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: 6ms * (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: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 8;
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);
429 static inline unsigned long
430 calc_delta_fair(unsigned long delta, struct sched_entity *se)
432 if (unlikely(se->load.weight != NICE_0_LOAD))
433 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
439 * The idea is to set a period in which each task runs once.
441 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
442 * this period because otherwise the slices get too small.
444 * p = (nr <= nl) ? l : l*nr/nl
446 static u64 __sched_period(unsigned long nr_running)
448 u64 period = sysctl_sched_latency;
449 unsigned long nr_latency = sched_nr_latency;
451 if (unlikely(nr_running > nr_latency)) {
452 period = sysctl_sched_min_granularity;
453 period *= nr_running;
460 * We calculate the wall-time slice from the period by taking a part
461 * proportional to the weight.
465 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
467 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
469 for_each_sched_entity(se) {
470 struct load_weight *load;
471 struct load_weight lw;
473 cfs_rq = cfs_rq_of(se);
474 load = &cfs_rq->load;
476 if (unlikely(!se->on_rq)) {
479 update_load_add(&lw, se->load.weight);
482 slice = calc_delta_mine(slice, se->load.weight, load);
488 * We calculate the vruntime slice of a to be inserted task
492 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
494 return calc_delta_fair(sched_slice(cfs_rq, se), se);
498 * Update the current task's runtime statistics. Skip current tasks that
499 * are not in our scheduling class.
502 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
503 unsigned long delta_exec)
505 unsigned long delta_exec_weighted;
507 schedstat_set(curr->statistics.exec_max,
508 max((u64)delta_exec, curr->statistics.exec_max));
510 curr->sum_exec_runtime += delta_exec;
511 schedstat_add(cfs_rq, exec_clock, delta_exec);
512 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
514 curr->vruntime += delta_exec_weighted;
515 update_min_vruntime(cfs_rq);
518 static void update_curr(struct cfs_rq *cfs_rq)
520 struct sched_entity *curr = cfs_rq->curr;
521 u64 now = rq_of(cfs_rq)->clock_task;
522 unsigned long delta_exec;
528 * Get the amount of time the current task was running
529 * since the last time we changed load (this cannot
530 * overflow on 32 bits):
532 delta_exec = (unsigned long)(now - curr->exec_start);
536 __update_curr(cfs_rq, curr, delta_exec);
537 curr->exec_start = now;
539 if (entity_is_task(curr)) {
540 struct task_struct *curtask = task_of(curr);
542 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
543 cpuacct_charge(curtask, delta_exec);
544 account_group_exec_runtime(curtask, delta_exec);
549 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
551 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
555 * Task is being enqueued - update stats:
557 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
560 * Are we enqueueing a waiting task? (for current tasks
561 * a dequeue/enqueue event is a NOP)
563 if (se != cfs_rq->curr)
564 update_stats_wait_start(cfs_rq, se);
568 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
570 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
571 rq_of(cfs_rq)->clock - se->statistics.wait_start));
572 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
573 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
574 rq_of(cfs_rq)->clock - se->statistics.wait_start);
575 #ifdef CONFIG_SCHEDSTATS
576 if (entity_is_task(se)) {
577 trace_sched_stat_wait(task_of(se),
578 rq_of(cfs_rq)->clock - se->statistics.wait_start);
581 schedstat_set(se->statistics.wait_start, 0);
585 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
588 * Mark the end of the wait period if dequeueing a
591 if (se != cfs_rq->curr)
592 update_stats_wait_end(cfs_rq, se);
596 * We are picking a new current task - update its stats:
599 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
602 * We are starting a new run period:
604 se->exec_start = rq_of(cfs_rq)->clock_task;
607 /**************************************************
608 * Scheduling class queueing methods:
611 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
613 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
615 cfs_rq->task_weight += weight;
619 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
625 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
627 update_load_add(&cfs_rq->load, se->load.weight);
628 if (!parent_entity(se))
629 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
630 if (entity_is_task(se)) {
631 add_cfs_task_weight(cfs_rq, se->load.weight);
632 list_add(&se->group_node, &cfs_rq->tasks);
634 cfs_rq->nr_running++;
638 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 update_load_sub(&cfs_rq->load, se->load.weight);
641 if (!parent_entity(se))
642 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
643 if (entity_is_task(se)) {
644 add_cfs_task_weight(cfs_rq, -se->load.weight);
645 list_del_init(&se->group_node);
647 cfs_rq->nr_running--;
650 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
651 static void update_cfs_load(struct cfs_rq *cfs_rq)
653 u64 period = sched_avg_period();
659 now = rq_of(cfs_rq)->clock;
660 delta = now - cfs_rq->load_stamp;
662 cfs_rq->load_stamp = now;
663 cfs_rq->load_period += delta;
664 cfs_rq->load_avg += delta * cfs_rq->load.weight;
666 while (cfs_rq->load_period > period) {
668 * Inline assembly required to prevent the compiler
669 * optimising this loop into a divmod call.
670 * See __iter_div_u64_rem() for another example of this.
672 asm("" : "+rm" (cfs_rq->load_period));
673 cfs_rq->load_period /= 2;
674 cfs_rq->load_avg /= 2;
678 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
679 unsigned long weight)
682 account_entity_dequeue(cfs_rq, se);
684 update_load_set(&se->load, weight);
687 account_entity_enqueue(cfs_rq, se);
690 static void update_cfs_shares(struct cfs_rq *cfs_rq)
692 struct task_group *tg;
693 struct sched_entity *se;
694 long load_weight, load, shares;
700 se = tg->se[cpu_of(rq_of(cfs_rq))];
704 load = cfs_rq->load.weight;
706 load_weight = atomic_read(&tg->load_weight);
707 load_weight -= cfs_rq->load_contribution;
710 shares = (tg->shares * load);
712 shares /= load_weight;
714 if (shares < MIN_SHARES)
716 if (shares > tg->shares)
719 reweight_entity(cfs_rq_of(se), se, shares);
721 #else /* CONFIG_FAIR_GROUP_SCHED */
722 static inline void update_cfs_load(struct cfs_rq *cfs_rq)
726 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
729 #endif /* CONFIG_FAIR_GROUP_SCHED */
731 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
733 #ifdef CONFIG_SCHEDSTATS
734 struct task_struct *tsk = NULL;
736 if (entity_is_task(se))
739 if (se->statistics.sleep_start) {
740 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
745 if (unlikely(delta > se->statistics.sleep_max))
746 se->statistics.sleep_max = delta;
748 se->statistics.sleep_start = 0;
749 se->statistics.sum_sleep_runtime += delta;
752 account_scheduler_latency(tsk, delta >> 10, 1);
753 trace_sched_stat_sleep(tsk, delta);
756 if (se->statistics.block_start) {
757 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
762 if (unlikely(delta > se->statistics.block_max))
763 se->statistics.block_max = delta;
765 se->statistics.block_start = 0;
766 se->statistics.sum_sleep_runtime += delta;
769 if (tsk->in_iowait) {
770 se->statistics.iowait_sum += delta;
771 se->statistics.iowait_count++;
772 trace_sched_stat_iowait(tsk, delta);
776 * Blocking time is in units of nanosecs, so shift by
777 * 20 to get a milliseconds-range estimation of the
778 * amount of time that the task spent sleeping:
780 if (unlikely(prof_on == SLEEP_PROFILING)) {
781 profile_hits(SLEEP_PROFILING,
782 (void *)get_wchan(tsk),
785 account_scheduler_latency(tsk, delta >> 10, 0);
791 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
793 #ifdef CONFIG_SCHED_DEBUG
794 s64 d = se->vruntime - cfs_rq->min_vruntime;
799 if (d > 3*sysctl_sched_latency)
800 schedstat_inc(cfs_rq, nr_spread_over);
805 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
807 u64 vruntime = cfs_rq->min_vruntime;
810 * The 'current' period is already promised to the current tasks,
811 * however the extra weight of the new task will slow them down a
812 * little, place the new task so that it fits in the slot that
813 * stays open at the end.
815 if (initial && sched_feat(START_DEBIT))
816 vruntime += sched_vslice(cfs_rq, se);
818 /* sleeps up to a single latency don't count. */
820 unsigned long thresh = sysctl_sched_latency;
823 * Halve their sleep time's effect, to allow
824 * for a gentler effect of sleepers:
826 if (sched_feat(GENTLE_FAIR_SLEEPERS))
832 /* ensure we never gain time by being placed backwards. */
833 vruntime = max_vruntime(se->vruntime, vruntime);
835 se->vruntime = vruntime;
839 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
842 * Update the normalized vruntime before updating min_vruntime
843 * through callig update_curr().
845 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
846 se->vruntime += cfs_rq->min_vruntime;
849 * Update run-time statistics of the 'current'.
852 update_cfs_load(cfs_rq);
853 account_entity_enqueue(cfs_rq, se);
854 update_cfs_shares(cfs_rq);
856 if (flags & ENQUEUE_WAKEUP) {
857 place_entity(cfs_rq, se, 0);
858 enqueue_sleeper(cfs_rq, se);
861 update_stats_enqueue(cfs_rq, se);
862 check_spread(cfs_rq, se);
863 if (se != cfs_rq->curr)
864 __enqueue_entity(cfs_rq, se);
868 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
870 if (!se || cfs_rq->last == se)
873 if (!se || cfs_rq->next == se)
877 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
879 for_each_sched_entity(se)
880 __clear_buddies(cfs_rq_of(se), se);
884 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
887 * Update run-time statistics of the 'current'.
891 update_stats_dequeue(cfs_rq, se);
892 if (flags & DEQUEUE_SLEEP) {
893 #ifdef CONFIG_SCHEDSTATS
894 if (entity_is_task(se)) {
895 struct task_struct *tsk = task_of(se);
897 if (tsk->state & TASK_INTERRUPTIBLE)
898 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
899 if (tsk->state & TASK_UNINTERRUPTIBLE)
900 se->statistics.block_start = rq_of(cfs_rq)->clock;
905 clear_buddies(cfs_rq, se);
907 if (se != cfs_rq->curr)
908 __dequeue_entity(cfs_rq, se);
910 update_cfs_load(cfs_rq);
911 account_entity_dequeue(cfs_rq, se);
912 update_min_vruntime(cfs_rq);
913 update_cfs_shares(cfs_rq);
916 * Normalize the entity after updating the min_vruntime because the
917 * update can refer to the ->curr item and we need to reflect this
918 * movement in our normalized position.
