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 * SCHED_OTHER wake-up granularity.
73 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
75 * This option delays the preemption effects of decoupled workloads
76 * and reduces their over-scheduling. Synchronous workloads will still
77 * have immediate wakeup/sleep latencies.
79 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
80 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
82 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
85 * The exponential sliding window over which load is averaged for shares
89 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
91 static const struct sched_class fair_sched_class;
93 /**************************************************************
94 * CFS operations on generic schedulable entities:
97 #ifdef CONFIG_FAIR_GROUP_SCHED
99 /* cpu runqueue to which this cfs_rq is attached */
100 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
105 /* An entity is a task if it doesn't "own" a runqueue */
106 #define entity_is_task(se) (!se->my_q)
108 static inline struct task_struct *task_of(struct sched_entity *se)
110 #ifdef CONFIG_SCHED_DEBUG
111 WARN_ON_ONCE(!entity_is_task(se));
113 return container_of(se, struct task_struct, se);
116 /* Walk up scheduling entities hierarchy */
117 #define for_each_sched_entity(se) \
118 for (; se; se = se->parent)
120 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
125 /* runqueue on which this entity is (to be) queued */
126 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
131 /* runqueue "owned" by this group */
132 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
137 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
138 * another cpu ('this_cpu')
140 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
142 return cfs_rq->tg->cfs_rq[this_cpu];
145 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
147 if (!cfs_rq->on_list) {
149 * Ensure we either appear before our parent (if already
150 * enqueued) or force our parent to appear after us when it is
151 * enqueued. The fact that we always enqueue bottom-up
152 * reduces this to two cases.
154 if (cfs_rq->tg->parent &&
155 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
156 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
157 &rq_of(cfs_rq)->leaf_cfs_rq_list);
159 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
160 &rq_of(cfs_rq)->leaf_cfs_rq_list);
167 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
169 if (cfs_rq->on_list) {
170 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
175 /* Iterate thr' all leaf cfs_rq's on a runqueue */
176 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
177 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
179 /* Do the two (enqueued) entities belong to the same group ? */
181 is_same_group(struct sched_entity *se, struct sched_entity *pse)
183 if (se->cfs_rq == pse->cfs_rq)
189 static inline struct sched_entity *parent_entity(struct sched_entity *se)
194 /* return depth at which a sched entity is present in the hierarchy */
195 static inline int depth_se(struct sched_entity *se)
199 for_each_sched_entity(se)
206 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
208 int se_depth, pse_depth;
211 * preemption test can be made between sibling entities who are in the
212 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
213 * both tasks until we find their ancestors who are siblings of common
217 /* First walk up until both entities are at same depth */
218 se_depth = depth_se(*se);
219 pse_depth = depth_se(*pse);
221 while (se_depth > pse_depth) {
223 *se = parent_entity(*se);
226 while (pse_depth > se_depth) {
228 *pse = parent_entity(*pse);
231 while (!is_same_group(*se, *pse)) {
232 *se = parent_entity(*se);
233 *pse = parent_entity(*pse);
237 #else /* !CONFIG_FAIR_GROUP_SCHED */
239 static inline struct task_struct *task_of(struct sched_entity *se)
241 return container_of(se, struct task_struct, se);
244 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
246 return container_of(cfs_rq, struct rq, cfs);
249 #define entity_is_task(se) 1
251 #define for_each_sched_entity(se) \
252 for (; se; se = NULL)
254 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
256 return &task_rq(p)->cfs;
259 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
261 struct task_struct *p = task_of(se);
262 struct rq *rq = task_rq(p);
267 /* runqueue "owned" by this group */
268 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
273 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
275 return &cpu_rq(this_cpu)->cfs;
278 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
282 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
286 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
287 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
290 is_same_group(struct sched_entity *se, struct sched_entity *pse)
295 static inline struct sched_entity *parent_entity(struct sched_entity *se)
301 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
305 #endif /* CONFIG_FAIR_GROUP_SCHED */
308 /**************************************************************
309 * Scheduling class tree data structure manipulation methods:
312 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
314 s64 delta = (s64)(vruntime - min_vruntime);
316 min_vruntime = vruntime;
321 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
323 s64 delta = (s64)(vruntime - min_vruntime);
325 min_vruntime = vruntime;
330 static inline int entity_before(struct sched_entity *a,
331 struct sched_entity *b)
333 return (s64)(a->vruntime - b->vruntime) < 0;
336 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
338 return se->vruntime - cfs_rq->min_vruntime;
341 static void update_min_vruntime(struct cfs_rq *cfs_rq)
343 u64 vruntime = cfs_rq->min_vruntime;
346 vruntime = cfs_rq->curr->vruntime;
348 if (cfs_rq->rb_leftmost) {
349 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
354 vruntime = se->vruntime;
356 vruntime = min_vruntime(vruntime, se->vruntime);
359 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
363 * Enqueue an entity into the rb-tree:
365 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
367 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
368 struct rb_node *parent = NULL;
369 struct sched_entity *entry;
370 s64 key = entity_key(cfs_rq, se);
374 * Find the right place in the rbtree:
378 entry = rb_entry(parent, struct sched_entity, run_node);
380 * We dont care about collisions. Nodes with
381 * the same key stay together.
