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>
24 * Targeted preemption latency for CPU-bound tasks:
25 * (default: 20ms, units: nanoseconds)
27 * NOTE: this latency value is not the same as the concept of
28 * 'timeslice length' - timeslices in CFS are of variable length
29 * and have no persistent notion like in traditional, time-slice
30 * based scheduling concepts.
32 * (to see the precise effective timeslice length of your workload,
33 * run vmstat and monitor the context-switches (cs) field)
35 const_debug unsigned int sysctl_sched_latency = 20000000ULL;
38 * After fork, child runs first. (default) If set to 0 then
39 * parent will (try to) run first.
41 const_debug unsigned int sysctl_sched_child_runs_first = 1;
44 * Minimal preemption granularity for CPU-bound tasks:
45 * (default: 2 msec, units: nanoseconds)
47 const_debug unsigned int sysctl_sched_nr_latency = 20;
50 * sys_sched_yield() compat mode
52 * This option switches the agressive yield implementation of the
53 * old scheduler back on.
55 unsigned int __read_mostly sysctl_sched_compat_yield;
58 * SCHED_BATCH wake-up granularity.
59 * (default: 10 msec, units: nanoseconds)
61 * This option delays the preemption effects of decoupled workloads
62 * and reduces their over-scheduling. Synchronous workloads will still
63 * have immediate wakeup/sleep latencies.
65 const_debug unsigned int sysctl_sched_batch_wakeup_granularity = 10000000UL;
68 * SCHED_OTHER wake-up granularity.
69 * (default: 10 msec, units: nanoseconds)
71 * This option delays the preemption effects of decoupled workloads
72 * and reduces their over-scheduling. Synchronous workloads will still
73 * have immediate wakeup/sleep latencies.
75 const_debug unsigned int sysctl_sched_wakeup_granularity = 10000000UL;
77 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
79 /**************************************************************
80 * CFS operations on generic schedulable entities:
83 #ifdef CONFIG_FAIR_GROUP_SCHED
85 /* cpu runqueue to which this cfs_rq is attached */
86 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
91 /* An entity is a task if it doesn't "own" a runqueue */
92 #define entity_is_task(se) (!se->my_q)
94 #else /* CONFIG_FAIR_GROUP_SCHED */
96 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
98 return container_of(cfs_rq, struct rq, cfs);
101 #define entity_is_task(se) 1
103 #endif /* CONFIG_FAIR_GROUP_SCHED */
105 static inline struct task_struct *task_of(struct sched_entity *se)
107 return container_of(se, struct task_struct, se);
111 /**************************************************************
112 * Scheduling class tree data structure manipulation methods:
115 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
117 s64 delta = (s64)(vruntime - min_vruntime);
119 min_vruntime = vruntime;
124 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
126 s64 delta = (s64)(vruntime - min_vruntime);
128 min_vruntime = vruntime;
133 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
135 return se->vruntime - cfs_rq->min_vruntime;
139 * Enqueue an entity into the rb-tree:
141 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
143 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
144 struct rb_node *parent = NULL;
145 struct sched_entity *entry;
146 s64 key = entity_key(cfs_rq, se);
150 * Find the right place in the rbtree:
154 entry = rb_entry(parent, struct sched_entity, run_node);
156 * We dont care about collisions. Nodes with
157 * the same key stay together.
159 if (key < entity_key(cfs_rq, entry)) {
160 link = &parent->rb_left;
162 link = &parent->rb_right;
168 * Maintain a cache of leftmost tree entries (it is frequently
172 cfs_rq->rb_leftmost = &se->run_node;
174 rb_link_node(&se->run_node, parent, link);
175 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
178 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
180 if (cfs_rq->rb_leftmost == &se->run_node)
181 cfs_rq->rb_leftmost = rb_next(&se->run_node);
183 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
186 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
188 return cfs_rq->rb_leftmost;
191 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
193 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
196 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
198 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
199 struct sched_entity *se = NULL;
200 struct rb_node *parent;
204 se = rb_entry(parent, struct sched_entity, run_node);
205 link = &parent->rb_right;
211 /**************************************************************
212 * Scheduling class statistics methods:
217 * The idea is to set a period in which each task runs once.
219 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
220 * this period because otherwise the slices get too small.
222 * p = (nr <= nl) ? l : l*nr/nl
224 static u64 __sched_period(unsigned long nr_running)
226 u64 period = sysctl_sched_latency;
227 unsigned long nr_latency = sysctl_sched_nr_latency;
229 if (unlikely(nr_running > nr_latency)) {
230 period *= nr_running;
231 do_div(period, nr_latency);
238 * We calculate the wall-time slice from the period by taking a part
239 * proportional to the weight.
