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>
24 #include <linux/sched.h>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency = 5000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 5000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 1000000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 1000000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 5;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
92 static const struct sched_class fair_sched_class;
94 /**************************************************************
95 * CFS operations on generic schedulable entities:
98 #ifdef CONFIG_FAIR_GROUP_SCHED
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
106 /* An entity is a task if it doesn't "own" a runqueue */
107 #define entity_is_task(se) (!se->my_q)
109 static inline struct task_struct *task_of(struct sched_entity *se)
111 #ifdef CONFIG_SCHED_DEBUG
112 WARN_ON_ONCE(!entity_is_task(se));
114 return container_of(se, struct task_struct, se);
117 /* Walk up scheduling entities hierarchy */
118 #define for_each_sched_entity(se) \
119 for (; se; se = se->parent)
121 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
126 /* runqueue on which this entity is (to be) queued */
127 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
132 /* runqueue "owned" by this group */
133 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
138 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
139 * another cpu ('this_cpu')
141 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
143 return cfs_rq->tg->cfs_rq[this_cpu];
146 /* Iterate thr' all leaf cfs_rq's on a runqueue */
147 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
148 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
150 /* Do the two (enqueued) entities belong to the same group ? */
152 is_same_group(struct sched_entity *se, struct sched_entity *pse)
154 if (se->cfs_rq == pse->cfs_rq)
160 static inline struct sched_entity *parent_entity(struct sched_entity *se)
165 /* return depth at which a sched entity is present in the hierarchy */
166 static inline int depth_se(struct sched_entity *se)
170 for_each_sched_entity(se)
177 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
179 int se_depth, pse_depth;
182 * preemption test can be made between sibling entities who are in the
183 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
184 * both tasks until we find their ancestors who are siblings of common
188 /* First walk up until both entities are at same depth */
189 se_depth = depth_se(*se);
190 pse_depth = depth_se(*pse);
192 while (se_depth > pse_depth) {
194 *se = parent_entity(*se);
197 while (pse_depth > se_depth) {
199 *pse = parent_entity(*pse);
202 while (!is_same_group(*se, *pse)) {
203 *se = parent_entity(*se);
204 *pse = parent_entity(*pse);
208 #else /* !CONFIG_FAIR_GROUP_SCHED */
210 static inline struct task_struct *task_of(struct sched_entity *se)
212 return container_of(se, struct task_struct, se);
215 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
217 return container_of(cfs_rq, struct rq, cfs);
220 #define entity_is_task(se) 1
222 #define for_each_sched_entity(se) \
223 for (; se; se = NULL)
225 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
227 return &task_rq(p)->cfs;
230 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
232 struct task_struct *p = task_of(se);
233 struct rq *rq = task_rq(p);
238 /* runqueue "owned" by this group */
239 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
244 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
246 return &cpu_rq(this_cpu)->cfs;
249 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
250 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
253 is_same_group(struct sched_entity *se, struct sched_entity *pse)
258 static inline struct sched_entity *parent_entity(struct sched_entity *se)
264 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
268 #endif /* CONFIG_FAIR_GROUP_SCHED */
271 /**************************************************************
272 * Scheduling class tree data structure manipulation methods:
275 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
277 s64 delta = (s64)(vruntime - min_vruntime);
279 min_vruntime = vruntime;
284 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
286 s64 delta = (s64)(vruntime - min_vruntime);
288 min_vruntime = vruntime;
293 static inline int entity_before(struct sched_entity *a,
294 struct sched_entity *b)
296 return (s64)(a->vruntime - b->vruntime) < 0;
299 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
301 return se->vruntime - cfs_rq->min_vruntime;
304 static void update_min_vruntime(struct cfs_rq *cfs_rq)
306 u64 vruntime = cfs_rq->min_vruntime;
309 vruntime = cfs_rq->curr->vruntime;
311 if (cfs_rq->rb_leftmost) {
312 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
317 vruntime = se->vruntime;
319 vruntime = min_vruntime(vruntime, se->vruntime);
322 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
326 * Enqueue an entity into the rb-tree:
328 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
330 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
331 struct rb_node *parent = NULL;
332 struct sched_entity *entry;
333 s64 key = entity_key(cfs_rq, se);
337 * Find the right place in the rbtree:
341 entry = rb_entry(parent, struct sched_entity, run_node);
343 * We dont care about collisions. Nodes with
344 * the same key stay together.
346 if (key < entity_key(cfs_rq, entry)) {
347 link = &parent->rb_left;
349 link = &parent->rb_right;
355 * Maintain a cache of leftmost tree entries (it is frequently
359 cfs_rq->rb_leftmost = &se->run_node;
361 rb_link_node(&se->run_node, parent, link);
362 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
365 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
367 if (cfs_rq->rb_leftmost == &se->run_node) {
368 struct rb_node *next_node;
370 next_node = rb_next(&se->run_node);
371 cfs_rq->rb_leftmost = next_node;
374 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
377 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
379 struct rb_node *left = cfs_rq->rb_leftmost;
384 return rb_entry(left, struct sched_entity, run_node);
387 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
389 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
394 return rb_entry(last, struct sched_entity, run_node);
397 /**************************************************************
398 * Scheduling class statistics methods:
401 #ifdef CONFIG_SCHED_DEBUG
402 int sched_proc_update_handler(struct ctl_table *table, int write,
403 void __user *buffer, size_t *lenp,
406 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
407 int factor = get_update_sysctl_factor();
412 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
413 sysctl_sched_min_granularity);
415 #define WRT_SYSCTL(name) \
416 (normalized_sysctl_##name = sysctl_##name / (factor))
417 WRT_SYSCTL(sched_min_granularity);
418 WRT_SYSCTL(sched_latency);
419 WRT_SYSCTL(sched_wakeup_granularity);
420 WRT_SYSCTL(sched_shares_ratelimit);
430 static inline unsigned long
431 calc_delta_fair(unsigned long delta, struct sched_entity *se)
433 if (unlikely(se->load.weight != NICE_0_LOAD))
434 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
440 * The idea is to set a period in which each task runs once.
442 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
443 * this period because otherwise the slices get too small.
445 * p = (nr <= nl) ? l : l*nr/nl
447 static u64 __sched_period(unsigned long nr_running)
449 u64 period = sysctl_sched_latency;
450 unsigned long nr_latency = sched_nr_latency;
452 if (unlikely(nr_running > nr_latency)) {
453 period = sysctl_sched_min_granularity;
454 period *= nr_running;
461 * We calculate the wall-time slice from the period by taking a part
462 * proportional to the weight.
466 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
468 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
470 for_each_sched_entity(se) {
471 struct load_weight *load;
472 struct load_weight lw;
474 cfs_rq = cfs_rq_of(se);
475 load = &cfs_rq->load;
477 if (unlikely(!se->on_rq)) {
480 update_load_add(&lw, se->load.weight);
483 slice = calc_delta_mine(slice, se->load.weight, load);
489 * We calculate the vruntime slice of a to be inserted task
493 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
495 return calc_delta_fair(sched_slice(cfs_rq, se), se);
499 * Update the current task's runtime statistics. Skip current tasks that
500 * are not in our scheduling class.
503 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
504 unsigned long delta_exec)
506 unsigned long delta_exec_weighted;
508 schedstat_set(curr->statistics.exec_max,
509 max((u64)delta_exec, curr->statistics.exec_max));
511 curr->sum_exec_runtime += delta_exec;
512 schedstat_add(cfs_rq, exec_clock, delta_exec);
513 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
515 curr->vruntime += delta_exec_weighted;
516 update_min_vruntime(cfs_rq);
519 static void update_curr(struct cfs_rq *cfs_rq)
521 struct sched_entity *curr = cfs_rq->curr;
522 u64 now = rq_of(cfs_rq)->clock;
523 unsigned long delta_exec;
529 * Get the amount of time the current task was running
530 * since the last time we changed load (this cannot
531 * overflow on 32 bits):
533 delta_exec = (unsigned long)(now - curr->exec_start);
537 __update_curr(cfs_rq, curr, delta_exec);
538 curr->exec_start = now;
540 if (entity_is_task(curr)) {
541 struct task_struct *curtask = task_of(curr);
543 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
544 cpuacct_charge(curtask, delta_exec);
545 account_group_exec_runtime(curtask, delta_exec);
550 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
552 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
556 * Task is being enqueued - update stats:
558 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
561 * Are we enqueueing a waiting task? (for current tasks
562 * a dequeue/enqueue event is a NOP)
564 if (se != cfs_rq->curr)
565 update_stats_wait_start(cfs_rq, se);
569 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
571 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
572 rq_of(cfs_rq)->clock - se->statistics.wait_start));
573 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
574 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
575 rq_of(cfs_rq)->clock - se->statistics.wait_start);
576 #ifdef CONFIG_SCHEDSTATS
577 if (entity_is_task(se)) {
578 trace_sched_stat_wait(task_of(se),
579 rq_of(cfs_rq)->clock - se->statistics.wait_start);
582 schedstat_set(se->statistics.wait_start, 0);
586 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
589 * Mark the end of the wait period if dequeueing a
592 if (se != cfs_rq->curr)
593 update_stats_wait_end(cfs_rq, se);
597 * We are picking a new current task - update its stats:
600 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
603 * We are starting a new run period:
605 se->exec_start = rq_of(cfs_rq)->clock;
608 /**************************************************
609 * Scheduling class queueing methods:
612 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
614 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
616 cfs_rq->task_weight += weight;
620 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
626 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 update_load_add(&cfs_rq->load, se->load.weight);
629 if (!parent_entity(se))
630 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
631 if (entity_is_task(se)) {
632 add_cfs_task_weight(cfs_rq, se->load.weight);
633 list_add(&se->group_node, &cfs_rq->tasks);
635 cfs_rq->nr_running++;
640 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
642 update_load_sub(&cfs_rq->load, se->load.weight);
643 if (!parent_entity(se))
644 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
645 if (entity_is_task(se)) {
646 add_cfs_task_weight(cfs_rq, -se->load.weight);
647 list_del_init(&se->group_node);
649 cfs_rq->nr_running--;
653 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
655 #ifdef CONFIG_SCHEDSTATS
656 struct task_struct *tsk = NULL;
658 if (entity_is_task(se))
661 if (se->statistics.sleep_start) {
662 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
667 if (unlikely(delta > se->statistics.sleep_max))
668 se->statistics.sleep_max = delta;
670 se->statistics.sleep_start = 0;
671 se->statistics.sum_sleep_runtime += delta;
674 account_scheduler_latency(tsk, delta >> 10, 1);
675 trace_sched_stat_sleep(tsk, delta);
678 if (se->statistics.block_start) {
679 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
684 if (unlikely(delta > se->statistics.block_max))
685 se->statistics.block_max = delta;
687 se->statistics.block_start = 0;
688 se->statistics.sum_sleep_runtime += delta;
691 if (tsk->in_iowait) {
692 se->statistics.iowait_sum += delta;
693 se->statistics.iowait_count++;
694 trace_sched_stat_iowait(tsk, delta);
698 * Blocking time is in units of nanosecs, so shift by
699 * 20 to get a milliseconds-range estimation of the
700 * amount of time that the task spent sleeping:
702 if (unlikely(prof_on == SLEEP_PROFILING)) {
703 profile_hits(SLEEP_PROFILING,
704 (void *)get_wchan(tsk),
707 account_scheduler_latency(tsk, delta >> 10, 0);
713 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
715 #ifdef CONFIG_SCHED_DEBUG
716 s64 d = se->vruntime - cfs_rq->min_vruntime;
721 if (d > 3*sysctl_sched_latency)
722 schedstat_inc(cfs_rq, nr_spread_over);
727 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
729 u64 vruntime = cfs_rq->min_vruntime;
732 * The 'current' period is already promised to the current tasks,
733 * however the extra weight of the new task will slow them down a
734 * little, place the new task so that it fits in the slot that
735 * stays open at the end.