383 if (key < entity_key(cfs_rq, entry)) {
384 link = &parent->rb_left;
386 link = &parent->rb_right;
392 * Maintain a cache of leftmost tree entries (it is frequently
396 cfs_rq->rb_leftmost = &se->run_node;
398 rb_link_node(&se->run_node, parent, link);
399 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
402 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
404 if (cfs_rq->rb_leftmost == &se->run_node) {
405 struct rb_node *next_node;
407 next_node = rb_next(&se->run_node);
408 cfs_rq->rb_leftmost = next_node;
411 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
414 static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
416 struct rb_node *left = cfs_rq->rb_leftmost;
421 return rb_entry(left, struct sched_entity, run_node);
424 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
426 struct rb_node *next = rb_next(&se->run_node);
431 return rb_entry(next, struct sched_entity, run_node);
434 #ifdef CONFIG_SCHED_DEBUG
435 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
437 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
442 return rb_entry(last, struct sched_entity, run_node);
445 /**************************************************************
446 * Scheduling class statistics methods:
449 int sched_proc_update_handler(struct ctl_table *table, int write,
450 void __user *buffer, size_t *lenp,
453 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
454 int factor = get_update_sysctl_factor();
459 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
460 sysctl_sched_min_granularity);
462 #define WRT_SYSCTL(name) \
463 (normalized_sysctl_##name = sysctl_##name / (factor))
464 WRT_SYSCTL(sched_min_granularity);
465 WRT_SYSCTL(sched_latency);
466 WRT_SYSCTL(sched_wakeup_granularity);
476 static inline unsigned long
477 calc_delta_fair(unsigned long delta, struct sched_entity *se)
479 if (unlikely(se->load.weight != NICE_0_LOAD))
480 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
486 * The idea is to set a period in which each task runs once.
488 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
489 * this period because otherwise the slices get too small.
491 * p = (nr <= nl) ? l : l*nr/nl
493 static u64 __sched_period(unsigned long nr_running)
495 u64 period = sysctl_sched_latency;
496 unsigned long nr_latency = sched_nr_latency;
498 if (unlikely(nr_running > nr_latency)) {
499 period = sysctl_sched_min_granularity;
500 period *= nr_running;
507 * We calculate the wall-time slice from the period by taking a part
508 * proportional to the weight.
512 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
514 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
516 for_each_sched_entity(se) {
517 struct load_weight *load;
518 struct load_weight lw;
520 cfs_rq = cfs_rq_of(se);
521 load = &cfs_rq->load;
523 if (unlikely(!se->on_rq)) {
526 update_load_add(&lw, se->load.weight);
529 slice = calc_delta_mine(slice, se->load.weight, load);
535 * We calculate the vruntime slice of a to be inserted task
539 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
541 return calc_delta_fair(sched_slice(cfs_rq, se), se);
544 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
545 static void update_cfs_shares(struct cfs_rq *cfs_rq);
548 * Update the current task's runtime statistics. Skip current tasks that
549 * are not in our scheduling class.
552 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
553 unsigned long delta_exec)
555 unsigned long delta_exec_weighted;
557 schedstat_set(curr->statistics.exec_max,
558 max((u64)delta_exec, curr->statistics.exec_max));
560 curr->sum_exec_runtime += delta_exec;
561 schedstat_add(cfs_rq, exec_clock, delta_exec);
562 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
564 curr->vruntime += delta_exec_weighted;
565 update_min_vruntime(cfs_rq);
567 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
568 cfs_rq->load_unacc_exec_time += delta_exec;
572 static void update_curr(struct cfs_rq *cfs_rq)
574 struct sched_entity *curr = cfs_rq->curr;
575 u64 now = rq_of(cfs_rq)->clock_task;
576 unsigned long delta_exec;
582 * Get the amount of time the current task was running
583 * since the last time we changed load (this cannot
584 * overflow on 32 bits):
586 delta_exec = (unsigned long)(now - curr->exec_start);
590 __update_curr(cfs_rq, curr, delta_exec);
591 curr->exec_start = now;
593 if (entity_is_task(curr)) {
594 struct task_struct *curtask = task_of(curr);
596 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
597 cpuacct_charge(curtask, delta_exec);
598 account_group_exec_runtime(curtask, delta_exec);
603 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
605 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
609 * Task is being enqueued - update stats:
611 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
614 * Are we enqueueing a waiting task? (for current tasks
615 * a dequeue/enqueue event is a NOP)
617 if (se != cfs_rq->curr)
618 update_stats_wait_start(cfs_rq, se);
622 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
624 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
625 rq_of(cfs_rq)->clock - se->statistics.wait_start));
626 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
627 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
628 rq_of(cfs_rq)->clock - se->statistics.wait_start);
629 #ifdef CONFIG_SCHEDSTATS
630 if (entity_is_task(se)) {
631 trace_sched_stat_wait(task_of(se),
632 rq_of(cfs_rq)->clock - se->statistics.wait_start);
635 schedstat_set(se->statistics.wait_start, 0);
639 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
642 * Mark the end of the wait period if dequeueing a
645 if (se != cfs_rq->curr)
646 update_stats_wait_end(cfs_rq, se);
650 * We are picking a new current task - update its stats:
653 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 * We are starting a new run period:
658 se->exec_start = rq_of(cfs_rq)->clock_task;
661 /**************************************************