243 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
245 u64 slice = __sched_period(cfs_rq->nr_running);
247 slice *= se->load.weight;
248 do_div(slice, cfs_rq->load.weight);
254 * We calculate the vruntime slice.
258 static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running)
260 u64 vslice = __sched_period(nr_running);
262 do_div(vslice, rq_weight);
267 static u64 sched_vslice(struct cfs_rq *cfs_rq)
269 return __sched_vslice(cfs_rq->load.weight, cfs_rq->nr_running);
272 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
274 return __sched_vslice(cfs_rq->load.weight + se->load.weight,
275 cfs_rq->nr_running + 1);
279 * Update the current task's runtime statistics. Skip current tasks that
280 * are not in our scheduling class.
283 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
284 unsigned long delta_exec)
286 unsigned long delta_exec_weighted;
289 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
291 curr->sum_exec_runtime += delta_exec;
292 schedstat_add(cfs_rq, exec_clock, delta_exec);
293 delta_exec_weighted = delta_exec;
294 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
295 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
298 curr->vruntime += delta_exec_weighted;
301 * maintain cfs_rq->min_vruntime to be a monotonic increasing
302 * value tracking the leftmost vruntime in the tree.
304 if (first_fair(cfs_rq)) {
305 vruntime = min_vruntime(curr->vruntime,
306 __pick_next_entity(cfs_rq)->vruntime);
308 vruntime = curr->vruntime;
310 cfs_rq->min_vruntime =
311 max_vruntime(cfs_rq->min_vruntime, vruntime);
314 static void update_curr(struct cfs_rq *cfs_rq)
316 struct sched_entity *curr = cfs_rq->curr;
317 u64 now = rq_of(cfs_rq)->clock;
318 unsigned long delta_exec;
324 * Get the amount of time the current task was running
325 * since the last time we changed load (this cannot
326 * overflow on 32 bits):
328 delta_exec = (unsigned long)(now - curr->exec_start);
330 __update_curr(cfs_rq, curr, delta_exec);
331 curr->exec_start = now;
335 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
337 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
341 * Task is being enqueued - update stats:
343 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
346 * Are we enqueueing a waiting task? (for current tasks
347 * a dequeue/enqueue event is a NOP)
349 if (se != cfs_rq->curr)
350 update_stats_wait_start(cfs_rq, se);
354 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
356 schedstat_set(se->wait_max, max(se->wait_max,
357 rq_of(cfs_rq)->clock - se->wait_start));
358 schedstat_set(se->wait_start, 0);
362 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
365 * Mark the end of the wait period if dequeueing a
368 if (se != cfs_rq->curr)
369 update_stats_wait_end(cfs_rq, se);
373 * We are picking a new current task - update its stats:
376 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
379 * We are starting a new run period:
381 se->exec_start = rq_of(cfs_rq)->clock;
384 /**************************************************
385 * Scheduling class queueing methods:
389 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
391 update_load_add(&cfs_rq->load, se->load.weight);
392 cfs_rq->nr_running++;
397 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
399 update_load_sub(&cfs_rq->load, se->load.weight);
400 cfs_rq->nr_running--;
404 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
406 #ifdef CONFIG_SCHEDSTATS
407 if (se->sleep_start) {
408 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
413 if (unlikely(delta > se->sleep_max))
414 se->sleep_max = delta;
417 se->sum_sleep_runtime += delta;
419 if (se->block_start) {
420 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
425 if (unlikely(delta > se->block_max))
426 se->block_max = delta;
429 se->sum_sleep_runtime += delta;
432 * Blocking time is in units of nanosecs, so shift by 20 to
433 * get a milliseconds-range estimation of the amount of
434 * time that the task spent sleeping:
436 if (unlikely(prof_on == SLEEP_PROFILING)) {
437 struct task_struct *tsk = task_of(se);
439 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
446 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
448 #ifdef CONFIG_SCHED_DEBUG
449 s64 d = se->vruntime - cfs_rq->min_vruntime;
454 if (d > 3*sysctl_sched_latency)
455 schedstat_inc(cfs_rq, nr_spread_over);
460 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
464 vruntime = cfs_rq->min_vruntime;
466 if (sched_feat(TREE_AVG)) {
467 struct sched_entity *last = __pick_last_entity(cfs_rq);
469 vruntime += last->vruntime;
472 } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running)
473 vruntime += sched_vslice(cfs_rq)/2;
475 if (initial && sched_feat(START_DEBIT))
476 vruntime += sched_vslice_add(cfs_rq, se);
479 if (sched_feat(NEW_FAIR_SLEEPERS) && entity_is_task(se) &&
480 task_of(se)->policy != SCHED_BATCH)
481 vruntime -= sysctl_sched_latency;
483 vruntime = max_t(s64, vruntime, se->vruntime);
486 se->vruntime = vruntime;
491 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
494 * Update run-time statistics of the 'current'.