920 if (!(flags & DEQUEUE_SLEEP))
921 se->vruntime -= cfs_rq->min_vruntime;
925 * Preempt the current task with a newly woken task if needed:
928 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
930 unsigned long ideal_runtime, delta_exec;
932 ideal_runtime = sched_slice(cfs_rq, curr);
933 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
934 if (delta_exec > ideal_runtime) {
935 resched_task(rq_of(cfs_rq)->curr);
937 * The current task ran long enough, ensure it doesn't get
938 * re-elected due to buddy favours.
940 clear_buddies(cfs_rq, curr);
945 * Ensure that a task that missed wakeup preemption by a
946 * narrow margin doesn't have to wait for a full slice.
947 * This also mitigates buddy induced latencies under load.
949 if (!sched_feat(WAKEUP_PREEMPT))
952 if (delta_exec < sysctl_sched_min_granularity)
955 if (cfs_rq->nr_running > 1) {
956 struct sched_entity *se = __pick_next_entity(cfs_rq);
957 s64 delta = curr->vruntime - se->vruntime;
959 if (delta > ideal_runtime)
960 resched_task(rq_of(cfs_rq)->curr);
965 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
967 /* 'current' is not kept within the tree. */
970 * Any task has to be enqueued before it get to execute on
971 * a CPU. So account for the time it spent waiting on the
974 update_stats_wait_end(cfs_rq, se);
975 __dequeue_entity(cfs_rq, se);
978 update_stats_curr_start(cfs_rq, se);
980 #ifdef CONFIG_SCHEDSTATS
982 * Track our maximum slice length, if the CPU's load is at
983 * least twice that of our own weight (i.e. dont track it
984 * when there are only lesser-weight tasks around):
986 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
987 se->statistics.slice_max = max(se->statistics.slice_max,
988 se->sum_exec_runtime - se->prev_sum_exec_runtime);
991 se->prev_sum_exec_runtime = se->sum_exec_runtime;
995 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
997 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
999 struct sched_entity *se = __pick_next_entity(cfs_rq);
1000 struct sched_entity *left = se;
1002 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1006 * Prefer last buddy, try to return the CPU to a preempted task.
1008 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1011 clear_buddies(cfs_rq, se);
1016 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1019 * If still on the runqueue then deactivate_task()
1020 * was not called and update_curr() has to be done:
1023 update_curr(cfs_rq);
1025 check_spread(cfs_rq, prev);
1027 update_stats_wait_start(cfs_rq, prev);
1028 /* Put 'current' back into the tree. */
1029 __enqueue_entity(cfs_rq, prev);
1031 cfs_rq->curr = NULL;
1035 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1038 * Update run-time statistics of the 'current'.
1040 update_curr(cfs_rq);
1042 #ifdef CONFIG_SCHED_HRTICK
1044 * queued ticks are scheduled to match the slice, so don't bother
1045 * validating it and just reschedule.
1048 resched_task(rq_of(cfs_rq)->curr);
1052 * don't let the period tick interfere with the hrtick preemption
1054 if (!sched_feat(DOUBLE_TICK) &&
1055 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1059 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1060 check_preempt_tick(cfs_rq, curr);
1063 /**************************************************
1064 * CFS operations on tasks:
1067 #ifdef CONFIG_SCHED_HRTICK
1068 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1070 struct sched_entity *se = &p->se;
1071 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1073 WARN_ON(task_rq(p) != rq);
1075 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1076 u64 slice = sched_slice(cfs_rq, se);
1077 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1078 s64 delta = slice - ran;
1087 * Don't schedule slices shorter than 10000ns, that just
1088 * doesn't make sense. Rely on vruntime for fairness.
1091 delta = max_t(s64, 10000LL, delta);
1093 hrtick_start(rq, delta);
1098 * called from enqueue/dequeue and updates the hrtick when the
1099 * current task is from our class and nr_running is low enough
1102 static void hrtick_update(struct rq *rq)
1104 struct task_struct *curr = rq->curr;
1106 if (curr->sched_class != &fair_sched_class)
1109 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1110 hrtick_start_fair(rq, curr);
1112 #else /* !CONFIG_SCHED_HRTICK */
1114 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1118 static inline void hrtick_update(struct rq *rq)
1124 * The enqueue_task method is called before nr_running is
1125 * increased. Here we update the fair scheduling stats and
1126 * then put the task into the rbtree:
1129 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1131 struct cfs_rq *cfs_rq;
1132 struct sched_entity *se = &p->se;
1134 for_each_sched_entity(se) {
1137 cfs_rq = cfs_rq_of(se);
1138 enqueue_entity(cfs_rq, se, flags);
1139 flags = ENQUEUE_WAKEUP;
1142 for_each_sched_entity(se) {
1143 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1145 update_cfs_load(cfs_rq);
1146 update_cfs_shares(cfs_rq);
1153 * The dequeue_task method is called before nr_running is
1154 * decreased. We remove the task from the rbtree and
1155 * update the fair scheduling stats:
1157 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1159 struct cfs_rq *cfs_rq;
1160 struct sched_entity *se = &p->se;
1162 for_each_sched_entity(se) {
1163 cfs_rq = cfs_rq_of(se);
1164 dequeue_entity(cfs_rq, se, flags);
1166 /* Don't dequeue parent if it has other entities besides us */
1167 if (cfs_rq->load.weight)
1169 flags |= DEQUEUE_SLEEP;
1172 for_each_sched_entity(se) {
1173 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1175 update_cfs_load(cfs_rq);
1176 update_cfs_shares(cfs_rq);
1183 * sched_yield() support is very simple - we dequeue and enqueue.
1185 * If compat_yield is turned on then we requeue to the end of the tree.
1187 static void yield_task_fair(struct rq *rq)
1189 struct task_struct *curr = rq->curr;
1190 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1191 struct sched_entity *rightmost, *se = &curr->se;
1194 * Are we the only task in the tree?
1196 if (unlikely(cfs_rq->nr_running == 1))
1199 clear_buddies(cfs_rq, se);
1201 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1202 update_rq_clock(rq);
1204 * Update run-time statistics of the 'current'.
1206 update_curr(cfs_rq);
1211 * Find the rightmost entry in the rbtree:
1213 rightmost = __pick_last_entity(cfs_rq);
1215 * Already in the rightmost position?
1217 if (unlikely(!rightmost || entity_before(rightmost, se)))
1221 * Minimally necessary key value to be last in the tree:
1222 * Upon rescheduling, sched_class::put_prev_task() will place
1223 * 'current' within the tree based on its new key value.
1225 se->vruntime = rightmost->vruntime + 1;
1230 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1232 struct sched_entity *se = &p->se;
1233 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1235 se->vruntime -= cfs_rq->min_vruntime;
1238 #ifdef CONFIG_FAIR_GROUP_SCHED
1240 * effective_load() calculates the load change as seen from the root_task_group
1242 * Adding load to a group doesn't make a group heavier, but can cause movement
1243 * of group shares between cpus. Assuming the shares were perfectly aligned one
1244 * can calculate the shift in shares.
1246 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1248 struct sched_entity *se = tg->se[cpu];
1253 for_each_sched_entity(se) {
1254 long S, rw, s, a, b;
1256 S = se->my_q->tg->shares;
1257 s = se->load.weight;
1258 rw = se->my_q->load.weight;
1269 * Assume the group is already running and will
1270 * thus already be accounted for in the weight.
1272 * That is, moving shares between CPUs, does not
1273 * alter the group weight.
1283 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1284 unsigned long wl, unsigned long wg)
1291 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1293 unsigned long this_load, load;
1294 int idx, this_cpu, prev_cpu;
1295 unsigned long tl_per_task;
1296 struct task_group *tg;
1297 unsigned long weight;
1301 this_cpu = smp_processor_id();
1302 prev_cpu = task_cpu(p);
1303 load = source_load(prev_cpu, idx);
1304 this_load = target_load(this_cpu, idx);
1307 * If sync wakeup then subtract the (maximum possible)
1308 * effect of the currently running task from the load
1309 * of the current CPU:
1313 tg = task_group(current);
1314 weight = current->se.load.weight;
1316 this_load += effective_load(tg, this_cpu, -weight, -weight);
1317 load += effective_load(tg, prev_cpu, 0, -weight);
1321 weight = p->se.load.weight;
1324 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1325 * due to the sync cause above having dropped this_load to 0, we'll
1326 * always have an imbalance, but there's really nothing you can do
1327 * about that, so that's good too.
1329 * Otherwise check if either cpus are near enough in load to allow this
1330 * task to be woken on this_cpu.
1333 unsigned long this_eff_load, prev_eff_load;
1335 this_eff_load = 100;
1336 this_eff_load *= power_of(prev_cpu);
1337 this_eff_load *= this_load +
1338 effective_load(tg, this_cpu, weight, weight);
1340 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1341 prev_eff_load *= power_of(this_cpu);
1342 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1344 balanced = this_eff_load <= prev_eff_load;
1350 * If the currently running task will sleep within
1351 * a reasonable amount of time then attract this newly
1354 if (sync && balanced)
1357 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1358 tl_per_task = cpu_avg_load_per_task(this_cpu);
1361 (this_load <= load &&
1362 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1364 * This domain has SD_WAKE_AFFINE and
1365 * p is cache cold in this domain, and
1366 * there is no bad imbalance.
