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 = 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
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: 2 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 2000000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 2000000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 3;
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. */
742 unsigned long thresh = sysctl_sched_latency;
745 * Halve their sleep time's effect, to allow
746 * for a gentler effect of sleepers:
748 if (sched_feat(GENTLE_FAIR_SLEEPERS))
754 /* ensure we never gain time by being placed backwards. */
755 vruntime = max_vruntime(se->vruntime, vruntime);
757 se->vruntime = vruntime;
761 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
764 * Update the normalized vruntime before updating min_vruntime
765 * through callig update_curr().
767 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
768 se->vruntime += cfs_rq->min_vruntime;
771 * Update run-time statistics of the 'current'.
774 account_entity_enqueue(cfs_rq, se);
776 if (flags & ENQUEUE_WAKEUP) {
777 place_entity(cfs_rq, se, 0);
778 enqueue_sleeper(cfs_rq, se);
781 update_stats_enqueue(cfs_rq, se);
782 check_spread(cfs_rq, se);
783 if (se != cfs_rq->curr)
784 __enqueue_entity(cfs_rq, se);
787 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
789 if (!se || cfs_rq->last == se)
792 if (!se || cfs_rq->next == se)
796 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 for_each_sched_entity(se)
799 __clear_buddies(cfs_rq_of(se), se);
803 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
806 * Update run-time statistics of the 'current'.
810 update_stats_dequeue(cfs_rq, se);
811 if (flags & DEQUEUE_SLEEP) {
812 #ifdef CONFIG_SCHEDSTATS
813 if (entity_is_task(se)) {
814 struct task_struct *tsk = task_of(se);
816 if (tsk->state & TASK_INTERRUPTIBLE)
817 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
818 if (tsk->state & TASK_UNINTERRUPTIBLE)
819 se->statistics.block_start = rq_of(cfs_rq)->clock;
824 clear_buddies(cfs_rq, se);
826 if (se != cfs_rq->curr)
827 __dequeue_entity(cfs_rq, se);
828 account_entity_dequeue(cfs_rq, se);
829 update_min_vruntime(cfs_rq);
832 * Normalize the entity after updating the min_vruntime because the
833 * update can refer to the ->curr item and we need to reflect this
834 * movement in our normalized position.
836 if (!(flags & DEQUEUE_SLEEP))
837 se->vruntime -= cfs_rq->min_vruntime;
841 * Preempt the current task with a newly woken task if needed:
844 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
846 unsigned long ideal_runtime, delta_exec;
848 ideal_runtime = sched_slice(cfs_rq, curr);
849 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
850 if (delta_exec > ideal_runtime) {
851 resched_task(rq_of(cfs_rq)->curr);
853 * The current task ran long enough, ensure it doesn't get
854 * re-elected due to buddy favours.
856 clear_buddies(cfs_rq, curr);
861 * Ensure that a task that missed wakeup preemption by a
862 * narrow margin doesn't have to wait for a full slice.
863 * This also mitigates buddy induced latencies under load.
865 if (!sched_feat(WAKEUP_PREEMPT))
868 if (delta_exec < sysctl_sched_min_granularity)
871 if (cfs_rq->nr_running > 1) {
872 struct sched_entity *se = __pick_next_entity(cfs_rq);
873 s64 delta = curr->vruntime - se->vruntime;
875 if (delta > ideal_runtime)
876 resched_task(rq_of(cfs_rq)->curr);
881 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
883 /* 'current' is not kept within the tree. */
886 * Any task has to be enqueued before it get to execute on
887 * a CPU. So account for the time it spent waiting on the
890 update_stats_wait_end(cfs_rq, se);
891 __dequeue_entity(cfs_rq, se);
894 update_stats_curr_start(cfs_rq, se);
896 #ifdef CONFIG_SCHEDSTATS
898 * Track our maximum slice length, if the CPU's load is at
899 * least twice that of our own weight (i.e. dont track it
900 * when there are only lesser-weight tasks around):
902 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
903 se->statistics.slice_max = max(se->statistics.slice_max,
904 se->sum_exec_runtime - se->prev_sum_exec_runtime);
907 se->prev_sum_exec_runtime = se->sum_exec_runtime;
911 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
913 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
915 struct sched_entity *se = __pick_next_entity(cfs_rq);
916 struct sched_entity *left = se;
918 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
922 * Prefer last buddy, try to return the CPU to a preempted task.
924 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
927 clear_buddies(cfs_rq, se);
932 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
935 * If still on the runqueue then deactivate_task()
936 * was not called and update_curr() has to be done:
941 check_spread(cfs_rq, prev);
943 update_stats_wait_start(cfs_rq, prev);
944 /* Put 'current' back into the tree. */
945 __enqueue_entity(cfs_rq, prev);
951 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
954 * Update run-time statistics of the 'current'.
958 #ifdef CONFIG_SCHED_HRTICK
960 * queued ticks are scheduled to match the slice, so don't bother
961 * validating it and just reschedule.
964 resched_task(rq_of(cfs_rq)->curr);
968 * don't let the period tick interfere with the hrtick preemption
970 if (!sched_feat(DOUBLE_TICK) &&
971 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
975 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
976 check_preempt_tick(cfs_rq, curr);
979 /**************************************************
980 * CFS operations on tasks:
983 #ifdef CONFIG_SCHED_HRTICK
984 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
986 struct sched_entity *se = &p->se;
987 struct cfs_rq *cfs_rq = cfs_rq_of(se);
989 WARN_ON(task_rq(p) != rq);
991 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
992 u64 slice = sched_slice(cfs_rq, se);
993 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
994 s64 delta = slice - ran;
1003 * Don't schedule slices shorter than 10000ns, that just
1004 * doesn't make sense. Rely on vruntime for fairness.
1007 delta = max_t(s64, 10000LL, delta);
1009 hrtick_start(rq, delta);
1014 * called from enqueue/dequeue and updates the hrtick when the
1015 * current task is from our class and nr_running is low enough
1018 static void hrtick_update(struct rq *rq)
1020 struct task_struct *curr = rq->curr;
1022 if (curr->sched_class != &fair_sched_class)
1025 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1026 hrtick_start_fair(rq, curr);
1028 #else /* !CONFIG_SCHED_HRTICK */
1030 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1034 static inline void hrtick_update(struct rq *rq)
1040 * The enqueue_task method is called before nr_running is
1041 * increased. Here we update the fair scheduling stats and
1042 * then put the task into the rbtree:
1045 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1047 struct cfs_rq *cfs_rq;
1048 struct sched_entity *se = &p->se;
1050 for_each_sched_entity(se) {
1053 cfs_rq = cfs_rq_of(se);
1054 enqueue_entity(cfs_rq, se, flags);
1055 flags = ENQUEUE_WAKEUP;
1062 * The dequeue_task method is called before nr_running is
1063 * decreased. We remove the task from the rbtree and
1064 * update the fair scheduling stats:
1066 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1068 struct cfs_rq *cfs_rq;
1069 struct sched_entity *se = &p->se;
1071 for_each_sched_entity(se) {
1072 cfs_rq = cfs_rq_of(se);
1073 dequeue_entity(cfs_rq, se, flags);
1074 /* Don't dequeue parent if it has other entities besides us */
1075 if (cfs_rq->load.weight)
1077 flags |= DEQUEUE_SLEEP;
1084 * sched_yield() support is very simple - we dequeue and enqueue.
1086 * If compat_yield is turned on then we requeue to the end of the tree.
1088 static void yield_task_fair(struct rq *rq)
1090 struct task_struct *curr = rq->curr;
1091 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1092 struct sched_entity *rightmost, *se = &curr->se;
1095 * Are we the only task in the tree?
1097 if (unlikely(cfs_rq->nr_running == 1))
1100 clear_buddies(cfs_rq, se);
1102 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1103 update_rq_clock(rq);
1105 * Update run-time statistics of the 'current'.
1107 update_curr(cfs_rq);
1112 * Find the rightmost entry in the rbtree:
1114 rightmost = __pick_last_entity(cfs_rq);
1116 * Already in the rightmost position?
1118 if (unlikely(!rightmost || entity_before(rightmost, se)))
1122 * Minimally necessary key value to be last in the tree:
1123 * Upon rescheduling, sched_class::put_prev_task() will place
1124 * 'current' within the tree based on its new key value.
1126 se->vruntime = rightmost->vruntime + 1;
1131 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1133 struct sched_entity *se = &p->se;
1134 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1136 se->vruntime -= cfs_rq->min_vruntime;
1139 #ifdef CONFIG_FAIR_GROUP_SCHED
1141 * effective_load() calculates the load change as seen from the root_task_group
1143 * Adding load to a group doesn't make a group heavier, but can cause movement
1144 * of group shares between cpus. Assuming the shares were perfectly aligned one
1145 * can calculate the shift in shares.
