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: 5ms * (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 = 5000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity = 1000000ULL;
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. If set to 0 (default) then
52 * parent will (try to) run first.
54 unsigned int sysctl_sched_child_runs_first __read_mostly;
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: 1 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 = 1000000UL;
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 #ifdef CONFIG_FAIR_GROUP_SCHED
84 /* cpu runqueue to which this cfs_rq is attached */
85 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
90 /* An entity is a task if it doesn't "own" a runqueue */
91 #define entity_is_task(se) (!se->my_q)
93 static inline struct task_struct *task_of(struct sched_entity *se)
95 #ifdef CONFIG_SCHED_DEBUG
96 WARN_ON_ONCE(!entity_is_task(se));
98 return container_of(se, struct task_struct, se);
101 /* Walk up scheduling entities hierarchy */
102 #define for_each_sched_entity(se) \
103 for (; se; se = se->parent)
105 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
110 /* runqueue on which this entity is (to be) queued */
111 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
116 /* runqueue "owned" by this group */
117 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
122 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
123 * another cpu ('this_cpu')
125 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
127 return cfs_rq->tg->cfs_rq[this_cpu];
130 /* Iterate thr' all leaf cfs_rq's on a runqueue */
131 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
132 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
134 /* Do the two (enqueued) entities belong to the same group ? */
136 is_same_group(struct sched_entity *se, struct sched_entity *pse)
138 if (se->cfs_rq == pse->cfs_rq)
144 static inline struct sched_entity *parent_entity(struct sched_entity *se)
149 /* return depth at which a sched entity is present in the hierarchy */
150 static inline int depth_se(struct sched_entity *se)
154 for_each_sched_entity(se)
161 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
163 int se_depth, pse_depth;
166 * preemption test can be made between sibling entities who are in the
167 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
168 * both tasks until we find their ancestors who are siblings of common
172 /* First walk up until both entities are at same depth */
173 se_depth = depth_se(*se);
174 pse_depth = depth_se(*pse);
176 while (se_depth > pse_depth) {
178 *se = parent_entity(*se);
181 while (pse_depth > se_depth) {
183 *pse = parent_entity(*pse);
186 while (!is_same_group(*se, *pse)) {
187 *se = parent_entity(*se);
188 *pse = parent_entity(*pse);
192 #else /* !CONFIG_FAIR_GROUP_SCHED */
194 static inline struct task_struct *task_of(struct sched_entity *se)
196 return container_of(se, struct task_struct, se);
199 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
201 return container_of(cfs_rq, struct rq, cfs);
204 #define entity_is_task(se) 1
206 #define for_each_sched_entity(se) \
207 for (; se; se = NULL)
209 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
211 return &task_rq(p)->cfs;
214 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
216 struct task_struct *p = task_of(se);
217 struct rq *rq = task_rq(p);
222 /* runqueue "owned" by this group */
223 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
228 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
230 return &cpu_rq(this_cpu)->cfs;
233 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
234 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
237 is_same_group(struct sched_entity *se, struct sched_entity *pse)
242 static inline struct sched_entity *parent_entity(struct sched_entity *se)
248 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
252 #endif /* CONFIG_FAIR_GROUP_SCHED */
255 /**************************************************************
256 * Scheduling class tree data structure manipulation methods:
259 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
261 s64 delta = (s64)(vruntime - min_vruntime);
263 min_vruntime = vruntime;
268 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
270 s64 delta = (s64)(vruntime - min_vruntime);
272 min_vruntime = vruntime;
277 static inline int entity_before(struct sched_entity *a,
278 struct sched_entity *b)
280 return (s64)(a->vruntime - b->vruntime) < 0;
283 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
285 return se->vruntime - cfs_rq->min_vruntime;
288 static void update_min_vruntime(struct cfs_rq *cfs_rq)
290 u64 vruntime = cfs_rq->min_vruntime;
293 vruntime = cfs_rq->curr->vruntime;
295 if (cfs_rq->rb_leftmost) {
296 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
301 vruntime = se->vruntime;
303 vruntime = min_vruntime(vruntime, se->vruntime);
306 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
310 * Enqueue an entity into the rb-tree:
312 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
314 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
315 struct rb_node *parent = NULL;
316 struct sched_entity *entry;
317 s64 key = entity_key(cfs_rq, se);
321 * Find the right place in the rbtree:
325 entry = rb_entry(parent, struct sched_entity, run_node);
327 * We dont care about collisions. Nodes with
328 * the same key stay together.
