DOUBLE_TICK) &&
872 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
873 return;
874 #endif
876 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
877 check_preempt_tick(cfs_rq, curr);
880 /**************************************************
881 * CFS operations on tasks:
884 #ifdef CONFIG_SCHED_HRTICK
885 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
887 struct sched_entity *se = &p->se;
888 struct cfs_rq *cfs_rq = cfs_rq_of(se);
890 WARN_ON(task_rq(p) != rq);
892 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
893 u64 slice = sched_slice(cfs_rq, se);
894 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
895 s64 delta = slice - ran;
897 if (delta < 0) {
898 if (rq->curr == p)
899 resched_task(p);
900 return;
904 * Don't schedule slices shorter than 10000ns, that just
905 * doesn't make sense. Rely on vruntime for fairness.
907 if (rq->curr != p)
908 delta = max_t(s64, 10000LL, delta);
910 hrtick_start(rq, delta);
915 * called from enqueue/dequeue and updates the hrtick when the
916 * current task is from our class and nr_running is low enough
917 * to matter.
919 static void hrtick_update(struct rq *rq)
921 struct task_struct *curr = rq->curr;
923 if (curr->sched_class != &fair_sched_class)
924 return;
926 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
927 hrtick_start_fair(rq, curr);
929 #else /* !CONFIG_SCHED_HRTICK */
930 static inline void
931 hrtick_start_fair(struct rq *rq, struct task_struct *p)
935 static inline void hrtick_update(struct rq *rq)
938 #endif
941 * The enqueue_task method is called before nr_running is
942 * increased. Here we update the fair scheduling stats and
943 * then put the task into the rbtree:
945 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
947 struct cfs_rq *cfs_rq;
948 struct sched_entity *se = &p->se;
950 for_each_sched_entity(se) {
951 if (se->on_rq)
952 break;
953 cfs_rq = cfs_rq_of(se);
954 enqueue_entity(cfs_rq, se, wakeup);
955 wakeup = 1;
958 hrtick_update(rq);
962 * The dequeue_task method is called before nr_running is
963 * decreased. We remove the task from the rbtree and
964 * update the fair scheduling stats:
966 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
968 struct cfs_rq *cfs_rq;
969 struct sched_entity *se = &p->se;
971 for_each_sched_entity(se) {
972 cfs_rq = cfs_rq_of(se);
973 dequeue_entity(cfs_rq, se, sleep);
974 /* Don't dequeue parent if it has other entities besides us */
975 if (cfs_rq->load.weight)
976 break;
977 sleep = 1;
980 hrtick_update(rq);
984 * sched_yield() support is very simple - we dequeue and enqueue.
986 * If compat_yield is turned on then we requeue to the end of the tree.
988 static void yield_task_fair(struct rq *rq)
990 struct task_struct *curr = rq->curr;
991 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
992 struct sched_entity *rightmost, *se = &curr->se;
995 * Are we the only task in the tree?
997 if (unlikely(cfs_rq->nr_running == 1))
998 return;
1000 clear_buddies(cfs_rq, se);
1002 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1003 update_rq_clock(rq);
1005 * Update run-time statistics of the 'current'.
1007 update_curr(cfs_rq);
1009 return;
1012 * Find the rightmost entry in the rbtree:
1014 rightmost = __pick_last_entity(cfs_rq);
1016 * Already in the rightmost position?
1018 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
1019 return;
1022 * Minimally necessary key value to be last in the tree:
1023 * Upon rescheduling, sched_class::put_prev_task() will place
1024 * 'current' within the tree based on its new key value.
1026 se->vruntime = rightmost->vruntime + 1;
1030 * wake_idle() will wake a task on an idle cpu if task->cpu is
1031 * not idle and an idle cpu is available. The span of cpus to
1032 * search starts with cpus closest then further out as needed,
1033 * so we always favor a closer, idle cpu.
1034 * Domains may include CPUs that are not usable for migration,
1035 * hence we need to mask them out (cpu_active_mask)
1037 * Returns the CPU we should wake onto.