1147 * The problem is that perfectly aligning the shares is rather expensive, hence
1148 * we try to avoid doing that too often - see update_shares(), which ratelimits
1151 * We compensate this by not only taking the current delta into account, but
1152 * also considering the delta between when the shares were last adjusted and
1155 * We still saw a performance dip, some tracing learned us that between
1156 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1157 * significantly. Therefore try to bias the error in direction of failing
1158 * the affine wakeup.
1161 static long effective_load(struct task_group *tg, int cpu,
1164 struct sched_entity *se = tg->se[cpu];
1170 * By not taking the decrease of shares on the other cpu into
1171 * account our error leans towards reducing the affine wakeups.
1173 if (!wl && sched_feat(ASYM_EFF_LOAD))
1176 for_each_sched_entity(se) {
1177 long S, rw, s, a, b;
1181 * Instead of using this increment, also add the difference
1182 * between when the shares were last updated and now.
1184 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1188 S = se->my_q->tg->shares;
1189 s = se->my_q->shares;
1190 rw = se->my_q->rq_weight;
1201 * Assume the group is already running and will
1202 * thus already be accounted for in the weight.
1204 * That is, moving shares between CPUs, does not
1205 * alter the group weight.
1215 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1216 unsigned long wl, unsigned long wg)
1223 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1225 unsigned long this_load, load;
1226 int idx, this_cpu, prev_cpu;
1227 unsigned long tl_per_task;
1228 struct task_group *tg;
1229 unsigned long weight;
1233 this_cpu = smp_processor_id();
1234 prev_cpu = task_cpu(p);
1235 load = source_load(prev_cpu, idx);
1236 this_load = target_load(this_cpu, idx);
1239 * If sync wakeup then subtract the (maximum possible)
1240 * effect of the currently running task from the load
1241 * of the current CPU:
1244 tg = task_group(current);
1245 weight = current->se.load.weight;
1247 this_load += effective_load(tg, this_cpu, -weight, -weight);
1248 load += effective_load(tg, prev_cpu, 0, -weight);
1252 weight = p->se.load.weight;
1255 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1256 * due to the sync cause above having dropped this_load to 0, we'll
1257 * always have an imbalance, but there's really nothing you can do
1258 * about that, so that's good too.
1260 * Otherwise check if either cpus are near enough in load to allow this
1261 * task to be woken on this_cpu.
1264 unsigned long this_eff_load, prev_eff_load;
1266 this_eff_load = 100;
1267 this_eff_load *= power_of(prev_cpu);
1268 this_eff_load *= this_load +
1269 effective_load(tg, this_cpu, weight, weight);
1271 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1272 prev_eff_load *= power_of(this_cpu);
1273 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1275 balanced = this_eff_load <= prev_eff_load;
1280 * If the currently running task will sleep within
1281 * a reasonable amount of time then attract this newly
1284 if (sync && balanced)
1287 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1288 tl_per_task = cpu_avg_load_per_task(this_cpu);
1291 (this_load <= load &&
1292 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1294 * This domain has SD_WAKE_AFFINE and
1295 * p is cache cold in this domain, and
1296 * there is no bad imbalance.
1298 schedstat_inc(sd, ttwu_move_affine);
1299 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1307 * find_idlest_group finds and returns the least busy CPU group within the
1310 static struct sched_group *
1311 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1312 int this_cpu, int load_idx)
1314 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1315 unsigned long min_load = ULONG_MAX, this_load = 0;
1316 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1319 unsigned long load, avg_load;
1323 /* Skip over this group if it has no CPUs allowed */
1324 if (!cpumask_intersects(sched_group_cpus(group),
1328 local_group = cpumask_test_cpu(this_cpu,
1329 sched_group_cpus(group));
1331 /* Tally up the load of all CPUs in the group */
1334 for_each_cpu(i, sched_group_cpus(group)) {
1335 /* Bias balancing toward cpus of our domain */
1337 load = source_load(i, load_idx);
1339 load = target_load(i, load_idx);
1344 /* Adjust by relative CPU power of the group */
1345 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1348 this_load = avg_load;
1350 } else if (avg_load < min_load) {
1351 min_load = avg_load;
1354 } while (group = group->next, group != sd->groups);
1356 if (!idlest || 100*this_load < imbalance*min_load)
1362 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1365 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1367 unsigned long load, min_load = ULONG_MAX;
1371 /* Traverse only the allowed CPUs */
1372 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1373 load = weighted_cpuload(i);
1375 if (load < min_load || (load == min_load && i == this_cpu)) {
1385 * Try and locate an idle CPU in the sched_domain.
1387 static int select_idle_sibling(struct task_struct *p, int target)
1389 int cpu = smp_processor_id();
1390 int prev_cpu = task_cpu(p);
1391 struct sched_domain *sd;
1395 * If the task is going to be woken-up on this cpu and if it is
1396 * already idle, then it is the right target.
1398 if (target == cpu && idle_cpu(cpu))
1402 * If the task is going to be woken-up on the cpu where it previously
1403 * ran and if it is currently idle, then it the right target.
1405 if (target == prev_cpu && idle_cpu(prev_cpu))
1409 * Otherwise, iterate the domains and find an elegible idle cpu.
1411 for_each_domain(target, sd) {
1412 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1415 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1423 * Lets stop looking for an idle sibling when we reached
1424 * the domain that spans the current cpu and prev_cpu.
1426 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1427 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1435 * sched_balance_self: balance the current task (running on cpu) in domains
1436 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1439 * Balance, ie. select the least loaded group.
1441 * Returns the target CPU number, or the same CPU if no balancing is needed.
1443 * preempt must be disabled.
1446 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1448 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1449 int cpu = smp_processor_id();
1450 int prev_cpu = task_cpu(p);
1452 int want_affine = 0;
1454 int sync = wake_flags & WF_SYNC;
1456 if (sd_flag & SD_BALANCE_WAKE) {
1457 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1462 for_each_domain(cpu, tmp) {
1463 if (!(tmp->flags & SD_LOAD_BALANCE))
1467 * If power savings logic is enabled for a domain, see if we
1468 * are not overloaded, if so, don't balance wider.
1470 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1471 unsigned long power = 0;
1472 unsigned long nr_running = 0;
1473 unsigned long capacity;
1476 for_each_cpu(i, sched_domain_span(tmp)) {
1477 power += power_of(i);
1478 nr_running += cpu_rq(i)->cfs.nr_running;
1481 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1483 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1486 if (nr_running < capacity)
1491 * If both cpu and prev_cpu are part of this domain,
1492 * cpu is a valid SD_WAKE_AFFINE target.
1494 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1495 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1500 if (!want_sd && !want_affine)
1503 if (!(tmp->flags & sd_flag))
1510 #ifdef CONFIG_FAIR_GROUP_SCHED
1511 if (sched_feat(LB_SHARES_UPDATE)) {
1513 * Pick the largest domain to update shares over
1516 if (affine_sd && (!tmp || affine_sd->span_weight > sd->span_weight))
1520 raw_spin_unlock(&rq->lock);
1522 raw_spin_lock(&rq->lock);
1528 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1529 return select_idle_sibling(p, cpu);
1531 return select_idle_sibling(p, prev_cpu);
1535 int load_idx = sd->forkexec_idx;
1536 struct sched_group *group;
1539 if (!(sd->flags & sd_flag)) {
1544 if (sd_flag & SD_BALANCE_WAKE)
1545 load_idx = sd->wake_idx;
1547 group = find_idlest_group(sd, p, cpu, load_idx);
1553 new_cpu = find_idlest_cpu(group, p, cpu);
1554 if (new_cpu == -1 || new_cpu == cpu) {
1555 /* Now try balancing at a lower domain level of cpu */
1560 /* Now try balancing at a lower domain level of new_cpu */
1562 weight = sd->span_weight;
1564 for_each_domain(cpu, tmp) {
1565 if (weight <= tmp->span_weight)
1567 if (tmp->flags & sd_flag)
1570 /* while loop will break here if sd == NULL */
1575 #endif /* CONFIG_SMP */
1577 static unsigned long
1578 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1580 unsigned long gran = sysctl_sched_wakeup_granularity;
1583 * Since its curr running now, convert the gran from real-time
1584 * to virtual-time in his units.
1586 * By using 'se' instead of 'curr' we penalize light tasks, so
1587 * they get preempted easier. That is, if 'se' < 'curr' then
1588 * the resulting gran will be larger, therefore penalizing the
1589 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1590 * be smaller, again penalizing the lighter task.
1592 * This is especially important for buddies when the leftmost
1593 * task is higher priority than the buddy.
1595 if (unlikely(se->load.weight != NICE_0_LOAD))
1596 gran = calc_delta_fair(gran, se);
1602 * Should 'se' preempt 'curr'.
