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;
38 unsigned int normalized_sysctl_sched_latency = 5000000ULL;
41 * Minimal preemption granularity for CPU-bound tasks:
42 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
44 unsigned int sysctl_sched_min_granularity = 1000000ULL;
45 unsigned int normalized_sysctl_sched_min_granularity = 1000000ULL;
48 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
50 static unsigned int sched_nr_latency = 5;
53 * After fork, child runs first. If set to 0 (default) then
54 * parent will (try to) run first.
56 unsigned int sysctl_sched_child_runs_first __read_mostly;
59 * sys_sched_yield() compat mode
61 * This option switches the agressive yield implementation of the
62 * old scheduler back on.
64 unsigned int __read_mostly sysctl_sched_compat_yield;
67 * SCHED_OTHER wake-up granularity.
68 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
70 * This option delays the preemption effects of decoupled workloads
71 * and reduces their over-scheduling. Synchronous workloads will still
72 * have immediate wakeup/sleep latencies.
74 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
75 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
77 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
79 static const 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 static inline struct task_struct *task_of(struct sched_entity *se)
98 #ifdef CONFIG_SCHED_DEBUG
99 WARN_ON_ONCE(!entity_is_task(se));
101 return container_of(se, struct task_struct, se);
104 /* Walk up scheduling entities hierarchy */
105 #define for_each_sched_entity(se) \
106 for (; se; se = se->parent)
108 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
113 /* runqueue on which this entity is (to be) queued */
114 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
119 /* runqueue "owned" by this group */
120 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
125 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
126 * another cpu ('this_cpu')
128 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
130 return cfs_rq->tg->cfs_rq[this_cpu];
133 /* Iterate thr' all leaf cfs_rq's on a runqueue */
134 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
135 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
137 /* Do the two (enqueued) entities belong to the same group ? */
139 is_same_group(struct sched_entity *se, struct sched_entity *pse)
141 if (se->cfs_rq == pse->cfs_rq)
147 static inline struct sched_entity *parent_entity(struct sched_entity *se)
152 /* return depth at which a sched entity is present in the hierarchy */
153 static inline int depth_se(struct sched_entity *se)
157 for_each_sched_entity(se)
164 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
166 int se_depth, pse_depth;
169 * preemption test can be made between sibling entities who are in the
170 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
171 * both tasks until we find their ancestors who are siblings of common
175 /* First walk up until both entities are at same depth */
176 se_depth = depth_se(*se);
177 pse_depth = depth_se(*pse);
179 while (se_depth > pse_depth) {
181 *se = parent_entity(*se);
184 while (pse_depth > se_depth) {
186 *pse = parent_entity(*pse);
189 while (!is_same_group(*se, *pse)) {
190 *se = parent_entity(*se);
191 *pse = parent_entity(*pse);
195 #else /* !CONFIG_FAIR_GROUP_SCHED */
197 static inline struct task_struct *task_of(struct sched_entity *se)
199 return container_of(se, struct task_struct, se);
202 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
204 return container_of(cfs_rq, struct rq, cfs);
207 #define entity_is_task(se) 1
209 #define for_each_sched_entity(se) \
210 for (; se; se = NULL)
212 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
214 return &task_rq(p)->cfs;
217 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
219 struct task_struct *p = task_of(se);
220 struct rq *rq = task_rq(p);
225 /* runqueue "owned" by this group */
226 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
231 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
233 return &cpu_rq(this_cpu)->cfs;
236 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
237 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
240 is_same_group(struct sched_entity *se, struct sched_entity *pse)
245 static inline struct sched_entity *parent_entity(struct sched_entity *se)
251 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
255 #endif /* CONFIG_FAIR_GROUP_SCHED */
258 /**************************************************************
259 * Scheduling class tree data structure manipulation methods:
262 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
264 s64 delta = (s64)(vruntime - min_vruntime);
266 min_vruntime = vruntime;
271 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
273 s64 delta = (s64)(vruntime - min_vruntime);
275 min_vruntime = vruntime;
280 static inline int entity_before(struct sched_entity *a,
281 struct sched_entity *b)
283 return (s64)(a->vruntime - b->vruntime) < 0;
286 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
288 return se->vruntime - cfs_rq->min_vruntime;
291 static void update_min_vruntime(struct cfs_rq *cfs_rq)
293 u64 vruntime = cfs_rq->min_vruntime;
296 vruntime = cfs_rq->curr->vruntime;
298 if (cfs_rq->rb_leftmost) {
299 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
304 vruntime = se->vruntime;
306 vruntime = min_vruntime(vruntime, se->vruntime);
309 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
313 * Enqueue an entity into the rb-tree:
315 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
317 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
318 struct rb_node *parent = NULL;
319 struct sched_entity *entry;
320 s64 key = entity_key(cfs_rq, se);
324 * Find the right place in the rbtree:
328 entry = rb_entry(parent, struct sched_entity, run_node);
330 * We dont care about collisions. Nodes with
331 * the same key stay together.