1039 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1040 static int wake_idle(int cpu, struct task_struct *p)
1042 struct sched_domain *sd;
1043 int i;
1044 unsigned int chosen_wakeup_cpu;
1045 int this_cpu;
1048 * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu
1049 * are idle and this is not a kernel thread and this task's affinity
1050 * allows it to be moved to preferred cpu, then just move!
1053 this_cpu = smp_processor_id();
1054 chosen_wakeup_cpu =
1055 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu;
1057 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP &&
1058 idle_cpu(cpu) && idle_cpu(this_cpu) &&
1059 p->mm && !(p->flags & PF_KTHREAD) &&
1060 cpu_isset(chosen_wakeup_cpu, p->cpus_allowed))
1061 return chosen_wakeup_cpu;
1064 * If it is idle, then it is the best cpu to run this task.
1066 * This cpu is also the best, if it has more than one task already.
1067 * Siblings must be also busy(in most cases) as they didn't already
1068 * pickup the extra load from this cpu and hence we need not check
1069 * sibling runqueue info. This will avoid the checks and cache miss
1070 * penalities associated with that.
1072 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
1073 return cpu;
1075 for_each_domain(cpu, sd) {
1076 if ((sd->flags & SD_WAKE_IDLE)
1077 || ((sd->flags & SD_WAKE_IDLE_FAR)
1078 && !task_hot(p, task_rq(p)->clock, sd))) {
1079 for_each_cpu_and(i, sched_domain_span(sd),
1080 &p->cpus_allowed) {
1081 if (cpu_active(i) && idle_cpu(i)) {
1082 if (i != task_cpu(p)) {
1083 schedstat_inc(p,
1084 se.nr_wakeups_idle);
1086 return i;
1089 } else {
1090 break;
1093 return cpu;
1095 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1096 static inline int wake_idle(int cpu, struct task_struct *p)
1098 return cpu;
1100 #endif
1102 #ifdef CONFIG_SMP
1104 #ifdef CONFIG_FAIR_GROUP_SCHED
1106 * effective_load() calculates the load change as seen from the root_task_group
1108 * Adding load to a group doesn't make a group heavier, but can cause movement
1109 * of group shares between cpus. Assuming the shares were perfectly aligned one
1110 * can calculate the shift in shares.
1112 * The problem is that perfectly aligning the shares is rather expensive, hence
1113 * we try to avoid doing that too often - see update_shares(), which ratelimits
1114 * this change.
1116 * We compensate this by not only taking the current delta into account, but
1117 * also considering the delta between when the shares were last adjusted and
1118 * now.
1120 * We still saw a performance dip, some tracing learned us that between
1121 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1122 * significantly. Therefore try to bias the error in direction of failing
1123 * the affine wakeup.
1126 static long effective_load(struct task_group *tg, int cpu,
1127 long wl, long wg)
1129 struct sched_entity *se = tg->se[cpu];
1131 if (!tg->parent)
1132 return wl;
1135 * By not taking the decrease of shares on the other cpu into
1136 * account our error leans towards reducing the affine wakeups.
1138 if (!wl && sched_feat(ASYM_EFF_LOAD))
1139 return wl;
1141 for_each_sched_entity(se) {
1142 long S, rw, s, a, b;
1143 long more_w;
1146 * Instead of using this increment, also add the difference
1147 * between when the shares were last updated and now.
1149 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1150 wl += more_w;
1151 wg += more_w;
1153 S = se->my_q->tg->shares;
1154 s = se->my_q->shares;
1155 rw = se->my_q->rq_weight;
1157 a = S*(rw + wl);
1158 b = S*rw + s*wg;
1160 wl = s*(a-b);
1162 if (likely(b))
1163 wl /= b;
1166 * Assume the group is already running and will
1167 * thus already be accounted for in the weight.
1169 * That is, moving shares between CPUs, does not
1170 * alter the group weight.