1616 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1618 s64 gran, vdiff = curr->vruntime - se->vruntime;
1623 gran = wakeup_gran(curr, se);
1630 static void set_last_buddy(struct sched_entity *se)
1632 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1633 for_each_sched_entity(se)
1634 cfs_rq_of(se)->last = se;
1638 static void set_next_buddy(struct sched_entity *se)
1640 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1641 for_each_sched_entity(se)
1642 cfs_rq_of(se)->next = se;
1647 * Preempt the current task with a newly woken task if needed:
1649 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1651 struct task_struct *curr = rq->curr;
1652 struct sched_entity *se = &curr->se, *pse = &p->se;
1653 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1654 int scale = cfs_rq->nr_running >= sched_nr_latency;
1656 if (unlikely(rt_prio(p->prio)))
1659 if (unlikely(p->sched_class != &fair_sched_class))
1662 if (unlikely(se == pse))
1665 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1666 set_next_buddy(pse);
1669 * We can come here with TIF_NEED_RESCHED already set from new task
1672 if (test_tsk_need_resched(curr))
1676 * Batch and idle tasks do not preempt (their preemption is driven by
1679 if (unlikely(p->policy != SCHED_NORMAL))
1682 /* Idle tasks are by definition preempted by everybody. */
1683 if (unlikely(curr->policy == SCHED_IDLE))
1686 if (!sched_feat(WAKEUP_PREEMPT))
1689 update_curr(cfs_rq);
1690 find_matching_se(&se, &pse);
1692 if (wakeup_preempt_entity(se, pse) == 1)
1700 * Only set the backward buddy when the current task is still
1701 * on the rq. This can happen when a wakeup gets interleaved
1702 * with schedule on the ->pre_schedule() or idle_balance()
1703 * point, either of which can * drop the rq lock.
1705 * Also, during early boot the idle thread is in the fair class,
1706 * for obvious reasons its a bad idea to schedule back to it.
1708 if (unlikely(!se->on_rq || curr == rq->idle))
1711 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1715 static struct task_struct *pick_next_task_fair(struct rq *rq)
1717 struct task_struct *p;
1718 struct cfs_rq *cfs_rq = &rq->cfs;
1719 struct sched_entity *se;
1721 if (!cfs_rq->nr_running)
1725 se = pick_next_entity(cfs_rq);
1726 set_next_entity(cfs_rq, se);
1727 cfs_rq = group_cfs_rq(se);
1731 hrtick_start_fair(rq, p);
1737 * Account for a descheduled task:
1739 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1741 struct sched_entity *se = &prev->se;
1742 struct cfs_rq *cfs_rq;
1744 for_each_sched_entity(se) {
1745 cfs_rq = cfs_rq_of(se);
1746 put_prev_entity(cfs_rq, se);
1751 /**************************************************
1752 * Fair scheduling class load-balancing methods:
1756 * pull_task - move a task from a remote runqueue to the local runqueue.
1757 * Both runqueues must be locked.
1759 static void pull_task(struct rq *src_rq, struct task_struct *p,
1760 struct rq *this_rq, int this_cpu)
1762 deactivate_task(src_rq, p, 0);
1763 set_task_cpu(p, this_cpu);
1764 activate_task(this_rq, p, 0);
1765 check_preempt_curr(this_rq, p, 0);
1769 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1772 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1773 struct sched_domain *sd, enum cpu_idle_type idle,
1776 int tsk_cache_hot = 0;
1778 * We do not migrate tasks that are:
1779 * 1) running (obviously), or
1780 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1781 * 3) are cache-hot on their current CPU.
1783 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1784 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1789 if (task_running(rq, p)) {
1790 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1795 * Aggressive migration if:
1796 * 1) task is cache cold, or
1797 * 2) too many balance attempts have failed.
1800 tsk_cache_hot = task_hot(p, rq->clock, sd);
1801 if (!tsk_cache_hot ||
1802 sd->nr_balance_failed > sd->cache_nice_tries) {
1803 #ifdef CONFIG_SCHEDSTATS
1804 if (tsk_cache_hot) {
1805 schedstat_inc(sd, lb_hot_gained[idle]);
1806 schedstat_inc(p, se.statistics.nr_forced_migrations);
1812 if (tsk_cache_hot) {
1813 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
1820 * move_one_task tries to move exactly one task from busiest to this_rq, as
1821 * part of active balancing operations within "domain".
1822 * Returns 1 if successful and 0 otherwise.
1824 * Called with both runqueues locked.
1827 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1828 struct sched_domain *sd, enum cpu_idle_type idle)
1830 struct task_struct *p, *n;
1831 struct cfs_rq *cfs_rq;
1834 for_each_leaf_cfs_rq(busiest, cfs_rq) {
1835 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1837 if (!can_migrate_task(p, busiest, this_cpu,
1841 pull_task(busiest, p, this_rq, this_cpu);
1843 * Right now, this is only the second place pull_task()
1844 * is called, so we can safely collect pull_task()
1845 * stats here rather than inside pull_task().
1847 schedstat_inc(sd, lb_gained[idle]);
1855 static unsigned long
1856 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1857 unsigned long max_load_move, struct sched_domain *sd,
1858 enum cpu_idle_type idle, int *all_pinned,
1859 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
1861 int loops = 0, pulled = 0, pinned = 0;
1862 long rem_load_move = max_load_move;
1863 struct task_struct *p, *n;
1865 if (max_load_move == 0)
1870 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
1871 if (loops++ > sysctl_sched_nr_migrate)
1874 if ((p->se.load.weight >> 1) > rem_load_move ||
1875 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
1878 pull_task(busiest, p, this_rq, this_cpu);
1880 rem_load_move -= p->se.load.weight;
1882 #ifdef CONFIG_PREEMPT
1884 * NEWIDLE balancing is a source of latency, so preemptible
1885 * kernels will stop after the first task is pulled to minimize
1886 * the critical section.
1888 if (idle == CPU_NEWLY_IDLE)
1893 * We only want to steal up to the prescribed amount of
1896 if (rem_load_move <= 0)
1899 if (p->prio < *this_best_prio)
1900 *this_best_prio = p->prio;
1904 * Right now, this is one of only two places pull_task() is called,
1905 * so we can safely collect pull_task() stats here rather than
1906 * inside pull_task().
1908 schedstat_add(sd, lb_gained[idle], pulled);
1911 *all_pinned = pinned;
1913 return max_load_move - rem_load_move;
1916 #ifdef CONFIG_FAIR_GROUP_SCHED
1917 static unsigned long
1918 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1919 unsigned long max_load_move,
1920 struct sched_domain *sd, enum cpu_idle_type idle,
1921 int *all_pinned, int *this_best_prio)
1923 long rem_load_move = max_load_move;
1924 int busiest_cpu = cpu_of(busiest);
1925 struct task_group *tg;
1928 update_h_load(busiest_cpu);
1930 list_for_each_entry_rcu(tg, &task_groups, list) {
1931 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1932 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1933 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1934 u64 rem_load, moved_load;
1939 if (!busiest_cfs_rq->task_weight)
1942 rem_load = (u64)rem_load_move * busiest_weight;
1943 rem_load = div_u64(rem_load, busiest_h_load + 1);
1945 moved_load = balance_tasks(this_rq, this_cpu, busiest,
1946 rem_load, sd, idle, all_pinned, this_best_prio,
1952 moved_load *= busiest_h_load;
1953 moved_load = div_u64(moved_load, busiest_weight + 1);
1955 rem_load_move -= moved_load;
1956 if (rem_load_move < 0)
1961 return max_load_move - rem_load_move;
1964 static unsigned long
1965 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1966 unsigned long max_load_move,
1967 struct sched_domain *sd, enum cpu_idle_type idle,
1968 int *all_pinned, int *this_best_prio)
1970 return balance_tasks(this_rq, this_cpu, busiest,
1971 max_load_move, sd, idle, all_pinned,
1972 this_best_prio, &busiest->cfs);
1977 * move_tasks tries to move up to max_load_move weighted load from busiest to
1978 * this_rq, as part of a balancing operation within domain "sd".
1979 * Returns 1 if successful and 0 otherwise.
1981 * Called with both runqueues locked.
1983 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1984 unsigned long max_load_move,
1985 struct sched_domain *sd, enum cpu_idle_type idle,
1988 unsigned long total_load_moved = 0, load_moved;
1989 int this_best_prio = this_rq->curr->prio;
1992 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
1993 max_load_move - total_load_moved,
1994 sd, idle, all_pinned, &this_best_prio);
1996 total_load_moved += load_moved;
1998 #ifdef CONFIG_PREEMPT
2000 * NEWIDLE balancing is a source of latency, so preemptible
2001 * kernels will stop after the first task is pulled to minimize
2002 * the critical section.
2004 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2007 if (raw_spin_is_contended(&this_rq->lock) ||
2008 raw_spin_is_contended(&busiest->lock))
2011 } while (load_moved && max_load_move > total_load_moved);
2013 return total_load_moved > 0;
2016 /********** Helpers for find_busiest_group ************************/
2018 * sd_lb_stats - Structure to store the statistics of a sched_domain
2019 * during load balancing.