737 if (initial && sched_feat(START_DEBIT))
738 vruntime += sched_vslice(cfs_rq, se);
740 /* sleeps up to a single latency don't count. */
741 if (!initial && sched_feat(FAIR_SLEEPERS)) {
742 unsigned long thresh = sysctl_sched_latency;
745 * Convert the sleeper threshold into virtual time.
746 * SCHED_IDLE is a special sub-class. We care about
747 * fairness only relative to other SCHED_IDLE tasks,
748 * all of which have the same weight.
750 if (sched_feat(NORMALIZED_SLEEPER) && (!entity_is_task(se) ||
751 task_of(se)->policy != SCHED_IDLE))
752 thresh = calc_delta_fair(thresh, se);
755 * Halve their sleep time's effect, to allow
756 * for a gentler effect of sleepers:
758 if (sched_feat(GENTLE_FAIR_SLEEPERS))
764 /* ensure we never gain time by being placed backwards. */
765 vruntime = max_vruntime(se->vruntime, vruntime);
767 se->vruntime = vruntime;
770 #define ENQUEUE_WAKEUP 1
771 #define ENQUEUE_MIGRATE 2
774 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
777 * Update the normalized vruntime before updating min_vruntime
778 * through callig update_curr().
780 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATE))
781 se->vruntime += cfs_rq->min_vruntime;
784 * Update run-time statistics of the 'current'.
787 account_entity_enqueue(cfs_rq, se);
789 if (flags & ENQUEUE_WAKEUP) {
790 place_entity(cfs_rq, se, 0);
791 enqueue_sleeper(cfs_rq, se);
794 update_stats_enqueue(cfs_rq, se);
795 check_spread(cfs_rq, se);
796 if (se != cfs_rq->curr)
797 __enqueue_entity(cfs_rq, se);
800 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
802 if (!se || cfs_rq->last == se)
805 if (!se || cfs_rq->next == se)
809 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
811 for_each_sched_entity(se)
812 __clear_buddies(cfs_rq_of(se), se);
816 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
819 * Update run-time statistics of the 'current'.
823 update_stats_dequeue(cfs_rq, se);
825 #ifdef CONFIG_SCHEDSTATS
826 if (entity_is_task(se)) {
827 struct task_struct *tsk = task_of(se);
829 if (tsk->state & TASK_INTERRUPTIBLE)
830 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
831 if (tsk->state & TASK_UNINTERRUPTIBLE)
832 se->statistics.block_start = rq_of(cfs_rq)->clock;
837 clear_buddies(cfs_rq, se);
839 if (se != cfs_rq->curr)
840 __dequeue_entity(cfs_rq, se);
841 account_entity_dequeue(cfs_rq, se);
842 update_min_vruntime(cfs_rq);
845 * Normalize the entity after updating the min_vruntime because the
846 * update can refer to the ->curr item and we need to reflect this
847 * movement in our normalized position.
850 se->vruntime -= cfs_rq->min_vruntime;
854 * Preempt the current task with a newly woken task if needed:
857 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
859 unsigned long ideal_runtime, delta_exec;
861 ideal_runtime = sched_slice(cfs_rq, curr);
862 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
863 if (delta_exec > ideal_runtime) {
864 resched_task(rq_of(cfs_rq)->curr);
866 * The current task ran long enough, ensure it doesn't get
867 * re-elected due to buddy favours.
869 clear_buddies(cfs_rq, curr);
874 * Ensure that a task that missed wakeup preemption by a
875 * narrow margin doesn't have to wait for a full slice.
876 * This also mitigates buddy induced latencies under load.
878 if (!sched_feat(WAKEUP_PREEMPT))
881 if (delta_exec < sysctl_sched_min_granularity)
884 if (cfs_rq->nr_running > 1) {
885 struct sched_entity *se = __pick_next_entity(cfs_rq);
886 s64 delta = curr->vruntime - se->vruntime;
888 if (delta > ideal_runtime)
889 resched_task(rq_of(cfs_rq)->curr);
894 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
896 /* 'current' is not kept within the tree. */
899 * Any task has to be enqueued before it get to execute on
900 * a CPU. So account for the time it spent waiting on the
903 update_stats_wait_end(cfs_rq, se);
904 __dequeue_entity(cfs_rq, se);
907 update_stats_curr_start(cfs_rq, se);
909 #ifdef CONFIG_SCHEDSTATS
911 * Track our maximum slice length, if the CPU's load is at
912 * least twice that of our own weight (i.e. dont track it
913 * when there are only lesser-weight tasks around):
915 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
916 se->statistics.slice_max = max(se->statistics.slice_max,
917 se->sum_exec_runtime - se->prev_sum_exec_runtime);
920 se->prev_sum_exec_runtime = se->sum_exec_runtime;
924 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
926 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
928 struct sched_entity *se = __pick_next_entity(cfs_rq);
929 struct sched_entity *left = se;
931 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
935 * Prefer last buddy, try to return the CPU to a preempted task.
937 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
940 clear_buddies(cfs_rq, se);
945 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
948 * If still on the runqueue then deactivate_task()
949 * was not called and update_curr() has to be done:
954 check_spread(cfs_rq, prev);
956 update_stats_wait_start(cfs_rq, prev);
957 /* Put 'current' back into the tree. */
958 __enqueue_entity(cfs_rq, prev);
964 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
967 * Update run-time statistics of the 'current'.
971 #ifdef CONFIG_SCHED_HRTICK
973 * queued ticks are scheduled to match the slice, so don't bother
974 * validating it and just reschedule.
977 resched_task(rq_of(cfs_rq)->curr);
981 * don't let the period tick interfere with the hrtick preemption
983 if (!sched_feat(DOUBLE_TICK) &&
984 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
988 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
989 check_preempt_tick(cfs_rq, curr);
992 /**************************************************
993 * CFS operations on tasks:
996 #ifdef CONFIG_SCHED_HRTICK
997 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
999 struct sched_entity *se = &p->se;
1000 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1002 WARN_ON(task_rq(p) != rq);
1004 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1005 u64 slice = sched_slice(cfs_rq, se);
1006 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1007 s64 delta = slice - ran;
1016 * Don't schedule slices shorter than 10000ns, that just
1017 * doesn't make sense. Rely on vruntime for fairness.
1020 delta = max_t(s64, 10000LL, delta);
1022 hrtick_start(rq, delta);
1027 * called from enqueue/dequeue and updates the hrtick when the
1028 * current task is from our class and nr_running is low enough
1031 static void hrtick_update(struct rq *rq)
1033 struct task_struct *curr = rq->curr;
1035 if (curr->sched_class != &fair_sched_class)
1038 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1039 hrtick_start_fair(rq, curr);
1041 #else /* !CONFIG_SCHED_HRTICK */
1043 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1047 static inline void hrtick_update(struct rq *rq)
1053 * The enqueue_task method is called before nr_running is
1054 * increased. Here we update the fair scheduling stats and
1055 * then put the task into the rbtree:
1058 enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1060 struct cfs_rq *cfs_rq;
1061 struct sched_entity *se = &p->se;
1065 flags |= ENQUEUE_WAKEUP;
1066 if (p->state == TASK_WAKING)
1067 flags |= ENQUEUE_MIGRATE;
1069 for_each_sched_entity(se) {
1072 cfs_rq = cfs_rq_of(se);
1073 enqueue_entity(cfs_rq, se, flags);
1074 flags = ENQUEUE_WAKEUP;
1081 * The dequeue_task method is called before nr_running is
1082 * decreased. We remove the task from the rbtree and
1083 * update the fair scheduling stats:
1085 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
1087 struct cfs_rq *cfs_rq;
1088 struct sched_entity *se = &p->se;
1090 for_each_sched_entity(se) {
1091 cfs_rq = cfs_rq_of(se);
1092 dequeue_entity(cfs_rq, se, sleep);
1093 /* Don't dequeue parent if it has other entities besides us */
1094 if (cfs_rq->load.weight)
1103 * sched_yield() support is very simple - we dequeue and enqueue.
1105 * If compat_yield is turned on then we requeue to the end of the tree.
1107 static void yield_task_fair(struct rq *rq)
1109 struct task_struct *curr = rq->curr;
1110 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1111 struct sched_entity *rightmost, *se = &curr->se;
1114 * Are we the only task in the tree?
