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 vslice *= NICE_0_LOAD;
263 do_div(vslice, rq_weight);
268 static u64 sched_vslice(struct cfs_rq *cfs_rq)
270 return __sched_vslice(cfs_rq->load.weight, cfs_rq->nr_running);
273 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
275 return __sched_vslice(cfs_rq->load.weight + se->load.weight,
276 cfs_rq->nr_running + 1);
280 * Update the current task's runtime statistics. Skip current tasks that
281 * are not in our scheduling class.
284 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
285 unsigned long delta_exec)
287 unsigned long delta_exec_weighted;
290 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
292 curr->sum_exec_runtime += delta_exec;
293 schedstat_add(cfs_rq, exec_clock, delta_exec);
294 delta_exec_weighted = delta_exec;
295 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
296 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
299 curr->vruntime += delta_exec_weighted;
302 * maintain cfs_rq->min_vruntime to be a monotonic increasing
303 * value tracking the leftmost vruntime in the tree.
305 if (first_fair(cfs_rq)) {
306 vruntime = min_vruntime(curr->vruntime,
307 __pick_next_entity(cfs_rq)->vruntime);
309 vruntime = curr->vruntime;
311 cfs_rq->min_vruntime =
312 max_vruntime(cfs_rq->min_vruntime, vruntime);
315 static void update_curr(struct cfs_rq *cfs_rq)
317 struct sched_entity *curr = cfs_rq->curr;
318 u64 now = rq_of(cfs_rq)->clock;
319 unsigned long delta_exec;
325 * Get the amount of time the current task was running
326 * since the last time we changed load (this cannot
327 * overflow on 32 bits):
329 delta_exec = (unsigned long)(now - curr->exec_start);
331 __update_curr(cfs_rq, curr, delta_exec);
332 curr->exec_start = now;
336 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
338 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
342 * Task is being enqueued - update stats:
344 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
347 * Are we enqueueing a waiting task? (for current tasks
348 * a dequeue/enqueue event is a NOP)
350 if (se != cfs_rq->curr)
351 update_stats_wait_start(cfs_rq, se);
355 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
357 schedstat_set(se->wait_max, max(se->wait_max,
358 rq_of(cfs_rq)->clock - se->wait_start));
359 schedstat_set(se->wait_start, 0);
363 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
366 * Mark the end of the wait period if dequeueing a
369 if (se != cfs_rq->curr)
370 update_stats_wait_end(cfs_rq, se);
374 * We are picking a new current task - update its stats:
377 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
380 * We are starting a new run period:
382 se->exec_start = rq_of(cfs_rq)->clock;
385 /**************************************************
386 * Scheduling class queueing methods:
390 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
392 update_load_add(&cfs_rq->load, se->load.weight);
393 cfs_rq->nr_running++;
398 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
400 update_load_sub(&cfs_rq->load, se->load.weight);
401 cfs_rq->nr_running--;
405 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
407 #ifdef CONFIG_SCHEDSTATS
408 if (se->sleep_start) {
409 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
414 if (unlikely(delta > se->sleep_max))
415 se->sleep_max = delta;
418 se->sum_sleep_runtime += delta;
420 if (se->block_start) {
421 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
426 if (unlikely(delta > se->block_max))
427 se->block_max = delta;
430 se->sum_sleep_runtime += delta;
433 * Blocking time is in units of nanosecs, so shift by 20 to
434 * get a milliseconds-range estimation of the amount of
435 * time that the task spent sleeping:
437 if (unlikely(prof_on == SLEEP_PROFILING)) {
438 struct task_struct *tsk = task_of(se);
440 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
447 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
449 #ifdef CONFIG_SCHED_DEBUG
450 s64 d = se->vruntime - cfs_rq->min_vruntime;
455 if (d > 3*sysctl_sched_latency)
456 schedstat_inc(cfs_rq, nr_spread_over);
461 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
465 vruntime = cfs_rq->min_vruntime;
467 if (sched_feat(TREE_AVG)) {
468 struct sched_entity *last = __pick_last_entity(cfs_rq);
470 vruntime += last->vruntime;
473 } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running)
474 vruntime += sched_vslice(cfs_rq)/2;
477 * The 'current' period is already promised to the current tasks,
478 * however the extra weight of the new task will slow them down a
479 * little, place the new task so that it fits in the slot that
480 * stays open at the end.