2021 struct sd_lb_stats {
2022 struct sched_group *busiest; /* Busiest group in this sd */
2023 struct sched_group *this; /* Local group in this sd */
2024 unsigned long total_load; /* Total load of all groups in sd */
2025 unsigned long total_pwr; /* Total power of all groups in sd */
2026 unsigned long avg_load; /* Average load across all groups in sd */
2028 /** Statistics of this group */
2029 unsigned long this_load;
2030 unsigned long this_load_per_task;
2031 unsigned long this_nr_running;
2033 /* Statistics of the busiest group */
2034 unsigned long max_load;
2035 unsigned long busiest_load_per_task;
2036 unsigned long busiest_nr_running;
2037 unsigned long busiest_group_capacity;
2039 int group_imb; /* Is there imbalance in this sd */
2040 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2041 int power_savings_balance; /* Is powersave balance needed for this sd */
2042 struct sched_group *group_min; /* Least loaded group in sd */
2043 struct sched_group *group_leader; /* Group which relieves group_min */
2044 unsigned long min_load_per_task; /* load_per_task in group_min */
2045 unsigned long leader_nr_running; /* Nr running of group_leader */
2046 unsigned long min_nr_running; /* Nr running of group_min */
2051 * sg_lb_stats - stats of a sched_group required for load_balancing
2053 struct sg_lb_stats {
2054 unsigned long avg_load; /*Avg load across the CPUs of the group */
2055 unsigned long group_load; /* Total load over the CPUs of the group */
2056 unsigned long sum_nr_running; /* Nr tasks running in the group */
2057 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2058 unsigned long group_capacity;
2059 int group_imb; /* Is there an imbalance in the group ? */
2063 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2064 * @group: The group whose first cpu is to be returned.
2066 static inline unsigned int group_first_cpu(struct sched_group *group)
2068 return cpumask_first(sched_group_cpus(group));
2072 * get_sd_load_idx - Obtain the load index for a given sched domain.
2073 * @sd: The sched_domain whose load_idx is to be obtained.
2074 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2076 static inline int get_sd_load_idx(struct sched_domain *sd,
2077 enum cpu_idle_type idle)
2083 load_idx = sd->busy_idx;
2086 case CPU_NEWLY_IDLE:
2087 load_idx = sd->newidle_idx;
2090 load_idx = sd->idle_idx;
2098 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2100 * init_sd_power_savings_stats - Initialize power savings statistics for
2101 * the given sched_domain, during load balancing.
2103 * @sd: Sched domain whose power-savings statistics are to be initialized.
2104 * @sds: Variable containing the statistics for sd.
2105 * @idle: Idle status of the CPU at which we're performing load-balancing.
2107 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2108 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2111 * Busy processors will not participate in power savings
2114 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2115 sds->power_savings_balance = 0;
2117 sds->power_savings_balance = 1;
2118 sds->min_nr_running = ULONG_MAX;
2119 sds->leader_nr_running = 0;
2124 * update_sd_power_savings_stats - Update the power saving stats for a
2125 * sched_domain while performing load balancing.
2127 * @group: sched_group belonging to the sched_domain under consideration.
2128 * @sds: Variable containing the statistics of the sched_domain
2129 * @local_group: Does group contain the CPU for which we're performing
2131 * @sgs: Variable containing the statistics of the group.
2133 static inline void update_sd_power_savings_stats(struct sched_group *group,
2134 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2137 if (!sds->power_savings_balance)
2141 * If the local group is idle or completely loaded
2142 * no need to do power savings balance at this domain
2144 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2145 !sds->this_nr_running))
2146 sds->power_savings_balance = 0;
2149 * If a group is already running at full capacity or idle,
2150 * don't include that group in power savings calculations
2152 if (!sds->power_savings_balance ||
2153 sgs->sum_nr_running >= sgs->group_capacity ||
2154 !sgs->sum_nr_running)
2158 * Calculate the group which has the least non-idle load.
2159 * This is the group from where we need to pick up the load
2162 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2163 (sgs->sum_nr_running == sds->min_nr_running &&
2164 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2165 sds->group_min = group;
2166 sds->min_nr_running = sgs->sum_nr_running;
2167 sds->min_load_per_task = sgs->sum_weighted_load /
2168 sgs->sum_nr_running;
2172 * Calculate the group which is almost near its
2173 * capacity but still has some space to pick up some load
2174 * from other group and save more power
2176 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2179 if (sgs->sum_nr_running > sds->leader_nr_running ||
2180 (sgs->sum_nr_running == sds->leader_nr_running &&
2181 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2182 sds->group_leader = group;
2183 sds->leader_nr_running = sgs->sum_nr_running;
2188 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2189 * @sds: Variable containing the statistics of the sched_domain
2190 * under consideration.
2191 * @this_cpu: Cpu at which we're currently performing load-balancing.
2192 * @imbalance: Variable to store the imbalance.
2195 * Check if we have potential to perform some power-savings balance.
2196 * If yes, set the busiest group to be the least loaded group in the
2197 * sched_domain, so that it's CPUs can be put to idle.
2199 * Returns 1 if there is potential to perform power-savings balance.
2202 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2203 int this_cpu, unsigned long *imbalance)
2205 if (!sds->power_savings_balance)
2208 if (sds->this != sds->group_leader ||
2209 sds->group_leader == sds->group_min)
2212 *imbalance = sds->min_load_per_task;
2213 sds->busiest = sds->group_min;
2218 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2219 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2220 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2225 static inline void update_sd_power_savings_stats(struct sched_group *group,
2226 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2231 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2232 int this_cpu, unsigned long *imbalance)
2236 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2239 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2241 return SCHED_LOAD_SCALE;
2244 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2246 return default_scale_freq_power(sd, cpu);
2249 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2251 unsigned long weight = sd->span_weight;
2252 unsigned long smt_gain = sd->smt_gain;
2259 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2261 return default_scale_smt_power(sd, cpu);
2264 unsigned long scale_rt_power(int cpu)
2266 struct rq *rq = cpu_rq(cpu);
2267 u64 total, available;
2269 sched_avg_update(rq);
2271 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2272 available = total - rq->rt_avg;
2274 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2275 total = SCHED_LOAD_SCALE;
2277 total >>= SCHED_LOAD_SHIFT;
2279 return div_u64(available, total);
2282 static void update_cpu_power(struct sched_domain *sd, int cpu)
2284 unsigned long weight = sd->span_weight;
2285 unsigned long power = SCHED_LOAD_SCALE;
2286 struct sched_group *sdg = sd->groups;
2288 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2289 if (sched_feat(ARCH_POWER))
2290 power *= arch_scale_smt_power(sd, cpu);
2292 power *= default_scale_smt_power(sd, cpu);
2294 power >>= SCHED_LOAD_SHIFT;
2297 sdg->cpu_power_orig = power;
2299 if (sched_feat(ARCH_POWER))
2300 power *= arch_scale_freq_power(sd, cpu);
2302 power *= default_scale_freq_power(sd, cpu);
2304 power >>= SCHED_LOAD_SHIFT;
2306 power *= scale_rt_power(cpu);
2307 power >>= SCHED_LOAD_SHIFT;
2312 cpu_rq(cpu)->cpu_power = power;
2313 sdg->cpu_power = power;
2316 static void update_group_power(struct sched_domain *sd, int cpu)
2318 struct sched_domain *child = sd->child;
2319 struct sched_group *group, *sdg = sd->groups;
2320 unsigned long power;
2323 update_cpu_power(sd, cpu);
2329 group = child->groups;
2331 power += group->cpu_power;
2332 group = group->next;
2333 } while (group != child->groups);
2335 sdg->cpu_power = power;
2339 * Try and fix up capacity for tiny siblings, this is needed when
2340 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2341 * which on its own isn't powerful enough.
2343 * See update_sd_pick_busiest() and check_asym_packing().
2346 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2349 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2351 if (sd->level != SD_LV_SIBLING)
2355 * If ~90% of the cpu_power is still there, we're good.
2357 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2364 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2365 * @sd: The sched_domain whose statistics are to be updated.
2366 * @group: sched_group whose statistics are to be updated.
2367 * @this_cpu: Cpu for which load balance is currently performed.
2368 * @idle: Idle status of this_cpu
2369 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2370 * @sd_idle: Idle status of the sched_domain containing group.
2371 * @local_group: Does group contain this_cpu.
2372 * @cpus: Set of cpus considered for load balancing.
2373 * @balance: Should we balance.
2374 * @sgs: variable to hold the statistics for this group.
2376 static inline void update_sg_lb_stats(struct sched_domain *sd,
2377 struct sched_group *group, int this_cpu,
2378 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2379 int local_group, const struct cpumask *cpus,
2380 int *balance, struct sg_lb_stats *sgs)
2382 unsigned long load, max_cpu_load, min_cpu_load;
2384 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2385 unsigned long avg_load_per_task = 0;
2388 balance_cpu = group_first_cpu(group);
2390 /* Tally up the load of all CPUs in the group */
2392 min_cpu_load = ~0UL;
2394 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2395 struct rq *rq = cpu_rq(i);
2397 if (*sd_idle && rq->nr_running)
2400 /* Bias balancing toward cpus of our domain */
2402 if (idle_cpu(i) && !first_idle_cpu) {
2407 load = target_load(i, load_idx);
2409 load = source_load(i, load_idx);
2410 if (load > max_cpu_load)
2411 max_cpu_load = load;
2412 if (min_cpu_load > load)
2413 min_cpu_load = load;
2416 sgs->group_load += load;
2417 sgs->sum_nr_running += rq->nr_running;
2418 sgs->sum_weighted_load += weighted_cpuload(i);
2423 * First idle cpu or the first cpu(busiest) in this sched group
2424 * is eligible for doing load balancing at this and above
2425 * domains. In the newly idle case, we will allow all the cpu's
2426 * to do the newly idle load balance.