1116 if (unlikely(cfs_rq->nr_running == 1))
1119 clear_buddies(cfs_rq, se);
1121 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1122 update_rq_clock(rq);
1124 * Update run-time statistics of the 'current'.
1126 update_curr(cfs_rq);
1131 * Find the rightmost entry in the rbtree:
1133 rightmost = __pick_last_entity(cfs_rq);
1135 * Already in the rightmost position?
1137 if (unlikely(!rightmost || entity_before(rightmost, se)))
1141 * Minimally necessary key value to be last in the tree:
1142 * Upon rescheduling, sched_class::put_prev_task() will place
1143 * 'current' within the tree based on its new key value.
1145 se->vruntime = rightmost->vruntime + 1;
1150 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1152 struct sched_entity *se = &p->se;
1153 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1155 se->vruntime -= cfs_rq->min_vruntime;
1158 #ifdef CONFIG_FAIR_GROUP_SCHED
1160 * effective_load() calculates the load change as seen from the root_task_group
1162 * Adding load to a group doesn't make a group heavier, but can cause movement
1163 * of group shares between cpus. Assuming the shares were perfectly aligned one
1164 * can calculate the shift in shares.
1166 * The problem is that perfectly aligning the shares is rather expensive, hence
1167 * we try to avoid doing that too often - see update_shares(), which ratelimits
1170 * We compensate this by not only taking the current delta into account, but
1171 * also considering the delta between when the shares were last adjusted and
1174 * We still saw a performance dip, some tracing learned us that between
1175 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1176 * significantly. Therefore try to bias the error in direction of failing
1177 * the affine wakeup.
1180 static long effective_load(struct task_group *tg, int cpu,
1183 struct sched_entity *se = tg->se[cpu];
1189 * By not taking the decrease of shares on the other cpu into
1190 * account our error leans towards reducing the affine wakeups.
1192 if (!wl && sched_feat(ASYM_EFF_LOAD))
1195 for_each_sched_entity(se) {
1196 long S, rw, s, a, b;
1200 * Instead of using this increment, also add the difference
1201 * between when the shares were last updated and now.
1203 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1207 S = se->my_q->tg->shares;
1208 s = se->my_q->shares;
1209 rw = se->my_q->rq_weight;
1220 * Assume the group is already running and will
1221 * thus already be accounted for in the weight.
1223 * That is, moving shares between CPUs, does not
1224 * alter the group weight.
1234 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1235 unsigned long wl, unsigned long wg)
1242 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1244 unsigned long this_load, load;
1245 int idx, this_cpu, prev_cpu;
1246 unsigned long tl_per_task;
1247 unsigned int imbalance;
1248 struct task_group *tg;
1249 unsigned long weight;
1253 this_cpu = smp_processor_id();
1254 prev_cpu = task_cpu(p);
1255 load = source_load(prev_cpu, idx);
1256 this_load = target_load(this_cpu, idx);
1259 * If sync wakeup then subtract the (maximum possible)
1260 * effect of the currently running task from the load
1261 * of the current CPU:
1264 tg = task_group(current);
1265 weight = current->se.load.weight;
1267 this_load += effective_load(tg, this_cpu, -weight, -weight);
1268 load += effective_load(tg, prev_cpu, 0, -weight);
1272 weight = p->se.load.weight;
1274 imbalance = 100 + (sd->imbalance_pct - 100) / 2;
1277 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1278 * due to the sync cause above having dropped this_load to 0, we'll
1279 * always have an imbalance, but there's really nothing you can do
1280 * about that, so that's good too.
1282 * Otherwise check if either cpus are near enough in load to allow this
1283 * task to be woken on this_cpu.
1285 balanced = !this_load ||
1286 100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
1287 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1290 * If the currently running task will sleep within
1291 * a reasonable amount of time then attract this newly
1294 if (sync && balanced)
1297 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1298 tl_per_task = cpu_avg_load_per_task(this_cpu);
1301 (this_load <= load &&
1302 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1304 * This domain has SD_WAKE_AFFINE and
1305 * p is cache cold in this domain, and
1306 * there is no bad imbalance.
1308 schedstat_inc(sd, ttwu_move_affine);
1309 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1317 * find_idlest_group finds and returns the least busy CPU group within the
1320 static struct sched_group *
1321 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1322 int this_cpu, int load_idx)
1324 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1325 unsigned long min_load = ULONG_MAX, this_load = 0;
1326 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1329 unsigned long load, avg_load;
1333 /* Skip over this group if it has no CPUs allowed */
1334 if (!cpumask_intersects(sched_group_cpus(group),
1338 local_group = cpumask_test_cpu(this_cpu,
1339 sched_group_cpus(group));
1341 /* Tally up the load of all CPUs in the group */
1344 for_each_cpu(i, sched_group_cpus(group)) {
1345 /* Bias balancing toward cpus of our domain */
1347 load = source_load(i, load_idx);
1349 load = target_load(i, load_idx);
1354 /* Adjust by relative CPU power of the group */
1355 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1358 this_load = avg_load;
1360 } else if (avg_load < min_load) {
1361 min_load = avg_load;
1364 } while (group = group->next, group != sd->groups);
1366 if (!idlest || 100*this_load < imbalance*min_load)
1372 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1375 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1377 unsigned long load, min_load = ULONG_MAX;
1381 /* Traverse only the allowed CPUs */
1382 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1383 load = weighted_cpuload(i);
1385 if (load < min_load || (load == min_load && i == this_cpu)) {
1395 * Try and locate an idle CPU in the sched_domain.
1398 select_idle_sibling(struct task_struct *p, struct sched_domain *sd, int target)
1400 int cpu = smp_processor_id();
1401 int prev_cpu = task_cpu(p);
1405 * If this domain spans both cpu and prev_cpu (see the SD_WAKE_AFFINE
1406 * test in select_task_rq_fair) and the prev_cpu is idle then that's
1407 * always a better target than the current cpu.
1409 if (target == cpu && !cpu_rq(prev_cpu)->cfs.nr_running)
1413 * Otherwise, iterate the domain and find an elegible idle cpu.
1415 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1416 if (!cpu_rq(i)->cfs.nr_running) {
1426 * sched_balance_self: balance the current task (running on cpu) in domains
1427 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1430 * Balance, ie. select the least loaded group.
1432 * Returns the target CPU number, or the same CPU if no balancing is needed.
1434 * preempt must be disabled.
1436 static int select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1438 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1439 int cpu = smp_processor_id();
1440 int prev_cpu = task_cpu(p);
1442 int want_affine = 0;
1444 int sync = wake_flags & WF_SYNC;
1446 if (sd_flag & SD_BALANCE_WAKE) {
1447 if (sched_feat(AFFINE_WAKEUPS) &&
1448 cpumask_test_cpu(cpu, &p->cpus_allowed))
1453 for_each_domain(cpu, tmp) {
1454 if (!(tmp->flags & SD_LOAD_BALANCE))
1458 * If power savings logic is enabled for a domain, see if we
1459 * are not overloaded, if so, don't balance wider.
1461 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1462 unsigned long power = 0;
1463 unsigned long nr_running = 0;
1464 unsigned long capacity;
1467 for_each_cpu(i, sched_domain_span(tmp)) {
1468 power += power_of(i);
1469 nr_running += cpu_rq(i)->cfs.nr_running;
1472 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1474 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1477 if (nr_running < capacity)
1482 * While iterating the domains looking for a spanning
1483 * WAKE_AFFINE domain, adjust the affine target to any idle cpu
1484 * in cache sharing domains along the way.
1490 * If both cpu and prev_cpu are part of this domain,
1491 * cpu is a valid SD_WAKE_AFFINE target.
1493 if (cpumask_test_cpu(prev_cpu, sched_domain_span(tmp)))
1497 * If there's an idle sibling in this domain, make that
1498 * the wake_affine target instead of the current cpu.
1500 if (tmp->flags & SD_SHARE_PKG_RESOURCES)
1501 target = select_idle_sibling(p, tmp, target);
1504 if (tmp->flags & SD_WAKE_AFFINE) {
1512 if (!want_sd && !want_affine)
1515 if (!(tmp->flags & sd_flag))
1522 if (sched_feat(LB_SHARES_UPDATE)) {
1524 * Pick the largest domain to update shares over
1527 if (affine_sd && (!tmp ||
1528 cpumask_weight(sched_domain_span(affine_sd)) >
1529 cpumask_weight(sched_domain_span(sd))))
1536 if (affine_sd && wake_affine(affine_sd, p, sync))
1540 int load_idx = sd->forkexec_idx;
1541 struct sched_group *group;
1544 if (!(sd->flags & sd_flag)) {
1549 if (sd_flag & SD_BALANCE_WAKE)
1550 load_idx = sd->wake_idx;
1552 group = find_idlest_group(sd, p, cpu, load_idx);
1558 new_cpu = find_idlest_cpu(group, p, cpu);
1559 if (new_cpu == -1 || new_cpu == cpu) {
1560 /* Now try balancing at a lower domain level of cpu */
1565 /* Now try balancing at a lower domain level of new_cpu */
1567 weight = cpumask_weight(sched_domain_span(sd));
1569 for_each_domain(cpu, tmp) {
1570 if (weight <= cpumask_weight(sched_domain_span(tmp)))
1572 if (tmp->flags & sd_flag)
1575 /* while loop will break here if sd == NULL */
1580 #endif /* CONFIG_SMP */
1582 static unsigned long
1583 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1585 unsigned long gran = sysctl_sched_wakeup_granularity;
1588 * Since its curr running now, convert the gran from real-time
1589 * to virtual-time in his units.
1591 if (sched_feat(ASYM_GRAN)) {
1593 * By using 'se' instead of 'curr' we penalize light tasks, so
1594 * they get preempted easier. That is, if 'se' < 'curr' then
1595 * the resulting gran will be larger, therefore penalizing the
1596 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1597 * be smaller, again penalizing the lighter task.
1599 * This is especially important for buddies when the leftmost
1600 * task is higher priority than the buddy.
1602 if (unlikely(se->load.weight != NICE_0_LOAD))
1603 gran = calc_delta_fair(gran, se);
1605 if (unlikely(curr->load.weight != NICE_0_LOAD))
1606 gran = calc_delta_fair(gran, curr);
1613 * Should 'se' preempt 'curr'.
