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 * (to see the precise effective timeslice length of your workload,
30 * run vmstat and monitor the context-switches field)
32 * On SMP systems the value of this is multiplied by the log2 of the
33 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
34 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
35 * Targeted preemption latency for CPU-bound tasks:
37 const_debug unsigned int sysctl_sched_latency = 20000000ULL;
40 * After fork, child runs first. (default) If set to 0 then
41 * parent will (try to) run first.
43 const_debug unsigned int sysctl_sched_child_runs_first = 1;
46 * Minimal preemption granularity for CPU-bound tasks:
47 * (default: 2 msec, units: nanoseconds)
49 unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL;
52 * sys_sched_yield() compat mode
54 * This option switches the agressive yield implementation of the
55 * old scheduler back on.
57 unsigned int __read_mostly sysctl_sched_compat_yield;
60 * SCHED_BATCH wake-up granularity.
61 * (default: 25 msec, units: nanoseconds)
63 * This option delays the preemption effects of decoupled workloads
64 * and reduces their over-scheduling. Synchronous workloads will still
65 * have immediate wakeup/sleep latencies.
67 const_debug unsigned int sysctl_sched_batch_wakeup_granularity = 25000000UL;
70 * SCHED_OTHER wake-up granularity.
71 * (default: 1 msec, units: nanoseconds)
73 * This option delays the preemption effects of decoupled workloads
74 * and reduces their over-scheduling. Synchronous workloads will still
75 * have immediate wakeup/sleep latencies.
77 const_debug unsigned int sysctl_sched_wakeup_granularity = 2000000UL;
79 extern struct sched_class fair_sched_class;
81 /**************************************************************
82 * CFS operations on generic schedulable entities:
85 #ifdef CONFIG_FAIR_GROUP_SCHED
87 /* cpu runqueue to which this cfs_rq is attached */
88 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
93 /* An entity is a task if it doesn't "own" a runqueue */
94 #define entity_is_task(se) (!se->my_q)
96 #else /* CONFIG_FAIR_GROUP_SCHED */
98 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
100 return container_of(cfs_rq, struct rq, cfs);
103 #define entity_is_task(se) 1
105 #endif /* CONFIG_FAIR_GROUP_SCHED */
107 static inline struct task_struct *task_of(struct sched_entity *se)
109 return container_of(se, struct task_struct, se);
113 /**************************************************************
114 * Scheduling class tree data structure manipulation methods:
118 max_vruntime(u64 min_vruntime, u64 vruntime)
120 s64 delta = (s64)(vruntime - min_vruntime);
122 min_vruntime = vruntime;
128 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
130 return se->vruntime - cfs_rq->min_vruntime;
134 * Enqueue an entity into the rb-tree:
137 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
139 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
140 struct rb_node *parent = NULL;
141 struct sched_entity *entry;
142 s64 key = entity_key(cfs_rq, se);
146 * Find the right place in the rbtree:
150 entry = rb_entry(parent, struct sched_entity, run_node);
152 * We dont care about collisions. Nodes with
153 * the same key stay together.
155 if (key < entity_key(cfs_rq, entry)) {
156 link = &parent->rb_left;
158 link = &parent->rb_right;
164 * Maintain a cache of leftmost tree entries (it is frequently
168 cfs_rq->rb_leftmost = &se->run_node;
170 rb_link_node(&se->run_node, parent, link);
171 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
175 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
177 if (cfs_rq->rb_leftmost == &se->run_node)
178 cfs_rq->rb_leftmost = rb_next(&se->run_node);
180 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
183 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
185 return cfs_rq->rb_leftmost;
188 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
190 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
193 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
195 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
196 struct sched_entity *se = NULL;
197 struct rb_node *parent;
201 se = rb_entry(parent, struct sched_entity, run_node);
202 link = &parent->rb_right;
208 /**************************************************************
209 * Scheduling class statistics methods:
212 static u64 __sched_period(unsigned long nr_running)
214 u64 period = sysctl_sched_latency;
215 unsigned long nr_latency =
216 sysctl_sched_latency / sysctl_sched_min_granularity;
218 if (unlikely(nr_running > nr_latency)) {
219 period *= nr_running;
220 do_div(period, nr_latency);
226 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
228 u64 period = __sched_period(cfs_rq->nr_running);
230 period *= se->load.weight;
231 do_div(period, cfs_rq->load.weight);
236 static u64 __sched_vslice(unsigned long nr_running)
238 u64 period = __sched_period(nr_running);
240 do_div(period, nr_running);
246 * Update the current task's runtime statistics. Skip current tasks that
247 * are not in our scheduling class.