333 if (key < entity_key(cfs_rq, entry)) {
334 link = &parent->rb_left;
336 link = &parent->rb_right;
342 * Maintain a cache of leftmost tree entries (it is frequently
346 cfs_rq->rb_leftmost = &se->run_node;
348 rb_link_node(&se->run_node, parent, link);
349 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
352 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
354 if (cfs_rq->rb_leftmost == &se->run_node) {
355 struct rb_node *next_node;
357 next_node = rb_next(&se->run_node);
358 cfs_rq->rb_leftmost = next_node;
361 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
364 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
366 struct rb_node *left = cfs_rq->rb_leftmost;
371 return rb_entry(left, struct sched_entity, run_node);
374 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
376 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
381 return rb_entry(last, struct sched_entity, run_node);
384 /**************************************************************
385 * Scheduling class statistics methods:
388 #ifdef CONFIG_SCHED_DEBUG
389 int sched_nr_latency_handler(struct ctl_table *table, int write,
390 void __user *buffer, size_t *lenp,
393 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
398 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
399 sysctl_sched_min_granularity);
408 static inline unsigned long
409 calc_delta_fair(unsigned long delta, struct sched_entity *se)
411 if (unlikely(se->load.weight != NICE_0_LOAD))
412 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
418 * The idea is to set a period in which each task runs once.
420 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
421 * this period because otherwise the slices get too small.
423 * p = (nr <= nl) ? l : l*nr/nl
425 static u64 __sched_period(unsigned long nr_running)
427 u64 period = sysctl_sched_latency;
428 unsigned long nr_latency = sched_nr_latency;
430 if (unlikely(nr_running > nr_latency)) {
431 period = sysctl_sched_min_granularity;
432 period *= nr_running;
439 * We calculate the wall-time slice from the period by taking a part
440 * proportional to the weight.
444 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
446 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
448 for_each_sched_entity(se) {
449 struct load_weight *load;
450 struct load_weight lw;
452 cfs_rq = cfs_rq_of(se);
453 load = &cfs_rq->load;
455 if (unlikely(!se->on_rq)) {
458 update_load_add(&lw, se->load.weight);
461 slice = calc_delta_mine(slice, se->load.weight, load);
467 * We calculate the vruntime slice of a to be inserted task
471 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
473 return calc_delta_fair(sched_slice(cfs_rq, se), se);
477 * Update the current task's runtime statistics. Skip current tasks that
478 * are not in our scheduling class.
481 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
482 unsigned long delta_exec)
484 unsigned long delta_exec_weighted;
486 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
488 curr->sum_exec_runtime += delta_exec;
489 schedstat_add(cfs_rq, exec_clock, delta_exec);
490 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
492 curr->vruntime += delta_exec_weighted;
493 update_min_vruntime(cfs_rq);
496 static void update_curr(struct cfs_rq *cfs_rq)
498 struct sched_entity *curr = cfs_rq->curr;
499 u64 now = rq_of(cfs_rq)->clock_task;
500 unsigned long delta_exec;
506 * Get the amount of time the current task was running
507 * since the last time we changed load (this cannot
508 * overflow on 32 bits):
510 delta_exec = (unsigned long)(now - curr->exec_start);
514 __update_curr(cfs_rq, curr, delta_exec);
515 curr->exec_start = now;
517 if (entity_is_task(curr)) {
518 struct task_struct *curtask = task_of(curr);
520 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
521 cpuacct_charge(curtask, delta_exec);
522 account_group_exec_runtime(curtask, delta_exec);
527 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
529 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
533 * Task is being enqueued - update stats:
535 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
538 * Are we enqueueing a waiting task? (for current tasks
539 * a dequeue/enqueue event is a NOP)
541 if (se != cfs_rq->curr)
542 update_stats_wait_start(cfs_rq, se);
546 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
548 schedstat_set(se->wait_max, max(se->wait_max,
549 rq_of(cfs_rq)->clock - se->wait_start));
550 schedstat_set(se->wait_count, se->wait_count + 1);
551 schedstat_set(se->wait_sum, se->wait_sum +
552 rq_of(cfs_rq)->clock - se->wait_start);
553 #ifdef CONFIG_SCHEDSTATS
554 if (entity_is_task(se)) {
555 trace_sched_stat_wait(task_of(se),
556 rq_of(cfs_rq)->clock - se->wait_start);
559 schedstat_set(se->wait_start, 0);
563 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
566 * Mark the end of the wait period if dequeueing a
569 if (se != cfs_rq->curr)
570 update_stats_wait_end(cfs_rq, se);
574 * We are picking a new current task - update its stats:
577 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
580 * We are starting a new run period:
582 se->exec_start = rq_of(cfs_rq)->clock_task;
585 /**************************************************
586 * Scheduling class queueing methods:
589 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
591 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
593 cfs_rq->task_weight += weight;
597 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
603 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
605 update_load_add(&cfs_rq->load, se->load.weight);
606 if (!parent_entity(se))
607 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
608 if (entity_is_task(se)) {
609 add_cfs_task_weight(cfs_rq, se->load.weight);
610 list_add(&se->group_node, &cfs_rq->tasks);
612 cfs_rq->nr_running++;
617 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
619 update_load_sub(&cfs_rq->load, se->load.weight);
620 if (!parent_entity(se))
621 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
622 if (entity_is_task(se)) {
623 add_cfs_task_weight(cfs_rq, -se->load.weight);
624 list_del_init(&se->group_node);
626 cfs_rq->nr_running--;
630 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
632 #ifdef CONFIG_SCHEDSTATS
633 struct task_struct *tsk = NULL;
635 if (entity_is_task(se))
638 if (se->sleep_start) {
639 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
644 if (unlikely(delta > se->sleep_max))
645 se->sleep_max = delta;
648 se->sum_sleep_runtime += delta;
651 account_scheduler_latency(tsk, delta >> 10, 1);
652 trace_sched_stat_sleep(tsk, delta);
655 if (se->block_start) {
656 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
661 if (unlikely(delta > se->block_max))
662 se->block_max = delta;
665 se->sum_sleep_runtime += delta;
668 if (tsk->in_iowait) {
669 se->iowait_sum += delta;
671 trace_sched_stat_iowait(tsk, delta);
675 * Blocking time is in units of nanosecs, so shift by
676 * 20 to get a milliseconds-range estimation of the
677 * amount of time that the task spent sleeping:
679 if (unlikely(prof_on == SLEEP_PROFILING)) {
680 profile_hits(SLEEP_PROFILING,
681 (void *)get_wchan(tsk),
684 account_scheduler_latency(tsk, delta >> 10, 0);
690 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
692 #ifdef CONFIG_SCHED_DEBUG
693 s64 d = se->vruntime - cfs_rq->min_vruntime;
698 if (d > 3*sysctl_sched_latency)
699 schedstat_inc(cfs_rq, nr_spread_over);
704 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
706 u64 vruntime = cfs_rq->min_vruntime;
709 * The 'current' period is already promised to the current tasks,
710 * however the extra weight of the new task will slow them down a
711 * little, place the new task so that it fits in the slot that
712 * stays open at the end.