1627 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1629 s64 gran, vdiff = curr->vruntime - se->vruntime;
1634 gran = wakeup_gran(curr, se);
1641 static void set_last_buddy(struct sched_entity *se)
1643 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1644 for_each_sched_entity(se)
1645 cfs_rq_of(se)->last = se;
1649 static void set_next_buddy(struct sched_entity *se)
1651 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1652 for_each_sched_entity(se)
1653 cfs_rq_of(se)->next = se;
1658 * Preempt the current task with a newly woken task if needed:
1660 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1662 struct task_struct *curr = rq->curr;
1663 struct sched_entity *se = &curr->se, *pse = &p->se;
1664 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1665 int sync = wake_flags & WF_SYNC;
1666 int scale = cfs_rq->nr_running >= sched_nr_latency;
1668 if (unlikely(rt_prio(p->prio)))
1671 if (unlikely(p->sched_class != &fair_sched_class))
1674 if (unlikely(se == pse))
1677 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1678 set_next_buddy(pse);
1681 * We can come here with TIF_NEED_RESCHED already set from new task
1684 if (test_tsk_need_resched(curr))
1688 * Batch and idle tasks do not preempt (their preemption is driven by
1691 if (unlikely(p->policy != SCHED_NORMAL))
1694 /* Idle tasks are by definition preempted by everybody. */
1695 if (unlikely(curr->policy == SCHED_IDLE))
1698 if (sched_feat(WAKEUP_SYNC) && sync)
1701 if (!sched_feat(WAKEUP_PREEMPT))
1704 update_curr(cfs_rq);
1705 find_matching_se(&se, &pse);
1707 if (wakeup_preempt_entity(se, pse) == 1)
1715 * Only set the backward buddy when the current task is still
1716 * on the rq. This can happen when a wakeup gets interleaved
1717 * with schedule on the ->pre_schedule() or idle_balance()
1718 * point, either of which can * drop the rq lock.
1720 * Also, during early boot the idle thread is in the fair class,
1721 * for obvious reasons its a bad idea to schedule back to it.
1723 if (unlikely(!se->on_rq || curr == rq->idle))
1726 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1730 static struct task_struct *pick_next_task_fair(struct rq *rq)
1732 struct task_struct *p;
1733 struct cfs_rq *cfs_rq = &rq->cfs;
1734 struct sched_entity *se;
1736 if (!cfs_rq->nr_running)
1740 se = pick_next_entity(cfs_rq);
1741 set_next_entity(cfs_rq, se);
1742 cfs_rq = group_cfs_rq(se);
1746 hrtick_start_fair(rq, p);
1752 * Account for a descheduled task:
1754 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1756 struct sched_entity *se = &prev->se;
1757 struct cfs_rq *cfs_rq;
1759 for_each_sched_entity(se) {
1760 cfs_rq = cfs_rq_of(se);
1761 put_prev_entity(cfs_rq, se);
1766 /**************************************************
1767 * Fair scheduling class load-balancing methods:
1771 * pull_task - move a task from a remote runqueue to the local runqueue.
1772 * Both runqueues must be locked.
1774 static void pull_task(struct rq *src_rq, struct task_struct *p,
1775 struct rq *this_rq, int this_cpu)
1777 deactivate_task(src_rq, p, 0);
1778 set_task_cpu(p, this_cpu);
1779 activate_task(this_rq, p, 0);
1780 check_preempt_curr(this_rq, p, 0);
1784 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1787 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1788 struct sched_domain *sd, enum cpu_idle_type idle,
1791 int tsk_cache_hot = 0;
1793 * We do not migrate tasks that are:
1794 * 1) running (obviously), or
1795 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1796 * 3) are cache-hot on their current CPU.
1798 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1799 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1804 if (task_running(rq, p)) {
1805 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1810 * Aggressive migration if:
1811 * 1) task is cache cold, or
1812 * 2) too many balance attempts have failed.
1815 tsk_cache_hot = task_hot(p, rq->clock, sd);
1816 if (!tsk_cache_hot ||
1817 sd->nr_balance_failed > sd->cache_nice_tries) {
1818 #ifdef CONFIG_SCHEDSTATS
1819 if (tsk_cache_hot) {
1820 schedstat_inc(sd, lb_hot_gained[idle]);
1821 schedstat_inc(p, se.statistics.nr_forced_migrations);
1827 if (tsk_cache_hot) {
1828 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
1835 * move_one_task tries to move exactly one task from busiest to this_rq, as
1836 * part of active balancing operations within "domain".
1837 * Returns 1 if successful and 0 otherwise.
1839 * Called with both runqueues locked.
1842 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1843 struct sched_domain *sd, enum cpu_idle_type idle)
1845 struct task_struct *p, *n;
1846 struct cfs_rq *cfs_rq;
1849 for_each_leaf_cfs_rq(busiest, cfs_rq) {
1850 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1852 if (!can_migrate_task(p, busiest, this_cpu,
1856 pull_task(busiest, p, this_rq, this_cpu);
1858 * Right now, this is only the second place pull_task()
1859 * is called, so we can safely collect pull_task()
1860 * stats here rather than inside pull_task().
1862 schedstat_inc(sd, lb_gained[idle]);
1870 static unsigned long
1871 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1872 unsigned long max_load_move, struct sched_domain *sd,
1873 enum cpu_idle_type idle, int *all_pinned,
1874 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
1876 int loops = 0, pulled = 0, pinned = 0;
1877 long rem_load_move = max_load_move;
1878 struct task_struct *p, *n;
1880 if (max_load_move == 0)
1885 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
1886 if (loops++ > sysctl_sched_nr_migrate)
1889 if ((p->se.load.weight >> 1) > rem_load_move ||
1890 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
1893 pull_task(busiest, p, this_rq, this_cpu);
1895 rem_load_move -= p->se.load.weight;
1897 #ifdef CONFIG_PREEMPT
1899 * NEWIDLE balancing is a source of latency, so preemptible
1900 * kernels will stop after the first task is pulled to minimize
1901 * the critical section.
1903 if (idle == CPU_NEWLY_IDLE)
1908 * We only want to steal up to the prescribed amount of
1911 if (rem_load_move <= 0)
1914 if (p->prio < *this_best_prio)
1915 *this_best_prio = p->prio;
1919 * Right now, this is one of only two places pull_task() is called,
1920 * so we can safely collect pull_task() stats here rather than
1921 * inside pull_task().
1923 schedstat_add(sd, lb_gained[idle], pulled);
1926 *all_pinned = pinned;
1928 return max_load_move - rem_load_move;
1931 #ifdef CONFIG_FAIR_GROUP_SCHED
1932 static unsigned long
1933 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1934 unsigned long max_load_move,
1935 struct sched_domain *sd, enum cpu_idle_type idle,
1936 int *all_pinned, int *this_best_prio)
1938 long rem_load_move = max_load_move;
1939 int busiest_cpu = cpu_of(busiest);
1940 struct task_group *tg;
1943 update_h_load(busiest_cpu);
1945 list_for_each_entry_rcu(tg, &task_groups, list) {
1946 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1947 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1948 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1949 u64 rem_load, moved_load;
1954 if (!busiest_cfs_rq->task_weight)
1957 rem_load = (u64)rem_load_move * busiest_weight;
1958 rem_load = div_u64(rem_load, busiest_h_load + 1);
1960 moved_load = balance_tasks(this_rq, this_cpu, busiest,
1961 rem_load, sd, idle, all_pinned, this_best_prio,
1967 moved_load *= busiest_h_load;
1968 moved_load = div_u64(moved_load, busiest_weight + 1);
1970 rem_load_move -= moved_load;
1971 if (rem_load_move < 0)
1976 return max_load_move - rem_load_move;
1979 static unsigned long
1980 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1981 unsigned long max_load_move,
1982 struct sched_domain *sd, enum cpu_idle_type idle,
1983 int *all_pinned, int *this_best_prio)
1985 return balance_tasks(this_rq, this_cpu, busiest,
1986 max_load_move, sd, idle, all_pinned,
1987 this_best_prio, &busiest->cfs);
1992 * move_tasks tries to move up to max_load_move weighted load from busiest to
1993 * this_rq, as part of a balancing operation within domain "sd".
1994 * Returns 1 if successful and 0 otherwise.
1996 * Called with both runqueues locked.
1998 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1999 unsigned long max_load_move,
2000 struct sched_domain *sd, enum cpu_idle_type idle,
2003 unsigned long total_load_moved = 0, load_moved;
2004 int this_best_prio = this_rq->curr->prio;
2007 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2008 max_load_move - total_load_moved,
2009 sd, idle, all_pinned, &this_best_prio);
2011 total_load_moved += load_moved;
2013 #ifdef CONFIG_PREEMPT
2015 * NEWIDLE balancing is a source of latency, so preemptible
2016 * kernels will stop after the first task is pulled to minimize
2017 * the critical section.
2019 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2022 if (raw_spin_is_contended(&this_rq->lock) ||
2023 raw_spin_is_contended(&busiest->lock))
2026 } while (load_moved && max_load_move > total_load_moved);
2028 return total_load_moved > 0;
2031 /********** Helpers for find_busiest_group ************************/
2033 * sd_lb_stats - Structure to store the statistics of a sched_domain
2034 * during load balancing.