482 if (initial && sched_feat(START_DEBIT))
483 vruntime += sched_vslice_add(cfs_rq, se);
486 /* sleeps upto a single latency don't count. */
487 if (sched_feat(NEW_FAIR_SLEEPERS) && entity_is_task(se) &&
488 task_of(se)->policy != SCHED_BATCH)
489 vruntime -= sysctl_sched_latency;
491 /* ensure we never gain time by being placed backwards. */
492 vruntime = max_vruntime(se->vruntime, vruntime);
495 se->vruntime = vruntime;
499 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
502 * Update run-time statistics of the 'current'.
507 place_entity(cfs_rq, se, 0);
508 enqueue_sleeper(cfs_rq, se);
511 update_stats_enqueue(cfs_rq, se);
512 check_spread(cfs_rq, se);
513 if (se != cfs_rq->curr)
514 __enqueue_entity(cfs_rq, se);
515 account_entity_enqueue(cfs_rq, se);
519 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
522 * Update run-time statistics of the 'current'.
526 update_stats_dequeue(cfs_rq, se);
528 se->peer_preempt = 0;
529 #ifdef CONFIG_SCHEDSTATS
530 if (entity_is_task(se)) {
531 struct task_struct *tsk = task_of(se);
533 if (tsk->state & TASK_INTERRUPTIBLE)
534 se->sleep_start = rq_of(cfs_rq)->clock;
535 if (tsk->state & TASK_UNINTERRUPTIBLE)
536 se->block_start = rq_of(cfs_rq)->clock;
541 if (se != cfs_rq->curr)
542 __dequeue_entity(cfs_rq, se);
543 account_entity_dequeue(cfs_rq, se);
547 * Preempt the current task with a newly woken task if needed:
550 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
552 unsigned long ideal_runtime, delta_exec;
554 ideal_runtime = sched_slice(cfs_rq, curr);
555 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
556 if (delta_exec > ideal_runtime ||
557 (sched_feat(PREEMPT_RESTRICT) && curr->peer_preempt))
558 resched_task(rq_of(cfs_rq)->curr);
559 curr->peer_preempt = 0;
563 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
565 /* 'current' is not kept within the tree. */
568 * Any task has to be enqueued before it get to execute on
569 * a CPU. So account for the time it spent waiting on the
572 update_stats_wait_end(cfs_rq, se);
573 __dequeue_entity(cfs_rq, se);
576 update_stats_curr_start(cfs_rq, se);
578 #ifdef CONFIG_SCHEDSTATS
580 * Track our maximum slice length, if the CPU's load is at
581 * least twice that of our own weight (i.e. dont track it
582 * when there are only lesser-weight tasks around):
584 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
585 se->slice_max = max(se->slice_max,
586 se->sum_exec_runtime - se->prev_sum_exec_runtime);
589 se->prev_sum_exec_runtime = se->sum_exec_runtime;
592 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
594 struct sched_entity *se = NULL;
596 if (first_fair(cfs_rq)) {
597 se = __pick_next_entity(cfs_rq);
598 set_next_entity(cfs_rq, se);
604 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
607 * If still on the runqueue then deactivate_task()
608 * was not called and update_curr() has to be done:
613 check_spread(cfs_rq, prev);
615 update_stats_wait_start(cfs_rq, prev);
616 /* Put 'current' back into the tree. */
617 __enqueue_entity(cfs_rq, prev);
622 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
625 * Update run-time statistics of the 'current'.
629 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
630 check_preempt_tick(cfs_rq, curr);
633 /**************************************************
634 * CFS operations on tasks:
637 #ifdef CONFIG_FAIR_GROUP_SCHED
639 /* Walk up scheduling entities hierarchy */
640 #define for_each_sched_entity(se) \
641 for (; se; se = se->parent)
643 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
648 /* runqueue on which this entity is (to be) queued */
649 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
654 /* runqueue "owned" by this group */
655 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
660 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
661 * another cpu ('this_cpu')
663 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
665 return cfs_rq->tg->cfs_rq[this_cpu];
668 /* Iterate thr' all leaf cfs_rq's on a runqueue */
669 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
670 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
672 /* Do the two (enqueued) entities belong to the same group ? */
674 is_same_group(struct sched_entity *se, struct sched_entity *pse)
676 if (se->cfs_rq == pse->cfs_rq)
682 static inline struct sched_entity *parent_entity(struct sched_entity *se)
687 #else /* CONFIG_FAIR_GROUP_SCHED */
689 #define for_each_sched_entity(se) \
690 for (; se; se = NULL)
692 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
694 return &task_rq(p)->cfs;
697 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
699 struct task_struct *p = task_of(se);
700 struct rq *rq = task_rq(p);
705 /* runqueue "owned" by this group */
706 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
711 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
713 return &cpu_rq(this_cpu)->cfs;
716 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
717 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
720 is_same_group(struct sched_entity *se, struct sched_entity *pse)
725 static inline struct sched_entity *parent_entity(struct sched_entity *se)
730 #endif /* CONFIG_FAIR_GROUP_SCHED */
733 * The enqueue_task method is called before nr_running is
734 * increased. Here we update the fair scheduling stats and
735 * then put the task into the rbtree:
737 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
739 struct cfs_rq *cfs_rq;
740 struct sched_entity *se = &p->se;
742 for_each_sched_entity(se) {
745 cfs_rq = cfs_rq_of(se);
746 enqueue_entity(cfs_rq, se, wakeup);
752 * The dequeue_task method is called before nr_running is
753 * decreased. We remove the task from the rbtree and
754 * update the fair scheduling stats:
756 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
758 struct cfs_rq *cfs_rq;
759 struct sched_entity *se = &p->se;
761 for_each_sched_entity(se) {
762 cfs_rq = cfs_rq_of(se);
763 dequeue_entity(cfs_rq, se, sleep);
764 /* Don't dequeue parent if it has other entities besides us */
765 if (cfs_rq->load.weight)
772 * sched_yield() support is very simple - we dequeue and enqueue.