714 if (initial && sched_feat(START_DEBIT))
715 vruntime += sched_vslice(cfs_rq, se);
717 /* sleeps up to a single latency don't count. */
718 if (!initial && sched_feat(FAIR_SLEEPERS)) {
719 unsigned long thresh = sysctl_sched_latency;
722 * Convert the sleeper threshold into virtual time.
723 * SCHED_IDLE is a special sub-class. We care about
724 * fairness only relative to other SCHED_IDLE tasks,
725 * all of which have the same weight.
727 if (sched_feat(NORMALIZED_SLEEPER) && (!entity_is_task(se) ||
728 task_of(se)->policy != SCHED_IDLE))
729 thresh = calc_delta_fair(thresh, se);
732 * Halve their sleep time's effect, to allow
733 * for a gentler effect of sleepers:
735 if (sched_feat(GENTLE_FAIR_SLEEPERS))
741 /* ensure we never gain time by being placed backwards. */
742 vruntime = max_vruntime(se->vruntime, vruntime);
744 se->vruntime = vruntime;
747 #define ENQUEUE_WAKEUP 1
748 #define ENQUEUE_MIGRATE 2
751 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
754 * Update the normalized vruntime before updating min_vruntime
755 * through callig update_curr().
757 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATE))
758 se->vruntime += cfs_rq->min_vruntime;
761 * Update run-time statistics of the 'current'.
764 account_entity_enqueue(cfs_rq, se);
766 if (flags & ENQUEUE_WAKEUP) {
767 place_entity(cfs_rq, se, 0);
768 enqueue_sleeper(cfs_rq, se);
771 update_stats_enqueue(cfs_rq, se);
772 check_spread(cfs_rq, se);
773 if (se != cfs_rq->curr)
774 __enqueue_entity(cfs_rq, se);
777 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
779 if (!se || cfs_rq->last == se)
782 if (!se || cfs_rq->next == se)
786 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
788 for_each_sched_entity(se)
789 __clear_buddies(cfs_rq_of(se), se);
793 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
796 * Update run-time statistics of the 'current'.
800 update_stats_dequeue(cfs_rq, se);
802 #ifdef CONFIG_SCHEDSTATS
803 if (entity_is_task(se)) {
804 struct task_struct *tsk = task_of(se);
806 if (tsk->state & TASK_INTERRUPTIBLE)
807 se->sleep_start = rq_of(cfs_rq)->clock;
808 if (tsk->state & TASK_UNINTERRUPTIBLE)
809 se->block_start = rq_of(cfs_rq)->clock;
814 clear_buddies(cfs_rq, se);
816 if (se != cfs_rq->curr)
817 __dequeue_entity(cfs_rq, se);
818 account_entity_dequeue(cfs_rq, se);
819 update_min_vruntime(cfs_rq);
822 * Normalize the entity after updating the min_vruntime because the
823 * update can refer to the ->curr item and we need to reflect this
824 * movement in our normalized position.
827 se->vruntime -= cfs_rq->min_vruntime;
831 * Preempt the current task with a newly woken task if needed:
834 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
836 unsigned long ideal_runtime, delta_exec;
838 ideal_runtime = sched_slice(cfs_rq, curr);
839 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
840 if (delta_exec > ideal_runtime) {
841 resched_task(rq_of(cfs_rq)->curr);
843 * The current task ran long enough, ensure it doesn't get
844 * re-elected due to buddy favours.
846 clear_buddies(cfs_rq, curr);
851 * Ensure that a task that missed wakeup preemption by a
852 * narrow margin doesn't have to wait for a full slice.
853 * This also mitigates buddy induced latencies under load.