250 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
251 unsigned long delta_exec)
253 unsigned long delta_exec_weighted;
254 u64 next_vruntime, min_vruntime;
256 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
258 curr->sum_exec_runtime += delta_exec;
259 schedstat_add(cfs_rq, exec_clock, delta_exec);
260 delta_exec_weighted = delta_exec;
261 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
262 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
265 curr->vruntime += delta_exec_weighted;
268 * maintain cfs_rq->min_vruntime to be a monotonic increasing
269 * value tracking the leftmost vruntime in the tree.
271 if (first_fair(cfs_rq)) {
272 next_vruntime = __pick_next_entity(cfs_rq)->vruntime;
274 /* min_vruntime() := !max_vruntime() */
275 min_vruntime = max_vruntime(curr->vruntime, next_vruntime);
276 if (min_vruntime == next_vruntime)
277 min_vruntime = curr->vruntime;
279 min_vruntime = next_vruntime;
281 min_vruntime = curr->vruntime;
283 cfs_rq->min_vruntime =
284 max_vruntime(cfs_rq->min_vruntime, min_vruntime);
287 static void update_curr(struct cfs_rq *cfs_rq)
289 struct sched_entity *curr = cfs_rq->curr;
290 u64 now = rq_of(cfs_rq)->clock;
291 unsigned long delta_exec;
297 * Get the amount of time the current task was running
298 * since the last time we changed load (this cannot
299 * overflow on 32 bits):
301 delta_exec = (unsigned long)(now - curr->exec_start);
303 __update_curr(cfs_rq, curr, delta_exec);
304 curr->exec_start = now;
308 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
310 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
313 static inline unsigned long
314 calc_weighted(unsigned long delta, struct sched_entity *se)
316 unsigned long weight = se->load.weight;
318 if (unlikely(weight != NICE_0_LOAD))
319 return (u64)delta * se->load.weight >> NICE_0_SHIFT;
325 * Task is being enqueued - update stats:
327 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
330 * Are we enqueueing a waiting task? (for current tasks
331 * a dequeue/enqueue event is a NOP)
333 if (se != cfs_rq->curr)
334 update_stats_wait_start(cfs_rq, se);
338 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
340 schedstat_set(se->wait_max, max(se->wait_max,
341 rq_of(cfs_rq)->clock - se->wait_start));
342 schedstat_set(se->wait_start, 0);
346 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
350 * Mark the end of the wait period if dequeueing a
353 if (se != cfs_rq->curr)
354 update_stats_wait_end(cfs_rq, se);
358 * We are picking a new current task - update its stats:
361 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
364 * We are starting a new run period:
366 se->exec_start = rq_of(cfs_rq)->clock;
370 * We are descheduling a task - update its stats:
373 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
378 /**************************************************
379 * Scheduling class queueing methods:
383 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
385 update_load_add(&cfs_rq->load, se->load.weight);
386 cfs_rq->nr_running++;
391 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
393 update_load_sub(&cfs_rq->load, se->load.weight);
394 cfs_rq->nr_running--;
398 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
400 #ifdef CONFIG_SCHEDSTATS
401 if (se->sleep_start) {
402 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
407 if (unlikely(delta > se->sleep_max))
408 se->sleep_max = delta;
411 se->sum_sleep_runtime += delta;
413 if (se->block_start) {
414 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
419 if (unlikely(delta > se->block_max))
420 se->block_max = delta;
423 se->sum_sleep_runtime += delta;
426 * Blocking time is in units of nanosecs, so shift by 20 to
427 * get a milliseconds-range estimation of the amount of
428 * time that the task spent sleeping:
430 if (unlikely(prof_on == SLEEP_PROFILING)) {
431 struct task_struct *tsk = task_of(se);
433 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
440 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
442 #ifdef CONFIG_SCHED_DEBUG
443 s64 d = se->vruntime - cfs_rq->min_vruntime;
448 if (d > 3*sysctl_sched_latency)
449 schedstat_inc(cfs_rq, nr_spread_over);
454 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
458 vruntime = cfs_rq->min_vruntime;
460 if (sched_feat(USE_TREE_AVG)) {
461 struct sched_entity *last = __pick_last_entity(cfs_rq);
463 vruntime += last->vruntime;
466 } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running)
467 vruntime += __sched_vslice(cfs_rq->nr_running)/2;
469 if (initial && sched_feat(START_DEBIT))
470 vruntime += __sched_vslice(cfs_rq->nr_running + 1);
473 if (sched_feat(NEW_FAIR_SLEEPERS))
474 vruntime -= sysctl_sched_latency;
476 vruntime = max_t(s64, vruntime, se->vruntime);
479 se->vruntime = vruntime;
484 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
487 * Update the fair clock.