330 if (key < entity_key(cfs_rq, entry)) {
331 link = &parent->rb_left;
333 link = &parent->rb_right;
339 * Maintain a cache of leftmost tree entries (it is frequently
343 cfs_rq->rb_leftmost = &se->run_node;
345 rb_link_node(&se->run_node, parent, link);
346 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
349 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
351 if (cfs_rq->rb_leftmost == &se->run_node) {
352 struct rb_node *next_node;
354 next_node = rb_next(&se->run_node);
355 cfs_rq->rb_leftmost = next_node;
358 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
361 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
363 struct rb_node *left = cfs_rq->rb_leftmost;
368 return rb_entry(left, struct sched_entity, run_node);
371 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
373 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
378 return rb_entry(last, struct sched_entity, run_node);
381 /**************************************************************
382 * Scheduling class statistics methods:
385 #ifdef CONFIG_SCHED_DEBUG
386 int sched_nr_latency_handler(struct ctl_table *table, int write,
387 struct file *filp, void __user *buffer, size_t *lenp,
390 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
395 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
396 sysctl_sched_min_granularity);
405 static inline unsigned long
406 calc_delta_fair(unsigned long delta, struct sched_entity *se)
408 if (unlikely(se->load.weight != NICE_0_LOAD))
409 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
415 * The idea is to set a period in which each task runs once.
417 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
418 * this period because otherwise the slices get too small.
420 * p = (nr <= nl) ? l : l*nr/nl
422 static u64 __sched_period(unsigned long nr_running)
424 u64 period = sysctl_sched_latency;
425 unsigned long nr_latency = sched_nr_latency;
427 if (unlikely(nr_running > nr_latency)) {
428 period = sysctl_sched_min_granularity;
429 period *= nr_running;
436 * We calculate the wall-time slice from the period by taking a part
437 * proportional to the weight.
441 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
443 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
445 for_each_sched_entity(se) {
446 struct load_weight *load;
447 struct load_weight lw;
449 cfs_rq = cfs_rq_of(se);
450 load = &cfs_rq->load;
452 if (unlikely(!se->on_rq)) {
455 update_load_add(&lw, se->load.weight);
458 slice = calc_delta_mine(slice, se->load.weight, load);
464 * We calculate the vruntime slice of a to be inserted task
468 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
470 return calc_delta_fair(sched_slice(cfs_rq, se), se);
474 * Update the current task's runtime statistics. Skip current tasks that
475 * are not in our scheduling class.
478 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
479 unsigned long delta_exec)
481 unsigned long delta_exec_weighted;
483 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
485 curr->sum_exec_runtime += delta_exec;
486 schedstat_add(cfs_rq, exec_clock, delta_exec);
487 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
488 curr->vruntime += delta_exec_weighted;
489 update_min_vruntime(cfs_rq);
492 static void update_curr(struct cfs_rq *cfs_rq)
494 struct sched_entity *curr = cfs_rq->curr;
495 u64 now = rq_of(cfs_rq)->clock;
496 unsigned long delta_exec;
502 * Get the amount of time the current task was running
503 * since the last time we changed load (this cannot
504 * overflow on 32 bits):
506 delta_exec = (unsigned long)(now - curr->exec_start);
510 __update_curr(cfs_rq, curr, delta_exec);
511 curr->exec_start = now;
513 if (entity_is_task(curr)) {
514 struct task_struct *curtask = task_of(curr);
516 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
517 cpuacct_charge(curtask, delta_exec);
518 account_group_exec_runtime(curtask, delta_exec);
523 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
529 * Task is being enqueued - update stats:
531 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
534 * Are we enqueueing a waiting task? (for current tasks
535 * a dequeue/enqueue event is a NOP)
537 if (se != cfs_rq->curr)
538 update_stats_wait_start(cfs_rq, se);
542 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
544 schedstat_set(se->wait_max, max(se->wait_max,
545 rq_of(cfs_rq)->clock - se->wait_start));
546 schedstat_set(se->wait_count, se->wait_count + 1);
547 schedstat_set(se->wait_sum, se->wait_sum +
548 rq_of(cfs_rq)->clock - se->wait_start);
549 #ifdef CONFIG_SCHEDSTATS
550 if (entity_is_task(se)) {
551 trace_sched_stat_wait(task_of(se),
552 rq_of(cfs_rq)->clock - se->wait_start);
555 schedstat_set(se->wait_start, 0);
559 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
562 * Mark the end of the wait period if dequeueing a
565 if (se != cfs_rq->curr)
566 update_stats_wait_end(cfs_rq, se);
570 * We are picking a new current task - update its stats:
573 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
576 * We are starting a new run period:
578 se->exec_start = rq_of(cfs_rq)->clock;