662 * Scheduling class queueing methods:
665 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
667 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
669 cfs_rq->task_weight += weight;
673 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
679 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
681 update_load_add(&cfs_rq->load, se->load.weight);
682 if (!parent_entity(se))
683 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
684 if (entity_is_task(se)) {
685 add_cfs_task_weight(cfs_rq, se->load.weight);
686 list_add(&se->group_node, &cfs_rq->tasks);
688 cfs_rq->nr_running++;
692 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
694 update_load_sub(&cfs_rq->load, se->load.weight);
695 if (!parent_entity(se))
696 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
697 if (entity_is_task(se)) {
698 add_cfs_task_weight(cfs_rq, -se->load.weight);
699 list_del_init(&se->group_node);
701 cfs_rq->nr_running--;
704 #ifdef CONFIG_FAIR_GROUP_SCHED
706 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
709 struct task_group *tg = cfs_rq->tg;
712 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
713 load_avg -= cfs_rq->load_contribution;
715 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
716 atomic_add(load_avg, &tg->load_weight);
717 cfs_rq->load_contribution += load_avg;
721 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
723 u64 period = sysctl_sched_shares_window;
725 unsigned long load = cfs_rq->load.weight;
727 if (cfs_rq->tg == &root_task_group)
730 now = rq_of(cfs_rq)->clock_task;
731 delta = now - cfs_rq->load_stamp;
733 /* truncate load history at 4 idle periods */
734 if (cfs_rq->load_stamp > cfs_rq->load_last &&
735 now - cfs_rq->load_last > 4 * period) {
736 cfs_rq->load_period = 0;
737 cfs_rq->load_avg = 0;
741 cfs_rq->load_stamp = now;
742 cfs_rq->load_unacc_exec_time = 0;
743 cfs_rq->load_period += delta;
745 cfs_rq->load_last = now;
746 cfs_rq->load_avg += delta * load;
749 /* consider updating load contribution on each fold or truncate */
750 if (global_update || cfs_rq->load_period > period
751 || !cfs_rq->load_period)
752 update_cfs_rq_load_contribution(cfs_rq, global_update);
754 while (cfs_rq->load_period > period) {
756 * Inline assembly required to prevent the compiler
757 * optimising this loop into a divmod call.
758 * See __iter_div_u64_rem() for another example of this.
760 asm("" : "+rm" (cfs_rq->load_period));
761 cfs_rq->load_period /= 2;
762 cfs_rq->load_avg /= 2;
765 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
766 list_del_leaf_cfs_rq(cfs_rq);
769 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
771 long load_weight, load, shares;
773 load = cfs_rq->load.weight;
775 load_weight = atomic_read(&tg->load_weight);
777 load_weight -= cfs_rq->load_contribution;
779 shares = (tg->shares * load);
781 shares /= load_weight;
783 if (shares < MIN_SHARES)
785 if (shares > tg->shares)
791 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
793 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
794 update_cfs_load(cfs_rq, 0);
795 update_cfs_shares(cfs_rq);
798 # else /* CONFIG_SMP */
799 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
803 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
808 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
811 # endif /* CONFIG_SMP */
812 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
813 unsigned long weight)
816 /* commit outstanding execution time */
817 if (cfs_rq->curr == se)
819 account_entity_dequeue(cfs_rq, se);
822 update_load_set(&se->load, weight);
825 account_entity_enqueue(cfs_rq, se);
828 static void update_cfs_shares(struct cfs_rq *cfs_rq)
830 struct task_group *tg;
831 struct sched_entity *se;
835 se = tg->se[cpu_of(rq_of(cfs_rq))];
839 if (likely(se->load.weight == tg->shares))
842 shares = calc_cfs_shares(cfs_rq, tg);
844 reweight_entity(cfs_rq_of(se), se, shares);
846 #else /* CONFIG_FAIR_GROUP_SCHED */
847 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
851 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
855 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
858 #endif /* CONFIG_FAIR_GROUP_SCHED */
860 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
862 #ifdef CONFIG_SCHEDSTATS
863 struct task_struct *tsk = NULL;
865 if (entity_is_task(se))
868 if (se->statistics.sleep_start) {
869 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
874 if (unlikely(delta > se->statistics.sleep_max))
875 se->statistics.sleep_max = delta;
877 se->statistics.sleep_start = 0;
878 se->statistics.sum_sleep_runtime += delta;
881 account_scheduler_latency(tsk, delta >> 10, 1);
882 trace_sched_stat_sleep(tsk, delta);
885 if (se->statistics.block_start) {
886 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
891 if (unlikely(delta > se->statistics.block_max))
892 se->statistics.block_max = delta;
894 se->statistics.block_start = 0;
895 se->statistics.sum_sleep_runtime += delta;
898 if (tsk->in_iowait) {
899 se->statistics.iowait_sum += delta;
900 se->statistics.iowait_count++;
901 trace_sched_stat_iowait(tsk, delta);
905 * Blocking time is in units of nanosecs, so shift by
906 * 20 to get a milliseconds-range estimation of the
907 * amount of time that the task spent sleeping:
909 if (unlikely(prof_on == SLEEP_PROFILING)) {
910 profile_hits(SLEEP_PROFILING,
911 (void *)get_wchan(tsk),
914 account_scheduler_latency(tsk, delta >> 10, 0);
920 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
922 #ifdef CONFIG_SCHED_DEBUG
923 s64 d = se->vruntime - cfs_rq->min_vruntime;
928 if (d > 3*sysctl_sched_latency)
929 schedstat_inc(cfs_rq, nr_spread_over);
934 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
936 u64 vruntime = cfs_rq->min_vruntime;
939 * The 'current' period is already promised to the current tasks,
940 * however the extra weight of the new task will slow them down a
941 * little, place the new task so that it fits in the slot that
942 * stays open at the end.