2036 struct sd_lb_stats {
2037 struct sched_group *busiest; /* Busiest group in this sd */
2038 struct sched_group *this; /* Local group in this sd */
2039 unsigned long total_load; /* Total load of all groups in sd */
2040 unsigned long total_pwr; /* Total power of all groups in sd */
2041 unsigned long avg_load; /* Average load across all groups in sd */
2043 /** Statistics of this group */
2044 unsigned long this_load;
2045 unsigned long this_load_per_task;
2046 unsigned long this_nr_running;
2048 /* Statistics of the busiest group */
2049 unsigned long max_load;
2050 unsigned long busiest_load_per_task;
2051 unsigned long busiest_nr_running;
2052 unsigned long busiest_group_capacity;
2054 int group_imb; /* Is there imbalance in this sd */
2055 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2056 int power_savings_balance; /* Is powersave balance needed for this sd */
2057 struct sched_group *group_min; /* Least loaded group in sd */
2058 struct sched_group *group_leader; /* Group which relieves group_min */
2059 unsigned long min_load_per_task; /* load_per_task in group_min */
2060 unsigned long leader_nr_running; /* Nr running of group_leader */
2061 unsigned long min_nr_running; /* Nr running of group_min */
2066 * sg_lb_stats - stats of a sched_group required for load_balancing
2068 struct sg_lb_stats {
2069 unsigned long avg_load; /*Avg load across the CPUs of the group */
2070 unsigned long group_load; /* Total load over the CPUs of the group */
2071 unsigned long sum_nr_running; /* Nr tasks running in the group */
2072 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2073 unsigned long group_capacity;
2074 int group_imb; /* Is there an imbalance in the group ? */
2078 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2079 * @group: The group whose first cpu is to be returned.
2081 static inline unsigned int group_first_cpu(struct sched_group *group)
2083 return cpumask_first(sched_group_cpus(group));
2087 * get_sd_load_idx - Obtain the load index for a given sched domain.
2088 * @sd: The sched_domain whose load_idx is to be obtained.
2089 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2091 static inline int get_sd_load_idx(struct sched_domain *sd,
2092 enum cpu_idle_type idle)
2098 load_idx = sd->busy_idx;
2101 case CPU_NEWLY_IDLE:
2102 load_idx = sd->newidle_idx;
2105 load_idx = sd->idle_idx;
2113 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2115 * init_sd_power_savings_stats - Initialize power savings statistics for
2116 * the given sched_domain, during load balancing.
2118 * @sd: Sched domain whose power-savings statistics are to be initialized.
2119 * @sds: Variable containing the statistics for sd.
2120 * @idle: Idle status of the CPU at which we're performing load-balancing.
2122 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2123 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2126 * Busy processors will not participate in power savings
2129 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2130 sds->power_savings_balance = 0;
2132 sds->power_savings_balance = 1;
2133 sds->min_nr_running = ULONG_MAX;
2134 sds->leader_nr_running = 0;
2139 * update_sd_power_savings_stats - Update the power saving stats for a
2140 * sched_domain while performing load balancing.
2142 * @group: sched_group belonging to the sched_domain under consideration.
2143 * @sds: Variable containing the statistics of the sched_domain
2144 * @local_group: Does group contain the CPU for which we're performing
2146 * @sgs: Variable containing the statistics of the group.
2148 static inline void update_sd_power_savings_stats(struct sched_group *group,
2149 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2152 if (!sds->power_savings_balance)
2156 * If the local group is idle or completely loaded
2157 * no need to do power savings balance at this domain
2159 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2160 !sds->this_nr_running))
2161 sds->power_savings_balance = 0;
2164 * If a group is already running at full capacity or idle,
2165 * don't include that group in power savings calculations
2167 if (!sds->power_savings_balance ||
2168 sgs->sum_nr_running >= sgs->group_capacity ||
2169 !sgs->sum_nr_running)
2173 * Calculate the group which has the least non-idle load.
2174 * This is the group from where we need to pick up the load
2177 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2178 (sgs->sum_nr_running == sds->min_nr_running &&
2179 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2180 sds->group_min = group;
2181 sds->min_nr_running = sgs->sum_nr_running;
2182 sds->min_load_per_task = sgs->sum_weighted_load /
2183 sgs->sum_nr_running;
2187 * Calculate the group which is almost near its
2188 * capacity but still has some space to pick up some load
2189 * from other group and save more power
2191 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2194 if (sgs->sum_nr_running > sds->leader_nr_running ||
2195 (sgs->sum_nr_running == sds->leader_nr_running &&
2196 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2197 sds->group_leader = group;
2198 sds->leader_nr_running = sgs->sum_nr_running;
2203 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2204 * @sds: Variable containing the statistics of the sched_domain
2205 * under consideration.
2206 * @this_cpu: Cpu at which we're currently performing load-balancing.
2207 * @imbalance: Variable to store the imbalance.
2210 * Check if we have potential to perform some power-savings balance.
2211 * If yes, set the busiest group to be the least loaded group in the
2212 * sched_domain, so that it's CPUs can be put to idle.
2214 * Returns 1 if there is potential to perform power-savings balance.
2217 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2218 int this_cpu, unsigned long *imbalance)
2220 if (!sds->power_savings_balance)
2223 if (sds->this != sds->group_leader ||
2224 sds->group_leader == sds->group_min)
2227 *imbalance = sds->min_load_per_task;
2228 sds->busiest = sds->group_min;
2233 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2234 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2235 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2240 static inline void update_sd_power_savings_stats(struct sched_group *group,
2241 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2246 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2247 int this_cpu, unsigned long *imbalance)
2251 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2254 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2256 return SCHED_LOAD_SCALE;
2259 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2261 return default_scale_freq_power(sd, cpu);
2264 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2266 unsigned long weight = cpumask_weight(sched_domain_span(sd));
2267 unsigned long smt_gain = sd->smt_gain;
2274 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2276 return default_scale_smt_power(sd, cpu);
2279 unsigned long scale_rt_power(int cpu)
2281 struct rq *rq = cpu_rq(cpu);
2282 u64 total, available;
2284 sched_avg_update(rq);
2286 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2287 available = total - rq->rt_avg;
2289 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2290 total = SCHED_LOAD_SCALE;
2292 total >>= SCHED_LOAD_SHIFT;
2294 return div_u64(available, total);
2297 static void update_cpu_power(struct sched_domain *sd, int cpu)
2299 unsigned long weight = cpumask_weight(sched_domain_span(sd));
2300 unsigned long power = SCHED_LOAD_SCALE;
2301 struct sched_group *sdg = sd->groups;
2303 if (sched_feat(ARCH_POWER))
2304 power *= arch_scale_freq_power(sd, cpu);
2306 power *= default_scale_freq_power(sd, cpu);
2308 power >>= SCHED_LOAD_SHIFT;
2310 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2311 if (sched_feat(ARCH_POWER))
2312 power *= arch_scale_smt_power(sd, cpu);
2314 power *= default_scale_smt_power(sd, cpu);
2316 power >>= SCHED_LOAD_SHIFT;
2319 power *= scale_rt_power(cpu);
2320 power >>= SCHED_LOAD_SHIFT;
2325 sdg->cpu_power = power;
2328 static void update_group_power(struct sched_domain *sd, int cpu)
2330 struct sched_domain *child = sd->child;
2331 struct sched_group *group, *sdg = sd->groups;
2332 unsigned long power;
2335 update_cpu_power(sd, cpu);
2341 group = child->groups;
2343 power += group->cpu_power;
2344 group = group->next;
2345 } while (group != child->groups);
2347 sdg->cpu_power = power;
2351 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2352 * @sd: The sched_domain whose statistics are to be updated.
2353 * @group: sched_group whose statistics are to be updated.
2354 * @this_cpu: Cpu for which load balance is currently performed.
2355 * @idle: Idle status of this_cpu
2356 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2357 * @sd_idle: Idle status of the sched_domain containing group.
2358 * @local_group: Does group contain this_cpu.
2359 * @cpus: Set of cpus considered for load balancing.
2360 * @balance: Should we balance.
2361 * @sgs: variable to hold the statistics for this group.
2363 static inline void update_sg_lb_stats(struct sched_domain *sd,
2364 struct sched_group *group, int this_cpu,
2365 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2366 int local_group, const struct cpumask *cpus,
2367 int *balance, struct sg_lb_stats *sgs)
2369 unsigned long load, max_cpu_load, min_cpu_load;
2371 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2372 unsigned long avg_load_per_task = 0;
2375 balance_cpu = group_first_cpu(group);
2377 /* Tally up the load of all CPUs in the group */
2379 min_cpu_load = ~0UL;
2381 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2382 struct rq *rq = cpu_rq(i);
2384 if (*sd_idle && rq->nr_running)
2387 /* Bias balancing toward cpus of our domain */
2389 if (idle_cpu(i) && !first_idle_cpu) {
2394 load = target_load(i, load_idx);
2396 load = source_load(i, load_idx);
2397 if (load > max_cpu_load)
2398 max_cpu_load = load;
2399 if (min_cpu_load > load)
2400 min_cpu_load = load;
2403 sgs->group_load += load;
2404 sgs->sum_nr_running += rq->nr_running;
2405 sgs->sum_weighted_load += weighted_cpuload(i);
2410 * First idle cpu or the first cpu(busiest) in this sched group
2411 * is eligible for doing load balancing at this and above
2412 * domains. In the newly idle case, we will allow all the cpu's
2413 * to do the newly idle load balance.
2415 if (idle != CPU_NEWLY_IDLE && local_group &&
2416 balance_cpu != this_cpu) {
2421 update_group_power(sd, this_cpu);
2423 /* Adjust by relative CPU power of the group */
2424 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2427 * Consider the group unbalanced when the imbalance is larger
2428 * than the average weight of two tasks.
2430 * APZ: with cgroup the avg task weight can vary wildly and
2431 * might not be a suitable number - should we keep a
2432 * normalized nr_running number somewhere that negates
2435 if (sgs->sum_nr_running)
2436 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2438 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
2441 sgs->group_capacity =
2442 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2446 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2447 * @sd: sched_domain whose statistics are to be updated.
2448 * @this_cpu: Cpu for which load balance is currently performed.
2449 * @idle: Idle status of this_cpu
2450 * @sd_idle: Idle status of the sched_domain containing group.
2451 * @cpus: Set of cpus considered for load balancing.
2452 * @balance: Should we balance.
2453 * @sds: variable to hold the statistics for this sched_domain.