1368 schedstat_inc(sd, ttwu_move_affine);
1369 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1377 * find_idlest_group finds and returns the least busy CPU group within the
1380 static struct sched_group *
1381 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1382 int this_cpu, int load_idx)
1384 struct sched_group *idlest = NULL, *group = sd->groups;
1385 unsigned long min_load = ULONG_MAX, this_load = 0;
1386 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1389 unsigned long load, avg_load;
1393 /* Skip over this group if it has no CPUs allowed */
1394 if (!cpumask_intersects(sched_group_cpus(group),
1398 local_group = cpumask_test_cpu(this_cpu,
1399 sched_group_cpus(group));
1401 /* Tally up the load of all CPUs in the group */
1404 for_each_cpu(i, sched_group_cpus(group)) {
1405 /* Bias balancing toward cpus of our domain */
1407 load = source_load(i, load_idx);
1409 load = target_load(i, load_idx);
1414 /* Adjust by relative CPU power of the group */
1415 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1418 this_load = avg_load;
1419 } else if (avg_load < min_load) {
1420 min_load = avg_load;
1423 } while (group = group->next, group != sd->groups);
1425 if (!idlest || 100*this_load < imbalance*min_load)
1431 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1434 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1436 unsigned long load, min_load = ULONG_MAX;
1440 /* Traverse only the allowed CPUs */
1441 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1442 load = weighted_cpuload(i);
1444 if (load < min_load || (load == min_load && i == this_cpu)) {
1454 * Try and locate an idle CPU in the sched_domain.
1456 static int select_idle_sibling(struct task_struct *p, int target)
1458 int cpu = smp_processor_id();
1459 int prev_cpu = task_cpu(p);
1460 struct sched_domain *sd;
1464 * If the task is going to be woken-up on this cpu and if it is
1465 * already idle, then it is the right target.
1467 if (target == cpu && idle_cpu(cpu))
1471 * If the task is going to be woken-up on the cpu where it previously
1472 * ran and if it is currently idle, then it the right target.
1474 if (target == prev_cpu && idle_cpu(prev_cpu))
1478 * Otherwise, iterate the domains and find an elegible idle cpu.
1480 for_each_domain(target, sd) {
1481 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1484 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1492 * Lets stop looking for an idle sibling when we reached
1493 * the domain that spans the current cpu and prev_cpu.
1495 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1496 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1504 * sched_balance_self: balance the current task (running on cpu) in domains
1505 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1508 * Balance, ie. select the least loaded group.
1510 * Returns the target CPU number, or the same CPU if no balancing is needed.
1512 * preempt must be disabled.
1515 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1517 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1518 int cpu = smp_processor_id();
1519 int prev_cpu = task_cpu(p);
1521 int want_affine = 0;
1523 int sync = wake_flags & WF_SYNC;
1525 if (sd_flag & SD_BALANCE_WAKE) {
1526 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1531 for_each_domain(cpu, tmp) {
1532 if (!(tmp->flags & SD_LOAD_BALANCE))
1536 * If power savings logic is enabled for a domain, see if we
1537 * are not overloaded, if so, don't balance wider.
1539 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1540 unsigned long power = 0;
1541 unsigned long nr_running = 0;
1542 unsigned long capacity;
1545 for_each_cpu(i, sched_domain_span(tmp)) {
1546 power += power_of(i);
1547 nr_running += cpu_rq(i)->cfs.nr_running;
1550 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1552 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1555 if (nr_running < capacity)
1560 * If both cpu and prev_cpu are part of this domain,
1561 * cpu is a valid SD_WAKE_AFFINE target.
1563 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1564 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1569 if (!want_sd && !want_affine)
1572 if (!(tmp->flags & sd_flag))
1580 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1581 return select_idle_sibling(p, cpu);
1583 return select_idle_sibling(p, prev_cpu);
1587 int load_idx = sd->forkexec_idx;
1588 struct sched_group *group;
1591 if (!(sd->flags & sd_flag)) {
1596 if (sd_flag & SD_BALANCE_WAKE)
1597 load_idx = sd->wake_idx;
1599 group = find_idlest_group(sd, p, cpu, load_idx);
1605 new_cpu = find_idlest_cpu(group, p, cpu);
1606 if (new_cpu == -1 || new_cpu == cpu) {
1607 /* Now try balancing at a lower domain level of cpu */
1612 /* Now try balancing at a lower domain level of new_cpu */
1614 weight = sd->span_weight;
1616 for_each_domain(cpu, tmp) {
1617 if (weight <= tmp->span_weight)
1619 if (tmp->flags & sd_flag)
1622 /* while loop will break here if sd == NULL */
1627 #endif /* CONFIG_SMP */
1629 static unsigned long
1630 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1632 unsigned long gran = sysctl_sched_wakeup_granularity;
1635 * Since its curr running now, convert the gran from real-time
1636 * to virtual-time in his units.
1638 * By using 'se' instead of 'curr' we penalize light tasks, so
1639 * they get preempted easier. That is, if 'se' < 'curr' then
1640 * the resulting gran will be larger, therefore penalizing the
1641 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1642 * be smaller, again penalizing the lighter task.
1644 * This is especially important for buddies when the leftmost
1645 * task is higher priority than the buddy.
1647 if (unlikely(se->load.weight != NICE_0_LOAD))
1648 gran = calc_delta_fair(gran, se);
1654 * Should 'se' preempt 'curr'.
1668 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1670 s64 gran, vdiff = curr->vruntime - se->vruntime;
1675 gran = wakeup_gran(curr, se);
1682 static void set_last_buddy(struct sched_entity *se)
1684 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1685 for_each_sched_entity(se)
1686 cfs_rq_of(se)->last = se;
1690 static void set_next_buddy(struct sched_entity *se)
1692 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1693 for_each_sched_entity(se)
1694 cfs_rq_of(se)->next = se;
1699 * Preempt the current task with a newly woken task if needed:
1701 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1703 struct task_struct *curr = rq->curr;
1704 struct sched_entity *se = &curr->se, *pse = &p->se;
1705 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1706 int scale = cfs_rq->nr_running >= sched_nr_latency;
1708 if (unlikely(rt_prio(p->prio)))
1711 if (unlikely(p->sched_class != &fair_sched_class))
1714 if (unlikely(se == pse))
1717 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1718 set_next_buddy(pse);
1721 * We can come here with TIF_NEED_RESCHED already set from new task
1724 if (test_tsk_need_resched(curr))
1728 * Batch and idle tasks do not preempt (their preemption is driven by
1731 if (unlikely(p->policy != SCHED_NORMAL))
1734 /* Idle tasks are by definition preempted by everybody. */
1735 if (unlikely(curr->policy == SCHED_IDLE))
1738 if (!sched_feat(WAKEUP_PREEMPT))
1741 update_curr(cfs_rq);
1742 find_matching_se(&se, &pse);
1744 if (wakeup_preempt_entity(se, pse) == 1)
1752 * Only set the backward buddy when the current task is still
1753 * on the rq. This can happen when a wakeup gets interleaved
1754 * with schedule on the ->pre_schedule() or idle_balance()
1755 * point, either of which can * drop the rq lock.
1757 * Also, during early boot the idle thread is in the fair class,
1758 * for obvious reasons its a bad idea to schedule back to it.
1760 if (unlikely(!se->on_rq || curr == rq->idle))
1763 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1767 static struct task_struct *pick_next_task_fair(struct rq *rq)
1769 struct task_struct *p;
1770 struct cfs_rq *cfs_rq = &rq->cfs;
1771 struct sched_entity *se;
1773 if (!cfs_rq->nr_running)
1777 se = pick_next_entity(cfs_rq);
1778 set_next_entity(cfs_rq, se);
1779 cfs_rq = group_cfs_rq(se);
1783 hrtick_start_fair(rq, p);
1789 * Account for a descheduled task:
1791 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1793 struct sched_entity *se = &prev->se;
1794 struct cfs_rq *cfs_rq;
1796 for_each_sched_entity(se) {
1797 cfs_rq = cfs_rq_of(se);
1798 put_prev_entity(cfs_rq, se);
1803 /**************************************************
1804 * Fair scheduling class load-balancing methods:
1808 * pull_task - move a task from a remote runqueue to the local runqueue.
1809 * Both runqueues must be locked.
1811 static void pull_task(struct rq *src_rq, struct task_struct *p,
1812 struct rq *this_rq, int this_cpu)
1814 deactivate_task(src_rq, p, 0);
1815 set_task_cpu(p, this_cpu);
1816 activate_task(this_rq, p, 0);
1817 check_preempt_curr(this_rq, p, 0);
1819 /* re-arm NEWIDLE balancing when moving tasks */
1820 src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost;
1821 this_rq->idle_stamp = 0;
1825 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1828 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1829 struct sched_domain *sd, enum cpu_idle_type idle,
1832 int tsk_cache_hot = 0;
1834 * We do not migrate tasks that are:
1835 * 1) running (obviously), or
1836 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1837 * 3) are cache-hot on their current CPU.
1839 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1840 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1845 if (task_running(rq, p)) {
1846 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1851 * Aggressive migration if:
1852 * 1) task is cache cold, or
1853 * 2) too many balance attempts have failed.
1856 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
1857 if (!tsk_cache_hot ||
1858 sd->nr_balance_failed > sd->cache_nice_tries) {
1859 #ifdef CONFIG_SCHEDSTATS
1860 if (tsk_cache_hot) {
1861 schedstat_inc(sd, lb_hot_gained[idle]);
1862 schedstat_inc(p, se.statistics.nr_forced_migrations);
1868 if (tsk_cache_hot) {
1869 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
1876 * move_one_task tries to move exactly one task from busiest to this_rq, as
1877 * part of active balancing operations within "domain".
1878 * Returns 1 if successful and 0 otherwise.
1880 * Called with both runqueues locked.
1883 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1884 struct sched_domain *sd, enum cpu_idle_type idle)
1886 struct task_struct *p, *n;
1887 struct cfs_rq *cfs_rq;
1890 for_each_leaf_cfs_rq(busiest, cfs_rq) {
1891 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1893 if (!can_migrate_task(p, busiest, this_cpu,
1897 pull_task(busiest, p, this_rq, this_cpu);
1899 * Right now, this is only the second place pull_task()
1900 * is called, so we can safely collect pull_task()
1901 * stats here rather than inside pull_task().