1172 wg = 0;
1175 return wl;
1178 #else
1180 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1181 unsigned long wl, unsigned long wg)
1183 return wl;
1186 #endif
1188 static int
1189 wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
1190 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1191 int idx, unsigned long load, unsigned long this_load,
1192 unsigned int imbalance)
1194 struct task_struct *curr = this_rq->curr;
1195 struct task_group *tg;
1196 unsigned long tl = this_load;
1197 unsigned long tl_per_task;
1198 unsigned long weight;
1199 int balanced;
1201 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1202 return 0;
1204 if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1205 p->se.avg_overlap > sysctl_sched_migration_cost))
1206 sync = 0;
1209 * If sync wakeup then subtract the (maximum possible)
1210 * effect of the currently running task from the load
1211 * of the current CPU:
1213 if (sync) {
1214 tg = task_group(current);
1215 weight = current->se.load.weight;
1217 tl += effective_load(tg, this_cpu, -weight, -weight);
1218 load += effective_load(tg, prev_cpu, 0, -weight);
1221 tg = task_group(p);
1222 weight = p->se.load.weight;
1224 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1225 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1228 * If the currently running task will sleep within
1229 * a reasonable amount of time then attract this newly
1230 * woken task:
1232 if (sync && balanced)
1233 return 1;
1235 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1236 tl_per_task = cpu_avg_load_per_task(this_cpu);
1238 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1239 tl_per_task)) {
1241 * This domain has SD_WAKE_AFFINE and
1242 * p is cache cold in this domain, and
1243 * there is no bad imbalance.
1245 schedstat_inc(this_sd, ttwu_move_affine);
1246 schedstat_inc(p, se.nr_wakeups_affine);
1248 return 1;
1250 return 0;
1253 static int select_task_rq_fair(struct task_struct *p, int sync)
1255 struct sched_domain *sd, *this_sd = NULL;
1256 int prev_cpu, this_cpu, new_cpu;
1257 unsigned long load, this_load;
1258 struct rq *this_rq;
1259 unsigned int imbalance;
1260 int idx;
1262 prev_cpu = task_cpu(p);
1263 this_cpu = smp_processor_id();
1264 this_rq = cpu_rq(this_cpu);
1265 new_cpu = prev_cpu;
1267 if (prev_cpu == this_cpu)
1268 goto out;
1270 * 'this_sd' is the first domain that both
1271 * this_cpu and prev_cpu are present in:
1273 for_each_domain(this_cpu, sd) {
1274 if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) {
1275 this_sd = sd;
1276 break;
1280 if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed)))
1281 goto out;
1284 * Check for affine wakeup and passive balancing possibilities.
1286 if (!this_sd)
1287 goto out;
1289 idx = this_sd->wake_idx;
1291 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1293 load = source_load(prev_cpu, idx);
1294 this_load = target_load(this_cpu, idx);
1296 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1297 load, this_load, imbalance))
1298 return this_cpu;
1301 * Start passive balancing when half the imbalance_pct
1302 * limit is reached.
1304 if (this_sd->flags & SD_WAKE_BALANCE) {
1305 if (imbalance*this_load <= 100*load) {
1306 schedstat_inc(this_sd, ttwu_move_balance);
1307 schedstat_inc(p, se.nr_wakeups_passive);
1308 return this_cpu;
1312 out:
1313 return wake_idle(new_cpu, p);
1315 #endif /* CONFIG_SMP */
1318 * Adaptive granularity
1320 * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1321 * with the limit of wakeup_gran -- when it never does a wakeup.
1323 * So the smaller avg_wakeup is the faster we want this task to preempt,
1324 * but we don't want to treat the preemptee unfairly and therefore allow it
1325 * to run for at least the amount of time we'd like to run.
1327 * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1329 * NOTE: we use *nr_running to scale with load, this nicely matches the
1330 * degrading latency on load.
1332 static unsigned long
1333 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1335 u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1336 u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1337 u64 gran = 0;
1339 if (this_run < expected_wakeup)
1340 gran = expected_wakeup - this_run;
1342 return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1345 static unsigned long
1346 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1348 unsigned long gran = sysctl_sched_wakeup_granularity;
1350 if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1351 gran = adaptive_gran(curr, se);
1354 * Since its curr running now, convert the gran from real-time
1355 * to virtual-time in his units.
1357 if (sched_feat(ASYM_GRAN)) {
1359 * By using 'se' instead of 'curr' we penalize light tasks, so
1360 * they get preempted easier. That is, if 'se' < 'curr' then
1361 * the resulting gran will be larger, therefore penalizing the
1362 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1363 * be smaller, again penalizing the lighter task.