581 /**************************************************
582 * Scheduling class queueing methods:
585 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
587 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
589 cfs_rq->task_weight += weight;
593 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
599 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
601 update_load_add(&cfs_rq->load, se->load.weight);
602 if (!parent_entity(se))
603 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
604 if (entity_is_task(se)) {
605 add_cfs_task_weight(cfs_rq, se->load.weight);
606 list_add(&se->group_node, &cfs_rq->tasks);
608 cfs_rq->nr_running++;
613 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
615 update_load_sub(&cfs_rq->load, se->load.weight);
616 if (!parent_entity(se))
617 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
618 if (entity_is_task(se)) {
619 add_cfs_task_weight(cfs_rq, -se->load.weight);
620 list_del_init(&se->group_node);
622 cfs_rq->nr_running--;
626 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 #ifdef CONFIG_SCHEDSTATS
629 struct task_struct *tsk = NULL;
631 if (entity_is_task(se))
634 if (se->sleep_start) {
635 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
640 if (unlikely(delta > se->sleep_max))
641 se->sleep_max = delta;
644 se->sum_sleep_runtime += delta;
647 account_scheduler_latency(tsk, delta >> 10, 1);
648 trace_sched_stat_sleep(tsk, delta);
651 if (se->block_start) {
652 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
657 if (unlikely(delta > se->block_max))
658 se->block_max = delta;
661 se->sum_sleep_runtime += delta;
664 if (tsk->in_iowait) {
665 se->iowait_sum += delta;
667 trace_sched_stat_iowait(tsk, delta);
671 * Blocking time is in units of nanosecs, so shift by
672 * 20 to get a milliseconds-range estimation of the
673 * amount of time that the task spent sleeping:
675 if (unlikely(prof_on == SLEEP_PROFILING)) {
676 profile_hits(SLEEP_PROFILING,
677 (void *)get_wchan(tsk),
680 account_scheduler_latency(tsk, delta >> 10, 0);
686 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
688 #ifdef CONFIG_SCHED_DEBUG
689 s64 d = se->vruntime - cfs_rq->min_vruntime;
694 if (d > 3*sysctl_sched_latency)
695 schedstat_inc(cfs_rq, nr_spread_over);
700 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
702 u64 vruntime = cfs_rq->min_vruntime;
705 * The 'current' period is already promised to the current tasks,
706 * however the extra weight of the new task will slow them down a
707 * little, place the new task so that it fits in the slot that
708 * stays open at the end.
710 if (initial && sched_feat(START_DEBIT))
711 vruntime += sched_vslice(cfs_rq, se);
714 /* sleeps upto a single latency don't count. */
715 if (sched_feat(FAIR_SLEEPERS)) {
716 unsigned long thresh = sysctl_sched_latency;
719 * Convert the sleeper threshold into virtual time.
720 * SCHED_IDLE is a special sub-class. We care about
721 * fairness only relative to other SCHED_IDLE tasks,
722 * all of which have the same weight.
724 if (sched_feat(NORMALIZED_SLEEPER) &&
725 (!entity_is_task(se) ||
726 task_of(se)->policy != SCHED_IDLE))
727 thresh = calc_delta_fair(thresh, se);
730 * Halve their sleep time's effect, to allow
731 * for a gentler effect of sleepers:
733 if (sched_feat(GENTLE_FAIR_SLEEPERS))
740 /* ensure we never gain time by being placed backwards. */
741 vruntime = max_vruntime(se->vruntime, vruntime);
743 se->vruntime = vruntime;
747 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
750 * Update run-time statistics of the 'current'.
753 account_entity_enqueue(cfs_rq, se);
756 place_entity(cfs_rq, se, 0);
757 enqueue_sleeper(cfs_rq, se);
760 update_stats_enqueue(cfs_rq, se);
761 check_spread(cfs_rq, se);
762 if (se != cfs_rq->curr)
763 __enqueue_entity(cfs_rq, se);
766 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 if (!se || cfs_rq->last == se)
771 if (!se || cfs_rq->next == se)
775 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
777 for_each_sched_entity(se)
778 __clear_buddies(cfs_rq_of(se), se);
782 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
785 * Update run-time statistics of the 'current'.
789 update_stats_dequeue(cfs_rq, se);
791 #ifdef CONFIG_SCHEDSTATS
792 if (entity_is_task(se)) {
793 struct task_struct *tsk = task_of(se);
795 if (tsk->state & TASK_INTERRUPTIBLE)
796 se->sleep_start = rq_of(cfs_rq)->clock;
797 if (tsk->state & TASK_UNINTERRUPTIBLE)
798 se->block_start = rq_of(cfs_rq)->clock;
803 clear_buddies(cfs_rq, se);
805 if (se != cfs_rq->curr)
806 __dequeue_entity(cfs_rq, se);
807 account_entity_dequeue(cfs_rq, se);
808 update_min_vruntime(cfs_rq);
812 * Preempt the current task with a newly woken task if needed:
815 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
817 unsigned long ideal_runtime, delta_exec;
819 ideal_runtime = sched_slice(cfs_rq, curr);
820 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
821 if (delta_exec > ideal_runtime) {
822 resched_task(rq_of(cfs_rq)->curr);
824 * The current task ran long enough, ensure it doesn't get
825 * re-elected due to buddy favours.