2455 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2456 enum cpu_idle_type idle, int *sd_idle,
2457 const struct cpumask *cpus, int *balance,
2458 struct sd_lb_stats *sds)
2460 struct sched_domain *child = sd->child;
2461 struct sched_group *group = sd->groups;
2462 struct sg_lb_stats sgs;
2463 int load_idx, prefer_sibling = 0;
2465 if (child && child->flags & SD_PREFER_SIBLING)
2468 init_sd_power_savings_stats(sd, sds, idle);
2469 load_idx = get_sd_load_idx(sd, idle);
2474 local_group = cpumask_test_cpu(this_cpu,
2475 sched_group_cpus(group));
2476 memset(&sgs, 0, sizeof(sgs));
2477 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
2478 local_group, cpus, balance, &sgs);
2480 if (local_group && !(*balance))
2483 sds->total_load += sgs.group_load;
2484 sds->total_pwr += group->cpu_power;
2487 * In case the child domain prefers tasks go to siblings
2488 * first, lower the group capacity to one so that we'll try
2489 * and move all the excess tasks away.
2492 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2495 sds->this_load = sgs.avg_load;
2497 sds->this_nr_running = sgs.sum_nr_running;
2498 sds->this_load_per_task = sgs.sum_weighted_load;
2499 } else if (sgs.avg_load > sds->max_load &&
2500 (sgs.sum_nr_running > sgs.group_capacity ||
2502 sds->max_load = sgs.avg_load;
2503 sds->busiest = group;
2504 sds->busiest_nr_running = sgs.sum_nr_running;
2505 sds->busiest_group_capacity = sgs.group_capacity;
2506 sds->busiest_load_per_task = sgs.sum_weighted_load;
2507 sds->group_imb = sgs.group_imb;
2510 update_sd_power_savings_stats(group, sds, local_group, &sgs);
2511 group = group->next;
2512 } while (group != sd->groups);
2516 * fix_small_imbalance - Calculate the minor imbalance that exists
2517 * amongst the groups of a sched_domain, during
2519 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2520 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2521 * @imbalance: Variable to store the imbalance.
2523 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2524 int this_cpu, unsigned long *imbalance)
2526 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2527 unsigned int imbn = 2;
2528 unsigned long scaled_busy_load_per_task;
2530 if (sds->this_nr_running) {
2531 sds->this_load_per_task /= sds->this_nr_running;
2532 if (sds->busiest_load_per_task >
2533 sds->this_load_per_task)
2536 sds->this_load_per_task =
2537 cpu_avg_load_per_task(this_cpu);
2539 scaled_busy_load_per_task = sds->busiest_load_per_task
2541 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2543 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2544 (scaled_busy_load_per_task * imbn)) {
2545 *imbalance = sds->busiest_load_per_task;
2550 * OK, we don't have enough imbalance to justify moving tasks,
2551 * however we may be able to increase total CPU power used by
2555 pwr_now += sds->busiest->cpu_power *
2556 min(sds->busiest_load_per_task, sds->max_load);
2557 pwr_now += sds->this->cpu_power *
2558 min(sds->this_load_per_task, sds->this_load);
2559 pwr_now /= SCHED_LOAD_SCALE;
2561 /* Amount of load we'd subtract */
2562 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2563 sds->busiest->cpu_power;
2564 if (sds->max_load > tmp)
2565 pwr_move += sds->busiest->cpu_power *
2566 min(sds->busiest_load_per_task, sds->max_load - tmp);
2568 /* Amount of load we'd add */
2569 if (sds->max_load * sds->busiest->cpu_power <
2570 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2571 tmp = (sds->max_load * sds->busiest->cpu_power) /
2572 sds->this->cpu_power;
2574 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2575 sds->this->cpu_power;
2576 pwr_move += sds->this->cpu_power *
2577 min(sds->this_load_per_task, sds->this_load + tmp);
2578 pwr_move /= SCHED_LOAD_SCALE;
2580 /* Move if we gain throughput */
2581 if (pwr_move > pwr_now)
2582 *imbalance = sds->busiest_load_per_task;
2586 * calculate_imbalance - Calculate the amount of imbalance present within the
2587 * groups of a given sched_domain during load balance.
2588 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2589 * @this_cpu: Cpu for which currently load balance is being performed.
2590 * @imbalance: The variable to store the imbalance.
2592 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2593 unsigned long *imbalance)
2595 unsigned long max_pull, load_above_capacity = ~0UL;
2597 sds->busiest_load_per_task /= sds->busiest_nr_running;
2598 if (sds->group_imb) {
2599 sds->busiest_load_per_task =
2600 min(sds->busiest_load_per_task, sds->avg_load);
2604 * In the presence of smp nice balancing, certain scenarios can have
2605 * max load less than avg load(as we skip the groups at or below
2606 * its cpu_power, while calculating max_load..)
2608 if (sds->max_load < sds->avg_load) {
2610 return fix_small_imbalance(sds, this_cpu, imbalance);
2613 if (!sds->group_imb) {
2615 * Don't want to pull so many tasks that a group would go idle.
2617 load_above_capacity = (sds->busiest_nr_running -
2618 sds->busiest_group_capacity);
2620 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2622 load_above_capacity /= sds->busiest->cpu_power;
2626 * We're trying to get all the cpus to the average_load, so we don't
2627 * want to push ourselves above the average load, nor do we wish to
2628 * reduce the max loaded cpu below the average load. At the same time,
2629 * we also don't want to reduce the group load below the group capacity
2630 * (so that we can implement power-savings policies etc). Thus we look
2631 * for the minimum possible imbalance.
2632 * Be careful of negative numbers as they'll appear as very large values
2633 * with unsigned longs.
2635 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2637 /* How much load to actually move to equalise the imbalance */
2638 *imbalance = min(max_pull * sds->busiest->cpu_power,
2639 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
2643 * if *imbalance is less than the average load per runnable task
2644 * there is no gaurantee that any tasks will be moved so we'll have
2645 * a think about bumping its value to force at least one task to be
2648 if (*imbalance < sds->busiest_load_per_task)
2649 return fix_small_imbalance(sds, this_cpu, imbalance);
2652 /******* find_busiest_group() helpers end here *********************/
2655 * find_busiest_group - Returns the busiest group within the sched_domain
2656 * if there is an imbalance. If there isn't an imbalance, and
2657 * the user has opted for power-savings, it returns a group whose
2658 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2659 * such a group exists.
2661 * Also calculates the amount of weighted load which should be moved
2662 * to restore balance.
2664 * @sd: The sched_domain whose busiest group is to be returned.
2665 * @this_cpu: The cpu for which load balancing is currently being performed.
2666 * @imbalance: Variable which stores amount of weighted load which should
2667 * be moved to restore balance/put a group to idle.
2668 * @idle: The idle status of this_cpu.
2669 * @sd_idle: The idleness of sd
2670 * @cpus: The set of CPUs under consideration for load-balancing.
2671 * @balance: Pointer to a variable indicating if this_cpu
2672 * is the appropriate cpu to perform load balancing at this_level.
2674 * Returns: - the busiest group if imbalance exists.
2675 * - If no imbalance and user has opted for power-savings balance,
2676 * return the least loaded group whose CPUs can be
2677 * put to idle by rebalancing its tasks onto our group.
2679 static struct sched_group *
2680 find_busiest_group(struct sched_domain *sd, int this_cpu,
2681 unsigned long *imbalance, enum cpu_idle_type idle,
2682 int *sd_idle, const struct cpumask *cpus, int *balance)
2684 struct sd_lb_stats sds;
2686 memset(&sds, 0, sizeof(sds));
2689 * Compute the various statistics relavent for load balancing at
2692 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
2695 /* Cases where imbalance does not exist from POV of this_cpu */
2696 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2698 * 2) There is no busy sibling group to pull from.
2699 * 3) This group is the busiest group.
2700 * 4) This group is more busy than the avg busieness at this
2702 * 5) The imbalance is within the specified limit.
2707 if (!sds.busiest || sds.busiest_nr_running == 0)
2710 if (sds.this_load >= sds.max_load)
2713 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
2715 if (sds.this_load >= sds.avg_load)
2718 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
2721 /* Looks like there is an imbalance. Compute it */
2722 calculate_imbalance(&sds, this_cpu, imbalance);
2727 * There is no obvious imbalance. But check if we can do some balancing
2730 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
2738 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2741 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
2742 unsigned long imbalance, const struct cpumask *cpus)
2744 struct rq *busiest = NULL, *rq;
2745 unsigned long max_load = 0;
2748 for_each_cpu(i, sched_group_cpus(group)) {
2749 unsigned long power = power_of(i);
2750 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
2753 if (!cpumask_test_cpu(i, cpus))
2757 wl = weighted_cpuload(i);
2760 * When comparing with imbalance, use weighted_cpuload()
2761 * which is not scaled with the cpu power.
2763 if (capacity && rq->nr_running == 1 && wl > imbalance)
2767 * For the load comparisons with the other cpu's, consider
2768 * the weighted_cpuload() scaled with the cpu power, so that
2769 * the load can be moved away from the cpu that is potentially
2770 * running at a lower capacity.
2772 wl = (wl * SCHED_LOAD_SCALE) / power;
2774 if (wl > max_load) {
2784 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2785 * so long as it is large enough.
2787 #define MAX_PINNED_INTERVAL 512
2789 /* Working cpumask for load_balance and load_balance_newidle. */
2790 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
2792 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle)
2794 if (idle == CPU_NEWLY_IDLE) {
2796 * The only task running in a non-idle cpu can be moved to this
2797 * cpu in an attempt to completely freeup the other CPU
2800 * The package power saving logic comes from
2801 * find_busiest_group(). If there are no imbalance, then
2802 * f_b_g() will return NULL. However when sched_mc={1,2} then
2803 * f_b_g() will select a group from which a running task may be
2804 * pulled to this cpu in order to make the other package idle.
2805 * If there is no opportunity to make a package idle and if
2806 * there are no imbalance, then f_b_g() will return NULL and no
2807 * action will be taken in load_balance_newidle().
2809 * Under normal task pull operation due to imbalance, there
2810 * will be more than one task in the source run queue and
2811 * move_tasks() will succeed. ld_moved will be true and this
2812 * active balance code will not be triggered.
2814 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2815 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2818 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
2822 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
2826 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2827 * tasks if there is an imbalance.