2428 if (idle != CPU_NEWLY_IDLE && local_group &&
2429 balance_cpu != this_cpu) {
2434 update_group_power(sd, this_cpu);
2436 /* Adjust by relative CPU power of the group */
2437 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2440 * Consider the group unbalanced when the imbalance is larger
2441 * than the average weight of two tasks.
2443 * APZ: with cgroup the avg task weight can vary wildly and
2444 * might not be a suitable number - should we keep a
2445 * normalized nr_running number somewhere that negates
2448 if (sgs->sum_nr_running)
2449 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2451 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
2454 sgs->group_capacity =
2455 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2456 if (!sgs->group_capacity)
2457 sgs->group_capacity = fix_small_capacity(sd, group);
2461 * update_sd_pick_busiest - return 1 on busiest group
2462 * @sd: sched_domain whose statistics are to be checked
2463 * @sds: sched_domain statistics
2464 * @sg: sched_group candidate to be checked for being the busiest
2465 * @sgs: sched_group statistics
2466 * @this_cpu: the current cpu
2468 * Determine if @sg is a busier group than the previously selected
2471 static bool update_sd_pick_busiest(struct sched_domain *sd,
2472 struct sd_lb_stats *sds,
2473 struct sched_group *sg,
2474 struct sg_lb_stats *sgs,
2477 if (sgs->avg_load <= sds->max_load)
2480 if (sgs->sum_nr_running > sgs->group_capacity)
2487 * ASYM_PACKING needs to move all the work to the lowest
2488 * numbered CPUs in the group, therefore mark all groups
2489 * higher than ourself as busy.
2491 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2492 this_cpu < group_first_cpu(sg)) {
2496 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2504 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2505 * @sd: sched_domain whose statistics are to be updated.
2506 * @this_cpu: Cpu for which load balance is currently performed.
2507 * @idle: Idle status of this_cpu
2508 * @sd_idle: Idle status of the sched_domain containing sg.
2509 * @cpus: Set of cpus considered for load balancing.
2510 * @balance: Should we balance.
2511 * @sds: variable to hold the statistics for this sched_domain.
2513 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2514 enum cpu_idle_type idle, int *sd_idle,
2515 const struct cpumask *cpus, int *balance,
2516 struct sd_lb_stats *sds)
2518 struct sched_domain *child = sd->child;
2519 struct sched_group *sg = sd->groups;
2520 struct sg_lb_stats sgs;
2521 int load_idx, prefer_sibling = 0;
2523 if (child && child->flags & SD_PREFER_SIBLING)
2526 init_sd_power_savings_stats(sd, sds, idle);
2527 load_idx = get_sd_load_idx(sd, idle);
2532 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2533 memset(&sgs, 0, sizeof(sgs));
2534 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2535 local_group, cpus, balance, &sgs);
2537 if (local_group && !(*balance))
2540 sds->total_load += sgs.group_load;
2541 sds->total_pwr += sg->cpu_power;
2544 * In case the child domain prefers tasks go to siblings
2545 * first, lower the sg capacity to one so that we'll try
2546 * and move all the excess tasks away.
2549 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2552 sds->this_load = sgs.avg_load;
2554 sds->this_nr_running = sgs.sum_nr_running;
2555 sds->this_load_per_task = sgs.sum_weighted_load;
2556 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2557 sds->max_load = sgs.avg_load;
2559 sds->busiest_nr_running = sgs.sum_nr_running;
2560 sds->busiest_group_capacity = sgs.group_capacity;
2561 sds->busiest_load_per_task = sgs.sum_weighted_load;
2562 sds->group_imb = sgs.group_imb;
2565 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2567 } while (sg != sd->groups);
2570 int __weak arch_sd_sibling_asym_packing(void)
2572 return 0*SD_ASYM_PACKING;
2576 * check_asym_packing - Check to see if the group is packed into the
2579 * This is primarily intended to used at the sibling level. Some
2580 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2581 * case of POWER7, it can move to lower SMT modes only when higher
2582 * threads are idle. When in lower SMT modes, the threads will
2583 * perform better since they share less core resources. Hence when we
2584 * have idle threads, we want them to be the higher ones.
2586 * This packing function is run on idle threads. It checks to see if
2587 * the busiest CPU in this domain (core in the P7 case) has a higher
2588 * CPU number than the packing function is being run on. Here we are
2589 * assuming lower CPU number will be equivalent to lower a SMT thread
2592 * Returns 1 when packing is required and a task should be moved to
2593 * this CPU. The amount of the imbalance is returned in *imbalance.
2595 * @sd: The sched_domain whose packing is to be checked.
2596 * @sds: Statistics of the sched_domain which is to be packed
2597 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2598 * @imbalance: returns amount of imbalanced due to packing.
2600 static int check_asym_packing(struct sched_domain *sd,
2601 struct sd_lb_stats *sds,
2602 int this_cpu, unsigned long *imbalance)
2606 if (!(sd->flags & SD_ASYM_PACKING))
2612 busiest_cpu = group_first_cpu(sds->busiest);
2613 if (this_cpu > busiest_cpu)
2616 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2622 * fix_small_imbalance - Calculate the minor imbalance that exists
2623 * amongst the groups of a sched_domain, during
2625 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2626 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2627 * @imbalance: Variable to store the imbalance.
2629 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2630 int this_cpu, unsigned long *imbalance)
2632 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2633 unsigned int imbn = 2;
2634 unsigned long scaled_busy_load_per_task;
2636 if (sds->this_nr_running) {
2637 sds->this_load_per_task /= sds->this_nr_running;
2638 if (sds->busiest_load_per_task >
2639 sds->this_load_per_task)
2642 sds->this_load_per_task =
2643 cpu_avg_load_per_task(this_cpu);
2645 scaled_busy_load_per_task = sds->busiest_load_per_task
2647 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2649 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2650 (scaled_busy_load_per_task * imbn)) {
2651 *imbalance = sds->busiest_load_per_task;
2656 * OK, we don't have enough imbalance to justify moving tasks,
2657 * however we may be able to increase total CPU power used by
2661 pwr_now += sds->busiest->cpu_power *
2662 min(sds->busiest_load_per_task, sds->max_load);
2663 pwr_now += sds->this->cpu_power *
2664 min(sds->this_load_per_task, sds->this_load);
2665 pwr_now /= SCHED_LOAD_SCALE;
2667 /* Amount of load we'd subtract */
2668 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2669 sds->busiest->cpu_power;
2670 if (sds->max_load > tmp)
2671 pwr_move += sds->busiest->cpu_power *
2672 min(sds->busiest_load_per_task, sds->max_load - tmp);
2674 /* Amount of load we'd add */
2675 if (sds->max_load * sds->busiest->cpu_power <
2676 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2677 tmp = (sds->max_load * sds->busiest->cpu_power) /
2678 sds->this->cpu_power;
2680 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2681 sds->this->cpu_power;
2682 pwr_move += sds->this->cpu_power *
2683 min(sds->this_load_per_task, sds->this_load + tmp);
2684 pwr_move /= SCHED_LOAD_SCALE;
2686 /* Move if we gain throughput */
2687 if (pwr_move > pwr_now)
2688 *imbalance = sds->busiest_load_per_task;
2692 * calculate_imbalance - Calculate the amount of imbalance present within the
2693 * groups of a given sched_domain during load balance.
2694 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2695 * @this_cpu: Cpu for which currently load balance is being performed.
2696 * @imbalance: The variable to store the imbalance.
2698 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2699 unsigned long *imbalance)
2701 unsigned long max_pull, load_above_capacity = ~0UL;
2703 sds->busiest_load_per_task /= sds->busiest_nr_running;
2704 if (sds->group_imb) {
2705 sds->busiest_load_per_task =
2706 min(sds->busiest_load_per_task, sds->avg_load);
2710 * In the presence of smp nice balancing, certain scenarios can have
2711 * max load less than avg load(as we skip the groups at or below
2712 * its cpu_power, while calculating max_load..)
2714 if (sds->max_load < sds->avg_load) {
2716 return fix_small_imbalance(sds, this_cpu, imbalance);
2719 if (!sds->group_imb) {
2721 * Don't want to pull so many tasks that a group would go idle.
2723 load_above_capacity = (sds->busiest_nr_running -
2724 sds->busiest_group_capacity);
2726 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2728 load_above_capacity /= sds->busiest->cpu_power;
2732 * We're trying to get all the cpus to the average_load, so we don't
2733 * want to push ourselves above the average load, nor do we wish to
2734 * reduce the max loaded cpu below the average load. At the same time,
2735 * we also don't want to reduce the group load below the group capacity
2736 * (so that we can implement power-savings policies etc). Thus we look
2737 * for the minimum possible imbalance.