827 clear_buddies(cfs_rq, curr);
832 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
834 /* 'current' is not kept within the tree. */
837 * Any task has to be enqueued before it get to execute on
838 * a CPU. So account for the time it spent waiting on the
841 update_stats_wait_end(cfs_rq, se);
842 __dequeue_entity(cfs_rq, se);
845 update_stats_curr_start(cfs_rq, se);
847 #ifdef CONFIG_SCHEDSTATS
849 * Track our maximum slice length, if the CPU's load is at
850 * least twice that of our own weight (i.e. dont track it
851 * when there are only lesser-weight tasks around):
853 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
854 se->slice_max = max(se->slice_max,
855 se->sum_exec_runtime - se->prev_sum_exec_runtime);
858 se->prev_sum_exec_runtime = se->sum_exec_runtime;
862 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
864 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
866 struct sched_entity *se = __pick_next_entity(cfs_rq);
868 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
871 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
877 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
880 * If still on the runqueue then deactivate_task()
881 * was not called and update_curr() has to be done:
886 check_spread(cfs_rq, prev);
888 update_stats_wait_start(cfs_rq, prev);
889 /* Put 'current' back into the tree. */
890 __enqueue_entity(cfs_rq, prev);
896 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
899 * Update run-time statistics of the 'current'.
903 #ifdef CONFIG_SCHED_HRTICK
905 * queued ticks are scheduled to match the slice, so don't bother
906 * validating it and just reschedule.
909 resched_task(rq_of(cfs_rq)->curr);
913 * don't let the period tick interfere with the hrtick preemption
915 if (!sched_feat(DOUBLE_TICK) &&
916 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
920 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
921 check_preempt_tick(cfs_rq, curr);
924 /**************************************************
925 * CFS operations on tasks:
928 #ifdef CONFIG_SCHED_HRTICK
929 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
931 struct sched_entity *se = &p->se;
932 struct cfs_rq *cfs_rq = cfs_rq_of(se);
934 WARN_ON(task_rq(p) != rq);
936 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
937 u64 slice = sched_slice(cfs_rq, se);
938 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
939 s64 delta = slice - ran;
948 * Don't schedule slices shorter than 10000ns, that just
949 * doesn't make sense. Rely on vruntime for fairness.
952 delta = max_t(s64, 10000LL, delta);
954 hrtick_start(rq, delta);
959 * called from enqueue/dequeue and updates the hrtick when the
960 * current task is from our class and nr_running is low enough
963 static void hrtick_update(struct rq *rq)
965 struct task_struct *curr = rq->curr;
967 if (curr->sched_class != &fair_sched_class)
970 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
971 hrtick_start_fair(rq, curr);
973 #else /* !CONFIG_SCHED_HRTICK */
975 hrtick_start_fair(struct rq *rq, struct task_struct *p)
979 static inline void hrtick_update(struct rq *rq)
985 * The enqueue_task method is called before nr_running is
986 * increased. Here we update the fair scheduling stats and
987 * then put the task into the rbtree:
989 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
991 struct cfs_rq *cfs_rq;
992 struct sched_entity *se = &p->se;
994 for_each_sched_entity(se) {
997 cfs_rq = cfs_rq_of(se);
998 enqueue_entity(cfs_rq, se, wakeup);
1006 * The dequeue_task method is called before nr_running is
1007 * decreased. We remove the task from the rbtree and
1008 * update the fair scheduling stats:
1010 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
1012 struct cfs_rq *cfs_rq;
1013 struct sched_entity *se = &p->se;
1015 for_each_sched_entity(se) {
1016 cfs_rq = cfs_rq_of(se);
1017 dequeue_entity(cfs_rq, se, sleep);
1018 /* Don't dequeue parent if it has other entities besides us */
1019 if (cfs_rq->load.weight)
1028 * sched_yield() support is very simple - we dequeue and enqueue.
1030 * If compat_yield is turned on then we requeue to the end of the tree.
1032 static void yield_task_fair(struct rq *rq)
1034 struct task_struct *curr = rq->curr;
1035 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1036 struct sched_entity *rightmost, *se = &curr->se;
1039 * Are we the only task in the tree?
1041 if (unlikely(cfs_rq->nr_running == 1))
1044 clear_buddies(cfs_rq, se);
1046 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1047 update_rq_clock(rq);
1049 * Update run-time statistics of the 'current'.
1051 update_curr(cfs_rq);
1056 * Find the rightmost entry in the rbtree:
1058 rightmost = __pick_last_entity(cfs_rq);
1060 * Already in the rightmost position?
1062 if (unlikely(!rightmost || entity_before(rightmost, se)))
1066 * Minimally necessary key value to be last in the tree:
1067 * Upon rescheduling, sched_class::put_prev_task() will place
1068 * 'current' within the tree based on its new key value.
1070 se->vruntime = rightmost->vruntime + 1;
1075 #ifdef CONFIG_FAIR_GROUP_SCHED
1077 * effective_load() calculates the load change as seen from the root_task_group
1079 * Adding load to a group doesn't make a group heavier, but can cause movement
1080 * of group shares between cpus. Assuming the shares were perfectly aligned one
1081 * can calculate the shift in shares.