1365 * This is especially important for buddies when the leftmost
1366 * task is higher priority than the buddy.
1368 if (unlikely(se->load.weight != NICE_0_LOAD))
1369 gran = calc_delta_fair(gran, se);
1370 } else {
1371 if (unlikely(curr->load.weight != NICE_0_LOAD))
1372 gran = calc_delta_fair(gran, curr);
1375 return gran;
1379 * Should 'se' preempt 'curr'.
1381 * |s1
1382 * |s2
1383 * |s3
1385 * |<--->|c
1387 * w(c, s1) = -1
1388 * w(c, s2) = 0
1389 * w(c, s3) = 1
1392 static int
1393 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1395 s64 gran, vdiff = curr->vruntime - se->vruntime;
1397 if (vdiff <= 0)
1398 return -1;
1400 gran = wakeup_gran(curr, se);
1401 if (vdiff > gran)
1402 return 1;
1404 return 0;
1407 static void set_last_buddy(struct sched_entity *se)
1409 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1410 for_each_sched_entity(se)
1411 cfs_rq_of(se)->last = se;
1415 static void set_next_buddy(struct sched_entity *se)
1417 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1418 for_each_sched_entity(se)
1419 cfs_rq_of(se)->next = se;
1424 * Preempt the current task with a newly woken task if needed:
1426 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
1428 struct task_struct *curr = rq->curr;
1429 struct sched_entity *se = &curr->se, *pse = &p->se;
1430 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1432 update_curr(cfs_rq);
1434 if (unlikely(rt_prio(p->prio))) {
1435 resched_task(curr);
1436 return;
1439 if (unlikely(p->sched_class != &fair_sched_class)){
1440 update_rq_clock(rq);
1441 update_curr(cfs_rq);
1442 resched_task(curr);
1443 return;
1447 if (unlikely(se == pse))
1448 return;
1451 * Only set the backward buddy when the current task is still on the
1452 * rq. This can happen when a wakeup gets interleaved with schedule on
1453 * the ->pre_schedule() or idle_balance() point, either of which can
1454 * drop the rq lock.
1456 * Also, during early boot the idle thread is in the fair class, for
1457 * obvious reasons its a bad idea to schedule back to the idle thread.
1459 if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1460 set_last_buddy(se);
1461 set_next_buddy(pse);
1464 * We can come here with TIF_NEED_RESCHED already set from new task
1465 * wake up path.
1467 if (test_tsk_need_resched(curr))
1468 return;
1471 * Batch and idle tasks do not preempt (their preemption is driven by
1472 * the tick):
1474 if (unlikely(p->policy != SCHED_NORMAL))
1475 return;
1477 /* Idle tasks are by definition preempted by everybody. */
1478 if (unlikely(curr->policy == SCHED_IDLE)) {
1479 resched_task(curr);
1480 return;
1483 if (!sched_feat(WAKEUP_PREEMPT))
1484 return;
1486 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1487 (se->avg_overlap < sysctl_sched_migration_cost &&
1488 pse->avg_overlap < sysctl_sched_migration_cost))) {
1489 resched_task(curr);
1490 return;
1493 find_matching_se(&se, &pse);
1495 while (se) {
1496 BUG_ON(!pse);
1498 if (wakeup_preempt_entity(se, pse) == 1) {
1499 resched_task(curr);
1500 break;
1503 se = parent_entity(se);
1504 pse = parent_entity(pse);
1508 static struct task_struct *pick_next_task_fair(struct rq *rq)
1510 struct task_struct *p;
1511 struct cfs_rq *cfs_rq = &rq->cfs;
1512 struct sched_entity *se;
1514 if (unlikely(!cfs_rq->nr_running))
1515 return NULL;
1517 do {
1518 se = pick_next_entity(cfs_rq);
1520 * If se was a buddy, clear it so that it will have to earn
1521 * the favour again.