944 if (initial && sched_feat(START_DEBIT))
945 vruntime += sched_vslice(cfs_rq, se);
947 /* sleeps up to a single latency don't count. */
949 unsigned long thresh = sysctl_sched_latency;
952 * Halve their sleep time's effect, to allow
953 * for a gentler effect of sleepers:
955 if (sched_feat(GENTLE_FAIR_SLEEPERS))
961 /* ensure we never gain time by being placed backwards. */
962 vruntime = max_vruntime(se->vruntime, vruntime);
964 se->vruntime = vruntime;
968 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
971 * Update the normalized vruntime before updating min_vruntime
972 * through callig update_curr().
974 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
975 se->vruntime += cfs_rq->min_vruntime;
978 * Update run-time statistics of the 'current'.
981 update_cfs_load(cfs_rq, 0);
982 account_entity_enqueue(cfs_rq, se);
983 update_cfs_shares(cfs_rq);
985 if (flags & ENQUEUE_WAKEUP) {
986 place_entity(cfs_rq, se, 0);
987 enqueue_sleeper(cfs_rq, se);
990 update_stats_enqueue(cfs_rq, se);
991 check_spread(cfs_rq, se);
992 if (se != cfs_rq->curr)
993 __enqueue_entity(cfs_rq, se);
996 if (cfs_rq->nr_running == 1)
997 list_add_leaf_cfs_rq(cfs_rq);
1000 static void __clear_buddies_last(struct sched_entity *se)
1002 for_each_sched_entity(se) {
1003 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1004 if (cfs_rq->last == se)
1005 cfs_rq->last = NULL;
1011 static void __clear_buddies_next(struct sched_entity *se)
1013 for_each_sched_entity(se) {
1014 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1015 if (cfs_rq->next == se)
1016 cfs_rq->next = NULL;
1022 static void __clear_buddies_skip(struct sched_entity *se)
1024 for_each_sched_entity(se) {
1025 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1026 if (cfs_rq->skip == se)
1027 cfs_rq->skip = NULL;
1033 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1035 if (cfs_rq->last == se)
1036 __clear_buddies_last(se);
1038 if (cfs_rq->next == se)
1039 __clear_buddies_next(se);
1041 if (cfs_rq->skip == se)
1042 __clear_buddies_skip(se);
1046 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1049 * Update run-time statistics of the 'current'.
1051 update_curr(cfs_rq);
1053 update_stats_dequeue(cfs_rq, se);
1054 if (flags & DEQUEUE_SLEEP) {
1055 #ifdef CONFIG_SCHEDSTATS
1056 if (entity_is_task(se)) {
1057 struct task_struct *tsk = task_of(se);
1059 if (tsk->state & TASK_INTERRUPTIBLE)
1060 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1061 if (tsk->state & TASK_UNINTERRUPTIBLE)
1062 se->statistics.block_start = rq_of(cfs_rq)->clock;
1067 clear_buddies(cfs_rq, se);
1069 if (se != cfs_rq->curr)
1070 __dequeue_entity(cfs_rq, se);
1072 update_cfs_load(cfs_rq, 0);
1073 account_entity_dequeue(cfs_rq, se);
1074 update_min_vruntime(cfs_rq);
1075 update_cfs_shares(cfs_rq);
1078 * Normalize the entity after updating the min_vruntime because the
1079 * update can refer to the ->curr item and we need to reflect this
1080 * movement in our normalized position.
1082 if (!(flags & DEQUEUE_SLEEP))
1083 se->vruntime -= cfs_rq->min_vruntime;
1087 * Preempt the current task with a newly woken task if needed:
1090 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1092 unsigned long ideal_runtime, delta_exec;
1094 ideal_runtime = sched_slice(cfs_rq, curr);
1095 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1096 if (delta_exec > ideal_runtime) {
1097 resched_task(rq_of(cfs_rq)->curr);
1099 * The current task ran long enough, ensure it doesn't get
1100 * re-elected due to buddy favours.
1102 clear_buddies(cfs_rq, curr);
1107 * Ensure that a task that missed wakeup preemption by a
1108 * narrow margin doesn't have to wait for a full slice.
1109 * This also mitigates buddy induced latencies under load.