492 /* se->vruntime += cfs_rq->min_vruntime; */
493 place_entity(cfs_rq, se, 0);
494 enqueue_sleeper(cfs_rq, se);
497 update_stats_enqueue(cfs_rq, se);
498 check_spread(cfs_rq, se);
499 if (se != cfs_rq->curr)
500 __enqueue_entity(cfs_rq, se);
501 account_entity_enqueue(cfs_rq, se);
505 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
507 update_stats_dequeue(cfs_rq, se);
509 #ifdef CONFIG_SCHEDSTATS
510 if (entity_is_task(se)) {
511 struct task_struct *tsk = task_of(se);
513 if (tsk->state & TASK_INTERRUPTIBLE)
514 se->sleep_start = rq_of(cfs_rq)->clock;
515 if (tsk->state & TASK_UNINTERRUPTIBLE)
516 se->block_start = rq_of(cfs_rq)->clock;
521 if (se != cfs_rq->curr)
522 __dequeue_entity(cfs_rq, se);
523 account_entity_dequeue(cfs_rq, se);
527 * Preempt the current task with a newly woken task if needed:
530 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
532 unsigned long ideal_runtime, delta_exec;
534 ideal_runtime = sched_slice(cfs_rq, curr);
535 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
536 if (delta_exec > ideal_runtime)
537 resched_task(rq_of(cfs_rq)->curr);
541 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 /* 'current' is not kept within the tree. */
546 * Any task has to be enqueued before it get to execute on
547 * a CPU. So account for the time it spent waiting on the
550 update_stats_wait_end(cfs_rq, se);
551 __dequeue_entity(cfs_rq, se);
554 update_stats_curr_start(cfs_rq, se);
556 #ifdef CONFIG_SCHEDSTATS
558 * Track our maximum slice length, if the CPU's load is at
559 * least twice that of our own weight (i.e. dont track it
560 * when there are only lesser-weight tasks around):
562 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
563 se->slice_max = max(se->slice_max,
564 se->sum_exec_runtime - se->prev_sum_exec_runtime);
567 se->prev_sum_exec_runtime = se->sum_exec_runtime;
570 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
572 struct sched_entity *se = __pick_next_entity(cfs_rq);
574 set_next_entity(cfs_rq, se);
579 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
582 * If still on the runqueue then deactivate_task()
583 * was not called and update_curr() has to be done:
588 update_stats_curr_end(cfs_rq, prev);
590 check_spread(cfs_rq, prev);
592 update_stats_wait_start(cfs_rq, prev);
593 /* Put 'current' back into the tree. */
594 __enqueue_entity(cfs_rq, prev);
599 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
602 * Update run-time statistics of the 'current'.
606 if (cfs_rq->nr_running > 1)
607 check_preempt_tick(cfs_rq, curr);
610 /**************************************************
611 * CFS operations on tasks:
614 #ifdef CONFIG_FAIR_GROUP_SCHED
616 /* Walk up scheduling entities hierarchy */
617 #define for_each_sched_entity(se) \
618 for (; se; se = se->parent)
620 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
625 /* runqueue on which this entity is (to be) queued */
626 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
631 /* runqueue "owned" by this group */
632 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
637 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
638 * another cpu ('this_cpu')
640 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
642 return cfs_rq->tg->cfs_rq[this_cpu];
645 /* Iterate thr' all leaf cfs_rq's on a runqueue */
646 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
647 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
649 /* Do the two (enqueued) tasks belong to the same group ? */
650 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
652 if (curr->se.cfs_rq == p->se.cfs_rq)
658 #else /* CONFIG_FAIR_GROUP_SCHED */
660 #define for_each_sched_entity(se) \
661 for (; se; se = NULL)
663 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
665 return &task_rq(p)->cfs;
668 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
670 struct task_struct *p = task_of(se);
671 struct rq *rq = task_rq(p);
676 /* runqueue "owned" by this group */
677 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
682 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
684 return &cpu_rq(this_cpu)->cfs;
687 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
688 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
690 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
695 #endif /* CONFIG_FAIR_GROUP_SCHED */
698 * The enqueue_task method is called before nr_running is
699 * increased. Here we update the fair scheduling stats and
700 * then put the task into the rbtree:
702 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
704 struct cfs_rq *cfs_rq;
705 struct sched_entity *se = &p->se;
707 for_each_sched_entity(se) {
710 cfs_rq = cfs_rq_of(se);
711 enqueue_entity(cfs_rq, se, wakeup);
716 * The dequeue_task method is called before nr_running is
717 * decreased. We remove the task from the rbtree and
718 * update the fair scheduling stats:
720 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
722 struct cfs_rq *cfs_rq;
723 struct sched_entity *se = &p->se;
725 for_each_sched_entity(se) {
726 cfs_rq = cfs_rq_of(se);
727 dequeue_entity(cfs_rq, se, sleep);
728 /* Don't dequeue parent if it has other entities besides us */
729 if (cfs_rq->load.weight)
735 * sched_yield() support is very simple - we dequeue and enqueue.