855 if (!sched_feat(WAKEUP_PREEMPT))
858 if (delta_exec < sysctl_sched_min_granularity)
861 if (cfs_rq->nr_running > 1) {
862 struct sched_entity *se = __pick_next_entity(cfs_rq);
863 s64 delta = curr->vruntime - se->vruntime;
865 if (delta > ideal_runtime)
866 resched_task(rq_of(cfs_rq)->curr);
871 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
873 /* 'current' is not kept within the tree. */
876 * Any task has to be enqueued before it get to execute on
877 * a CPU. So account for the time it spent waiting on the
880 update_stats_wait_end(cfs_rq, se);
881 __dequeue_entity(cfs_rq, se);
884 update_stats_curr_start(cfs_rq, se);
886 #ifdef CONFIG_SCHEDSTATS
888 * Track our maximum slice length, if the CPU's load is at
889 * least twice that of our own weight (i.e. dont track it
890 * when there are only lesser-weight tasks around):
892 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
893 se->slice_max = max(se->slice_max,
894 se->sum_exec_runtime - se->prev_sum_exec_runtime);
897 se->prev_sum_exec_runtime = se->sum_exec_runtime;
901 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
903 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
905 struct sched_entity *se = __pick_next_entity(cfs_rq);
906 struct sched_entity *left = se;
908 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
912 * Prefer last buddy, try to return the CPU to a preempted task.
914 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
917 clear_buddies(cfs_rq, se);
922 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
925 * If still on the runqueue then deactivate_task()
926 * was not called and update_curr() has to be done:
931 check_spread(cfs_rq, prev);
933 update_stats_wait_start(cfs_rq, prev);
934 /* Put 'current' back into the tree. */
935 __enqueue_entity(cfs_rq, prev);
941 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
944 * Update run-time statistics of the 'current'.
948 #ifdef CONFIG_SCHED_HRTICK
950 * queued ticks are scheduled to match the slice, so don't bother
951 * validating it and just reschedule.
954 resched_task(rq_of(cfs_rq)->curr);
958 * don't let the period tick interfere with the hrtick preemption
960 if (!sched_feat(DOUBLE_TICK) &&
961 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
965 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
966 check_preempt_tick(cfs_rq, curr);
969 /**************************************************
970 * CFS operations on tasks:
973 #ifdef CONFIG_SCHED_HRTICK
974 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
976 struct sched_entity *se = &p->se;
977 struct cfs_rq *cfs_rq = cfs_rq_of(se);
979 WARN_ON(task_rq(p) != rq);
981 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
982 u64 slice = sched_slice(cfs_rq, se);
983 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
984 s64 delta = slice - ran;
993 * Don't schedule slices shorter than 10000ns, that just
994 * doesn't make sense. Rely on vruntime for fairness.
997 delta = max_t(s64, 10000LL, delta);
999 hrtick_start(rq, delta);
1004 * called from enqueue/dequeue and updates the hrtick when the
1005 * current task is from our class and nr_running is low enough
1008 static void hrtick_update(struct rq *rq)
1010 struct task_struct *curr = rq->curr;
1012 if (curr->sched_class != &fair_sched_class)
1015 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1016 hrtick_start_fair(rq, curr);
1018 #else /* !CONFIG_SCHED_HRTICK */
1020 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1024 static inline void hrtick_update(struct rq *rq)
1030 * The enqueue_task method is called before nr_running is
1031 * increased. Here we update the fair scheduling stats and
1032 * then put the task into the rbtree:
1035 enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1037 struct cfs_rq *cfs_rq;
1038 struct sched_entity *se = &p->se;
1042 flags |= ENQUEUE_WAKEUP;
1043 if (p->state == TASK_WAKING)
1044 flags |= ENQUEUE_MIGRATE;
1046 for_each_sched_entity(se) {
1049 cfs_rq = cfs_rq_of(se);
1050 enqueue_entity(cfs_rq, se, flags);
1051 flags = ENQUEUE_WAKEUP;
1058 * The dequeue_task method is called before nr_running is
1059 * decreased. We remove the task from the rbtree and
1060 * update the fair scheduling stats:
1062 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
1064 struct cfs_rq *cfs_rq;
1065 struct sched_entity *se = &p->se;
1067 for_each_sched_entity(se) {
1068 cfs_rq = cfs_rq_of(se);
1069 dequeue_entity(cfs_rq, se, sleep);
1070 /* Don't dequeue parent if it has other entities besides us */
1071 if (cfs_rq->load.weight)
1080 * sched_yield() support is very simple - we dequeue and enqueue.
1082 * If compat_yield is turned on then we requeue to the end of the tree.
1084 static void yield_task_fair(struct rq *rq)
1086 struct task_struct *curr = rq->curr;
1087 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1088 struct sched_entity *rightmost, *se = &curr->se;
1091 * Are we the only task in the tree?
1093 if (unlikely(cfs_rq->nr_running == 1))
1096 clear_buddies(cfs_rq, se);
1098 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1099 update_rq_clock(rq);
1101 * Update run-time statistics of the 'current'.
1103 update_curr(cfs_rq);
1108 * Find the rightmost entry in the rbtree:
1110 rightmost = __pick_last_entity(cfs_rq);
1112 * Already in the rightmost position?
1114 if (unlikely(!rightmost || entity_before(rightmost, se)))
1118 * Minimally necessary key value to be last in the tree:
1119 * Upon rescheduling, sched_class::put_prev_task() will place
1120 * 'current' within the tree based on its new key value.
1122 se->vruntime = rightmost->vruntime + 1;
1127 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1129 struct sched_entity *se = &p->se;
1130 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1132 se->vruntime -= cfs_rq->min_vruntime;
1135 #ifdef CONFIG_FAIR_GROUP_SCHED
1137 * effective_load() calculates the load change as seen from the root_task_group
1139 * Adding load to a group doesn't make a group heavier, but can cause movement
1140 * of group shares between cpus. Assuming the shares were perfectly aligned one
1141 * can calculate the shift in shares.