1111 if (!sched_feat(WAKEUP_PREEMPT))
1114 if (delta_exec < sysctl_sched_min_granularity)
1117 if (cfs_rq->nr_running > 1) {
1118 struct sched_entity *se = __pick_first_entity(cfs_rq);
1119 s64 delta = curr->vruntime - se->vruntime;
1124 if (delta > ideal_runtime)
1125 resched_task(rq_of(cfs_rq)->curr);
1130 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1132 /* 'current' is not kept within the tree. */
1135 * Any task has to be enqueued before it get to execute on
1136 * a CPU. So account for the time it spent waiting on the
1139 update_stats_wait_end(cfs_rq, se);
1140 __dequeue_entity(cfs_rq, se);
1143 update_stats_curr_start(cfs_rq, se);
1145 #ifdef CONFIG_SCHEDSTATS
1147 * Track our maximum slice length, if the CPU's load is at
1148 * least twice that of our own weight (i.e. dont track it
1149 * when there are only lesser-weight tasks around):
1151 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1152 se->statistics.slice_max = max(se->statistics.slice_max,
1153 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1156 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1160 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1163 * Pick the next process, keeping these things in mind, in this order:
1164 * 1) keep things fair between processes/task groups
1165 * 2) pick the "next" process, since someone really wants that to run
1166 * 3) pick the "last" process, for cache locality
1167 * 4) do not run the "skip" process, if something else is available
1169 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1171 struct sched_entity *se = __pick_first_entity(cfs_rq);
1172 struct sched_entity *left = se;
1175 * Avoid running the skip buddy, if running something else can
1176 * be done without getting too unfair.
1178 if (cfs_rq->skip == se) {
1179 struct sched_entity *second = __pick_next_entity(se);
1180 if (second && wakeup_preempt_entity(second, left) < 1)
1185 * Prefer last buddy, try to return the CPU to a preempted task.
1187 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1191 * Someone really wants this to run. If it's not unfair, run it.
1193 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1196 clear_buddies(cfs_rq, se);
1201 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1204 * If still on the runqueue then deactivate_task()
1205 * was not called and update_curr() has to be done:
1208 update_curr(cfs_rq);
1210 check_spread(cfs_rq, prev);
1212 update_stats_wait_start(cfs_rq, prev);
1213 /* Put 'current' back into the tree. */
1214 __enqueue_entity(cfs_rq, prev);
1216 cfs_rq->curr = NULL;
1220 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1223 * Update run-time statistics of the 'current'.
1225 update_curr(cfs_rq);
1228 * Update share accounting for long-running entities.
1230 update_entity_shares_tick(cfs_rq);
1232 #ifdef CONFIG_SCHED_HRTICK
1234 * queued ticks are scheduled to match the slice, so don't bother
1235 * validating it and just reschedule.
1238 resched_task(rq_of(cfs_rq)->curr);
1242 * don't let the period tick interfere with the hrtick preemption
1244 if (!sched_feat(DOUBLE_TICK) &&
1245 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1249 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1250 check_preempt_tick(cfs_rq, curr);
1253 /**************************************************
1254 * CFS operations on tasks:
1257 #ifdef CONFIG_SCHED_HRTICK
1258 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1260 struct sched_entity *se = &p->se;
1261 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1263 WARN_ON(task_rq(p) != rq);
1265 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1266 u64 slice = sched_slice(cfs_rq, se);
1267 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1268 s64 delta = slice - ran;
1277 * Don't schedule slices shorter than 10000ns, that just
1278 * doesn't make sense. Rely on vruntime for fairness.
1281 delta = max_t(s64, 10000LL, delta);
1283 hrtick_start(rq, delta);
1288 * called from enqueue/dequeue and updates the hrtick when the
1289 * current task is from our class and nr_running is low enough
1292 static void hrtick_update(struct rq *rq)
1294 struct task_struct *curr = rq->curr;
1296 if (curr->sched_class != &fair_sched_class)
1299 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1300 hrtick_start_fair(rq, curr);
1302 #else /* !CONFIG_SCHED_HRTICK */
1304 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1308 static inline void hrtick_update(struct rq *rq)
1314 * The enqueue_task method is called before nr_running is
1315 * increased. Here we update the fair scheduling stats and
1316 * then put the task into the rbtree:
1319 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1321 struct cfs_rq *cfs_rq;
1322 struct sched_entity *se = &p->se;
1324 for_each_sched_entity(se) {
1327 cfs_rq = cfs_rq_of(se);
1328 enqueue_entity(cfs_rq, se, flags);
1329 flags = ENQUEUE_WAKEUP;
1332 for_each_sched_entity(se) {
1333 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1335 update_cfs_load(cfs_rq, 0);
1336 update_cfs_shares(cfs_rq);
1343 * The dequeue_task method is called before nr_running is
1344 * decreased. We remove the task from the rbtree and
1345 * update the fair scheduling stats:
1347 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1349 struct cfs_rq *cfs_rq;
1350 struct sched_entity *se = &p->se;
1352 for_each_sched_entity(se) {
1353 cfs_rq = cfs_rq_of(se);
1354 dequeue_entity(cfs_rq, se, flags);
1356 /* Don't dequeue parent if it has other entities besides us */
1357 if (cfs_rq->load.weight)
1359 flags |= DEQUEUE_SLEEP;
1362 for_each_sched_entity(se) {
1363 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1365 update_cfs_load(cfs_rq, 0);
1366 update_cfs_shares(cfs_rq);
1374 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1376 struct sched_entity *se = &p->se;
1377 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1379 se->vruntime -= cfs_rq->min_vruntime;
1382 #ifdef CONFIG_FAIR_GROUP_SCHED
1384 * effective_load() calculates the load change as seen from the root_task_group
1386 * Adding load to a group doesn't make a group heavier, but can cause movement
1387 * of group shares between cpus. Assuming the shares were perfectly aligned one
1388 * can calculate the shift in shares.