737 * If compat_yield is turned on then we requeue to the end of the tree.
739 static void yield_task_fair(struct rq *rq)
741 struct cfs_rq *cfs_rq = task_cfs_rq(rq->curr);
742 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
743 struct sched_entity *rightmost, *se = &rq->curr->se;
744 struct rb_node *parent;
747 * Are we the only task in the tree?
749 if (unlikely(cfs_rq->nr_running == 1))
752 if (likely(!sysctl_sched_compat_yield)) {
753 __update_rq_clock(rq);
755 * Dequeue and enqueue the task to update its
756 * position within the tree:
758 dequeue_entity(cfs_rq, se, 0);
759 enqueue_entity(cfs_rq, se, 0);
764 * Find the rightmost entry in the rbtree:
768 link = &parent->rb_right;
771 rightmost = rb_entry(parent, struct sched_entity, run_node);
773 * Already in the rightmost position?
775 if (unlikely(rightmost == se))
779 * Minimally necessary key value to be last in the tree:
781 se->vruntime = rightmost->vruntime + 1;
783 if (cfs_rq->rb_leftmost == &se->run_node)
784 cfs_rq->rb_leftmost = rb_next(&se->run_node);
786 * Relink the task to the rightmost position:
788 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
789 rb_link_node(&se->run_node, parent, link);
790 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
794 * Preempt the current task with a newly woken task if needed:
796 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
798 struct task_struct *curr = rq->curr;
799 struct cfs_rq *cfs_rq = task_cfs_rq(curr), *pcfs_rq;
800 struct sched_entity *se = &curr->se, *pse = &p->se;
802 if (unlikely(rt_prio(p->prio))) {
809 for_each_sched_entity(se) {
810 cfs_rq = cfs_rq_of(se);
811 pcfs_rq = cfs_rq_of(pse);
813 if (cfs_rq == pcfs_rq) {
814 s64 delta = se->vruntime - pse->vruntime;
816 if (delta > (s64)sysctl_sched_wakeup_granularity)
820 #ifdef CONFIG_FAIR_GROUP_SCHED
826 static struct task_struct *pick_next_task_fair(struct rq *rq)
828 struct cfs_rq *cfs_rq = &rq->cfs;
829 struct sched_entity *se;
831 if (unlikely(!cfs_rq->nr_running))
835 se = pick_next_entity(cfs_rq);
836 cfs_rq = group_cfs_rq(se);
843 * Account for a descheduled task:
845 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
847 struct sched_entity *se = &prev->se;
848 struct cfs_rq *cfs_rq;
850 for_each_sched_entity(se) {
851 cfs_rq = cfs_rq_of(se);
852 put_prev_entity(cfs_rq, se);
856 /**************************************************
857 * Fair scheduling class load-balancing methods:
861 * Load-balancing iterator. Note: while the runqueue stays locked
862 * during the whole iteration, the current task might be
863 * dequeued so the iterator has to be dequeue-safe. Here we
864 * achieve that by always pre-iterating before returning
867 static inline struct task_struct *
868 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
870 struct task_struct *p;
875 p = rb_entry(curr, struct task_struct, se.run_node);
876 cfs_rq->rb_load_balance_curr = rb_next(curr);
881 static struct task_struct *load_balance_start_fair(void *arg)
883 struct cfs_rq *cfs_rq = arg;
885 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
888 static struct task_struct *load_balance_next_fair(void *arg)
890 struct cfs_rq *cfs_rq = arg;
892 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
895 #ifdef CONFIG_FAIR_GROUP_SCHED
896 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
898 struct sched_entity *curr;
899 struct task_struct *p;
901 if (!cfs_rq->nr_running)
906 curr = __pick_next_entity(cfs_rq);
915 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
916 unsigned long max_nr_move, unsigned long max_load_move,
917 struct sched_domain *sd, enum cpu_idle_type idle,
918 int *all_pinned, int *this_best_prio)
920 struct cfs_rq *busy_cfs_rq;