1083 * The problem is that perfectly aligning the shares is rather expensive, hence
1084 * we try to avoid doing that too often - see update_shares(), which ratelimits
1087 * We compensate this by not only taking the current delta into account, but
1088 * also considering the delta between when the shares were last adjusted and
1091 * We still saw a performance dip, some tracing learned us that between
1092 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1093 * significantly. Therefore try to bias the error in direction of failing
1094 * the affine wakeup.
1097 static long effective_load(struct task_group *tg, int cpu,
1100 struct sched_entity *se = tg->se[cpu];
1106 * By not taking the decrease of shares on the other cpu into
1107 * account our error leans towards reducing the affine wakeups.
1109 if (!wl && sched_feat(ASYM_EFF_LOAD))
1112 for_each_sched_entity(se) {
1113 long S, rw, s, a, b;
1117 * Instead of using this increment, also add the difference
1118 * between when the shares were last updated and now.
1120 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1124 S = se->my_q->tg->shares;
1125 s = se->my_q->shares;
1126 rw = se->my_q->rq_weight;
1137 * Assume the group is already running and will
1138 * thus already be accounted for in the weight.
1140 * That is, moving shares between CPUs, does not
1141 * alter the group weight.
1151 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1152 unsigned long wl, unsigned long wg)
1159 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1161 struct task_struct *curr = current;
1162 unsigned long this_load, load;
1163 int idx, this_cpu, prev_cpu;
1164 unsigned long tl_per_task;
1165 unsigned int imbalance;
1166 struct task_group *tg;
1167 unsigned long weight;
1171 this_cpu = smp_processor_id();
1172 prev_cpu = task_cpu(p);
1173 load = source_load(prev_cpu, idx);
1174 this_load = target_load(this_cpu, idx);
1177 if (sched_feat(SYNC_LESS) &&
1178 (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1179 p->se.avg_overlap > sysctl_sched_migration_cost))
1182 if (sched_feat(SYNC_MORE) &&
1183 (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1184 p->se.avg_overlap < sysctl_sched_migration_cost))
1189 * If sync wakeup then subtract the (maximum possible)
1190 * effect of the currently running task from the load
1191 * of the current CPU:
1194 tg = task_group(current);
1195 weight = current->se.load.weight;
1197 this_load += effective_load(tg, this_cpu, -weight, -weight);
1198 load += effective_load(tg, prev_cpu, 0, -weight);
1202 weight = p->se.load.weight;
1204 imbalance = 100 + (sd->imbalance_pct - 100) / 2;
1207 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1208 * due to the sync cause above having dropped this_load to 0, we'll
1209 * always have an imbalance, but there's really nothing you can do
1210 * about that, so that's good too.
1212 * Otherwise check if either cpus are near enough in load to allow this
1213 * task to be woken on this_cpu.
1215 balanced = !this_load ||
1216 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
1217 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1220 * If the currently running task will sleep within
1221 * a reasonable amount of time then attract this newly
1224 if (sync && balanced)
1227 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1228 tl_per_task = cpu_avg_load_per_task(this_cpu);
1231 (this_load <= load &&
1232 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1234 * This domain has SD_WAKE_AFFINE and
1235 * p is cache cold in this domain, and
1236 * there is no bad imbalance.
1238 schedstat_inc(sd, ttwu_move_affine);
1239 schedstat_inc(p, se.nr_wakeups_affine);
1247 * find_idlest_group finds and returns the least busy CPU group within the
1250 static struct sched_group *
1251 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1252 int this_cpu, int load_idx)
1254 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1255 unsigned long min_load = ULONG_MAX, this_load = 0;
1256 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1259 unsigned long load, avg_load;
1263 /* Skip over this group if it has no CPUs allowed */
1264 if (!cpumask_intersects(sched_group_cpus(group),
1268 local_group = cpumask_test_cpu(this_cpu,
1269 sched_group_cpus(group));
1271 /* Tally up the load of all CPUs in the group */
1274 for_each_cpu(i, sched_group_cpus(group)) {
1275 /* Bias balancing toward cpus of our domain */
1277 load = source_load(i, load_idx);
1279 load = target_load(i, load_idx);
1284 /* Adjust by relative CPU power of the group */
1285 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1288 this_load = avg_load;
1290 } else if (avg_load < min_load) {
1291 min_load = avg_load;
1294 } while (group = group->next, group != sd->groups);
1296 if (!idlest || 100*this_load < imbalance*min_load)
1302 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1305 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1307 unsigned long load, min_load = ULONG_MAX;
1311 /* Traverse only the allowed CPUs */
1312 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1313 load = weighted_cpuload(i);
1315 if (load < min_load || (load == min_load && i == this_cpu)) {
1325 * sched_balance_self: balance the current task (running on cpu) in domains
1326 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1329 * Balance, ie. select the least loaded group.
1331 * Returns the target CPU number, or the same CPU if no balancing is needed.