1390 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1392 struct sched_entity *se = tg->se[cpu];
1397 for_each_sched_entity(se) {
1401 w = se->my_q->load.weight;
1403 /* use this cpu's instantaneous contribution */
1404 lw = atomic_read(&tg->load_weight);
1405 lw -= se->my_q->load_contribution;
1410 if (lw > 0 && wl < lw)
1411 wl = (wl * tg->shares) / lw;
1415 /* zero point is MIN_SHARES */
1416 if (wl < MIN_SHARES)
1418 wl -= se->load.weight;
1427 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1428 unsigned long wl, unsigned long wg)
1435 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1437 s64 this_load, load;
1438 int idx, this_cpu, prev_cpu;
1439 unsigned long tl_per_task;
1440 struct task_group *tg;
1441 unsigned long weight;
1445 this_cpu = smp_processor_id();
1446 prev_cpu = task_cpu(p);
1447 load = source_load(prev_cpu, idx);
1448 this_load = target_load(this_cpu, idx);
1451 * If sync wakeup then subtract the (maximum possible)
1452 * effect of the currently running task from the load
1453 * of the current CPU:
1457 tg = task_group(current);
1458 weight = current->se.load.weight;
1460 this_load += effective_load(tg, this_cpu, -weight, -weight);
1461 load += effective_load(tg, prev_cpu, 0, -weight);
1465 weight = p->se.load.weight;
1468 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1469 * due to the sync cause above having dropped this_load to 0, we'll
1470 * always have an imbalance, but there's really nothing you can do
1471 * about that, so that's good too.
1473 * Otherwise check if either cpus are near enough in load to allow this
1474 * task to be woken on this_cpu.
1476 if (this_load > 0) {
1477 s64 this_eff_load, prev_eff_load;
1479 this_eff_load = 100;
1480 this_eff_load *= power_of(prev_cpu);
1481 this_eff_load *= this_load +
1482 effective_load(tg, this_cpu, weight, weight);
1484 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1485 prev_eff_load *= power_of(this_cpu);
1486 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1488 balanced = this_eff_load <= prev_eff_load;
1494 * If the currently running task will sleep within
1495 * a reasonable amount of time then attract this newly
1498 if (sync && balanced)
1501 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1502 tl_per_task = cpu_avg_load_per_task(this_cpu);
1505 (this_load <= load &&
1506 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1508 * This domain has SD_WAKE_AFFINE and
1509 * p is cache cold in this domain, and
1510 * there is no bad imbalance.
1512 schedstat_inc(sd, ttwu_move_affine);
1513 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1521 * find_idlest_group finds and returns the least busy CPU group within the
1524 static struct sched_group *
1525 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1526 int this_cpu, int load_idx)
1528 struct sched_group *idlest = NULL, *group = sd->groups;
1529 unsigned long min_load = ULONG_MAX, this_load = 0;
1530 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1533 unsigned long load, avg_load;
1537 /* Skip over this group if it has no CPUs allowed */
1538 if (!cpumask_intersects(sched_group_cpus(group),
1542 local_group = cpumask_test_cpu(this_cpu,
1543 sched_group_cpus(group));
1545 /* Tally up the load of all CPUs in the group */
1548 for_each_cpu(i, sched_group_cpus(group)) {
1549 /* Bias balancing toward cpus of our domain */
1551 load = source_load(i, load_idx);
1553 load = target_load(i, load_idx);
1558 /* Adjust by relative CPU power of the group */
1559 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1562 this_load = avg_load;
1563 } else if (avg_load < min_load) {
1564 min_load = avg_load;
1567 } while (group = group->next, group != sd->groups);
1569 if (!idlest || 100*this_load < imbalance*min_load)
1575 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1578 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1580 unsigned long load, min_load = ULONG_MAX;
1584 /* Traverse only the allowed CPUs */
1585 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1586 load = weighted_cpuload(i);
1588 if (load < min_load || (load == min_load && i == this_cpu)) {
1598 * Try and locate an idle CPU in the sched_domain.
1600 static int select_idle_sibling(struct task_struct *p, int target)
1602 int cpu = smp_processor_id();
1603 int prev_cpu = task_cpu(p);
1604 struct sched_domain *sd;
1608 * If the task is going to be woken-up on this cpu and if it is
1609 * already idle, then it is the right target.
1611 if (target == cpu && idle_cpu(cpu))
1615 * If the task is going to be woken-up on the cpu where it previously
1616 * ran and if it is currently idle, then it the right target.
1618 if (target == prev_cpu && idle_cpu(prev_cpu))
1622 * Otherwise, iterate the domains and find an elegible idle cpu.
1624 for_each_domain(target, sd) {
1625 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1628 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1636 * Lets stop looking for an idle sibling when we reached
1637 * the domain that spans the current cpu and prev_cpu.
1639 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1640 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1648 * sched_balance_self: balance the current task (running on cpu) in domains
1649 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1652 * Balance, ie. select the least loaded group.
1654 * Returns the target CPU number, or the same CPU if no balancing is needed.
1656 * preempt must be disabled.