2738 * Be careful of negative numbers as they'll appear as very large values
2739 * with unsigned longs.
2741 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2743 /* How much load to actually move to equalise the imbalance */
2744 *imbalance = min(max_pull * sds->busiest->cpu_power,
2745 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
2749 * if *imbalance is less than the average load per runnable task
2750 * there is no gaurantee that any tasks will be moved so we'll have
2751 * a think about bumping its value to force at least one task to be
2754 if (*imbalance < sds->busiest_load_per_task)
2755 return fix_small_imbalance(sds, this_cpu, imbalance);
2758 /******* find_busiest_group() helpers end here *********************/
2761 * find_busiest_group - Returns the busiest group within the sched_domain
2762 * if there is an imbalance. If there isn't an imbalance, and
2763 * the user has opted for power-savings, it returns a group whose
2764 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2765 * such a group exists.
2767 * Also calculates the amount of weighted load which should be moved
2768 * to restore balance.
2770 * @sd: The sched_domain whose busiest group is to be returned.
2771 * @this_cpu: The cpu for which load balancing is currently being performed.
2772 * @imbalance: Variable which stores amount of weighted load which should
2773 * be moved to restore balance/put a group to idle.
2774 * @idle: The idle status of this_cpu.
2775 * @sd_idle: The idleness of sd
2776 * @cpus: The set of CPUs under consideration for load-balancing.
2777 * @balance: Pointer to a variable indicating if this_cpu
2778 * is the appropriate cpu to perform load balancing at this_level.
2780 * Returns: - the busiest group if imbalance exists.
2781 * - If no imbalance and user has opted for power-savings balance,
2782 * return the least loaded group whose CPUs can be
2783 * put to idle by rebalancing its tasks onto our group.
2785 static struct sched_group *
2786 find_busiest_group(struct sched_domain *sd, int this_cpu,
2787 unsigned long *imbalance, enum cpu_idle_type idle,
2788 int *sd_idle, const struct cpumask *cpus, int *balance)
2790 struct sd_lb_stats sds;
2792 memset(&sds, 0, sizeof(sds));
2795 * Compute the various statistics relavent for load balancing at
2798 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
2801 /* Cases where imbalance does not exist from POV of this_cpu */
2802 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2804 * 2) There is no busy sibling group to pull from.
2805 * 3) This group is the busiest group.
2806 * 4) This group is more busy than the avg busieness at this
2808 * 5) The imbalance is within the specified limit.
2813 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
2814 check_asym_packing(sd, &sds, this_cpu, imbalance))
2817 if (!sds.busiest || sds.busiest_nr_running == 0)
2820 if (sds.this_load >= sds.max_load)
2823 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
2825 if (sds.this_load >= sds.avg_load)
2828 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
2831 /* Looks like there is an imbalance. Compute it */
2832 calculate_imbalance(&sds, this_cpu, imbalance);
2837 * There is no obvious imbalance. But check if we can do some balancing
2840 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
2848 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2851 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
2852 enum cpu_idle_type idle, unsigned long imbalance,
2853 const struct cpumask *cpus)
2855 struct rq *busiest = NULL, *rq;
2856 unsigned long max_load = 0;
2859 for_each_cpu(i, sched_group_cpus(group)) {
2860 unsigned long power = power_of(i);
2861 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
2865 capacity = fix_small_capacity(sd, group);
2867 if (!cpumask_test_cpu(i, cpus))
2871 wl = weighted_cpuload(i);
2874 * When comparing with imbalance, use weighted_cpuload()
2875 * which is not scaled with the cpu power.
2877 if (capacity && rq->nr_running == 1 && wl > imbalance)
2881 * For the load comparisons with the other cpu's, consider
2882 * the weighted_cpuload() scaled with the cpu power, so that
2883 * the load can be moved away from the cpu that is potentially
2884 * running at a lower capacity.
2886 wl = (wl * SCHED_LOAD_SCALE) / power;
2888 if (wl > max_load) {
2898 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2899 * so long as it is large enough.
2901 #define MAX_PINNED_INTERVAL 512
2903 /* Working cpumask for load_balance and load_balance_newidle. */
2904 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
2906 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
2907 int busiest_cpu, int this_cpu)
2909 if (idle == CPU_NEWLY_IDLE) {
2912 * ASYM_PACKING needs to force migrate tasks from busy but
2913 * higher numbered CPUs in order to pack all tasks in the
2914 * lowest numbered CPUs.
2916 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
2920 * The only task running in a non-idle cpu can be moved to this
2921 * cpu in an attempt to completely freeup the other CPU
2924 * The package power saving logic comes from
2925 * find_busiest_group(). If there are no imbalance, then
2926 * f_b_g() will return NULL. However when sched_mc={1,2} then
2927 * f_b_g() will select a group from which a running task may be
2928 * pulled to this cpu in order to make the other package idle.
2929 * If there is no opportunity to make a package idle and if
2930 * there are no imbalance, then f_b_g() will return NULL and no
2931 * action will be taken in load_balance_newidle().
2933 * Under normal task pull operation due to imbalance, there
2934 * will be more than one task in the source run queue and
2935 * move_tasks() will succeed. ld_moved will be true and this
2936 * active balance code will not be triggered.
2938 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
2939 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2942 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
2946 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
2949 static int active_load_balance_cpu_stop(void *data);
2952 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2953 * tasks if there is an imbalance.
2955 static int load_balance(int this_cpu, struct rq *this_rq,
2956 struct sched_domain *sd, enum cpu_idle_type idle,
2959 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
2960 struct sched_group *group;
2961 unsigned long imbalance;
2963 unsigned long flags;
2964 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
2966 cpumask_copy(cpus, cpu_active_mask);
2969 * When power savings policy is enabled for the parent domain, idle
2970 * sibling can pick up load irrespective of busy siblings. In this case,
2971 * let the state of idle sibling percolate up as CPU_IDLE, instead of
2972 * portraying it as CPU_NOT_IDLE.
2974 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
2975 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
2978 schedstat_inc(sd, lb_count[idle]);
2982 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
2989 schedstat_inc(sd, lb_nobusyg[idle]);
2993 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
2995 schedstat_inc(sd, lb_nobusyq[idle]);
2999 BUG_ON(busiest == this_rq);
3001 schedstat_add(sd, lb_imbalance[idle], imbalance);
3004 if (busiest->nr_running > 1) {
3006 * Attempt to move tasks. If find_busiest_group has found
3007 * an imbalance but busiest->nr_running <= 1, the group is
3008 * still unbalanced. ld_moved simply stays zero, so it is
3009 * correctly treated as an imbalance.
3011 local_irq_save(flags);
3012 double_rq_lock(this_rq, busiest);
3013 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3014 imbalance, sd, idle, &all_pinned);
3015 double_rq_unlock(this_rq, busiest);
3016 local_irq_restore(flags);
3019 * some other cpu did the load balance for us.
3021 if (ld_moved && this_cpu != smp_processor_id())
3022 resched_cpu(this_cpu);
3024 /* All tasks on this runqueue were pinned by CPU affinity */
3025 if (unlikely(all_pinned)) {
3026 cpumask_clear_cpu(cpu_of(busiest), cpus);
3027 if (!cpumask_empty(cpus))
3034 schedstat_inc(sd, lb_failed[idle]);
3035 sd->nr_balance_failed++;
3037 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3039 raw_spin_lock_irqsave(&busiest->lock, flags);
3041 /* don't kick the active_load_balance_cpu_stop,
3042 * if the curr task on busiest cpu can't be
3045 if (!cpumask_test_cpu(this_cpu,
3046 &busiest->curr->cpus_allowed)) {
3047 raw_spin_unlock_irqrestore(&busiest->lock,
3050 goto out_one_pinned;
3054 * ->active_balance synchronizes accesses to
3055 * ->active_balance_work. Once set, it's cleared
3056 * only after active load balance is finished.
3058 if (!busiest->active_balance) {
3059 busiest->active_balance = 1;
3060 busiest->push_cpu = this_cpu;
3063 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3066 stop_one_cpu_nowait(cpu_of(busiest),
3067 active_load_balance_cpu_stop, busiest,
3068 &busiest->active_balance_work);
3071 * We've kicked active balancing, reset the failure
3074 sd->nr_balance_failed = sd->cache_nice_tries+1;
3077 sd->nr_balance_failed = 0;
3079 if (likely(!active_balance)) {
3080 /* We were unbalanced, so reset the balancing interval */
3081 sd->balance_interval = sd->min_interval;
3084 * If we've begun active balancing, start to back off. This
3085 * case may not be covered by the all_pinned logic if there
3086 * is only 1 task on the busy runqueue (because we don't call
3089 if (sd->balance_interval < sd->max_interval)
3090 sd->balance_interval *= 2;
3093 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3094 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3100 schedstat_inc(sd, lb_balanced[idle]);
3102 sd->nr_balance_failed = 0;
3105 /* tune up the balancing interval */
3106 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3107 (sd->balance_interval < sd->max_interval))
3108 sd->balance_interval *= 2;
3110 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3111 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3122 * idle_balance is called by schedule() if this_cpu is about to become
3123 * idle. Attempts to pull tasks from other CPUs.