499 place_entity(cfs_rq, se, 0);
500 enqueue_sleeper(cfs_rq, se);
503 update_stats_enqueue(cfs_rq, se);
504 check_spread(cfs_rq, se);
505 if (se != cfs_rq->curr)
506 __enqueue_entity(cfs_rq, se);
507 account_entity_enqueue(cfs_rq, se);
511 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
514 * Update run-time statistics of the 'current'.
518 update_stats_dequeue(cfs_rq, se);
520 se->peer_preempt = 0;
521 #ifdef CONFIG_SCHEDSTATS
522 if (entity_is_task(se)) {
523 struct task_struct *tsk = task_of(se);
525 if (tsk->state & TASK_INTERRUPTIBLE)
526 se->sleep_start = rq_of(cfs_rq)->clock;
527 if (tsk->state & TASK_UNINTERRUPTIBLE)
528 se->block_start = rq_of(cfs_rq)->clock;
533 if (se != cfs_rq->curr)
534 __dequeue_entity(cfs_rq, se);
535 account_entity_dequeue(cfs_rq, se);
539 * Preempt the current task with a newly woken task if needed:
542 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
544 unsigned long ideal_runtime, delta_exec;
546 ideal_runtime = sched_slice(cfs_rq, curr);
547 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
548 if (delta_exec > ideal_runtime ||
549 (sched_feat(PREEMPT_RESTRICT) && curr->peer_preempt))
550 resched_task(rq_of(cfs_rq)->curr);
551 curr->peer_preempt = 0;
555 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
557 /* 'current' is not kept within the tree. */
560 * Any task has to be enqueued before it get to execute on
561 * a CPU. So account for the time it spent waiting on the
564 update_stats_wait_end(cfs_rq, se);
565 __dequeue_entity(cfs_rq, se);
568 update_stats_curr_start(cfs_rq, se);
570 #ifdef CONFIG_SCHEDSTATS
572 * Track our maximum slice length, if the CPU's load is at
573 * least twice that of our own weight (i.e. dont track it
574 * when there are only lesser-weight tasks around):
576 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
577 se->slice_max = max(se->slice_max,
578 se->sum_exec_runtime - se->prev_sum_exec_runtime);
581 se->prev_sum_exec_runtime = se->sum_exec_runtime;
584 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
586 struct sched_entity *se = NULL;
588 if (first_fair(cfs_rq)) {
589 se = __pick_next_entity(cfs_rq);
590 set_next_entity(cfs_rq, se);
596 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
599 * If still on the runqueue then deactivate_task()
600 * was not called and update_curr() has to be done:
605 check_spread(cfs_rq, prev);
607 update_stats_wait_start(cfs_rq, prev);
608 /* Put 'current' back into the tree. */
609 __enqueue_entity(cfs_rq, prev);
614 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
617 * Update run-time statistics of the 'current'.