1333 * preempt must be disabled.
1335 static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1337 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1338 int cpu = smp_processor_id();
1339 int prev_cpu = task_cpu(p);
1341 int want_affine = 0;
1343 int sync = wake_flags & WF_SYNC;
1345 if (sd_flag & SD_BALANCE_WAKE) {
1346 if (sched_feat(AFFINE_WAKEUPS))
1352 for_each_domain(cpu, tmp) {
1354 * If power savings logic is enabled for a domain, see if we
1355 * are not overloaded, if so, don't balance wider.
1357 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1358 unsigned long power = 0;
1359 unsigned long nr_running = 0;
1360 unsigned long capacity;
1363 for_each_cpu(i, sched_domain_span(tmp)) {
1364 power += power_of(i);
1365 nr_running += cpu_rq(i)->cfs.nr_running;
1368 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1370 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1373 if (nr_running < capacity)
1377 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1378 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1384 if (!want_sd && !want_affine)
1387 if (!(tmp->flags & sd_flag))
1394 if (sched_feat(LB_SHARES_UPDATE)) {
1396 * Pick the largest domain to update shares over
1399 if (affine_sd && (!tmp ||
1400 cpumask_weight(sched_domain_span(affine_sd)) >
1401 cpumask_weight(sched_domain_span(sd))))
1408 if (affine_sd && wake_affine(affine_sd, p, sync)) {
1414 int load_idx = sd->forkexec_idx;
1415 struct sched_group *group;
1418 if (!(sd->flags & sd_flag)) {
1423 if (sd_flag & SD_BALANCE_WAKE)
1424 load_idx = sd->wake_idx;
1426 group = find_idlest_group(sd, p, cpu, load_idx);
1432 new_cpu = find_idlest_cpu(group, p, cpu);
1433 if (new_cpu == -1 || new_cpu == cpu) {
1434 /* Now try balancing at a lower domain level of cpu */
1439 /* Now try balancing at a lower domain level of new_cpu */
1441 weight = cpumask_weight(sched_domain_span(sd));
1443 for_each_domain(cpu, tmp) {
1444 if (weight <= cpumask_weight(sched_domain_span(tmp)))
1446 if (tmp->flags & sd_flag)
1449 /* while loop will break here if sd == NULL */
1456 #endif /* CONFIG_SMP */
1459 * Adaptive granularity
1461 * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1462 * with the limit of wakeup_gran -- when it never does a wakeup.
1464 * So the smaller avg_wakeup is the faster we want this task to preempt,
1465 * but we don't want to treat the preemptee unfairly and therefore allow it
1466 * to run for at least the amount of time we'd like to run.
1468 * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1470 * NOTE: we use *nr_running to scale with load, this nicely matches the
1471 * degrading latency on load.
1473 static unsigned long
1474 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1476 u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1477 u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1480 if (this_run < expected_wakeup)
1481 gran = expected_wakeup - this_run;
1483 return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1486 static unsigned long
1487 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1489 unsigned long gran = sysctl_sched_wakeup_granularity;
1491 if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1492 gran = adaptive_gran(curr, se);
1495 * Since its curr running now, convert the gran from real-time
1496 * to virtual-time in his units.
1498 if (sched_feat(ASYM_GRAN)) {
1500 * By using 'se' instead of 'curr' we penalize light tasks, so
1501 * they get preempted easier. That is, if 'se' < 'curr' then
1502 * the resulting gran will be larger, therefore penalizing the
1503 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1504 * be smaller, again penalizing the lighter task.
1506 * This is especially important for buddies when the leftmost
1507 * task is higher priority than the buddy.
1509 if (unlikely(se->load.weight != NICE_0_LOAD))
1510 gran = calc_delta_fair(gran, se);
1512 if (unlikely(curr->load.weight != NICE_0_LOAD))
1513 gran = calc_delta_fair(gran, curr);
1520 * Should 'se' preempt 'curr'.
1534 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1536 s64 gran, vdiff = curr->vruntime - se->vruntime;
1541 gran = wakeup_gran(curr, se);
1548 static void set_last_buddy(struct sched_entity *se)
1550 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1551 for_each_sched_entity(se)
1552 cfs_rq_of(se)->last = se;
1556 static void set_next_buddy(struct sched_entity *se)
1558 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1559 for_each_sched_entity(se)
1560 cfs_rq_of(se)->next = se;
1565 * Preempt the current task with a newly woken task if needed:
1567 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1569 struct task_struct *curr = rq->curr;
1570 struct sched_entity *se = &curr->se, *pse = &p->se;
1571 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1572 int sync = wake_flags & WF_SYNC;
1574 update_curr(cfs_rq);
1576 if (unlikely(rt_prio(p->prio))) {
1581 if (unlikely(p->sched_class != &fair_sched_class))
1584 if (unlikely(se == pse))
1588 * Only set the backward buddy when the current task is still on the
1589 * rq. This can happen when a wakeup gets interleaved with schedule on
1590 * the ->pre_schedule() or idle_balance() point, either of which can
1593 * Also, during early boot the idle thread is in the fair class, for
1594 * obvious reasons its a bad idea to schedule back to the idle thread.