3125 static void idle_balance(int this_cpu, struct rq *this_rq)
3127 struct sched_domain *sd;
3128 int pulled_task = 0;
3129 unsigned long next_balance = jiffies + HZ;
3131 this_rq->idle_stamp = this_rq->clock;
3133 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3137 * Drop the rq->lock, but keep IRQ/preempt disabled.
3139 raw_spin_unlock(&this_rq->lock);
3141 for_each_domain(this_cpu, sd) {
3142 unsigned long interval;
3145 if (!(sd->flags & SD_LOAD_BALANCE))
3148 if (sd->flags & SD_BALANCE_NEWIDLE) {
3149 /* If we've pulled tasks over stop searching: */
3150 pulled_task = load_balance(this_cpu, this_rq,
3151 sd, CPU_NEWLY_IDLE, &balance);
3154 interval = msecs_to_jiffies(sd->balance_interval);
3155 if (time_after(next_balance, sd->last_balance + interval))
3156 next_balance = sd->last_balance + interval;
3158 this_rq->idle_stamp = 0;
3163 raw_spin_lock(&this_rq->lock);
3165 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3167 * We are going idle. next_balance may be set based on
3168 * a busy processor. So reset next_balance.
3170 this_rq->next_balance = next_balance;
3175 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3176 * running tasks off the busiest CPU onto idle CPUs. It requires at
3177 * least 1 task to be running on each physical CPU where possible, and
3178 * avoids physical / logical imbalances.
3180 static int active_load_balance_cpu_stop(void *data)
3182 struct rq *busiest_rq = data;
3183 int busiest_cpu = cpu_of(busiest_rq);
3184 int target_cpu = busiest_rq->push_cpu;
3185 struct rq *target_rq = cpu_rq(target_cpu);
3186 struct sched_domain *sd;
3188 raw_spin_lock_irq(&busiest_rq->lock);
3190 /* make sure the requested cpu hasn't gone down in the meantime */
3191 if (unlikely(busiest_cpu != smp_processor_id() ||
3192 !busiest_rq->active_balance))
3195 /* Is there any task to move? */
3196 if (busiest_rq->nr_running <= 1)
3200 * This condition is "impossible", if it occurs
3201 * we need to fix it. Originally reported by
3202 * Bjorn Helgaas on a 128-cpu setup.
3204 BUG_ON(busiest_rq == target_rq);
3206 /* move a task from busiest_rq to target_rq */
3207 double_lock_balance(busiest_rq, target_rq);
3209 /* Search for an sd spanning us and the target CPU. */
3210 for_each_domain(target_cpu, sd) {
3211 if ((sd->flags & SD_LOAD_BALANCE) &&
3212 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3217 schedstat_inc(sd, alb_count);
3219 if (move_one_task(target_rq, target_cpu, busiest_rq,
3221 schedstat_inc(sd, alb_pushed);
3223 schedstat_inc(sd, alb_failed);
3225 double_unlock_balance(busiest_rq, target_rq);
3227 busiest_rq->active_balance = 0;
3228 raw_spin_unlock_irq(&busiest_rq->lock);
3234 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3236 static void trigger_sched_softirq(void *data)
3238 raise_softirq_irqoff(SCHED_SOFTIRQ);
3241 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3243 csd->func = trigger_sched_softirq;
3250 * idle load balancing details
3251 * - One of the idle CPUs nominates itself as idle load_balancer, while
3253 * - This idle load balancer CPU will also go into tickless mode when
3254 * it is idle, just like all other idle CPUs
3255 * - When one of the busy CPUs notice that there may be an idle rebalancing
3256 * needed, they will kick the idle load balancer, which then does idle
3257 * load balancing for all the idle CPUs.
3260 atomic_t load_balancer;
3261 atomic_t first_pick_cpu;
3262 atomic_t second_pick_cpu;
3263 cpumask_var_t idle_cpus_mask;
3264 cpumask_var_t grp_idle_mask;
3265 unsigned long next_balance; /* in jiffy units */
3266 } nohz ____cacheline_aligned;
3268 int get_nohz_load_balancer(void)
3270 return atomic_read(&nohz.load_balancer);
3273 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3275 * lowest_flag_domain - Return lowest sched_domain containing flag.
3276 * @cpu: The cpu whose lowest level of sched domain is to
3278 * @flag: The flag to check for the lowest sched_domain
3279 * for the given cpu.
3281 * Returns the lowest sched_domain of a cpu which contains the given flag.
3283 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3285 struct sched_domain *sd;
3287 for_each_domain(cpu, sd)
3288 if (sd && (sd->flags & flag))
3295 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3296 * @cpu: The cpu whose domains we're iterating over.
3297 * @sd: variable holding the value of the power_savings_sd
3299 * @flag: The flag to filter the sched_domains to be iterated.
3301 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3302 * set, starting from the lowest sched_domain to the highest.
3304 #define for_each_flag_domain(cpu, sd, flag) \
3305 for (sd = lowest_flag_domain(cpu, flag); \
3306 (sd && (sd->flags & flag)); sd = sd->parent)
3309 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3310 * @ilb_group: group to be checked for semi-idleness
3312 * Returns: 1 if the group is semi-idle. 0 otherwise.
3314 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3315 * and atleast one non-idle CPU. This helper function checks if the given
3316 * sched_group is semi-idle or not.
3318 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3320 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3321 sched_group_cpus(ilb_group));
3324 * A sched_group is semi-idle when it has atleast one busy cpu
3325 * and atleast one idle cpu.
3327 if (cpumask_empty(nohz.grp_idle_mask))
3330 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3336 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3337 * @cpu: The cpu which is nominating a new idle_load_balancer.
3339 * Returns: Returns the id of the idle load balancer if it exists,
3340 * Else, returns >= nr_cpu_ids.
3342 * This algorithm picks the idle load balancer such that it belongs to a
3343 * semi-idle powersavings sched_domain. The idea is to try and avoid
3344 * completely idle packages/cores just for the purpose of idle load balancing
3345 * when there are other idle cpu's which are better suited for that job.
3347 static int find_new_ilb(int cpu)
3349 struct sched_domain *sd;
3350 struct sched_group *ilb_group;
3353 * Have idle load balancer selection from semi-idle packages only
3354 * when power-aware load balancing is enabled
3356 if (!(sched_smt_power_savings || sched_mc_power_savings))
3360 * Optimize for the case when we have no idle CPUs or only one
3361 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3363 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3366 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3367 ilb_group = sd->groups;
3370 if (is_semi_idle_group(ilb_group))
3371 return cpumask_first(nohz.grp_idle_mask);
3373 ilb_group = ilb_group->next;
3375 } while (ilb_group != sd->groups);
3381 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3382 static inline int find_new_ilb(int call_cpu)
3389 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3390 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3391 * CPU (if there is one).
3393 static void nohz_balancer_kick(int cpu)
3397 nohz.next_balance++;
3399 ilb_cpu = get_nohz_load_balancer();
3401 if (ilb_cpu >= nr_cpu_ids) {
3402 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3403 if (ilb_cpu >= nr_cpu_ids)
3407 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3408 struct call_single_data *cp;
3410 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3411 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3412 __smp_call_function_single(ilb_cpu, cp, 0);
3418 * This routine will try to nominate the ilb (idle load balancing)
3419 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3420 * load balancing on behalf of all those cpus.
3422 * When the ilb owner becomes busy, we will not have new ilb owner until some
3423 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3424 * idle load balancing by kicking one of the idle CPUs.
3426 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3427 * ilb owner CPU in future (when there is a need for idle load balancing on
3428 * behalf of all idle CPUs).
3430 void select_nohz_load_balancer(int stop_tick)
3432 int cpu = smp_processor_id();
3435 if (!cpu_active(cpu)) {
3436 if (atomic_read(&nohz.load_balancer) != cpu)
3440 * If we are going offline and still the leader,
3443 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3450 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3452 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3453 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3454 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3455 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3457 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3460 /* make me the ilb owner */
3461 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3466 * Check to see if there is a more power-efficient
3469 new_ilb = find_new_ilb(cpu);
3470 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3471 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3472 resched_cpu(new_ilb);
3478 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3481 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3483 if (atomic_read(&nohz.load_balancer) == cpu)
3484 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3492 static DEFINE_SPINLOCK(balancing);
3495 * It checks each scheduling domain to see if it is due to be balanced,
3496 * and initiates a balancing operation if so.
3498 * Balancing parameters are set up in arch_init_sched_domains.