1143 * The problem is that perfectly aligning the shares is rather expensive, hence
1144 * we try to avoid doing that too often - see update_shares(), which ratelimits
1147 * We compensate this by not only taking the current delta into account, but
1148 * also considering the delta between when the shares were last adjusted and
1151 * We still saw a performance dip, some tracing learned us that between
1152 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1153 * significantly. Therefore try to bias the error in direction of failing
1154 * the affine wakeup.
1157 static long effective_load(struct task_group *tg, int cpu,
1160 struct sched_entity *se = tg->se[cpu];
1166 * By not taking the decrease of shares on the other cpu into
1167 * account our error leans towards reducing the affine wakeups.
1169 if (!wl && sched_feat(ASYM_EFF_LOAD))
1172 for_each_sched_entity(se) {
1173 long S, rw, s, a, b;
1177 * Instead of using this increment, also add the difference
1178 * between when the shares were last updated and now.
1180 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1184 S = se->my_q->tg->shares;
1185 s = se->my_q->shares;
1186 rw = se->my_q->rq_weight;
1197 * Assume the group is already running and will
1198 * thus already be accounted for in the weight.
1200 * That is, moving shares between CPUs, does not
1201 * alter the group weight.
1211 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1212 unsigned long wl, unsigned long wg)
1219 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1221 struct task_struct *curr = current;
1222 unsigned long this_load, load;
1223 int idx, this_cpu, prev_cpu;
1224 unsigned long tl_per_task;
1225 unsigned int imbalance;
1226 struct task_group *tg;
1227 unsigned long weight;
1231 this_cpu = smp_processor_id();
1232 prev_cpu = task_cpu(p);
1233 load = source_load(prev_cpu, idx);
1234 this_load = target_load(this_cpu, idx);
1237 if (sched_feat(SYNC_LESS) &&
1238 (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1239 p->se.avg_overlap > sysctl_sched_migration_cost))
1242 if (sched_feat(SYNC_MORE) &&
1243 (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1244 p->se.avg_overlap < sysctl_sched_migration_cost))
1249 * If sync wakeup then subtract the (maximum possible)
1250 * effect of the currently running task from the load
1251 * of the current CPU:
1255 tg = task_group(current);
1256 weight = current->se.load.weight;
1258 this_load += effective_load(tg, this_cpu, -weight, -weight);
1259 load += effective_load(tg, prev_cpu, 0, -weight);
1263 weight = p->se.load.weight;
1265 imbalance = 100 + (sd->imbalance_pct - 100) / 2;
1268 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1269 * due to the sync cause above having dropped this_load to 0, we'll
1270 * always have an imbalance, but there's really nothing you can do
1271 * about that, so that's good too.
1273 * Otherwise check if either cpus are near enough in load to allow this
1274 * task to be woken on this_cpu.
1276 balanced = !this_load ||
1277 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
1278 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1282 * If the currently running task will sleep within
1283 * a reasonable amount of time then attract this newly
1286 if (sync && balanced)
1289 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1290 tl_per_task = cpu_avg_load_per_task(this_cpu);
1293 (this_load <= load &&
1294 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1296 * This domain has SD_WAKE_AFFINE and
1297 * p is cache cold in this domain, and
1298 * there is no bad imbalance.
1300 schedstat_inc(sd, ttwu_move_affine);
1301 schedstat_inc(p, se.nr_wakeups_affine);
1309 * find_idlest_group finds and returns the least busy CPU group within the
1312 static struct sched_group *
1313 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1314 int this_cpu, int load_idx)
1316 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1317 unsigned long min_load = ULONG_MAX, this_load = 0;
1318 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1321 unsigned long load, avg_load;
1325 /* Skip over this group if it has no CPUs allowed */
1326 if (!cpumask_intersects(sched_group_cpus(group),
1330 local_group = cpumask_test_cpu(this_cpu,
1331 sched_group_cpus(group));
1333 /* Tally up the load of all CPUs in the group */
1336 for_each_cpu(i, sched_group_cpus(group)) {
1337 /* Bias balancing toward cpus of our domain */
1339 load = source_load(i, load_idx);
1341 load = target_load(i, load_idx);
1346 /* Adjust by relative CPU power of the group */
1347 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1350 this_load = avg_load;
1352 } else if (avg_load < min_load) {
1353 min_load = avg_load;
1356 } while (group = group->next, group != sd->groups);
1358 if (!idlest || 100*this_load < imbalance*min_load)
1364 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1367 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1369 unsigned long load, min_load = ULONG_MAX;
1373 /* Traverse only the allowed CPUs */
1374 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1375 load = weighted_cpuload(i);
1377 if (load < min_load || (load == min_load && i == this_cpu)) {
1387 * Try and locate an idle CPU in the sched_domain.
1389 static int select_idle_sibling(struct task_struct *p, int target)
1391 int cpu = smp_processor_id();
1392 int prev_cpu = task_cpu(p);
1393 struct sched_domain *sd;
1397 * If the task is going to be woken-up on this cpu and if it is
1398 * already idle, then it is the right target.
1400 if (target == cpu && idle_cpu(cpu))
1404 * If the task is going to be woken-up on the cpu where it previously
1405 * ran and if it is currently idle, then it the right target.
1407 if (target == prev_cpu && idle_cpu(prev_cpu))
1411 * Otherwise, iterate the domains and find an elegible idle cpu.
1413 for_each_domain(target, sd) {
1414 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1417 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1425 * Lets stop looking for an idle sibling when we reached
1426 * the domain that spans the current cpu and prev_cpu.