2829 static int load_balance(int this_cpu, struct rq *this_rq,
2830 struct sched_domain *sd, enum cpu_idle_type idle,
2833 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2834 struct sched_group *group;
2835 unsigned long imbalance;
2837 unsigned long flags;
2838 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
2840 cpumask_copy(cpus, cpu_active_mask);
2843 * When power savings policy is enabled for the parent domain, idle
2844 * sibling can pick up load irrespective of busy siblings. In this case,
2845 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2846 * portraying it as CPU_NOT_IDLE.
2848 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2849 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2852 schedstat_inc(sd, lb_count[idle]);
2856 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2863 schedstat_inc(sd, lb_nobusyg[idle]);
2867 busiest = find_busiest_queue(group, idle, imbalance, cpus);
2869 schedstat_inc(sd, lb_nobusyq[idle]);
2873 BUG_ON(busiest == this_rq);
2875 schedstat_add(sd, lb_imbalance[idle], imbalance);
2878 if (busiest->nr_running > 1) {
2880 * Attempt to move tasks. If find_busiest_group has found
2881 * an imbalance but busiest->nr_running <= 1, the group is
2882 * still unbalanced. ld_moved simply stays zero, so it is
2883 * correctly treated as an imbalance.
2885 local_irq_save(flags);
2886 double_rq_lock(this_rq, busiest);
2887 ld_moved = move_tasks(this_rq, this_cpu, busiest,
2888 imbalance, sd, idle, &all_pinned);
2889 double_rq_unlock(this_rq, busiest);
2890 local_irq_restore(flags);
2893 * some other cpu did the load balance for us.
2895 if (ld_moved && this_cpu != smp_processor_id())
2896 resched_cpu(this_cpu);
2898 /* All tasks on this runqueue were pinned by CPU affinity */
2899 if (unlikely(all_pinned)) {
2900 cpumask_clear_cpu(cpu_of(busiest), cpus);
2901 if (!cpumask_empty(cpus))
2908 schedstat_inc(sd, lb_failed[idle]);
2909 sd->nr_balance_failed++;
2911 if (need_active_balance(sd, sd_idle, idle)) {
2912 raw_spin_lock_irqsave(&busiest->lock, flags);
2914 /* don't kick the migration_thread, if the curr
2915 * task on busiest cpu can't be moved to this_cpu
2917 if (!cpumask_test_cpu(this_cpu,
2918 &busiest->curr->cpus_allowed)) {
2919 raw_spin_unlock_irqrestore(&busiest->lock,
2922 goto out_one_pinned;
2925 if (!busiest->active_balance) {
2926 busiest->active_balance = 1;
2927 busiest->push_cpu = this_cpu;
2930 raw_spin_unlock_irqrestore(&busiest->lock, flags);
2932 wake_up_process(busiest->migration_thread);
2935 * We've kicked active balancing, reset the failure
2938 sd->nr_balance_failed = sd->cache_nice_tries+1;
2941 sd->nr_balance_failed = 0;
2943 if (likely(!active_balance)) {
2944 /* We were unbalanced, so reset the balancing interval */
2945 sd->balance_interval = sd->min_interval;
2948 * If we've begun active balancing, start to back off. This
2949 * case may not be covered by the all_pinned logic if there
2950 * is only 1 task on the busy runqueue (because we don't call
2953 if (sd->balance_interval < sd->max_interval)
2954 sd->balance_interval *= 2;
2957 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2958 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2964 schedstat_inc(sd, lb_balanced[idle]);
2966 sd->nr_balance_failed = 0;
2969 /* tune up the balancing interval */
2970 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
2971 (sd->balance_interval < sd->max_interval))
2972 sd->balance_interval *= 2;
2974 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2975 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2986 * idle_balance is called by schedule() if this_cpu is about to become
2987 * idle. Attempts to pull tasks from other CPUs.
2989 static void idle_balance(int this_cpu, struct rq *this_rq)
2991 struct sched_domain *sd;
2992 int pulled_task = 0;
2993 unsigned long next_balance = jiffies + HZ;
2995 this_rq->idle_stamp = this_rq->clock;
2997 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3001 * Drop the rq->lock, but keep IRQ/preempt disabled.
3003 raw_spin_unlock(&this_rq->lock);
3005 for_each_domain(this_cpu, sd) {
3006 unsigned long interval;
3009 if (!(sd->flags & SD_LOAD_BALANCE))
3012 if (sd->flags & SD_BALANCE_NEWIDLE) {
3013 /* If we've pulled tasks over stop searching: */
3014 pulled_task = load_balance(this_cpu, this_rq,
3015 sd, CPU_NEWLY_IDLE, &balance);
3018 interval = msecs_to_jiffies(sd->balance_interval);
3019 if (time_after(next_balance, sd->last_balance + interval))
3020 next_balance = sd->last_balance + interval;
3022 this_rq->idle_stamp = 0;
3027 raw_spin_lock(&this_rq->lock);
3029 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3031 * We are going idle. next_balance may be set based on
3032 * a busy processor. So reset next_balance.
3034 this_rq->next_balance = next_balance;
3039 * active_load_balance is run by migration threads. It pushes running tasks
3040 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3041 * running on each physical CPU where possible, and avoids physical /
3042 * logical imbalances.
3044 * Called with busiest_rq locked.
3046 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3048 int target_cpu = busiest_rq->push_cpu;
3049 struct sched_domain *sd;
3050 struct rq *target_rq;
3052 /* Is there any task to move? */
3053 if (busiest_rq->nr_running <= 1)
3056 target_rq = cpu_rq(target_cpu);
3059 * This condition is "impossible", if it occurs
3060 * we need to fix it. Originally reported by
3061 * Bjorn Helgaas on a 128-cpu setup.
3063 BUG_ON(busiest_rq == target_rq);
3065 /* move a task from busiest_rq to target_rq */
3066 double_lock_balance(busiest_rq, target_rq);
3067 update_rq_clock(busiest_rq);
3068 update_rq_clock(target_rq);
3070 /* Search for an sd spanning us and the target CPU. */
3071 for_each_domain(target_cpu, sd) {
3072 if ((sd->flags & SD_LOAD_BALANCE) &&
3073 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3078 schedstat_inc(sd, alb_count);
3080 if (move_one_task(target_rq, target_cpu, busiest_rq,
3082 schedstat_inc(sd, alb_pushed);
3084 schedstat_inc(sd, alb_failed);
3086 double_unlock_balance(busiest_rq, target_rq);
3091 atomic_t load_balancer;
3092 cpumask_var_t cpu_mask;
3093 cpumask_var_t ilb_grp_nohz_mask;
3094 } nohz ____cacheline_aligned = {
3095 .load_balancer = ATOMIC_INIT(-1),
3098 int get_nohz_load_balancer(void)
3100 return atomic_read(&nohz.load_balancer);
3103 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3105 * lowest_flag_domain - Return lowest sched_domain containing flag.
3106 * @cpu: The cpu whose lowest level of sched domain is to
3108 * @flag: The flag to check for the lowest sched_domain
3109 * for the given cpu.
3111 * Returns the lowest sched_domain of a cpu which contains the given flag.
3113 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3115 struct sched_domain *sd;
3117 for_each_domain(cpu, sd)
3118 if (sd && (sd->flags & flag))
3125 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3126 * @cpu: The cpu whose domains we're iterating over.
3127 * @sd: variable holding the value of the power_savings_sd
3129 * @flag: The flag to filter the sched_domains to be iterated.
3131 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3132 * set, starting from the lowest sched_domain to the highest.
3134 #define for_each_flag_domain(cpu, sd, flag) \
3135 for (sd = lowest_flag_domain(cpu, flag); \
3136 (sd && (sd->flags & flag)); sd = sd->parent)
3139 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3140 * @ilb_group: group to be checked for semi-idleness
3142 * Returns: 1 if the group is semi-idle. 0 otherwise.
3144 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3145 * and atleast one non-idle CPU. This helper function checks if the given
3146 * sched_group is semi-idle or not.
3148 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3150 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
3151 sched_group_cpus(ilb_group));
3154 * A sched_group is semi-idle when it has atleast one busy cpu
3155 * and atleast one idle cpu.
3157 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
3160 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
3166 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3167 * @cpu: The cpu which is nominating a new idle_load_balancer.
3169 * Returns: Returns the id of the idle load balancer if it exists,
3170 * Else, returns >= nr_cpu_ids.
3172 * This algorithm picks the idle load balancer such that it belongs to a
3173 * semi-idle powersavings sched_domain. The idea is to try and avoid
3174 * completely idle packages/cores just for the purpose of idle load balancing
3175 * when there are other idle cpu's which are better suited for that job.
3177 static int find_new_ilb(int cpu)
3179 struct sched_domain *sd;
3180 struct sched_group *ilb_group;
3183 * Have idle load balancer selection from semi-idle packages only
3184 * when power-aware load balancing is enabled
3186 if (!(sched_smt_power_savings || sched_mc_power_savings))
3190 * Optimize for the case when we have no idle CPUs or only one
3191 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3193 if (cpumask_weight(nohz.cpu_mask) < 2)
3196 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3197 ilb_group = sd->groups;
3200 if (is_semi_idle_group(ilb_group))
3201 return cpumask_first(nohz.ilb_grp_nohz_mask);
3203 ilb_group = ilb_group->next;
3205 } while (ilb_group != sd->groups);
3209 return cpumask_first(nohz.cpu_mask);
3211 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3212 static inline int find_new_ilb(int call_cpu)
3214 return cpumask_first(nohz.cpu_mask);
3219 * This routine will try to nominate the ilb (idle load balancing)
3220 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3221 * load balancing on behalf of all those cpus. If all the cpus in the system
3222 * go into this tickless mode, then there will be no ilb owner (as there is
3223 * no need for one) and all the cpus will sleep till the next wakeup event
3226 * For the ilb owner, tick is not stopped. And this tick will be used
3227 * for idle load balancing. ilb owner will still be part of
3230 * While stopping the tick, this cpu will become the ilb owner if there
3231 * is no other owner. And will be the owner till that cpu becomes busy
3232 * or if all cpus in the system stop their ticks at which point
3233 * there is no need for ilb owner.