1523 __clear_buddies(cfs_rq, se);
1524 set_next_entity(cfs_rq, se);
1525 cfs_rq = group_cfs_rq(se);
1526 } while (cfs_rq);
1528 p = task_of(se);
1529 hrtick_start_fair(rq, p);
1531 return p;
1535 * Account for a descheduled task:
1537 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1539 struct sched_entity *se = &prev->se;
1540 struct cfs_rq *cfs_rq;
1542 for_each_sched_entity(se) {
1543 cfs_rq = cfs_rq_of(se);
1544 put_prev_entity(cfs_rq, se);
1548 #ifdef CONFIG_SMP
1549 /**************************************************
1550 * Fair scheduling class load-balancing methods:
1554 * Load-balancing iterator. Note: while the runqueue stays locked
1555 * during the whole iteration, the current task might be
1556 * dequeued so the iterator has to be dequeue-safe. Here we
1557 * achieve that by always pre-iterating before returning
1558 * the current task:
1560 static struct task_struct *
1561 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1563 struct task_struct *p = NULL;
1564 struct sched_entity *se;
1566 if (next == &cfs_rq->tasks)
1567 return NULL;
1569 se = list_entry(next, struct sched_entity, group_node);
1570 p = task_of(se);
1571 cfs_rq->balance_iterator = next->next;
1573 return p;
1576 static struct task_struct *load_balance_start_fair(void *arg)
1578 struct cfs_rq *cfs_rq = arg;
1580 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1583 static struct task_struct *load_balance_next_fair(void *arg)
1585 struct cfs_rq *cfs_rq = arg;
1587 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1590 static unsigned long
1591 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1592 unsigned long max_load_move, struct sched_domain *sd,
1593 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1594 struct cfs_rq *cfs_rq)
1596 struct rq_iterator cfs_rq_iterator;
1598 cfs_rq_iterator.start = load_balance_start_fair;
1599 cfs_rq_iterator.next = load_balance_next_fair;
1600 cfs_rq_iterator.arg = cfs_rq;
1602 return balance_tasks(this_rq, this_cpu, busiest,
1603 max_load_move, sd, idle, all_pinned,
1604 this_best_prio, &cfs_rq_iterator);
1607 #ifdef CONFIG_FAIR_GROUP_SCHED
1608 static unsigned long
1609 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1610 unsigned long max_load_move,
1611 struct sched_domain *sd, enum cpu_idle_type idle,
1612 int *all_pinned, int *this_best_prio)
1614 long rem_load_move = max_load_move;
1615 int busiest_cpu = cpu_of(busiest);
1616 struct task_group *tg;
1618 rcu_read_lock();
1619 update_h_load(busiest_cpu);
1621 list_for_each_entry_rcu(tg, &task_groups, list) {
1622 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1623 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1624 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1625 u64 rem_load, moved_load;
1628 * empty group
1630 if (!busiest_cfs_rq->task_weight)
1631 continue;
1633 rem_load = (u64)rem_load_move * busiest_weight;
1634 rem_load = div_u64(rem_load, busiest_h_load + 1);
1636 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1637 rem_load, sd, idle, all_pinned, this_best_prio,
1638 tg->cfs_rq[busiest_cpu]);
1640 if (!moved_load)
1641 continue;
1643 moved_load *= busiest_h_load;
1644 moved_load = div_u64(moved_load, busiest_weight + 1);
1646 rem_load_move -= moved_load;
1647 if (rem_load_move < 0)
1648 break;
1650 rcu_read_unlock();
1652 return max_load_move - rem_load_move;
1654 #else
1655 static unsigned long
1656 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1657 unsigned long max_load_move,
1658 struct sched_domain *sd, enum cpu_idle_type idle,
1659 int *all_pinned, int *this_best_prio)
1661 return __load_balance_fair(this_rq, this_cpu, busiest,
1662 max_load_move, sd, idle, all_pinned,
1663 this_best_prio, &busiest->cfs);
1665 #endif
1667 static int
1668 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1669 struct sched_domain *sd, enum cpu_idle_type idle)
1671 struct cfs_rq *busy_cfs_rq;
1672 struct rq_iterator cfs_rq_iterator;
1674 cfs_rq_iterator.start = load_balance_start_fair;
1675 cfs_rq_iterator.next = load_balance_next_fair;
1677 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1679 * pass busy_cfs_rq argument into
1680 * load_balance_[start|next]_fair iterators
1682 cfs_rq_iterator.arg = busy_cfs_rq;
1683 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1684 &cfs_rq_iterator))
1685 return 1;
1688 return 0;
1690 #endif /* CONFIG_SMP */
1693 * scheduler tick hitting a task of our scheduling class:
1695 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1697 struct cfs_rq *cfs_rq;
1698 struct sched_entity *se = &curr->se;
1700 for_each_sched_entity(se) {
1701 cfs_rq = cfs_rq_of(se);
1702 entity_tick(cfs_rq, se, queued);
1707 * Share the fairness runtime between parent and child, thus the
1708 * total amount of pressure for CPU stays equal - new tasks
1709 * get a chance to run but frequent forkers are not allowed to
1710 * monopolize the CPU. Note: the parent runqueue is locked,
1711 * the child is not running yet.