621 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
622 check_preempt_tick(cfs_rq, curr);
625 /**************************************************
626 * CFS operations on tasks:
629 #ifdef CONFIG_FAIR_GROUP_SCHED
631 /* Walk up scheduling entities hierarchy */
632 #define for_each_sched_entity(se) \
633 for (; se; se = se->parent)
635 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
640 /* runqueue on which this entity is (to be) queued */
641 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
646 /* runqueue "owned" by this group */
647 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
652 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
653 * another cpu ('this_cpu')
655 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
657 return cfs_rq->tg->cfs_rq[this_cpu];
660 /* Iterate thr' all leaf cfs_rq's on a runqueue */
661 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
662 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
664 /* Do the two (enqueued) entities belong to the same group ? */
666 is_same_group(struct sched_entity *se, struct sched_entity *pse)
668 if (se->cfs_rq == pse->cfs_rq)
674 static inline struct sched_entity *parent_entity(struct sched_entity *se)
679 #else /* CONFIG_FAIR_GROUP_SCHED */
681 #define for_each_sched_entity(se) \
682 for (; se; se = NULL)
684 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
686 return &task_rq(p)->cfs;
689 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
691 struct task_struct *p = task_of(se);
692 struct rq *rq = task_rq(p);
697 /* runqueue "owned" by this group */
698 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
703 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
705 return &cpu_rq(this_cpu)->cfs;
708 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
709 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
712 is_same_group(struct sched_entity *se, struct sched_entity *pse)
717 static inline struct sched_entity *parent_entity(struct sched_entity *se)
722 #endif /* CONFIG_FAIR_GROUP_SCHED */
725 * The enqueue_task method is called before nr_running is
726 * increased. Here we update the fair scheduling stats and
727 * then put the task into the rbtree:
729 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
731 struct cfs_rq *cfs_rq;
732 struct sched_entity *se = &p->se;
734 for_each_sched_entity(se) {
737 cfs_rq = cfs_rq_of(se);
738 enqueue_entity(cfs_rq, se, wakeup);
744 * The dequeue_task method is called before nr_running is
745 * decreased. We remove the task from the rbtree and
746 * update the fair scheduling stats:
748 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
750 struct cfs_rq *cfs_rq;
751 struct sched_entity *se = &p->se;
753 for_each_sched_entity(se) {
754 cfs_rq = cfs_rq_of(se);
755 dequeue_entity(cfs_rq, se, sleep);
756 /* Don't dequeue parent if it has other entities besides us */
757 if (cfs_rq->load.weight)
764 * sched_yield() support is very simple - we dequeue and enqueue.
766 * If compat_yield is turned on then we requeue to the end of the tree.
768 static void yield_task_fair(struct rq *rq)
770 struct cfs_rq *cfs_rq = task_cfs_rq(rq->curr);
771 struct sched_entity *rightmost, *se = &rq->curr->se;
774 * Are we the only task in the tree?
776 if (unlikely(cfs_rq->nr_running == 1))
779 if (likely(!sysctl_sched_compat_yield)) {
780 __update_rq_clock(rq);
782 * Update run-time statistics of the 'current'.
789 * Find the rightmost entry in the rbtree:
791 rightmost = __pick_last_entity(cfs_rq);
793 * Already in the rightmost position?
795 if (unlikely(rightmost->vruntime < se->vruntime))
799 * Minimally necessary key value to be last in the tree:
800 * Upon rescheduling, sched_class::put_prev_task() will place
801 * 'current' within the tree based on its new key value.
803 se->vruntime = rightmost->vruntime + 1;
807 * Preempt the current task with a newly woken task if needed:
809 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
811 struct task_struct *curr = rq->curr;
812 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
813 struct sched_entity *se = &curr->se, *pse = &p->se;
816 if (unlikely(rt_prio(p->prio))) {
823 * Batch tasks do not preempt (their preemption is driven by
826 if (unlikely(p->policy == SCHED_BATCH))
829 if (sched_feat(WAKEUP_PREEMPT)) {
830 while (!is_same_group(se, pse)) {
831 se = parent_entity(se);
832 pse = parent_entity(pse);
835 delta = se->vruntime - pse->vruntime;
836 gran = sysctl_sched_wakeup_granularity;
837 if (unlikely(se->load.weight != NICE_0_LOAD))
838 gran = calc_delta_fair(gran, &se->load);
841 int now = !sched_feat(PREEMPT_RESTRICT);
843 if (now || p->prio < curr->prio || !se->peer_preempt++)
849 static struct task_struct *pick_next_task_fair(struct rq *rq)
851 struct cfs_rq *cfs_rq = &rq->cfs;
852 struct sched_entity *se;
854 if (unlikely(!cfs_rq->nr_running))
858 se = pick_next_entity(cfs_rq);
859 cfs_rq = group_cfs_rq(se);
866 * Account for a descheduled task:
868 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
870 struct sched_entity *se = &prev->se;
871 struct cfs_rq *cfs_rq;
873 for_each_sched_entity(se) {
874 cfs_rq = cfs_rq_of(se);
875 put_prev_entity(cfs_rq, se);