3235 * When the ilb owner becomes busy, it nominates another owner, during the
3236 * next busy scheduler_tick()
3238 int select_nohz_load_balancer(int stop_tick)
3240 int cpu = smp_processor_id();
3243 cpu_rq(cpu)->in_nohz_recently = 1;
3245 if (!cpu_active(cpu)) {
3246 if (atomic_read(&nohz.load_balancer) != cpu)
3250 * If we are going offline and still the leader,
3253 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3259 cpumask_set_cpu(cpu, nohz.cpu_mask);
3261 /* time for ilb owner also to sleep */
3262 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
3263 if (atomic_read(&nohz.load_balancer) == cpu)
3264 atomic_set(&nohz.load_balancer, -1);
3268 if (atomic_read(&nohz.load_balancer) == -1) {
3269 /* make me the ilb owner */
3270 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3272 } else if (atomic_read(&nohz.load_balancer) == cpu) {
3275 if (!(sched_smt_power_savings ||
3276 sched_mc_power_savings))
3279 * Check to see if there is a more power-efficient
3282 new_ilb = find_new_ilb(cpu);
3283 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3284 atomic_set(&nohz.load_balancer, -1);
3285 resched_cpu(new_ilb);
3291 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3294 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3296 if (atomic_read(&nohz.load_balancer) == cpu)
3297 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3304 static DEFINE_SPINLOCK(balancing);
3307 * It checks each scheduling domain to see if it is due to be balanced,
3308 * and initiates a balancing operation if so.
3310 * Balancing parameters are set up in arch_init_sched_domains.
3312 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3315 struct rq *rq = cpu_rq(cpu);
3316 unsigned long interval;
3317 struct sched_domain *sd;
3318 /* Earliest time when we have to do rebalance again */
3319 unsigned long next_balance = jiffies + 60*HZ;
3320 int update_next_balance = 0;
3323 for_each_domain(cpu, sd) {
3324 if (!(sd->flags & SD_LOAD_BALANCE))
3327 interval = sd->balance_interval;
3328 if (idle != CPU_IDLE)
3329 interval *= sd->busy_factor;
3331 /* scale ms to jiffies */
3332 interval = msecs_to_jiffies(interval);
3333 if (unlikely(!interval))
3335 if (interval > HZ*NR_CPUS/10)
3336 interval = HZ*NR_CPUS/10;
3338 need_serialize = sd->flags & SD_SERIALIZE;
3340 if (need_serialize) {
3341 if (!spin_trylock(&balancing))
3345 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3346 if (load_balance(cpu, rq, sd, idle, &balance)) {
3348 * We've pulled tasks over so either we're no
3349 * longer idle, or one of our SMT siblings is
3352 idle = CPU_NOT_IDLE;
3354 sd->last_balance = jiffies;
3357 spin_unlock(&balancing);
3359 if (time_after(next_balance, sd->last_balance + interval)) {
3360 next_balance = sd->last_balance + interval;
3361 update_next_balance = 1;
3365 * Stop the load balance at this level. There is another
3366 * CPU in our sched group which is doing load balancing more
3374 * next_balance will be updated only when there is a need.
3375 * When the cpu is attached to null domain for ex, it will not be
3378 if (likely(update_next_balance))
3379 rq->next_balance = next_balance;
3383 * run_rebalance_domains is triggered when needed from the scheduler tick.
3384 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3385 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3387 static void run_rebalance_domains(struct softirq_action *h)
3389 int this_cpu = smp_processor_id();
3390 struct rq *this_rq = cpu_rq(this_cpu);
3391 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3392 CPU_IDLE : CPU_NOT_IDLE;
3394 rebalance_domains(this_cpu, idle);
3398 * If this cpu is the owner for idle load balancing, then do the
3399 * balancing on behalf of the other idle cpus whose ticks are
3402 if (this_rq->idle_at_tick &&
3403 atomic_read(&nohz.load_balancer) == this_cpu) {
3407 for_each_cpu(balance_cpu, nohz.cpu_mask) {
3408 if (balance_cpu == this_cpu)
3412 * If this cpu gets work to do, stop the load balancing
3413 * work being done for other cpus. Next load
3414 * balancing owner will pick it up.
3419 rebalance_domains(balance_cpu, CPU_IDLE);
3421 rq = cpu_rq(balance_cpu);
3422 if (time_after(this_rq->next_balance, rq->next_balance))
3423 this_rq->next_balance = rq->next_balance;
3429 static inline int on_null_domain(int cpu)
3431 return !rcu_dereference(cpu_rq(cpu)->sd);
3435 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3437 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3438 * idle load balancing owner or decide to stop the periodic load balancing,
3439 * if the whole system is idle.
3441 static inline void trigger_load_balance(struct rq *rq, int cpu)
3445 * If we were in the nohz mode recently and busy at the current
3446 * scheduler tick, then check if we need to nominate new idle
3449 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3450 rq->in_nohz_recently = 0;
3452 if (atomic_read(&nohz.load_balancer) == cpu) {
3453 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3454 atomic_set(&nohz.load_balancer, -1);
3457 if (atomic_read(&nohz.load_balancer) == -1) {
3458 int ilb = find_new_ilb(cpu);
3460 if (ilb < nr_cpu_ids)
3466 * If this cpu is idle and doing idle load balancing for all the
3467 * cpus with ticks stopped, is it time for that to stop?
3469 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3470 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3476 * If this cpu is idle and the idle load balancing is done by
3477 * someone else, then no need raise the SCHED_SOFTIRQ
3479 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3480 cpumask_test_cpu(cpu, nohz.cpu_mask))
3483 /* Don't need to rebalance while attached to NULL domain */
3484 if (time_after_eq(jiffies, rq->next_balance) &&
3485 likely(!on_null_domain(cpu)))
3486 raise_softirq(SCHED_SOFTIRQ);
3489 static void rq_online_fair(struct rq *rq)
3494 static void rq_offline_fair(struct rq *rq)
3499 #else /* CONFIG_SMP */
3502 * on UP we do not need to balance between CPUs:
3504 static inline void idle_balance(int cpu, struct rq *rq)
3508 #endif /* CONFIG_SMP */
3511 * scheduler tick hitting a task of our scheduling class:
3513 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
3515 struct cfs_rq *cfs_rq;
3516 struct sched_entity *se = &curr->se;
3518 for_each_sched_entity(se) {
3519 cfs_rq = cfs_rq_of(se);
3520 entity_tick(cfs_rq, se, queued);
3525 * called on fork with the child task as argument from the parent's context
3526 * - child not yet on the tasklist
3527 * - preemption disabled
3529 static void task_fork_fair(struct task_struct *p)
3531 struct cfs_rq *cfs_rq = task_cfs_rq(current);
3532 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
3533 int this_cpu = smp_processor_id();
3534 struct rq *rq = this_rq();
3535 unsigned long flags;
3537 raw_spin_lock_irqsave(&rq->lock, flags);
3539 if (unlikely(task_cpu(p) != this_cpu))
3540 __set_task_cpu(p, this_cpu);
3542 update_curr(cfs_rq);
3545 se->vruntime = curr->vruntime;
3546 place_entity(cfs_rq, se, 1);
3548 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
3550 * Upon rescheduling, sched_class::put_prev_task() will place
3551 * 'current' within the tree based on its new key value.
3553 swap(curr->vruntime, se->vruntime);
3554 resched_task(rq->curr);
3557 se->vruntime -= cfs_rq->min_vruntime;
3559 raw_spin_unlock_irqrestore(&rq->lock, flags);
3563 * Priority of the task has changed. Check to see if we preempt
3566 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
3567 int oldprio, int running)
3570 * Reschedule if we are currently running on this runqueue and
3571 * our priority decreased, or if we are not currently running on
3572 * this runqueue and our priority is higher than the current's
3575 if (p->prio > oldprio)
3576 resched_task(rq->curr);
3578 check_preempt_curr(rq, p, 0);
3582 * We switched to the sched_fair class.
3584 static void switched_to_fair(struct rq *rq, struct task_struct *p,
3588 * We were most likely switched from sched_rt, so
3589 * kick off the schedule if running, otherwise just see
3590 * if we can still preempt the current task.
3593 resched_task(rq->curr);
3595 check_preempt_curr(rq, p, 0);
3598 /* Account for a task changing its policy or group.
3600 * This routine is mostly called to set cfs_rq->curr field when a task
3601 * migrates between groups/classes.
3603 static void set_curr_task_fair(struct rq *rq)
3605 struct sched_entity *se = &rq->curr->se;
3607 for_each_sched_entity(se)
3608 set_next_entity(cfs_rq_of(se), se);
3611 #ifdef CONFIG_FAIR_GROUP_SCHED
3612 static void moved_group_fair(struct task_struct *p, int on_rq)
3614 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3616 update_curr(cfs_rq);
3618 place_entity(cfs_rq, &p->se, 1);
3622 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
3624 struct sched_entity *se = &task->se;
3625 unsigned int rr_interval = 0;
3628 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
3631 if (rq->cfs.load.weight)
3632 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
3638 * All the scheduling class methods:
3640 static const struct sched_class fair_sched_class = {
3641 .next = &idle_sched_class,
3642 .enqueue_task = enqueue_task_fair,
3643 .dequeue_task = dequeue_task_fair,
3644 .yield_task = yield_task_fair,
3646 .check_preempt_curr = check_preempt_wakeup,
3648 .pick_next_task = pick_next_task_fair,
3649 .put_prev_task = put_prev_task_fair,
3652 .select_task_rq = select_task_rq_fair,
3654 .rq_online = rq_online_fair,
3655 .rq_offline = rq_offline_fair,
3657 .task_waking = task_waking_fair,
3660 .set_curr_task = set_curr_task_fair,
3661 .task_tick = task_tick_fair,
3662 .task_fork = task_fork_fair,
3664 .prio_changed = prio_changed_fair,
3665 .switched_to = switched_to_fair,
3667 .get_rr_interval = get_rr_interval_fair,
3669 #ifdef CONFIG_FAIR_GROUP_SCHED
3670 .moved_group = moved_group_fair,
3674 #ifdef CONFIG_SCHED_DEBUG
3675 static void print_cfs_stats(struct seq_file *m, int cpu)
3677 struct cfs_rq *cfs_rq;
3680 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
3681 print_cfs_rq(m, cpu, cfs_rq);