774 * If compat_yield is turned on then we requeue to the end of the tree.
776 static void yield_task_fair(struct rq *rq)
778 struct cfs_rq *cfs_rq = task_cfs_rq(rq->curr);
779 struct sched_entity *rightmost, *se = &rq->curr->se;
782 * Are we the only task in the tree?
784 if (unlikely(cfs_rq->nr_running == 1))
787 if (likely(!sysctl_sched_compat_yield)) {
788 __update_rq_clock(rq);
790 * Update run-time statistics of the 'current'.
797 * Find the rightmost entry in the rbtree:
799 rightmost = __pick_last_entity(cfs_rq);
801 * Already in the rightmost position?
803 if (unlikely(rightmost->vruntime < se->vruntime))
807 * Minimally necessary key value to be last in the tree:
808 * Upon rescheduling, sched_class::put_prev_task() will place
809 * 'current' within the tree based on its new key value.
811 se->vruntime = rightmost->vruntime + 1;
815 * Preempt the current task with a newly woken task if needed:
817 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
819 struct task_struct *curr = rq->curr;
820 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
821 struct sched_entity *se = &curr->se, *pse = &p->se;
824 if (unlikely(rt_prio(p->prio))) {
831 * Batch tasks do not preempt (their preemption is driven by
834 if (unlikely(p->policy == SCHED_BATCH))
837 if (sched_feat(WAKEUP_PREEMPT)) {
838 while (!is_same_group(se, pse)) {
839 se = parent_entity(se);
840 pse = parent_entity(pse);
843 delta = se->vruntime - pse->vruntime;
844 gran = sysctl_sched_wakeup_granularity;
845 if (unlikely(se->load.weight != NICE_0_LOAD))
846 gran = calc_delta_fair(gran, &se->load);
849 int now = !sched_feat(PREEMPT_RESTRICT);
851 if (now || p->prio < curr->prio || !se->peer_preempt++)
857 static struct task_struct *pick_next_task_fair(struct rq *rq)
859 struct cfs_rq *cfs_rq = &rq->cfs;
860 struct sched_entity *se;
862 if (unlikely(!cfs_rq->nr_running))
866 se = pick_next_entity(cfs_rq);
867 cfs_rq = group_cfs_rq(se);
874 * Account for a descheduled task:
876 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
878 struct sched_entity *se = &prev->se;
879 struct cfs_rq *cfs_rq;
881 for_each_sched_entity(se) {
882 cfs_rq = cfs_rq_of(se);
883 put_prev_entity(cfs_rq, se);
888 /**************************************************
889 * Fair scheduling class load-balancing methods:
893 * Load-balancing iterator. Note: while the runqueue stays locked
894 * during the whole iteration, the current task might be
895 * dequeued so the iterator has to be dequeue-safe. Here we
896 * achieve that by always pre-iterating before returning
899 static struct task_struct *
900 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
902 struct task_struct *p;
907 p = rb_entry(curr, struct task_struct, se.run_node);
908 cfs_rq->rb_load_balance_curr = rb_next(curr);
913 static struct task_struct *load_balance_start_fair(void *arg)
915 struct cfs_rq *cfs_rq = arg;
917 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
920 static struct task_struct *load_balance_next_fair(void *arg)
922 struct cfs_rq *cfs_rq = arg;
924 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
927 #ifdef CONFIG_FAIR_GROUP_SCHED
928 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
930 struct sched_entity *curr;
931 struct task_struct *p;
933 if (!cfs_rq->nr_running)
938 curr = __pick_next_entity(cfs_rq);
947 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
948 unsigned long max_load_move,
949 struct sched_domain *sd, enum cpu_idle_type idle,
950 int *all_pinned, int *this_best_prio)
952 struct cfs_rq *busy_cfs_rq;
953 long rem_load_move = max_load_move;
954 struct rq_iterator cfs_rq_iterator;
956 cfs_rq_iterator.start = load_balance_start_fair;
957 cfs_rq_iterator.next = load_balance_next_fair;
959 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
960 #ifdef CONFIG_FAIR_GROUP_SCHED