1428 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1429 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1437 * sched_balance_self: balance the current task (running on cpu) in domains
1438 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1441 * Balance, ie. select the least loaded group.
1443 * Returns the target CPU number, or the same CPU if no balancing is needed.
1445 * preempt must be disabled.
1448 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1450 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1451 int cpu = smp_processor_id();
1452 int prev_cpu = task_cpu(p);
1454 int want_affine = 0;
1456 int sync = wake_flags & WF_SYNC;
1458 if (sd_flag & SD_BALANCE_WAKE) {
1459 if (sched_feat(AFFINE_WAKEUPS) &&
1460 cpumask_test_cpu(cpu, &p->cpus_allowed))
1465 for_each_domain(cpu, tmp) {
1466 if (!(tmp->flags & SD_LOAD_BALANCE))
1470 * If power savings logic is enabled for a domain, see if we
1471 * are not overloaded, if so, don't balance wider.
1473 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1474 unsigned long power = 0;
1475 unsigned long nr_running = 0;
1476 unsigned long capacity;
1479 for_each_cpu(i, sched_domain_span(tmp)) {
1480 power += power_of(i);
1481 nr_running += cpu_rq(i)->cfs.nr_running;
1484 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1486 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1489 if (nr_running < capacity)
1494 * If both cpu and prev_cpu are part of this domain,
1495 * cpu is a valid SD_WAKE_AFFINE target.
1497 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1498 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1503 if (!want_sd && !want_affine)
1506 if (!(tmp->flags & sd_flag))
1513 #ifdef CONFIG_FAIR_GROUP_SCHED
1514 if (sched_feat(LB_SHARES_UPDATE)) {
1516 * Pick the largest domain to update shares over
1519 if (affine_sd && (!tmp || affine_sd->span_weight > sd->span_weight))
1523 spin_unlock(&rq->lock);
1525 spin_lock(&rq->lock);
1531 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1532 return select_idle_sibling(p, cpu);
1534 return select_idle_sibling(p, prev_cpu);
1538 int load_idx = sd->forkexec_idx;
1539 struct sched_group *group;
1542 if (!(sd->flags & sd_flag)) {
1547 if (sd_flag & SD_BALANCE_WAKE)
1548 load_idx = sd->wake_idx;
1550 group = find_idlest_group(sd, p, cpu, load_idx);
1556 new_cpu = find_idlest_cpu(group, p, cpu);
1557 if (new_cpu == -1 || new_cpu == cpu) {
1558 /* Now try balancing at a lower domain level of cpu */
1563 /* Now try balancing at a lower domain level of new_cpu */
1565 weight = sd->span_weight;
1567 for_each_domain(cpu, tmp) {
1568 if (weight <= tmp->span_weight)
1570 if (tmp->flags & sd_flag)
1573 /* while loop will break here if sd == NULL */
1578 #endif /* CONFIG_SMP */
1581 * Adaptive granularity
1583 * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1584 * with the limit of wakeup_gran -- when it never does a wakeup.
1586 * So the smaller avg_wakeup is the faster we want this task to preempt,
1587 * but we don't want to treat the preemptee unfairly and therefore allow it
1588 * to run for at least the amount of time we'd like to run.
1590 * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1592 * NOTE: we use *nr_running to scale with load, this nicely matches the
1593 * degrading latency on load.
1595 static unsigned long
1596 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1598 u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1599 u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1602 if (this_run < expected_wakeup)
1603 gran = expected_wakeup - this_run;
1605 return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1608 static unsigned long
1609 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1611 unsigned long gran = sysctl_sched_wakeup_granularity;
1613 if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1614 gran = adaptive_gran(curr, se);
1617 * Since its curr running now, convert the gran from real-time
1618 * to virtual-time in his units.
1620 if (sched_feat(ASYM_GRAN)) {
1622 * By using 'se' instead of 'curr' we penalize light tasks, so
1623 * they get preempted easier. That is, if 'se' < 'curr' then
1624 * the resulting gran will be larger, therefore penalizing the
1625 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1626 * be smaller, again penalizing the lighter task.
1628 * This is especially important for buddies when the leftmost
1629 * task is higher priority than the buddy.
1631 if (unlikely(se->load.weight != NICE_0_LOAD))
1632 gran = calc_delta_fair(gran, se);
1634 if (unlikely(curr->load.weight != NICE_0_LOAD))
1635 gran = calc_delta_fair(gran, curr);
1642 * Should 'se' preempt 'curr'.