879 /**************************************************
880 * Fair scheduling class load-balancing methods:
884 * Load-balancing iterator. Note: while the runqueue stays locked
885 * during the whole iteration, the current task might be
886 * dequeued so the iterator has to be dequeue-safe. Here we
887 * achieve that by always pre-iterating before returning
890 static struct task_struct *
891 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
893 struct task_struct *p;
898 p = rb_entry(curr, struct task_struct, se.run_node);
899 cfs_rq->rb_load_balance_curr = rb_next(curr);
904 static struct task_struct *load_balance_start_fair(void *arg)
906 struct cfs_rq *cfs_rq = arg;
908 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
911 static struct task_struct *load_balance_next_fair(void *arg)
913 struct cfs_rq *cfs_rq = arg;
915 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
918 #ifdef CONFIG_FAIR_GROUP_SCHED
919 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
921 struct sched_entity *curr;
922 struct task_struct *p;
924 if (!cfs_rq->nr_running)
929 curr = __pick_next_entity(cfs_rq);
938 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
939 unsigned long max_nr_move, unsigned long max_load_move,
940 struct sched_domain *sd, enum cpu_idle_type idle,
941 int *all_pinned, int *this_best_prio)
943 struct cfs_rq *busy_cfs_rq;
944 unsigned long load_moved, total_nr_moved = 0, nr_moved;
945 long rem_load_move = max_load_move;
946 struct rq_iterator cfs_rq_iterator;
948 cfs_rq_iterator.start = load_balance_start_fair;
949 cfs_rq_iterator.next = load_balance_next_fair;
951 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
952 #ifdef CONFIG_FAIR_GROUP_SCHED
953 struct cfs_rq *this_cfs_rq;
955 unsigned long maxload;
957 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
959 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
960 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
964 /* Don't pull more than imbalance/2 */
966 maxload = min(rem_load_move, imbalance);
968 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
970 # define maxload rem_load_move
972 /* pass busy_cfs_rq argument into
973 * load_balance_[start|next]_fair iterators
975 cfs_rq_iterator.arg = busy_cfs_rq;
976 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
977 max_nr_move, maxload, sd, idle, all_pinned,
978 &load_moved, this_best_prio, &cfs_rq_iterator);
980 total_nr_moved += nr_moved;
981 max_nr_move -= nr_moved;
982 rem_load_move -= load_moved;
984 if (max_nr_move <= 0 || rem_load_move <= 0)
988 return max_load_move - rem_load_move;
992 * scheduler tick hitting a task of our scheduling class:
994 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
996 struct cfs_rq *cfs_rq;
997 struct sched_entity *se = &curr->se;
999 for_each_sched_entity(se) {
1000 cfs_rq = cfs_rq_of(se);
1001 entity_tick(cfs_rq, se);
1005 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1008 * Share the fairness runtime between parent and child, thus the
1009 * total amount of pressure for CPU stays equal - new tasks
1010 * get a chance to run but frequent forkers are not allowed to
1011 * monopolize the CPU. Note: the parent runqueue is locked,
1012 * the child is not running yet.
1014 static void task_new_fair(struct rq *rq, struct task_struct *p)
1016 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1017 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1018 int this_cpu = smp_processor_id();
1020 sched_info_queued(p);
1022 update_curr(cfs_rq);
1023 place_entity(cfs_rq, se, 1);
1025 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1026 curr->vruntime < se->vruntime) {
1028 * Upon rescheduling, sched_class::put_prev_task() will place
1029 * 'current' within the tree based on its new key value.
1031 swap(curr->vruntime, se->vruntime);
1034 update_stats_enqueue(cfs_rq, se);
1035 check_spread(cfs_rq, se);
1036 check_spread(cfs_rq, curr);
1037 __enqueue_entity(cfs_rq, se);
1038 account_entity_enqueue(cfs_rq, se);
1039 se->peer_preempt = 0;
1040 resched_task(rq->curr);
1043 /* Account for a task changing its policy or group.
1045 * This routine is mostly called to set cfs_rq->curr field when a task
1046 * migrates between groups/classes.
1048 static void set_curr_task_fair(struct rq *rq)
1050 struct sched_entity *se = &rq->curr->se;
1052 for_each_sched_entity(se)
1053 set_next_entity(cfs_rq_of(se), se);
1057 * All the scheduling class methods:
1059 static const struct sched_class fair_sched_class = {
1060 .next = &idle_sched_class,
1061 .enqueue_task = enqueue_task_fair,
1062 .dequeue_task = dequeue_task_fair,
1063 .yield_task = yield_task_fair,
1065 .check_preempt_curr = check_preempt_wakeup,
1067 .pick_next_task = pick_next_task_fair,
1068 .put_prev_task = put_prev_task_fair,
1070 .load_balance = load_balance_fair,
1072 .set_curr_task = set_curr_task_fair,
1073 .task_tick = task_tick_fair,
1074 .task_new = task_new_fair,
1077 #ifdef CONFIG_SCHED_DEBUG
1078 static void print_cfs_stats(struct seq_file *m, int cpu)
1080 struct cfs_rq *cfs_rq;
1082 #ifdef CONFIG_FAIR_GROUP_SCHED
1083 print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs);
1085 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1086 print_cfs_rq(m, cpu, cfs_rq);