1713 static void task_new_fair(struct rq *rq, struct task_struct *p)
1715 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1716 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1717 int this_cpu = smp_processor_id();
1719 sched_info_queued(p);
1721 update_curr(cfs_rq);
1722 place_entity(cfs_rq, se, 1);
1724 /* 'curr' will be NULL if the child belongs to a different group */
1725 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1726 curr && curr->vruntime < se->vruntime) {
1728 * Upon rescheduling, sched_class::put_prev_task() will place
1729 * 'current' within the tree based on its new key value.
1731 swap(curr->vruntime, se->vruntime);
1732 resched_task(rq->curr);
1735 enqueue_task_fair(rq, p, 0);
1739 * Priority of the task has changed. Check to see if we preempt
1740 * the current task.
1742 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1743 int oldprio, int running)
1746 * Reschedule if we are currently running on this runqueue and
1747 * our priority decreased, or if we are not currently running on
1748 * this runqueue and our priority is higher than the current's
1750 if (running) {
1751 if (p->prio > oldprio)
1752 resched_task(rq->curr);
1753 } else
1754 check_preempt_curr(rq, p, 0);
1758 * We switched to the sched_fair class.
1760 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1761 int running)
1764 * We were most likely switched from sched_rt, so
1765 * kick off the schedule if running, otherwise just see
1766 * if we can still preempt the current task.
1768 if (running)
1769 resched_task(rq->curr);
1770 else
1771 check_preempt_curr(rq, p, 0);
1774 /* Account for a task changing its policy or group.
1776 * This routine is mostly called to set cfs_rq->curr field when a task
1777 * migrates between groups/classes.
1779 static void set_curr_task_fair(struct rq *rq)
1781 struct sched_entity *se = &rq->curr->se;
1783 for_each_sched_entity(se)
1784 set_next_entity(cfs_rq_of(se), se);
1787 #ifdef CONFIG_FAIR_GROUP_SCHED
1788 static void moved_group_fair(struct task_struct *p)
1790 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1792 update_curr(cfs_rq);
1793 place_entity(cfs_rq, &p->se, 1);
1795 #endif
1798 * All the scheduling class methods:
1800 static const struct sched_class fair_sched_class = {
1801 .next = &idle_sched_class,
1802 .enqueue_task = enqueue_task_fair,
1803 .dequeue_task = dequeue_task_fair,
1804 .yield_task = yield_task_fair,
1806 .check_preempt_curr = check_preempt_wakeup,
1808 .pick_next_task = pick_next_task_fair,
1809 .put_prev_task = put_prev_task_fair,
1811 #ifdef CONFIG_SMP
1812 .select_task_rq = select_task_rq_fair,
1814 .load_balance = load_balance_fair,
1815 .move_one_task = move_one_task_fair,
1816 #endif
1818 .set_curr_task = set_curr_task_fair,
1819 .task_tick = task_tick_fair,
1820 .task_new = task_new_fair,
1822 .prio_changed = prio_changed_fair,
1823 .switched_to = switched_to_fair,
1825 #ifdef CONFIG_FAIR_GROUP_SCHED
1826 .moved_group = moved_group_fair,
1827 #endif
1830 #ifdef CONFIG_SCHED_DEBUG
1831 static void print_cfs_stats(struct seq_file *m, int cpu)
1833 struct cfs_rq *cfs_rq;
1835 rcu_read_lock();
1836 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1837 print_cfs_rq(m, cpu, cfs_rq);
1838 rcu_read_unlock();
1840 #endif