1903 schedstat_inc(sd, lb_gained[idle]);
1911 static unsigned long
1912 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1913 unsigned long max_load_move, struct sched_domain *sd,
1914 enum cpu_idle_type idle, int *all_pinned,
1915 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
1917 int loops = 0, pulled = 0, pinned = 0;
1918 long rem_load_move = max_load_move;
1919 struct task_struct *p, *n;
1921 if (max_load_move == 0)
1926 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
1927 if (loops++ > sysctl_sched_nr_migrate)
1930 if ((p->se.load.weight >> 1) > rem_load_move ||
1931 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
1934 pull_task(busiest, p, this_rq, this_cpu);
1936 rem_load_move -= p->se.load.weight;
1938 #ifdef CONFIG_PREEMPT
1940 * NEWIDLE balancing is a source of latency, so preemptible
1941 * kernels will stop after the first task is pulled to minimize
1942 * the critical section.
1944 if (idle == CPU_NEWLY_IDLE)
1949 * We only want to steal up to the prescribed amount of
1952 if (rem_load_move <= 0)
1955 if (p->prio < *this_best_prio)
1956 *this_best_prio = p->prio;
1960 * Right now, this is one of only two places pull_task() is called,
1961 * so we can safely collect pull_task() stats here rather than
1962 * inside pull_task().
1964 schedstat_add(sd, lb_gained[idle], pulled);
1967 *all_pinned = pinned;
1969 return max_load_move - rem_load_move;
1972 #ifdef CONFIG_FAIR_GROUP_SCHED
1973 static unsigned long
1974 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1975 unsigned long max_load_move,
1976 struct sched_domain *sd, enum cpu_idle_type idle,
1977 int *all_pinned, int *this_best_prio)
1979 long rem_load_move = max_load_move;
1980 int busiest_cpu = cpu_of(busiest);
1981 struct task_group *tg;
1984 update_h_load(busiest_cpu);
1986 list_for_each_entry_rcu(tg, &task_groups, list) {
1987 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1988 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1989 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1990 u64 rem_load, moved_load;
1995 if (!busiest_cfs_rq->task_weight)
1998 rem_load = (u64)rem_load_move * busiest_weight;
1999 rem_load = div_u64(rem_load, busiest_h_load + 1);
2001 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2002 rem_load, sd, idle, all_pinned, this_best_prio,
2008 moved_load *= busiest_h_load;
2009 moved_load = div_u64(moved_load, busiest_weight + 1);
2011 rem_load_move -= moved_load;
2012 if (rem_load_move < 0)
2017 return max_load_move - rem_load_move;
2020 static unsigned long
2021 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2022 unsigned long max_load_move,
2023 struct sched_domain *sd, enum cpu_idle_type idle,
2024 int *all_pinned, int *this_best_prio)
2026 return balance_tasks(this_rq, this_cpu, busiest,
2027 max_load_move, sd, idle, all_pinned,
2028 this_best_prio, &busiest->cfs);
2033 * move_tasks tries to move up to max_load_move weighted load from busiest to
2034 * this_rq, as part of a balancing operation within domain "sd".
2035 * Returns 1 if successful and 0 otherwise.
2037 * Called with both runqueues locked.
2039 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2040 unsigned long max_load_move,
2041 struct sched_domain *sd, enum cpu_idle_type idle,
2044 unsigned long total_load_moved = 0, load_moved;
2045 int this_best_prio = this_rq->curr->prio;
2048 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2049 max_load_move - total_load_moved,
2050 sd, idle, all_pinned, &this_best_prio);
2052 total_load_moved += load_moved;
2054 #ifdef CONFIG_PREEMPT
2056 * NEWIDLE balancing is a source of latency, so preemptible
2057 * kernels will stop after the first task is pulled to minimize
2058 * the critical section.
2060 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2063 if (raw_spin_is_contended(&this_rq->lock) ||
2064 raw_spin_is_contended(&busiest->lock))
2067 } while (load_moved && max_load_move > total_load_moved);
2069 return total_load_moved > 0;
2072 /********** Helpers for find_busiest_group ************************/
2074 * sd_lb_stats - Structure to store the statistics of a sched_domain
2075 * during load balancing.
2077 struct sd_lb_stats {
2078 struct sched_group *busiest; /* Busiest group in this sd */
2079 struct sched_group *this; /* Local group in this sd */
2080 unsigned long total_load; /* Total load of all groups in sd */
2081 unsigned long total_pwr; /* Total power of all groups in sd */
2082 unsigned long avg_load; /* Average load across all groups in sd */
2084 /** Statistics of this group */
2085 unsigned long this_load;
2086 unsigned long this_load_per_task;
2087 unsigned long this_nr_running;
2088 unsigned long this_has_capacity;
2090 /* Statistics of the busiest group */
2091 unsigned long max_load;
2092 unsigned long busiest_load_per_task;
2093 unsigned long busiest_nr_running;
2094 unsigned long busiest_group_capacity;
2095 unsigned long busiest_has_capacity;
2097 int group_imb; /* Is there imbalance in this sd */
2098 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2099 int power_savings_balance; /* Is powersave balance needed for this sd */
2100 struct sched_group *group_min; /* Least loaded group in sd */
2101 struct sched_group *group_leader; /* Group which relieves group_min */
2102 unsigned long min_load_per_task; /* load_per_task in group_min */
2103 unsigned long leader_nr_running; /* Nr running of group_leader */
2104 unsigned long min_nr_running; /* Nr running of group_min */
2109 * sg_lb_stats - stats of a sched_group required for load_balancing
2111 struct sg_lb_stats {
2112 unsigned long avg_load; /*Avg load across the CPUs of the group */
2113 unsigned long group_load; /* Total load over the CPUs of the group */
2114 unsigned long sum_nr_running; /* Nr tasks running in the group */
2115 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2116 unsigned long group_capacity;
2117 int group_imb; /* Is there an imbalance in the group ? */
2118 int group_has_capacity; /* Is there extra capacity in the group? */
2122 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2123 * @group: The group whose first cpu is to be returned.
2125 static inline unsigned int group_first_cpu(struct sched_group *group)
2127 return cpumask_first(sched_group_cpus(group));
2131 * get_sd_load_idx - Obtain the load index for a given sched domain.
2132 * @sd: The sched_domain whose load_idx is to be obtained.
2133 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2135 static inline int get_sd_load_idx(struct sched_domain *sd,
2136 enum cpu_idle_type idle)
2142 load_idx = sd->busy_idx;
2145 case CPU_NEWLY_IDLE:
2146 load_idx = sd->newidle_idx;
2149 load_idx = sd->idle_idx;
2157 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2159 * init_sd_power_savings_stats - Initialize power savings statistics for
2160 * the given sched_domain, during load balancing.
2162 * @sd: Sched domain whose power-savings statistics are to be initialized.
2163 * @sds: Variable containing the statistics for sd.
2164 * @idle: Idle status of the CPU at which we're performing load-balancing.
2166 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2167 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2170 * Busy processors will not participate in power savings
2173 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2174 sds->power_savings_balance = 0;
2176 sds->power_savings_balance = 1;
2177 sds->min_nr_running = ULONG_MAX;
2178 sds->leader_nr_running = 0;
2183 * update_sd_power_savings_stats - Update the power saving stats for a
2184 * sched_domain while performing load balancing.
2186 * @group: sched_group belonging to the sched_domain under consideration.
2187 * @sds: Variable containing the statistics of the sched_domain
2188 * @local_group: Does group contain the CPU for which we're performing
2190 * @sgs: Variable containing the statistics of the group.
2192 static inline void update_sd_power_savings_stats(struct sched_group *group,
2193 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2196 if (!sds->power_savings_balance)
2200 * If the local group is idle or completely loaded
2201 * no need to do power savings balance at this domain
2203 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2204 !sds->this_nr_running))
2205 sds->power_savings_balance = 0;
2208 * If a group is already running at full capacity or idle,
2209 * don't include that group in power savings calculations
2211 if (!sds->power_savings_balance ||
2212 sgs->sum_nr_running >= sgs->group_capacity ||
2213 !sgs->sum_nr_running)
2217 * Calculate the group which has the least non-idle load.
2218 * This is the group from where we need to pick up the load
2221 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2222 (sgs->sum_nr_running == sds->min_nr_running &&
2223 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2224 sds->group_min = group;
2225 sds->min_nr_running = sgs->sum_nr_running;
2226 sds->min_load_per_task = sgs->sum_weighted_load /
2227 sgs->sum_nr_running;
2231 * Calculate the group which is almost near its
2232 * capacity but still has some space to pick up some load
2233 * from other group and save more power
2235 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2238 if (sgs->sum_nr_running > sds->leader_nr_running ||
2239 (sgs->sum_nr_running == sds->leader_nr_running &&
2240 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2241 sds->group_leader = group;
2242 sds->leader_nr_running = sgs->sum_nr_running;
2247 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2248 * @sds: Variable containing the statistics of the sched_domain
2249 * under consideration.
2250 * @this_cpu: Cpu at which we're currently performing load-balancing.
2251 * @imbalance: Variable to store the imbalance.
2254 * Check if we have potential to perform some power-savings balance.
2255 * If yes, set the busiest group to be the least loaded group in the
2256 * sched_domain, so that it's CPUs can be put to idle.
2258 * Returns 1 if there is potential to perform power-savings balance.