961 struct cfs_rq *this_cfs_rq;
963 unsigned long maxload;
965 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
967 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
968 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
972 /* Don't pull more than imbalance/2 */
974 maxload = min(rem_load_move, imbalance);
976 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
978 # define maxload rem_load_move
981 * pass busy_cfs_rq argument into
982 * load_balance_[start|next]_fair iterators
984 cfs_rq_iterator.arg = busy_cfs_rq;
985 rem_load_move -= balance_tasks(this_rq, this_cpu, busiest,
986 maxload, sd, idle, all_pinned,
990 if (rem_load_move <= 0)
994 return max_load_move - rem_load_move;
998 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
999 struct sched_domain *sd, enum cpu_idle_type idle)
1001 struct cfs_rq *busy_cfs_rq;
1002 struct rq_iterator cfs_rq_iterator;
1004 cfs_rq_iterator.start = load_balance_start_fair;
1005 cfs_rq_iterator.next = load_balance_next_fair;
1007 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1009 * pass busy_cfs_rq argument into
1010 * load_balance_[start|next]_fair iterators
1012 cfs_rq_iterator.arg = busy_cfs_rq;
1013 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1023 * scheduler tick hitting a task of our scheduling class:
1025 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1027 struct cfs_rq *cfs_rq;
1028 struct sched_entity *se = &curr->se;
1030 for_each_sched_entity(se) {
1031 cfs_rq = cfs_rq_of(se);
1032 entity_tick(cfs_rq, se);
1036 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1039 * Share the fairness runtime between parent and child, thus the
1040 * total amount of pressure for CPU stays equal - new tasks
1041 * get a chance to run but frequent forkers are not allowed to
1042 * monopolize the CPU. Note: the parent runqueue is locked,
1043 * the child is not running yet.
1045 static void task_new_fair(struct rq *rq, struct task_struct *p)
1047 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1048 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1049 int this_cpu = smp_processor_id();
1051 sched_info_queued(p);
1053 update_curr(cfs_rq);
1054 place_entity(cfs_rq, se, 1);
1056 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1057 curr->vruntime < se->vruntime) {
1059 * Upon rescheduling, sched_class::put_prev_task() will place
1060 * 'current' within the tree based on its new key value.
1062 swap(curr->vruntime, se->vruntime);
1065 se->peer_preempt = 0;
1066 enqueue_task_fair(rq, p, 0);
1067 resched_task(rq->curr);
1070 /* Account for a task changing its policy or group.
1072 * This routine is mostly called to set cfs_rq->curr field when a task
1073 * migrates between groups/classes.
1075 static void set_curr_task_fair(struct rq *rq)
1077 struct sched_entity *se = &rq->curr->se;
1079 for_each_sched_entity(se)
1080 set_next_entity(cfs_rq_of(se), se);
1084 * All the scheduling class methods:
1086 static const struct sched_class fair_sched_class = {
1087 .next = &idle_sched_class,
1088 .enqueue_task = enqueue_task_fair,
1089 .dequeue_task = dequeue_task_fair,
1090 .yield_task = yield_task_fair,
1092 .check_preempt_curr = check_preempt_wakeup,
1094 .pick_next_task = pick_next_task_fair,
1095 .put_prev_task = put_prev_task_fair,
1098 .load_balance = load_balance_fair,
1099 .move_one_task = move_one_task_fair,
1102 .set_curr_task = set_curr_task_fair,
1103 .task_tick = task_tick_fair,
1104 .task_new = task_new_fair,
1107 #ifdef CONFIG_SCHED_DEBUG
1108 static void print_cfs_stats(struct seq_file *m, int cpu)
1110 struct cfs_rq *cfs_rq;
1112 #ifdef CONFIG_FAIR_GROUP_SCHED
1113 print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs);
1115 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1116 print_cfs_rq(m, cpu, cfs_rq);