1656 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1658 s64 gran, vdiff = curr->vruntime - se->vruntime;
1663 gran = wakeup_gran(curr, se);
1670 static void set_last_buddy(struct sched_entity *se)
1672 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1673 for_each_sched_entity(se)
1674 cfs_rq_of(se)->last = se;
1678 static void set_next_buddy(struct sched_entity *se)
1680 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1681 for_each_sched_entity(se)
1682 cfs_rq_of(se)->next = se;
1687 * Preempt the current task with a newly woken task if needed:
1689 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1691 struct task_struct *curr = rq->curr;
1692 struct sched_entity *se = &curr->se, *pse = &p->se;
1693 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1694 int sync = wake_flags & WF_SYNC;
1695 int scale = cfs_rq->nr_running >= sched_nr_latency;
1697 update_curr(cfs_rq);
1699 if (unlikely(rt_prio(p->prio))) {
1704 if (unlikely(p->sched_class != &fair_sched_class))
1707 if (unlikely(se == pse))
1710 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1711 set_next_buddy(pse);
1714 * We can come here with TIF_NEED_RESCHED already set from new task
1717 if (test_tsk_need_resched(curr))
1721 * Batch and idle tasks do not preempt (their preemption is driven by
1724 if (unlikely(p->policy != SCHED_NORMAL))
1727 /* Idle tasks are by definition preempted by everybody. */
1728 if (unlikely(curr->policy == SCHED_IDLE)) {
1733 if ((sched_feat(WAKEUP_SYNC) && sync) ||
1734 (sched_feat(WAKEUP_OVERLAP) &&
1735 (se->avg_overlap < sysctl_sched_migration_cost &&
1736 pse->avg_overlap < sysctl_sched_migration_cost))) {
1741 if (sched_feat(WAKEUP_RUNNING)) {
1742 if (pse->avg_running < se->avg_running) {
1743 set_next_buddy(pse);
1749 if (!sched_feat(WAKEUP_PREEMPT))
1752 find_matching_se(&se, &pse);
1756 if (wakeup_preempt_entity(se, pse) == 1) {
1759 * Only set the backward buddy when the current task is still
1760 * on the rq. This can happen when a wakeup gets interleaved
1761 * with schedule on the ->pre_schedule() or idle_balance()
1762 * point, either of which can * drop the rq lock.
1764 * Also, during early boot the idle thread is in the fair class,
1765 * for obvious reasons its a bad idea to schedule back to it.
1767 if (unlikely(!se->on_rq || curr == rq->idle))
1769 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1774 static struct task_struct *pick_next_task_fair(struct rq *rq)
1776 struct task_struct *p;
1777 struct cfs_rq *cfs_rq = &rq->cfs;
1778 struct sched_entity *se;
1780 if (unlikely(!cfs_rq->nr_running))
1784 se = pick_next_entity(cfs_rq);
1785 set_next_entity(cfs_rq, se);
1786 cfs_rq = group_cfs_rq(se);
1790 hrtick_start_fair(rq, p);
1796 * Account for a descheduled task:
1798 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1800 struct sched_entity *se = &prev->se;
1801 struct cfs_rq *cfs_rq;
1803 for_each_sched_entity(se) {
1804 cfs_rq = cfs_rq_of(se);
1805 put_prev_entity(cfs_rq, se);
1810 /**************************************************
1811 * Fair scheduling class load-balancing methods:
1815 * Load-balancing iterator. Note: while the runqueue stays locked
1816 * during the whole iteration, the current task might be
1817 * dequeued so the iterator has to be dequeue-safe. Here we
1818 * achieve that by always pre-iterating before returning
1821 static struct task_struct *
1822 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1824 struct task_struct *p = NULL;
1825 struct sched_entity *se;
1827 if (next == &cfs_rq->tasks)
1830 se = list_entry(next, struct sched_entity, group_node);
1832 cfs_rq->balance_iterator = next->next;
1837 static struct task_struct *load_balance_start_fair(void *arg)
1839 struct cfs_rq *cfs_rq = arg;
1841 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1844 static struct task_struct *load_balance_next_fair(void *arg)
1846 struct cfs_rq *cfs_rq = arg;
1848 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1851 static unsigned long
1852 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1853 unsigned long max_load_move, struct sched_domain *sd,
1854 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1855 struct cfs_rq *cfs_rq)
1857 struct rq_iterator cfs_rq_iterator;
1859 cfs_rq_iterator.start = load_balance_start_fair;
1860 cfs_rq_iterator.next = load_balance_next_fair;
1861 cfs_rq_iterator.arg = cfs_rq;
1863 return balance_tasks(this_rq, this_cpu, busiest,
1864 max_load_move, sd, idle, all_pinned,
1865 this_best_prio, &cfs_rq_iterator);
1868 #ifdef CONFIG_FAIR_GROUP_SCHED
1869 static unsigned long
1870 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1871 unsigned long max_load_move,
1872 struct sched_domain *sd, enum cpu_idle_type idle,
1873 int *all_pinned, int *this_best_prio)
1875 long rem_load_move = max_load_move;
1876 int busiest_cpu = cpu_of(busiest);
1877 struct task_group *tg;
1880 update_h_load(busiest_cpu);
1882 list_for_each_entry_rcu(tg, &task_groups, list) {
1883 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1884 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1885 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1886 u64 rem_load, moved_load;
1891 if (!busiest_cfs_rq->task_weight)
1894 rem_load = (u64)rem_load_move * busiest_weight;
1895 rem_load = div_u64(rem_load, busiest_h_load + 1);
1897 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1898 rem_load, sd, idle, all_pinned, this_best_prio,
1899 tg->cfs_rq[busiest_cpu]);
1904 moved_load *= busiest_h_load;
1905 moved_load = div_u64(moved_load, busiest_weight + 1);
1907 rem_load_move -= moved_load;
1908 if (rem_load_move < 0)
1913 return max_load_move - rem_load_move;
1916 static unsigned long