1659 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1661 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1662 int cpu = smp_processor_id();
1663 int prev_cpu = task_cpu(p);
1665 int want_affine = 0;
1667 int sync = wake_flags & WF_SYNC;
1669 if (sd_flag & SD_BALANCE_WAKE) {
1670 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1675 for_each_domain(cpu, tmp) {
1676 if (!(tmp->flags & SD_LOAD_BALANCE))
1680 * If power savings logic is enabled for a domain, see if we
1681 * are not overloaded, if so, don't balance wider.
1683 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1684 unsigned long power = 0;
1685 unsigned long nr_running = 0;
1686 unsigned long capacity;
1689 for_each_cpu(i, sched_domain_span(tmp)) {
1690 power += power_of(i);
1691 nr_running += cpu_rq(i)->cfs.nr_running;
1694 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1696 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1699 if (nr_running < capacity)
1704 * If both cpu and prev_cpu are part of this domain,
1705 * cpu is a valid SD_WAKE_AFFINE target.
1707 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1708 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1713 if (!want_sd && !want_affine)
1716 if (!(tmp->flags & sd_flag))
1724 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1725 return select_idle_sibling(p, cpu);
1727 return select_idle_sibling(p, prev_cpu);
1731 int load_idx = sd->forkexec_idx;
1732 struct sched_group *group;
1735 if (!(sd->flags & sd_flag)) {
1740 if (sd_flag & SD_BALANCE_WAKE)
1741 load_idx = sd->wake_idx;
1743 group = find_idlest_group(sd, p, cpu, load_idx);
1749 new_cpu = find_idlest_cpu(group, p, cpu);
1750 if (new_cpu == -1 || new_cpu == cpu) {
1751 /* Now try balancing at a lower domain level of cpu */
1756 /* Now try balancing at a lower domain level of new_cpu */
1758 weight = sd->span_weight;
1760 for_each_domain(cpu, tmp) {
1761 if (weight <= tmp->span_weight)
1763 if (tmp->flags & sd_flag)
1766 /* while loop will break here if sd == NULL */
1771 #endif /* CONFIG_SMP */
1773 static unsigned long
1774 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1776 unsigned long gran = sysctl_sched_wakeup_granularity;
1779 * Since its curr running now, convert the gran from real-time
1780 * to virtual-time in his units.
1782 * By using 'se' instead of 'curr' we penalize light tasks, so
1783 * they get preempted easier. That is, if 'se' < 'curr' then
1784 * the resulting gran will be larger, therefore penalizing the
1785 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1786 * be smaller, again penalizing the lighter task.
1788 * This is especially important for buddies when the leftmost
1789 * task is higher priority than the buddy.
1791 if (unlikely(se->load.weight != NICE_0_LOAD))
1792 gran = calc_delta_fair(gran, se);
1798 * Should 'se' preempt 'curr'.
1812 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1814 s64 gran, vdiff = curr->vruntime - se->vruntime;
1819 gran = wakeup_gran(curr, se);
1826 static void set_last_buddy(struct sched_entity *se)
1828 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1829 for_each_sched_entity(se)
1830 cfs_rq_of(se)->last = se;
1834 static void set_next_buddy(struct sched_entity *se)
1836 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1837 for_each_sched_entity(se)
1838 cfs_rq_of(se)->next = se;
1842 static void set_skip_buddy(struct sched_entity *se)
1844 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1845 for_each_sched_entity(se)
1846 cfs_rq_of(se)->skip = se;
1851 * Preempt the current task with a newly woken task if needed:
1853 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1855 struct task_struct *curr = rq->curr;
1856 struct sched_entity *se = &curr->se, *pse = &p->se;
1857 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1858 int scale = cfs_rq->nr_running >= sched_nr_latency;
1860 if (unlikely(se == pse))
1863 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1864 set_next_buddy(pse);
1867 * We can come here with TIF_NEED_RESCHED already set from new task
1870 if (test_tsk_need_resched(curr))
1874 * Batch and idle tasks do not preempt (their preemption is driven by
1877 if (unlikely(p->policy != SCHED_NORMAL))
1880 /* Idle tasks are by definition preempted by everybody. */
1881 if (unlikely(curr->policy == SCHED_IDLE))
1884 if (!sched_feat(WAKEUP_PREEMPT))
1887 update_curr(cfs_rq);
1888 find_matching_se(&se, &pse);
1890 if (wakeup_preempt_entity(se, pse) == 1)
1898 * Only set the backward buddy when the current task is still
1899 * on the rq. This can happen when a wakeup gets interleaved
1900 * with schedule on the ->pre_schedule() or idle_balance()
1901 * point, either of which can * drop the rq lock.
1903 * Also, during early boot the idle thread is in the fair class,
1904 * for obvious reasons its a bad idea to schedule back to it.
1906 if (unlikely(!se->on_rq || curr == rq->idle))
1909 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1913 static struct task_struct *pick_next_task_fair(struct rq *rq)
1915 struct task_struct *p;
1916 struct cfs_rq *cfs_rq = &rq->cfs;
1917 struct sched_entity *se;
1919 if (!cfs_rq->nr_running)
1923 se = pick_next_entity(cfs_rq);
1924 set_next_entity(cfs_rq, se);
1925 cfs_rq = group_cfs_rq(se);
1929 hrtick_start_fair(rq, p);
1935 * Account for a descheduled task:
1937 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1939 struct sched_entity *se = &prev->se;
1940 struct cfs_rq *cfs_rq;
1942 for_each_sched_entity(se) {
1943 cfs_rq = cfs_rq_of(se);
1944 put_prev_entity(cfs_rq, se);
1949 * sched_yield() is very simple
1951 * The magic of dealing with the ->skip buddy is in pick_next_entity.