2261 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2262 int this_cpu, unsigned long *imbalance)
2264 if (!sds->power_savings_balance)
2267 if (sds->this != sds->group_leader ||
2268 sds->group_leader == sds->group_min)
2271 *imbalance = sds->min_load_per_task;
2272 sds->busiest = sds->group_min;
2277 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2278 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2279 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2284 static inline void update_sd_power_savings_stats(struct sched_group *group,
2285 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2290 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2291 int this_cpu, unsigned long *imbalance)
2295 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2298 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2300 return SCHED_LOAD_SCALE;
2303 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2305 return default_scale_freq_power(sd, cpu);
2308 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2310 unsigned long weight = sd->span_weight;
2311 unsigned long smt_gain = sd->smt_gain;
2318 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2320 return default_scale_smt_power(sd, cpu);
2323 unsigned long scale_rt_power(int cpu)
2325 struct rq *rq = cpu_rq(cpu);
2326 u64 total, available;
2328 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2330 if (unlikely(total < rq->rt_avg)) {
2331 /* Ensures that power won't end up being negative */
2334 available = total - rq->rt_avg;
2337 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2338 total = SCHED_LOAD_SCALE;
2340 total >>= SCHED_LOAD_SHIFT;
2342 return div_u64(available, total);
2345 static void update_cpu_power(struct sched_domain *sd, int cpu)
2347 unsigned long weight = sd->span_weight;
2348 unsigned long power = SCHED_LOAD_SCALE;
2349 struct sched_group *sdg = sd->groups;
2351 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2352 if (sched_feat(ARCH_POWER))
2353 power *= arch_scale_smt_power(sd, cpu);
2355 power *= default_scale_smt_power(sd, cpu);
2357 power >>= SCHED_LOAD_SHIFT;
2360 sdg->cpu_power_orig = power;
2362 if (sched_feat(ARCH_POWER))
2363 power *= arch_scale_freq_power(sd, cpu);
2365 power *= default_scale_freq_power(sd, cpu);
2367 power >>= SCHED_LOAD_SHIFT;
2369 power *= scale_rt_power(cpu);
2370 power >>= SCHED_LOAD_SHIFT;
2375 cpu_rq(cpu)->cpu_power = power;
2376 sdg->cpu_power = power;
2379 static void update_group_power(struct sched_domain *sd, int cpu)
2381 struct sched_domain *child = sd->child;
2382 struct sched_group *group, *sdg = sd->groups;
2383 unsigned long power;
2386 update_cpu_power(sd, cpu);
2392 group = child->groups;
2394 power += group->cpu_power;
2395 group = group->next;
2396 } while (group != child->groups);
2398 sdg->cpu_power = power;
2402 * Try and fix up capacity for tiny siblings, this is needed when
2403 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2404 * which on its own isn't powerful enough.
2406 * See update_sd_pick_busiest() and check_asym_packing().
2409 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2412 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2414 if (sd->level != SD_LV_SIBLING)
2418 * If ~90% of the cpu_power is still there, we're good.
2420 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2427 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2428 * @sd: The sched_domain whose statistics are to be updated.
2429 * @group: sched_group whose statistics are to be updated.
2430 * @this_cpu: Cpu for which load balance is currently performed.
2431 * @idle: Idle status of this_cpu
2432 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2433 * @sd_idle: Idle status of the sched_domain containing group.
2434 * @local_group: Does group contain this_cpu.
2435 * @cpus: Set of cpus considered for load balancing.
2436 * @balance: Should we balance.
2437 * @sgs: variable to hold the statistics for this group.
2439 static inline void update_sg_lb_stats(struct sched_domain *sd,
2440 struct sched_group *group, int this_cpu,
2441 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2442 int local_group, const struct cpumask *cpus,
2443 int *balance, struct sg_lb_stats *sgs)
2445 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2447 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2448 unsigned long avg_load_per_task = 0;
2451 balance_cpu = group_first_cpu(group);
2453 /* Tally up the load of all CPUs in the group */
2455 min_cpu_load = ~0UL;
2458 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2459 struct rq *rq = cpu_rq(i);
2461 if (*sd_idle && rq->nr_running)
2464 /* Bias balancing toward cpus of our domain */
2466 if (idle_cpu(i) && !first_idle_cpu) {
2471 load = target_load(i, load_idx);
2473 load = source_load(i, load_idx);
2474 if (load > max_cpu_load) {
2475 max_cpu_load = load;
2476 max_nr_running = rq->nr_running;
2478 if (min_cpu_load > load)
2479 min_cpu_load = load;
2482 sgs->group_load += load;
2483 sgs->sum_nr_running += rq->nr_running;
2484 sgs->sum_weighted_load += weighted_cpuload(i);
2489 * First idle cpu or the first cpu(busiest) in this sched group
2490 * is eligible for doing load balancing at this and above
2491 * domains. In the newly idle case, we will allow all the cpu's
2492 * to do the newly idle load balance.
2494 if (idle != CPU_NEWLY_IDLE && local_group) {
2495 if (balance_cpu != this_cpu) {
2499 update_group_power(sd, this_cpu);
2502 /* Adjust by relative CPU power of the group */
2503 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2506 * Consider the group unbalanced when the imbalance is larger
2507 * than the average weight of two tasks.
2509 * APZ: with cgroup the avg task weight can vary wildly and
2510 * might not be a suitable number - should we keep a
2511 * normalized nr_running number somewhere that negates
2514 if (sgs->sum_nr_running)
2515 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2517 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2520 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2521 if (!sgs->group_capacity)
2522 sgs->group_capacity = fix_small_capacity(sd, group);
2524 if (sgs->group_capacity > sgs->sum_nr_running)
2525 sgs->group_has_capacity = 1;
2529 * update_sd_pick_busiest - return 1 on busiest group
2530 * @sd: sched_domain whose statistics are to be checked
2531 * @sds: sched_domain statistics
2532 * @sg: sched_group candidate to be checked for being the busiest
2533 * @sgs: sched_group statistics
2534 * @this_cpu: the current cpu
2536 * Determine if @sg is a busier group than the previously selected
2539 static bool update_sd_pick_busiest(struct sched_domain *sd,
2540 struct sd_lb_stats *sds,
2541 struct sched_group *sg,
2542 struct sg_lb_stats *sgs,
2545 if (sgs->avg_load <= sds->max_load)
2548 if (sgs->sum_nr_running > sgs->group_capacity)
2555 * ASYM_PACKING needs to move all the work to the lowest
2556 * numbered CPUs in the group, therefore mark all groups
2557 * higher than ourself as busy.
2559 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2560 this_cpu < group_first_cpu(sg)) {
2564 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2572 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2573 * @sd: sched_domain whose statistics are to be updated.
2574 * @this_cpu: Cpu for which load balance is currently performed.
2575 * @idle: Idle status of this_cpu
2576 * @sd_idle: Idle status of the sched_domain containing sg.
2577 * @cpus: Set of cpus considered for load balancing.
2578 * @balance: Should we balance.
2579 * @sds: variable to hold the statistics for this sched_domain.
2581 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2582 enum cpu_idle_type idle, int *sd_idle,
2583 const struct cpumask *cpus, int *balance,
2584 struct sd_lb_stats *sds)
2586 struct sched_domain *child = sd->child;
2587 struct sched_group *sg = sd->groups;
2588 struct sg_lb_stats sgs;
2589 int load_idx, prefer_sibling = 0;
2591 if (child && child->flags & SD_PREFER_SIBLING)
2594 init_sd_power_savings_stats(sd, sds, idle);
2595 load_idx = get_sd_load_idx(sd, idle);
2600 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2601 memset(&sgs, 0, sizeof(sgs));
2602 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2603 local_group, cpus, balance, &sgs);
2605 if (local_group && !(*balance))
2608 sds->total_load += sgs.group_load;
2609 sds->total_pwr += sg->cpu_power;
2612 * In case the child domain prefers tasks go to siblings
2613 * first, lower the sg capacity to one so that we'll try
2614 * and move all the excess tasks away. We lower the capacity
2615 * of a group only if the local group has the capacity to fit
2616 * these excess tasks, i.e. nr_running < group_capacity. The
2617 * extra check prevents the case where you always pull from the
2618 * heaviest group when it is already under-utilized (possible
2619 * with a large weight task outweighs the tasks on the system).
2621 if (prefer_sibling && !local_group && sds->this_has_capacity)
2622 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2625 sds->this_load = sgs.avg_load;
2627 sds->this_nr_running = sgs.sum_nr_running;
2628 sds->this_load_per_task = sgs.sum_weighted_load;
2629 sds->this_has_capacity = sgs.group_has_capacity;
2630 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2631 sds->max_load = sgs.avg_load;
2633 sds->busiest_nr_running = sgs.sum_nr_running;
2634 sds->busiest_group_capacity = sgs.group_capacity;
2635 sds->busiest_load_per_task = sgs.sum_weighted_load;
2636 sds->busiest_has_capacity = sgs.group_has_capacity;
2637 sds->group_imb = sgs.group_imb;
2640 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2642 } while (sg != sd->groups);
2645 int __weak arch_sd_sibling_asym_packing(void)
2647 return 0*SD_ASYM_PACKING;
2651 * check_asym_packing - Check to see if the group is packed into the
2654 * This is primarily intended to used at the sibling level. Some
2655 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2656 * case of POWER7, it can move to lower SMT modes only when higher
2657 * threads are idle. When in lower SMT modes, the threads will
2658 * perform better since they share less core resources. Hence when we
2659 * have idle threads, we want them to be the higher ones.
2661 * This packing function is run on idle threads. It checks to see if
2662 * the busiest CPU in this domain (core in the P7 case) has a higher
2663 * CPU number than the packing function is being run on. Here we are
2664 * assuming lower CPU number will be equivalent to lower a SMT thread
2667 * Returns 1 when packing is required and a task should be moved to
2668 * this CPU. The amount of the imbalance is returned in *imbalance.
2670 * @sd: The sched_domain whose packing is to be checked.
2671 * @sds: Statistics of the sched_domain which is to be packed
2672 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2673 * @imbalance: returns amount of imbalanced due to packing.
2675 static int check_asym_packing(struct sched_domain *sd,
2676 struct sd_lb_stats *sds,
2677 int this_cpu, unsigned long *imbalance)
2681 if (!(sd->flags & SD_ASYM_PACKING))
2687 busiest_cpu = group_first_cpu(sds->busiest);
2688 if (this_cpu > busiest_cpu)
2691 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2697 * fix_small_imbalance - Calculate the minor imbalance that exists
2698 * amongst the groups of a sched_domain, during
2700 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2701 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2702 * @imbalance: Variable to store the imbalance.