1917 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1918 unsigned long max_load_move,
1919 struct sched_domain *sd, enum cpu_idle_type idle,
1920 int *all_pinned, int *this_best_prio)
1922 return __load_balance_fair(this_rq, this_cpu, busiest,
1923 max_load_move, sd, idle, all_pinned,
1924 this_best_prio, &busiest->cfs);
1929 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1930 struct sched_domain *sd, enum cpu_idle_type idle)
1932 struct cfs_rq *busy_cfs_rq;
1933 struct rq_iterator cfs_rq_iterator;
1935 cfs_rq_iterator.start = load_balance_start_fair;
1936 cfs_rq_iterator.next = load_balance_next_fair;
1938 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1940 * pass busy_cfs_rq argument into
1941 * load_balance_[start|next]_fair iterators
1943 cfs_rq_iterator.arg = busy_cfs_rq;
1944 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1952 static void rq_online_fair(struct rq *rq)
1957 static void rq_offline_fair(struct rq *rq)
1962 #endif /* CONFIG_SMP */
1965 * scheduler tick hitting a task of our scheduling class:
1967 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1969 struct cfs_rq *cfs_rq;
1970 struct sched_entity *se = &curr->se;
1972 for_each_sched_entity(se) {
1973 cfs_rq = cfs_rq_of(se);
1974 entity_tick(cfs_rq, se, queued);
1979 * called on fork with the child task as argument from the parent's context
1980 * - child not yet on the tasklist
1981 * - preemption disabled
1983 static void task_fork_fair(struct task_struct *p)
1985 struct cfs_rq *cfs_rq = task_cfs_rq(current);
1986 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1987 int this_cpu = smp_processor_id();
1988 struct rq *rq = this_rq();
1989 unsigned long flags;
1991 spin_lock_irqsave(&rq->lock, flags);
1993 update_rq_clock(rq);
1995 if (unlikely(task_cpu(p) != this_cpu)) {
1997 __set_task_cpu(p, this_cpu);
2001 update_curr(cfs_rq);
2004 se->vruntime = curr->vruntime;
2005 place_entity(cfs_rq, se, 1);
2007 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
2009 * Upon rescheduling, sched_class::put_prev_task() will place
2010 * 'current' within the tree based on its new key value.
2012 swap(curr->vruntime, se->vruntime);
2013 resched_task(rq->curr);
2016 se->vruntime -= cfs_rq->min_vruntime;
2018 spin_unlock_irqrestore(&rq->lock, flags);
2022 * Priority of the task has changed. Check to see if we preempt
2025 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
2026 int oldprio, int running)
2029 * Reschedule if we are currently running on this runqueue and
2030 * our priority decreased, or if we are not currently running on
2031 * this runqueue and our priority is higher than the current's
2034 if (p->prio > oldprio)
2035 resched_task(rq->curr);
2037 check_preempt_curr(rq, p, 0);
2041 * We switched to the sched_fair class.
2043 static void switched_to_fair(struct rq *rq, struct task_struct *p,
2047 * We were most likely switched from sched_rt, so
2048 * kick off the schedule if running, otherwise just see
2049 * if we can still preempt the current task.
2052 resched_task(rq->curr);
2054 check_preempt_curr(rq, p, 0);
2057 /* Account for a task changing its policy or group.
2059 * This routine is mostly called to set cfs_rq->curr field when a task
2060 * migrates between groups/classes.
2062 static void set_curr_task_fair(struct rq *rq)
2064 struct sched_entity *se = &rq->curr->se;
2066 for_each_sched_entity(se)
2067 set_next_entity(cfs_rq_of(se), se);
2070 #ifdef CONFIG_FAIR_GROUP_SCHED
2071 static void task_move_group_fair(struct task_struct *p, int on_rq)
2074 * If the task was not on the rq at the time of this cgroup movement
2075 * it must have been asleep, sleeping tasks keep their ->vruntime
2076 * absolute on their old rq until wakeup (needed for the fair sleeper
2077 * bonus in place_entity()).
2079 * If it was on the rq, we've just 'preempted' it, which does convert
2080 * ->vruntime to a relative base.
2082 * Make sure both cases convert their relative position when migrating
2083 * to another cgroup's rq. This does somewhat interfere with the
2084 * fair sleeper stuff for the first placement, but who cares.
2087 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
2088 set_task_rq(p, task_cpu(p));
2090 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
2094 unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
2096 struct sched_entity *se = &task->se;
2097 unsigned int rr_interval = 0;
2100 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
2103 if (rq->cfs.load.weight)
2104 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
2110 * All the scheduling class methods:
2112 static const struct sched_class fair_sched_class = {
2113 .next = &idle_sched_class,
2114 .enqueue_task = enqueue_task_fair,
2115 .dequeue_task = dequeue_task_fair,
2116 .yield_task = yield_task_fair,
2118 .check_preempt_curr = check_preempt_wakeup,
2120 .pick_next_task = pick_next_task_fair,
2121 .put_prev_task = put_prev_task_fair,
2124 .select_task_rq = select_task_rq_fair,
2126 .load_balance = load_balance_fair,
2127 .move_one_task = move_one_task_fair,
2128 .rq_online = rq_online_fair,
2129 .rq_offline = rq_offline_fair,
2131 .task_waking = task_waking_fair,
2134 .set_curr_task = set_curr_task_fair,
2135 .task_tick = task_tick_fair,
2136 .task_fork = task_fork_fair,
2138 .prio_changed = prio_changed_fair,
2139 .switched_to = switched_to_fair,
2141 .get_rr_interval = get_rr_interval_fair,
2143 #ifdef CONFIG_FAIR_GROUP_SCHED
2144 .task_move_group = task_move_group_fair,
2148 #ifdef CONFIG_SCHED_DEBUG
2149 static void print_cfs_stats(struct seq_file *m, int cpu)
2151 struct cfs_rq *cfs_rq;
2154 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
2155 print_cfs_rq(m, cpu, cfs_rq);