3500 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3503 struct rq *rq = cpu_rq(cpu);
3504 unsigned long interval;
3505 struct sched_domain *sd;
3506 /* Earliest time when we have to do rebalance again */
3507 unsigned long next_balance = jiffies + 60*HZ;
3508 int update_next_balance = 0;
3511 for_each_domain(cpu, sd) {
3512 if (!(sd->flags & SD_LOAD_BALANCE))
3515 interval = sd->balance_interval;
3516 if (idle != CPU_IDLE)
3517 interval *= sd->busy_factor;
3519 /* scale ms to jiffies */
3520 interval = msecs_to_jiffies(interval);
3521 if (unlikely(!interval))
3523 if (interval > HZ*NR_CPUS/10)
3524 interval = HZ*NR_CPUS/10;
3526 need_serialize = sd->flags & SD_SERIALIZE;
3528 if (need_serialize) {
3529 if (!spin_trylock(&balancing))
3533 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3534 if (load_balance(cpu, rq, sd, idle, &balance)) {
3536 * We've pulled tasks over so either we're no
3537 * longer idle, or one of our SMT siblings is
3540 idle = CPU_NOT_IDLE;
3542 sd->last_balance = jiffies;
3545 spin_unlock(&balancing);
3547 if (time_after(next_balance, sd->last_balance + interval)) {
3548 next_balance = sd->last_balance + interval;
3549 update_next_balance = 1;
3553 * Stop the load balance at this level. There is another
3554 * CPU in our sched group which is doing load balancing more
3562 * next_balance will be updated only when there is a need.
3563 * When the cpu is attached to null domain for ex, it will not be
3566 if (likely(update_next_balance))
3567 rq->next_balance = next_balance;
3572 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3573 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3575 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3577 struct rq *this_rq = cpu_rq(this_cpu);
3581 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3584 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3585 if (balance_cpu == this_cpu)
3589 * If this cpu gets work to do, stop the load balancing
3590 * work being done for other cpus. Next load
3591 * balancing owner will pick it up.
3593 if (need_resched()) {
3594 this_rq->nohz_balance_kick = 0;
3598 raw_spin_lock_irq(&this_rq->lock);
3599 update_cpu_load(this_rq);
3600 raw_spin_unlock_irq(&this_rq->lock);
3602 rebalance_domains(balance_cpu, CPU_IDLE);
3604 rq = cpu_rq(balance_cpu);
3605 if (time_after(this_rq->next_balance, rq->next_balance))
3606 this_rq->next_balance = rq->next_balance;
3608 nohz.next_balance = this_rq->next_balance;
3609 this_rq->nohz_balance_kick = 0;
3613 * Current heuristic for kicking the idle load balancer
3614 * - first_pick_cpu is the one of the busy CPUs. It will kick
3615 * idle load balancer when it has more than one process active. This
3616 * eliminates the need for idle load balancing altogether when we have
3617 * only one running process in the system (common case).
3618 * - If there are more than one busy CPU, idle load balancer may have
3619 * to run for active_load_balance to happen (i.e., two busy CPUs are
3620 * SMT or core siblings and can run better if they move to different
3621 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3622 * which will kick idle load balancer as soon as it has any load.
3624 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3626 unsigned long now = jiffies;
3628 int first_pick_cpu, second_pick_cpu;
3630 if (time_before(now, nohz.next_balance))
3633 if (!rq->nr_running)
3636 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3637 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3639 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3640 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3643 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3644 if (ret == nr_cpu_ids || ret == cpu) {
3645 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3646 if (rq->nr_running > 1)
3649 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3650 if (ret == nr_cpu_ids || ret == cpu) {
3658 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3662 * run_rebalance_domains is triggered when needed from the scheduler tick.
3663 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3665 static void run_rebalance_domains(struct softirq_action *h)
3667 int this_cpu = smp_processor_id();
3668 struct rq *this_rq = cpu_rq(this_cpu);
3669 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3670 CPU_IDLE : CPU_NOT_IDLE;
3672 rebalance_domains(this_cpu, idle);
3675 * If this cpu has a pending nohz_balance_kick, then do the
3676 * balancing on behalf of the other idle cpus whose ticks are
3679 nohz_idle_balance(this_cpu, idle);
3682 static inline int on_null_domain(int cpu)
3684 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3688 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3690 static inline void trigger_load_balance(struct rq *rq, int cpu)
3692 /* Don't need to rebalance while attached to NULL domain */
3693 if (time_after_eq(jiffies, rq->next_balance) &&
3694 likely(!on_null_domain(cpu)))
3695 raise_softirq(SCHED_SOFTIRQ);
3697 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
3698 nohz_balancer_kick(cpu);
3702 static void rq_online_fair(struct rq *rq)
3707 static void rq_offline_fair(struct rq *rq)
3712 #else /* CONFIG_SMP */
3715 * on UP we do not need to balance between CPUs:
3717 static inline void idle_balance(int cpu, struct rq *rq)
3721 #endif /* CONFIG_SMP */
3724 * scheduler tick hitting a task of our scheduling class:
3726 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
3728 struct cfs_rq *cfs_rq;
3729 struct sched_entity *se = &curr->se;
3731 for_each_sched_entity(se) {
3732 cfs_rq = cfs_rq_of(se);
3733 entity_tick(cfs_rq, se, queued);
3738 * called on fork with the child task as argument from the parent's context
3739 * - child not yet on the tasklist
3740 * - preemption disabled
3742 static void task_fork_fair(struct task_struct *p)
3744 struct cfs_rq *cfs_rq = task_cfs_rq(current);
3745 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
3746 int this_cpu = smp_processor_id();
3747 struct rq *rq = this_rq();
3748 unsigned long flags;
3750 raw_spin_lock_irqsave(&rq->lock, flags);
3752 if (unlikely(task_cpu(p) != this_cpu))
3753 __set_task_cpu(p, this_cpu);
3755 update_curr(cfs_rq);
3758 se->vruntime = curr->vruntime;
3759 place_entity(cfs_rq, se, 1);
3761 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
3763 * Upon rescheduling, sched_class::put_prev_task() will place
3764 * 'current' within the tree based on its new key value.
3766 swap(curr->vruntime, se->vruntime);
3767 resched_task(rq->curr);
3770 se->vruntime -= cfs_rq->min_vruntime;
3772 raw_spin_unlock_irqrestore(&rq->lock, flags);
3776 * Priority of the task has changed. Check to see if we preempt
3779 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
3780 int oldprio, int running)
3783 * Reschedule if we are currently running on this runqueue and
3784 * our priority decreased, or if we are not currently running on
3785 * this runqueue and our priority is higher than the current's
3788 if (p->prio > oldprio)
3789 resched_task(rq->curr);
3791 check_preempt_curr(rq, p, 0);
3795 * We switched to the sched_fair class.
3797 static void switched_to_fair(struct rq *rq, struct task_struct *p,
3801 * We were most likely switched from sched_rt, so
3802 * kick off the schedule if running, otherwise just see
3803 * if we can still preempt the current task.
3806 resched_task(rq->curr);
3808 check_preempt_curr(rq, p, 0);
3811 /* Account for a task changing its policy or group.
3813 * This routine is mostly called to set cfs_rq->curr field when a task
3814 * migrates between groups/classes.
3816 static void set_curr_task_fair(struct rq *rq)
3818 struct sched_entity *se = &rq->curr->se;
3820 for_each_sched_entity(se)
3821 set_next_entity(cfs_rq_of(se), se);
3824 #ifdef CONFIG_FAIR_GROUP_SCHED
3825 static void moved_group_fair(struct task_struct *p, int on_rq)
3827 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3829 update_curr(cfs_rq);
3831 place_entity(cfs_rq, &p->se, 1);
3835 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
3837 struct sched_entity *se = &task->se;
3838 unsigned int rr_interval = 0;
3841 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
3844 if (rq->cfs.load.weight)
3845 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
3851 * All the scheduling class methods:
3853 static const struct sched_class fair_sched_class = {
3854 .next = &idle_sched_class,
3855 .enqueue_task = enqueue_task_fair,
3856 .dequeue_task = dequeue_task_fair,
3857 .yield_task = yield_task_fair,
3859 .check_preempt_curr = check_preempt_wakeup,
3861 .pick_next_task = pick_next_task_fair,
3862 .put_prev_task = put_prev_task_fair,
3865 .select_task_rq = select_task_rq_fair,
3867 .rq_online = rq_online_fair,
3868 .rq_offline = rq_offline_fair,
3870 .task_waking = task_waking_fair,
3873 .set_curr_task = set_curr_task_fair,
3874 .task_tick = task_tick_fair,
3875 .task_fork = task_fork_fair,
3877 .prio_changed = prio_changed_fair,
3878 .switched_to = switched_to_fair,
3880 .get_rr_interval = get_rr_interval_fair,
3882 #ifdef CONFIG_FAIR_GROUP_SCHED
3883 .moved_group = moved_group_fair,
3887 #ifdef CONFIG_SCHED_DEBUG
3888 static void print_cfs_stats(struct seq_file *m, int cpu)
3890 struct cfs_rq *cfs_rq;
3893 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
3894 print_cfs_rq(m, cpu, cfs_rq);