2704 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2705 int this_cpu, unsigned long *imbalance)
2707 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2708 unsigned int imbn = 2;
2709 unsigned long scaled_busy_load_per_task;
2711 if (sds->this_nr_running) {
2712 sds->this_load_per_task /= sds->this_nr_running;
2713 if (sds->busiest_load_per_task >
2714 sds->this_load_per_task)
2717 sds->this_load_per_task =
2718 cpu_avg_load_per_task(this_cpu);
2720 scaled_busy_load_per_task = sds->busiest_load_per_task
2722 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2724 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2725 (scaled_busy_load_per_task * imbn)) {
2726 *imbalance = sds->busiest_load_per_task;
2731 * OK, we don't have enough imbalance to justify moving tasks,
2732 * however we may be able to increase total CPU power used by
2736 pwr_now += sds->busiest->cpu_power *
2737 min(sds->busiest_load_per_task, sds->max_load);
2738 pwr_now += sds->this->cpu_power *
2739 min(sds->this_load_per_task, sds->this_load);
2740 pwr_now /= SCHED_LOAD_SCALE;
2742 /* Amount of load we'd subtract */
2743 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2744 sds->busiest->cpu_power;
2745 if (sds->max_load > tmp)
2746 pwr_move += sds->busiest->cpu_power *
2747 min(sds->busiest_load_per_task, sds->max_load - tmp);
2749 /* Amount of load we'd add */
2750 if (sds->max_load * sds->busiest->cpu_power <
2751 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2752 tmp = (sds->max_load * sds->busiest->cpu_power) /
2753 sds->this->cpu_power;
2755 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2756 sds->this->cpu_power;
2757 pwr_move += sds->this->cpu_power *
2758 min(sds->this_load_per_task, sds->this_load + tmp);
2759 pwr_move /= SCHED_LOAD_SCALE;
2761 /* Move if we gain throughput */
2762 if (pwr_move > pwr_now)
2763 *imbalance = sds->busiest_load_per_task;
2767 * calculate_imbalance - Calculate the amount of imbalance present within the
2768 * groups of a given sched_domain during load balance.
2769 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2770 * @this_cpu: Cpu for which currently load balance is being performed.
2771 * @imbalance: The variable to store the imbalance.
2773 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2774 unsigned long *imbalance)
2776 unsigned long max_pull, load_above_capacity = ~0UL;
2778 sds->busiest_load_per_task /= sds->busiest_nr_running;
2779 if (sds->group_imb) {
2780 sds->busiest_load_per_task =
2781 min(sds->busiest_load_per_task, sds->avg_load);
2785 * In the presence of smp nice balancing, certain scenarios can have
2786 * max load less than avg load(as we skip the groups at or below
2787 * its cpu_power, while calculating max_load..)
2789 if (sds->max_load < sds->avg_load) {
2791 return fix_small_imbalance(sds, this_cpu, imbalance);
2794 if (!sds->group_imb) {
2796 * Don't want to pull so many tasks that a group would go idle.
2798 load_above_capacity = (sds->busiest_nr_running -
2799 sds->busiest_group_capacity);
2801 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2803 load_above_capacity /= sds->busiest->cpu_power;
2807 * We're trying to get all the cpus to the average_load, so we don't
2808 * want to push ourselves above the average load, nor do we wish to
2809 * reduce the max loaded cpu below the average load. At the same time,
2810 * we also don't want to reduce the group load below the group capacity
2811 * (so that we can implement power-savings policies etc). Thus we look
2812 * for the minimum possible imbalance.
2813 * Be careful of negative numbers as they'll appear as very large values
2814 * with unsigned longs.
2816 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2818 /* How much load to actually move to equalise the imbalance */
2819 *imbalance = min(max_pull * sds->busiest->cpu_power,
2820 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
2824 * if *imbalance is less than the average load per runnable task
2825 * there is no gaurantee that any tasks will be moved so we'll have
2826 * a think about bumping its value to force at least one task to be
2829 if (*imbalance < sds->busiest_load_per_task)
2830 return fix_small_imbalance(sds, this_cpu, imbalance);
2834 /******* find_busiest_group() helpers end here *********************/
2837 * find_busiest_group - Returns the busiest group within the sched_domain
2838 * if there is an imbalance. If there isn't an imbalance, and
2839 * the user has opted for power-savings, it returns a group whose
2840 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2841 * such a group exists.
2843 * Also calculates the amount of weighted load which should be moved
2844 * to restore balance.
2846 * @sd: The sched_domain whose busiest group is to be returned.
2847 * @this_cpu: The cpu for which load balancing is currently being performed.
2848 * @imbalance: Variable which stores amount of weighted load which should
2849 * be moved to restore balance/put a group to idle.
2850 * @idle: The idle status of this_cpu.
2851 * @sd_idle: The idleness of sd
2852 * @cpus: The set of CPUs under consideration for load-balancing.
2853 * @balance: Pointer to a variable indicating if this_cpu
2854 * is the appropriate cpu to perform load balancing at this_level.
2856 * Returns: - the busiest group if imbalance exists.
2857 * - If no imbalance and user has opted for power-savings balance,
2858 * return the least loaded group whose CPUs can be
2859 * put to idle by rebalancing its tasks onto our group.
2861 static struct sched_group *
2862 find_busiest_group(struct sched_domain *sd, int this_cpu,
2863 unsigned long *imbalance, enum cpu_idle_type idle,
2864 int *sd_idle, const struct cpumask *cpus, int *balance)
2866 struct sd_lb_stats sds;
2868 memset(&sds, 0, sizeof(sds));
2871 * Compute the various statistics relavent for load balancing at
2874 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
2877 /* Cases where imbalance does not exist from POV of this_cpu */
2878 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2880 * 2) There is no busy sibling group to pull from.
2881 * 3) This group is the busiest group.
2882 * 4) This group is more busy than the avg busieness at this
2884 * 5) The imbalance is within the specified limit.
2886 * Note: when doing newidle balance, if the local group has excess
2887 * capacity (i.e. nr_running < group_capacity) and the busiest group
2888 * does not have any capacity, we force a load balance to pull tasks
2889 * to the local group. In this case, we skip past checks 3, 4 and 5.
2894 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
2895 check_asym_packing(sd, &sds, this_cpu, imbalance))
2898 if (!sds.busiest || sds.busiest_nr_running == 0)
2901 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
2902 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
2903 !sds.busiest_has_capacity)
2906 if (sds.this_load >= sds.max_load)
2909 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
2911 if (sds.this_load >= sds.avg_load)
2914 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
2918 /* Looks like there is an imbalance. Compute it */
2919 calculate_imbalance(&sds, this_cpu, imbalance);
2924 * There is no obvious imbalance. But check if we can do some balancing
2927 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
2935 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2938 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
2939 enum cpu_idle_type idle, unsigned long imbalance,
2940 const struct cpumask *cpus)
2942 struct rq *busiest = NULL, *rq;
2943 unsigned long max_load = 0;
2946 for_each_cpu(i, sched_group_cpus(group)) {
2947 unsigned long power = power_of(i);
2948 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
2952 capacity = fix_small_capacity(sd, group);
2954 if (!cpumask_test_cpu(i, cpus))
2958 wl = weighted_cpuload(i);
2961 * When comparing with imbalance, use weighted_cpuload()
2962 * which is not scaled with the cpu power.
2964 if (capacity && rq->nr_running == 1 && wl > imbalance)
2968 * For the load comparisons with the other cpu's, consider
2969 * the weighted_cpuload() scaled with the cpu power, so that
2970 * the load can be moved away from the cpu that is potentially
2971 * running at a lower capacity.
2973 wl = (wl * SCHED_LOAD_SCALE) / power;
2975 if (wl > max_load) {
2985 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2986 * so long as it is large enough.
2988 #define MAX_PINNED_INTERVAL 512
2990 /* Working cpumask for load_balance and load_balance_newidle. */
2991 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
2993 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
2994 int busiest_cpu, int this_cpu)
2996 if (idle == CPU_NEWLY_IDLE) {
2999 * ASYM_PACKING needs to force migrate tasks from busy but
3000 * higher numbered CPUs in order to pack all tasks in the
3001 * lowest numbered CPUs.
3003 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3007 * The only task running in a non-idle cpu can be moved to this
3008 * cpu in an attempt to completely freeup the other CPU
3011 * The package power saving logic comes from
3012 * find_busiest_group(). If there are no imbalance, then
3013 * f_b_g() will return NULL. However when sched_mc={1,2} then
3014 * f_b_g() will select a group from which a running task may be
3015 * pulled to this cpu in order to make the other package idle.
3016 * If there is no opportunity to make a package idle and if
3017 * there are no imbalance, then f_b_g() will return NULL and no
3018 * action will be taken in load_balance_newidle().
3020 * Under normal task pull operation due to imbalance, there
3021 * will be more than one task in the source run queue and
3022 * move_tasks() will succeed. ld_moved will be true and this
3023 * active balance code will not be triggered.
3025 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3026 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3029 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3033 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3036 static int active_load_balance_cpu_stop(void *data);
3039 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3040 * tasks if there is an imbalance.
3042 static int load_balance(int this_cpu, struct rq *this_rq,
3043 struct sched_domain *sd, enum cpu_idle_type idle,
3046 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3047 struct sched_group *group;
3048 unsigned long imbalance;
3050 unsigned long flags;
3051 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3053 cpumask_copy(cpus, cpu_active_mask);
3056 * When power savings policy is enabled for the parent domain, idle
3057 * sibling can pick up load irrespective of busy siblings. In this case,
3058 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3059 * portraying it as CPU_NOT_IDLE.
3061 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3062 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3065 schedstat_inc(sd, lb_count[idle]);
3068 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3075 schedstat_inc(sd, lb_nobusyg[idle]);
3079 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3081 schedstat_inc(sd, lb_nobusyq[idle]);
3085 BUG_ON(busiest == this_rq);
3087 schedstat_add(sd, lb_imbalance[idle], imbalance);
3090 if (busiest->nr_running > 1) {
3092 * Attempt to move tasks. If find_busiest_group has found
3093 * an imbalance but busiest->nr_running <= 1, the group is
3094 * still unbalanced. ld_moved simply stays zero, so it is
3095 * correctly treated as an imbalance.
3097 local_irq_save(flags);
3098 double_rq_lock(this_rq, busiest);
3099 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3100 imbalance, sd, idle, &all_pinned);
3101 double_rq_unlock(this_rq, busiest);
3102 local_irq_restore(flags);
3105 * some other cpu did the load balance for us.
3107 if (ld_moved && this_cpu != smp_processor_id())
3108 resched_cpu(this_cpu);
3110 /* All tasks on this runqueue were pinned by CPU affinity */
3111 if (unlikely(all_pinned)) {
3112 cpumask_clear_cpu(cpu_of(busiest), cpus);
3113 if (!cpumask_empty(cpus))
3120 schedstat_inc(sd, lb_failed[idle]);
3122 * Increment the failure counter only on periodic balance.