921 unsigned long load_moved, total_nr_moved = 0, nr_moved;
922 long rem_load_move = max_load_move;
923 struct rq_iterator cfs_rq_iterator;
925 cfs_rq_iterator.start = load_balance_start_fair;
926 cfs_rq_iterator.next = load_balance_next_fair;
928 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
929 #ifdef CONFIG_FAIR_GROUP_SCHED
930 struct cfs_rq *this_cfs_rq;
932 unsigned long maxload;
934 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
936 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
937 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
941 /* Don't pull more than imbalance/2 */
943 maxload = min(rem_load_move, imbalance);
945 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
947 # define maxload rem_load_move
949 /* pass busy_cfs_rq argument into
950 * load_balance_[start|next]_fair iterators
952 cfs_rq_iterator.arg = busy_cfs_rq;
953 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
954 max_nr_move, maxload, sd, idle, all_pinned,
955 &load_moved, this_best_prio, &cfs_rq_iterator);
957 total_nr_moved += nr_moved;
958 max_nr_move -= nr_moved;
959 rem_load_move -= load_moved;
961 if (max_nr_move <= 0 || rem_load_move <= 0)
965 return max_load_move - rem_load_move;
969 * scheduler tick hitting a task of our scheduling class:
971 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
973 struct cfs_rq *cfs_rq;
974 struct sched_entity *se = &curr->se;
976 for_each_sched_entity(se) {
977 cfs_rq = cfs_rq_of(se);
978 entity_tick(cfs_rq, se);
982 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
985 * Share the fairness runtime between parent and child, thus the
986 * total amount of pressure for CPU stays equal - new tasks
987 * get a chance to run but frequent forkers are not allowed to
988 * monopolize the CPU. Note: the parent runqueue is locked,
989 * the child is not running yet.
991 static void task_new_fair(struct rq *rq, struct task_struct *p)
993 struct cfs_rq *cfs_rq = task_cfs_rq(p);
994 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
996 sched_info_queued(p);
999 place_entity(cfs_rq, se, 1);
1001 if (sysctl_sched_child_runs_first &&
1002 curr->vruntime < se->vruntime) {
1004 * Upon rescheduling, sched_class::put_prev_task() will place
1005 * 'current' within the tree based on its new key value.
1007 swap(curr->vruntime, se->vruntime);
1010 update_stats_enqueue(cfs_rq, se);
1011 check_spread(cfs_rq, se);
1012 check_spread(cfs_rq, curr);
1013 __enqueue_entity(cfs_rq, se);
1014 account_entity_enqueue(cfs_rq, se);
1015 resched_task(rq->curr);
1018 /* Account for a task changing its policy or group.
1020 * This routine is mostly called to set cfs_rq->curr field when a task
1021 * migrates between groups/classes.
1023 static void set_curr_task_fair(struct rq *rq)
1025 struct sched_entity *se = &rq->curr->se;
1027 for_each_sched_entity(se)
1028 set_next_entity(cfs_rq_of(se), se);
1032 * All the scheduling class methods:
1034 struct sched_class fair_sched_class __read_mostly = {
1035 .enqueue_task = enqueue_task_fair,
1036 .dequeue_task = dequeue_task_fair,
1037 .yield_task = yield_task_fair,
1039 .check_preempt_curr = check_preempt_wakeup,
1041 .pick_next_task = pick_next_task_fair,
1042 .put_prev_task = put_prev_task_fair,
1044 .load_balance = load_balance_fair,
1046 .set_curr_task = set_curr_task_fair,
1047 .task_tick = task_tick_fair,
1048 .task_new = task_new_fair,
1051 #ifdef CONFIG_SCHED_DEBUG
1052 static void print_cfs_stats(struct seq_file *m, int cpu)
1054 struct cfs_rq *cfs_rq;
1056 #ifdef CONFIG_FAIR_GROUP_SCHED
1057 print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs);
1059 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1060 print_cfs_rq(m, cpu, cfs_rq);