1596 if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1598 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK))
1599 set_next_buddy(pse);
1602 * We can come here with TIF_NEED_RESCHED already set from new task
1605 if (test_tsk_need_resched(curr))
1609 * Batch and idle tasks do not preempt (their preemption is driven by
1612 if (unlikely(p->policy != SCHED_NORMAL))
1615 /* Idle tasks are by definition preempted by everybody. */
1616 if (unlikely(curr->policy == SCHED_IDLE)) {
1621 if ((sched_feat(WAKEUP_SYNC) && sync) ||
1622 (sched_feat(WAKEUP_OVERLAP) &&
1623 (se->avg_overlap < sysctl_sched_migration_cost &&
1624 pse->avg_overlap < sysctl_sched_migration_cost))) {
1629 if (sched_feat(WAKEUP_RUNNING)) {
1630 if (pse->avg_running < se->avg_running) {
1631 set_next_buddy(pse);
1637 if (!sched_feat(WAKEUP_PREEMPT))
1640 find_matching_se(&se, &pse);
1644 if (wakeup_preempt_entity(se, pse) == 1)
1648 static struct task_struct *pick_next_task_fair(struct rq *rq)
1650 struct task_struct *p;
1651 struct cfs_rq *cfs_rq = &rq->cfs;
1652 struct sched_entity *se;
1654 if (unlikely(!cfs_rq->nr_running))
1658 se = pick_next_entity(cfs_rq);
1660 * If se was a buddy, clear it so that it will have to earn
1663 * If se was not a buddy, clear the buddies because neither
1664 * was elegible to run, let them earn it again.
1666 * IOW. unconditionally clear buddies.
1668 __clear_buddies(cfs_rq, NULL);
1669 set_next_entity(cfs_rq, se);
1670 cfs_rq = group_cfs_rq(se);
1674 hrtick_start_fair(rq, p);
1680 * Account for a descheduled task:
1682 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1684 struct sched_entity *se = &prev->se;
1685 struct cfs_rq *cfs_rq;
1687 for_each_sched_entity(se) {
1688 cfs_rq = cfs_rq_of(se);
1689 put_prev_entity(cfs_rq, se);
1694 /**************************************************
1695 * Fair scheduling class load-balancing methods:
1699 * Load-balancing iterator. Note: while the runqueue stays locked
1700 * during the whole iteration, the current task might be
1701 * dequeued so the iterator has to be dequeue-safe. Here we
1702 * achieve that by always pre-iterating before returning
1705 static struct task_struct *
1706 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1708 struct task_struct *p = NULL;
1709 struct sched_entity *se;
1711 if (next == &cfs_rq->tasks)
1714 se = list_entry(next, struct sched_entity, group_node);
1716 cfs_rq->balance_iterator = next->next;
1721 static struct task_struct *load_balance_start_fair(void *arg)
1723 struct cfs_rq *cfs_rq = arg;
1725 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1728 static struct task_struct *load_balance_next_fair(void *arg)
1730 struct cfs_rq *cfs_rq = arg;
1732 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1735 static unsigned long
1736 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1737 unsigned long max_load_move, struct sched_domain *sd,
1738 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1739 struct cfs_rq *cfs_rq)
1741 struct rq_iterator cfs_rq_iterator;
1743 cfs_rq_iterator.start = load_balance_start_fair;
1744 cfs_rq_iterator.next = load_balance_next_fair;
1745 cfs_rq_iterator.arg = cfs_rq;
1747 return balance_tasks(this_rq, this_cpu, busiest,
1748 max_load_move, sd, idle, all_pinned,
1749 this_best_prio, &cfs_rq_iterator);
1752 #ifdef CONFIG_FAIR_GROUP_SCHED
1753 static unsigned long
1754 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1755 unsigned long max_load_move,
1756 struct sched_domain *sd, enum cpu_idle_type idle,
1757 int *all_pinned, int *this_best_prio)
1759 long rem_load_move = max_load_move;
1760 int busiest_cpu = cpu_of(busiest);
1761 struct task_group *tg;
1764 update_h_load(busiest_cpu);
1766 list_for_each_entry_rcu(tg, &task_groups, list) {
1767 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1768 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1769 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1770 u64 rem_load, moved_load;
1775 if (!busiest_cfs_rq->task_weight)
1778 rem_load = (u64)rem_load_move * busiest_weight;
1779 rem_load = div_u64(rem_load, busiest_h_load + 1);
1781 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1782 rem_load, sd, idle, all_pinned, this_best_prio,
1783 tg->cfs_rq[busiest_cpu]);
1788 moved_load *= busiest_h_load;
1789 moved_load = div_u64(moved_load, busiest_weight + 1);
1791 rem_load_move -= moved_load;
1792 if (rem_load_move < 0)
1797 return max_load_move - rem_load_move;
1800 static unsigned long