3123 * We do not want newidle balance, which can be very
3124 * frequent, pollute the failure counter causing
3125 * excessive cache_hot migrations and active balances.
3127 if (idle != CPU_NEWLY_IDLE)
3128 sd->nr_balance_failed++;
3130 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3132 raw_spin_lock_irqsave(&busiest->lock, flags);
3134 /* don't kick the active_load_balance_cpu_stop,
3135 * if the curr task on busiest cpu can't be
3138 if (!cpumask_test_cpu(this_cpu,
3139 &busiest->curr->cpus_allowed)) {
3140 raw_spin_unlock_irqrestore(&busiest->lock,
3143 goto out_one_pinned;
3147 * ->active_balance synchronizes accesses to
3148 * ->active_balance_work. Once set, it's cleared
3149 * only after active load balance is finished.
3151 if (!busiest->active_balance) {
3152 busiest->active_balance = 1;
3153 busiest->push_cpu = this_cpu;
3156 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3159 stop_one_cpu_nowait(cpu_of(busiest),
3160 active_load_balance_cpu_stop, busiest,
3161 &busiest->active_balance_work);
3164 * We've kicked active balancing, reset the failure
3167 sd->nr_balance_failed = sd->cache_nice_tries+1;
3170 sd->nr_balance_failed = 0;
3172 if (likely(!active_balance)) {
3173 /* We were unbalanced, so reset the balancing interval */
3174 sd->balance_interval = sd->min_interval;
3177 * If we've begun active balancing, start to back off. This
3178 * case may not be covered by the all_pinned logic if there
3179 * is only 1 task on the busy runqueue (because we don't call
3182 if (sd->balance_interval < sd->max_interval)
3183 sd->balance_interval *= 2;
3186 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3187 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3193 schedstat_inc(sd, lb_balanced[idle]);
3195 sd->nr_balance_failed = 0;
3198 /* tune up the balancing interval */
3199 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3200 (sd->balance_interval < sd->max_interval))
3201 sd->balance_interval *= 2;
3203 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3204 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3213 * idle_balance is called by schedule() if this_cpu is about to become
3214 * idle. Attempts to pull tasks from other CPUs.
3216 static void idle_balance(int this_cpu, struct rq *this_rq)
3218 struct sched_domain *sd;
3219 int pulled_task = 0;
3220 unsigned long next_balance = jiffies + HZ;
3222 this_rq->idle_stamp = this_rq->clock;
3224 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3228 * Drop the rq->lock, but keep IRQ/preempt disabled.
3230 raw_spin_unlock(&this_rq->lock);
3232 for_each_domain(this_cpu, sd) {
3233 unsigned long interval;
3236 if (!(sd->flags & SD_LOAD_BALANCE))
3239 if (sd->flags & SD_BALANCE_NEWIDLE) {
3240 /* If we've pulled tasks over stop searching: */
3241 pulled_task = load_balance(this_cpu, this_rq,
3242 sd, CPU_NEWLY_IDLE, &balance);
3245 interval = msecs_to_jiffies(sd->balance_interval);
3246 if (time_after(next_balance, sd->last_balance + interval))
3247 next_balance = sd->last_balance + interval;
3252 raw_spin_lock(&this_rq->lock);
3254 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3256 * We are going idle. next_balance may be set based on
3257 * a busy processor. So reset next_balance.
3259 this_rq->next_balance = next_balance;
3264 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3265 * running tasks off the busiest CPU onto idle CPUs. It requires at
3266 * least 1 task to be running on each physical CPU where possible, and
3267 * avoids physical / logical imbalances.
3269 static int active_load_balance_cpu_stop(void *data)
3271 struct rq *busiest_rq = data;
3272 int busiest_cpu = cpu_of(busiest_rq);
3273 int target_cpu = busiest_rq->push_cpu;
3274 struct rq *target_rq = cpu_rq(target_cpu);
3275 struct sched_domain *sd;
3277 raw_spin_lock_irq(&busiest_rq->lock);
3279 /* make sure the requested cpu hasn't gone down in the meantime */
3280 if (unlikely(busiest_cpu != smp_processor_id() ||
3281 !busiest_rq->active_balance))
3284 /* Is there any task to move? */
3285 if (busiest_rq->nr_running <= 1)
3289 * This condition is "impossible", if it occurs
3290 * we need to fix it. Originally reported by
3291 * Bjorn Helgaas on a 128-cpu setup.
3293 BUG_ON(busiest_rq == target_rq);
3295 /* move a task from busiest_rq to target_rq */
3296 double_lock_balance(busiest_rq, target_rq);
3298 /* Search for an sd spanning us and the target CPU. */
3299 for_each_domain(target_cpu, sd) {
3300 if ((sd->flags & SD_LOAD_BALANCE) &&
3301 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3306 schedstat_inc(sd, alb_count);
3308 if (move_one_task(target_rq, target_cpu, busiest_rq,
3310 schedstat_inc(sd, alb_pushed);
3312 schedstat_inc(sd, alb_failed);
3314 double_unlock_balance(busiest_rq, target_rq);
3316 busiest_rq->active_balance = 0;
3317 raw_spin_unlock_irq(&busiest_rq->lock);
3323 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3325 static void trigger_sched_softirq(void *data)
3327 raise_softirq_irqoff(SCHED_SOFTIRQ);
3330 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3332 csd->func = trigger_sched_softirq;
3339 * idle load balancing details
3340 * - One of the idle CPUs nominates itself as idle load_balancer, while
3342 * - This idle load balancer CPU will also go into tickless mode when
3343 * it is idle, just like all other idle CPUs
3344 * - When one of the busy CPUs notice that there may be an idle rebalancing
3345 * needed, they will kick the idle load balancer, which then does idle
3346 * load balancing for all the idle CPUs.
3349 atomic_t load_balancer;
3350 atomic_t first_pick_cpu;
3351 atomic_t second_pick_cpu;
3352 cpumask_var_t idle_cpus_mask;
3353 cpumask_var_t grp_idle_mask;
3354 unsigned long next_balance; /* in jiffy units */
3355 } nohz ____cacheline_aligned;
3357 int get_nohz_load_balancer(void)
3359 return atomic_read(&nohz.load_balancer);
3362 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3364 * lowest_flag_domain - Return lowest sched_domain containing flag.
3365 * @cpu: The cpu whose lowest level of sched domain is to
3367 * @flag: The flag to check for the lowest sched_domain
3368 * for the given cpu.
3370 * Returns the lowest sched_domain of a cpu which contains the given flag.
3372 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3374 struct sched_domain *sd;
3376 for_each_domain(cpu, sd)
3377 if (sd && (sd->flags & flag))
3384 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3385 * @cpu: The cpu whose domains we're iterating over.
3386 * @sd: variable holding the value of the power_savings_sd
3388 * @flag: The flag to filter the sched_domains to be iterated.
3390 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3391 * set, starting from the lowest sched_domain to the highest.
3393 #define for_each_flag_domain(cpu, sd, flag) \
3394 for (sd = lowest_flag_domain(cpu, flag); \
3395 (sd && (sd->flags & flag)); sd = sd->parent)
3398 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3399 * @ilb_group: group to be checked for semi-idleness
3401 * Returns: 1 if the group is semi-idle. 0 otherwise.
3403 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3404 * and atleast one non-idle CPU. This helper function checks if the given
3405 * sched_group is semi-idle or not.
3407 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3409 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3410 sched_group_cpus(ilb_group));
3413 * A sched_group is semi-idle when it has atleast one busy cpu
3414 * and atleast one idle cpu.
3416 if (cpumask_empty(nohz.grp_idle_mask))
3419 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3425 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3426 * @cpu: The cpu which is nominating a new idle_load_balancer.
3428 * Returns: Returns the id of the idle load balancer if it exists,
3429 * Else, returns >= nr_cpu_ids.
3431 * This algorithm picks the idle load balancer such that it belongs to a
3432 * semi-idle powersavings sched_domain. The idea is to try and avoid
3433 * completely idle packages/cores just for the purpose of idle load balancing
3434 * when there are other idle cpu's which are better suited for that job.
3436 static int find_new_ilb(int cpu)
3438 struct sched_domain *sd;
3439 struct sched_group *ilb_group;
3442 * Have idle load balancer selection from semi-idle packages only
3443 * when power-aware load balancing is enabled
3445 if (!(sched_smt_power_savings || sched_mc_power_savings))
3449 * Optimize for the case when we have no idle CPUs or only one
3450 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3452 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3455 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3456 ilb_group = sd->groups;
3459 if (is_semi_idle_group(ilb_group))
3460 return cpumask_first(nohz.grp_idle_mask);
3462 ilb_group = ilb_group->next;
3464 } while (ilb_group != sd->groups);
3470 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3471 static inline int find_new_ilb(int call_cpu)
3478 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3479 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3480 * CPU (if there is one).
3482 static void nohz_balancer_kick(int cpu)
3486 nohz.next_balance++;
3488 ilb_cpu = get_nohz_load_balancer();
3490 if (ilb_cpu >= nr_cpu_ids) {
3491 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3492 if (ilb_cpu >= nr_cpu_ids)
3496 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3497 struct call_single_data *cp;
3499 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3500 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3501 __smp_call_function_single(ilb_cpu, cp, 0);
3507 * This routine will try to nominate the ilb (idle load balancing)
3508 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3509 * load balancing on behalf of all those cpus.
3511 * When the ilb owner becomes busy, we will not have new ilb owner until some
3512 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3513 * idle load balancing by kicking one of the idle CPUs.
3515 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3516 * ilb owner CPU in future (when there is a need for idle load balancing on
3517 * behalf of all idle CPUs).
3519 void select_nohz_load_balancer(int stop_tick)
3521 int cpu = smp_processor_id();
3524 if (!cpu_active(cpu)) {
3525 if (atomic_read(&nohz.load_balancer) != cpu)
3529 * If we are going offline and still the leader,
3532 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3539 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3541 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3542 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3543 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3544 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3546 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3549 /* make me the ilb owner */
3550 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3555 * Check to see if there is a more power-efficient
3558 new_ilb = find_new_ilb(cpu);
3559 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3560 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3561 resched_cpu(new_ilb);
3567 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3570 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3572 if (atomic_read(&nohz.load_balancer) == cpu)
3573 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3581 static DEFINE_SPINLOCK(balancing);
3584 * It checks each scheduling domain to see if it is due to be balanced,
3585 * and initiates a balancing operation if so.