1801 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1802 unsigned long max_load_move,
1803 struct sched_domain *sd, enum cpu_idle_type idle,
1804 int *all_pinned, int *this_best_prio)
1806 return __load_balance_fair(this_rq, this_cpu, busiest,
1807 max_load_move, sd, idle, all_pinned,
1808 this_best_prio, &busiest->cfs);
1813 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1814 struct sched_domain *sd, enum cpu_idle_type idle)
1816 struct cfs_rq *busy_cfs_rq;
1817 struct rq_iterator cfs_rq_iterator;
1819 cfs_rq_iterator.start = load_balance_start_fair;
1820 cfs_rq_iterator.next = load_balance_next_fair;
1822 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1824 * pass busy_cfs_rq argument into
1825 * load_balance_[start|next]_fair iterators
1827 cfs_rq_iterator.arg = busy_cfs_rq;
1828 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1835 #endif /* CONFIG_SMP */
1838 * scheduler tick hitting a task of our scheduling class:
1840 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1842 struct cfs_rq *cfs_rq;
1843 struct sched_entity *se = &curr->se;
1845 for_each_sched_entity(se) {
1846 cfs_rq = cfs_rq_of(se);
1847 entity_tick(cfs_rq, se, queued);
1852 * Share the fairness runtime between parent and child, thus the
1853 * total amount of pressure for CPU stays equal - new tasks
1854 * get a chance to run but frequent forkers are not allowed to
1855 * monopolize the CPU. Note: the parent runqueue is locked,
1856 * the child is not running yet.
1858 static void task_new_fair(struct rq *rq, struct task_struct *p)
1860 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1861 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1862 int this_cpu = smp_processor_id();
1864 sched_info_queued(p);
1866 update_curr(cfs_rq);
1868 se->vruntime = curr->vruntime;
1869 place_entity(cfs_rq, se, 1);
1871 /* 'curr' will be NULL if the child belongs to a different group */
1872 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1873 curr && entity_before(curr, se)) {
1875 * Upon rescheduling, sched_class::put_prev_task() will place
1876 * 'current' within the tree based on its new key value.
1878 swap(curr->vruntime, se->vruntime);
1879 resched_task(rq->curr);
1882 enqueue_task_fair(rq, p, 0);
1886 * Priority of the task has changed. Check to see if we preempt
1889 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1890 int oldprio, int running)
1893 * Reschedule if we are currently running on this runqueue and
1894 * our priority decreased, or if we are not currently running on
1895 * this runqueue and our priority is higher than the current's
1898 if (p->prio > oldprio)
1899 resched_task(rq->curr);
1901 check_preempt_curr(rq, p, 0);
1905 * We switched to the sched_fair class.
1907 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1911 * We were most likely switched from sched_rt, so
1912 * kick off the schedule if running, otherwise just see
1913 * if we can still preempt the current task.
1916 resched_task(rq->curr);
1918 check_preempt_curr(rq, p, 0);
1921 /* Account for a task changing its policy or group.
1923 * This routine is mostly called to set cfs_rq->curr field when a task
1924 * migrates between groups/classes.
1926 static void set_curr_task_fair(struct rq *rq)
1928 struct sched_entity *se = &rq->curr->se;
1930 for_each_sched_entity(se)
1931 set_next_entity(cfs_rq_of(se), se);
1934 #ifdef CONFIG_FAIR_GROUP_SCHED
1935 static void moved_group_fair(struct task_struct *p)
1937 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1939 update_curr(cfs_rq);
1940 place_entity(cfs_rq, &p->se, 1);
1945 * All the scheduling class methods:
1947 static const struct sched_class fair_sched_class = {
1948 .next = &idle_sched_class,
1949 .enqueue_task = enqueue_task_fair,
1950 .dequeue_task = dequeue_task_fair,
1951 .yield_task = yield_task_fair,
1953 .check_preempt_curr = check_preempt_wakeup,
1955 .pick_next_task = pick_next_task_fair,
1956 .put_prev_task = put_prev_task_fair,
1959 .select_task_rq = select_task_rq_fair,
1961 .load_balance = load_balance_fair,
1962 .move_one_task = move_one_task_fair,
1965 .set_curr_task = set_curr_task_fair,
1966 .task_tick = task_tick_fair,
1967 .task_new = task_new_fair,
1969 .prio_changed = prio_changed_fair,
1970 .switched_to = switched_to_fair,
1972 #ifdef CONFIG_FAIR_GROUP_SCHED
1973 .moved_group = moved_group_fair,
1977 #ifdef CONFIG_SCHED_DEBUG
1978 static void print_cfs_stats(struct seq_file *m, int cpu)
1980 struct cfs_rq *cfs_rq;
1983 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1984 print_cfs_rq(m, cpu, cfs_rq);