3587 * Balancing parameters are set up in arch_init_sched_domains.
3589 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3592 struct rq *rq = cpu_rq(cpu);
3593 unsigned long interval;
3594 struct sched_domain *sd;
3595 /* Earliest time when we have to do rebalance again */
3596 unsigned long next_balance = jiffies + 60*HZ;
3597 int update_next_balance = 0;
3602 for_each_domain(cpu, sd) {
3603 if (!(sd->flags & SD_LOAD_BALANCE))
3606 interval = sd->balance_interval;
3607 if (idle != CPU_IDLE)
3608 interval *= sd->busy_factor;
3610 /* scale ms to jiffies */
3611 interval = msecs_to_jiffies(interval);
3612 if (unlikely(!interval))
3614 if (interval > HZ*NR_CPUS/10)
3615 interval = HZ*NR_CPUS/10;
3617 need_serialize = sd->flags & SD_SERIALIZE;
3619 if (need_serialize) {
3620 if (!spin_trylock(&balancing))
3624 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3625 if (load_balance(cpu, rq, sd, idle, &balance)) {
3627 * We've pulled tasks over so either we're no
3628 * longer idle, or one of our SMT siblings is
3631 idle = CPU_NOT_IDLE;
3633 sd->last_balance = jiffies;
3636 spin_unlock(&balancing);
3638 if (time_after(next_balance, sd->last_balance + interval)) {
3639 next_balance = sd->last_balance + interval;
3640 update_next_balance = 1;
3644 * Stop the load balance at this level. There is another
3645 * CPU in our sched group which is doing load balancing more
3653 * next_balance will be updated only when there is a need.
3654 * When the cpu is attached to null domain for ex, it will not be
3657 if (likely(update_next_balance))
3658 rq->next_balance = next_balance;
3663 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3664 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3666 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3668 struct rq *this_rq = cpu_rq(this_cpu);
3672 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3675 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3676 if (balance_cpu == this_cpu)
3680 * If this cpu gets work to do, stop the load balancing
3681 * work being done for other cpus. Next load
3682 * balancing owner will pick it up.
3684 if (need_resched()) {
3685 this_rq->nohz_balance_kick = 0;
3689 raw_spin_lock_irq(&this_rq->lock);
3690 update_rq_clock(this_rq);
3691 update_cpu_load(this_rq);
3692 raw_spin_unlock_irq(&this_rq->lock);
3694 rebalance_domains(balance_cpu, CPU_IDLE);
3696 rq = cpu_rq(balance_cpu);
3697 if (time_after(this_rq->next_balance, rq->next_balance))
3698 this_rq->next_balance = rq->next_balance;
3700 nohz.next_balance = this_rq->next_balance;
3701 this_rq->nohz_balance_kick = 0;
3705 * Current heuristic for kicking the idle load balancer
3706 * - first_pick_cpu is the one of the busy CPUs. It will kick
3707 * idle load balancer when it has more than one process active. This
3708 * eliminates the need for idle load balancing altogether when we have
3709 * only one running process in the system (common case).
3710 * - If there are more than one busy CPU, idle load balancer may have
3711 * to run for active_load_balance to happen (i.e., two busy CPUs are
3712 * SMT or core siblings and can run better if they move to different
3713 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3714 * which will kick idle load balancer as soon as it has any load.
3716 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3718 unsigned long now = jiffies;
3720 int first_pick_cpu, second_pick_cpu;
3722 if (time_before(now, nohz.next_balance))
3725 if (rq->idle_at_tick)
3728 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3729 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3731 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3732 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3735 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3736 if (ret == nr_cpu_ids || ret == cpu) {
3737 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3738 if (rq->nr_running > 1)
3741 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3742 if (ret == nr_cpu_ids || ret == cpu) {
3750 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3754 * run_rebalance_domains is triggered when needed from the scheduler tick.
3755 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3757 static void run_rebalance_domains(struct softirq_action *h)
3759 int this_cpu = smp_processor_id();
3760 struct rq *this_rq = cpu_rq(this_cpu);
3761 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3762 CPU_IDLE : CPU_NOT_IDLE;
3764 rebalance_domains(this_cpu, idle);
3767 * If this cpu has a pending nohz_balance_kick, then do the
3768 * balancing on behalf of the other idle cpus whose ticks are
3771 nohz_idle_balance(this_cpu, idle);
3774 static inline int on_null_domain(int cpu)
3776 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3780 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3782 static inline void trigger_load_balance(struct rq *rq, int cpu)
3784 /* Don't need to rebalance while attached to NULL domain */
3785 if (time_after_eq(jiffies, rq->next_balance) &&
3786 likely(!on_null_domain(cpu)))
3787 raise_softirq(SCHED_SOFTIRQ);
3789 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
3790 nohz_balancer_kick(cpu);
3794 static void rq_online_fair(struct rq *rq)
3799 static void rq_offline_fair(struct rq *rq)
3804 #else /* CONFIG_SMP */
3807 * on UP we do not need to balance between CPUs:
3809 static inline void idle_balance(int cpu, struct rq *rq)
3813 #endif /* CONFIG_SMP */
3816 * scheduler tick hitting a task of our scheduling class:
3818 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
3820 struct cfs_rq *cfs_rq;
3821 struct sched_entity *se = &curr->se;
3823 for_each_sched_entity(se) {
3824 cfs_rq = cfs_rq_of(se);
3825 entity_tick(cfs_rq, se, queued);
3830 * called on fork with the child task as argument from the parent's context
3831 * - child not yet on the tasklist
3832 * - preemption disabled
3834 static void task_fork_fair(struct task_struct *p)
3836 struct cfs_rq *cfs_rq = task_cfs_rq(current);
3837 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
3838 int this_cpu = smp_processor_id();
3839 struct rq *rq = this_rq();
3840 unsigned long flags;
3842 raw_spin_lock_irqsave(&rq->lock, flags);
3844 update_rq_clock(rq);
3846 if (unlikely(task_cpu(p) != this_cpu)) {
3848 __set_task_cpu(p, this_cpu);
3852 update_curr(cfs_rq);
3855 se->vruntime = curr->vruntime;
3856 place_entity(cfs_rq, se, 1);
3858 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
3860 * Upon rescheduling, sched_class::put_prev_task() will place
3861 * 'current' within the tree based on its new key value.
3863 swap(curr->vruntime, se->vruntime);
3864 resched_task(rq->curr);
3867 se->vruntime -= cfs_rq->min_vruntime;
3869 raw_spin_unlock_irqrestore(&rq->lock, flags);
3873 * Priority of the task has changed. Check to see if we preempt
3876 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
3877 int oldprio, int running)
3880 * Reschedule if we are currently running on this runqueue and
3881 * our priority decreased, or if we are not currently running on
3882 * this runqueue and our priority is higher than the current's
3885 if (p->prio > oldprio)
3886 resched_task(rq->curr);
3888 check_preempt_curr(rq, p, 0);
3892 * We switched to the sched_fair class.
3894 static void switched_to_fair(struct rq *rq, struct task_struct *p,
3898 * We were most likely switched from sched_rt, so
3899 * kick off the schedule if running, otherwise just see
3900 * if we can still preempt the current task.
3903 resched_task(rq->curr);
3905 check_preempt_curr(rq, p, 0);
3908 /* Account for a task changing its policy or group.
3910 * This routine is mostly called to set cfs_rq->curr field when a task
3911 * migrates between groups/classes.
3913 static void set_curr_task_fair(struct rq *rq)
3915 struct sched_entity *se = &rq->curr->se;
3917 for_each_sched_entity(se)
3918 set_next_entity(cfs_rq_of(se), se);
3921 #ifdef CONFIG_FAIR_GROUP_SCHED
3922 static void task_move_group_fair(struct task_struct *p, int on_rq)
3925 * If the task was not on the rq at the time of this cgroup movement
3926 * it must have been asleep, sleeping tasks keep their ->vruntime
3927 * absolute on their old rq until wakeup (needed for the fair sleeper
3928 * bonus in place_entity()).
3930 * If it was on the rq, we've just 'preempted' it, which does convert
3931 * ->vruntime to a relative base.
3933 * Make sure both cases convert their relative position when migrating
3934 * to another cgroup's rq. This does somewhat interfere with the
3935 * fair sleeper stuff for the first placement, but who cares.
3938 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
3939 set_task_rq(p, task_cpu(p));
3941 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
3945 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
3947 struct sched_entity *se = &task->se;
3948 unsigned int rr_interval = 0;
3951 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
3954 if (rq->cfs.load.weight)
3955 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
3961 * All the scheduling class methods:
3963 static const struct sched_class fair_sched_class = {
3964 .next = &idle_sched_class,
3965 .enqueue_task = enqueue_task_fair,
3966 .dequeue_task = dequeue_task_fair,
3967 .yield_task = yield_task_fair,
3969 .check_preempt_curr = check_preempt_wakeup,
3971 .pick_next_task = pick_next_task_fair,
3972 .put_prev_task = put_prev_task_fair,
3975 .select_task_rq = select_task_rq_fair,
3977 .rq_online = rq_online_fair,
3978 .rq_offline = rq_offline_fair,
3980 .task_waking = task_waking_fair,
3983 .set_curr_task = set_curr_task_fair,
3984 .task_tick = task_tick_fair,
3985 .task_fork = task_fork_fair,
3987 .prio_changed = prio_changed_fair,
3988 .switched_to = switched_to_fair,
3990 .get_rr_interval = get_rr_interval_fair,
3992 #ifdef CONFIG_FAIR_GROUP_SCHED
3993 .task_move_group = task_move_group_fair,
3997 #ifdef CONFIG_SCHED_DEBUG
3998 static void print_cfs_stats(struct seq_file *m, int cpu)
4000 struct cfs_rq *cfs_rq;
4003 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4004 print_cfs_rq(m, cpu, cfs_rq);