1953 static void yield_task_fair(struct rq *rq)
1955 struct task_struct *curr = rq->curr;
1956 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1957 struct sched_entity *se = &curr->se;
1960 * Are we the only task in the tree?
1962 if (unlikely(rq->nr_running == 1))
1965 clear_buddies(cfs_rq, se);
1967 if (curr->policy != SCHED_BATCH) {
1968 update_rq_clock(rq);
1970 * Update run-time statistics of the 'current'.
1972 update_curr(cfs_rq);
1979 /**************************************************
1980 * Fair scheduling class load-balancing methods:
1984 * pull_task - move a task from a remote runqueue to the local runqueue.
1985 * Both runqueues must be locked.
1987 static void pull_task(struct rq *src_rq, struct task_struct *p,
1988 struct rq *this_rq, int this_cpu)
1990 deactivate_task(src_rq, p, 0);
1991 set_task_cpu(p, this_cpu);
1992 activate_task(this_rq, p, 0);
1993 check_preempt_curr(this_rq, p, 0);
1997 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2000 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2001 struct sched_domain *sd, enum cpu_idle_type idle,
2004 int tsk_cache_hot = 0;
2006 * We do not migrate tasks that are:
2007 * 1) running (obviously), or
2008 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2009 * 3) are cache-hot on their current CPU.
2011 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2012 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2017 if (task_running(rq, p)) {
2018 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2023 * Aggressive migration if:
2024 * 1) task is cache cold, or
2025 * 2) too many balance attempts have failed.
2028 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2029 if (!tsk_cache_hot ||
2030 sd->nr_balance_failed > sd->cache_nice_tries) {
2031 #ifdef CONFIG_SCHEDSTATS
2032 if (tsk_cache_hot) {
2033 schedstat_inc(sd, lb_hot_gained[idle]);
2034 schedstat_inc(p, se.statistics.nr_forced_migrations);
2040 if (tsk_cache_hot) {
2041 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2048 * move_one_task tries to move exactly one task from busiest to this_rq, as
2049 * part of active balancing operations within "domain".
2050 * Returns 1 if successful and 0 otherwise.
2052 * Called with both runqueues locked.
2055 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2056 struct sched_domain *sd, enum cpu_idle_type idle)
2058 struct task_struct *p, *n;
2059 struct cfs_rq *cfs_rq;
2062 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2063 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2065 if (!can_migrate_task(p, busiest, this_cpu,
2069 pull_task(busiest, p, this_rq, this_cpu);
2071 * Right now, this is only the second place pull_task()
2072 * is called, so we can safely collect pull_task()
2073 * stats here rather than inside pull_task().
2075 schedstat_inc(sd, lb_gained[idle]);
2083 static unsigned long
2084 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2085 unsigned long max_load_move, struct sched_domain *sd,
2086 enum cpu_idle_type idle, int *all_pinned,
2087 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
2089 int loops = 0, pulled = 0, pinned = 0;
2090 long rem_load_move = max_load_move;
2091 struct task_struct *p, *n;
2093 if (max_load_move == 0)
2098 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2099 if (loops++ > sysctl_sched_nr_migrate)
2102 if ((p->se.load.weight >> 1) > rem_load_move ||
2103 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
2106 pull_task(busiest, p, this_rq, this_cpu);
2108 rem_load_move -= p->se.load.weight;
2110 #ifdef CONFIG_PREEMPT
2112 * NEWIDLE balancing is a source of latency, so preemptible
2113 * kernels will stop after the first task is pulled to minimize
2114 * the critical section.
2116 if (idle == CPU_NEWLY_IDLE)
2121 * We only want to steal up to the prescribed amount of
2124 if (rem_load_move <= 0)
2127 if (p->prio < *this_best_prio)
2128 *this_best_prio = p->prio;
2132 * Right now, this is one of only two places pull_task() is called,
2133 * so we can safely collect pull_task() stats here rather than
2134 * inside pull_task().
2136 schedstat_add(sd, lb_gained[idle], pulled);
2139 *all_pinned = pinned;
2141 return max_load_move - rem_load_move;
2144 #ifdef CONFIG_FAIR_GROUP_SCHED
2146 * update tg->load_weight by folding this cpu's load_avg
2148 static int update_shares_cpu(struct task_group *tg, int cpu)
2150 struct cfs_rq *cfs_rq;
2151 unsigned long flags;
2158 cfs_rq = tg->cfs_rq[cpu];
2160 raw_spin_lock_irqsave(&rq->lock, flags);
2162 update_rq_clock(rq);
2163 update_cfs_load(cfs_rq, 1);
2166 * We need to update shares after updating tg->load_weight in
2167 * order to adjust the weight of groups with long running tasks.
2169 update_cfs_shares(cfs_rq);
2171 raw_spin_unlock_irqrestore(&rq->lock, flags);
2176 static void update_shares(int cpu)
2178 struct cfs_rq *cfs_rq;
2179 struct rq *rq = cpu_rq(cpu);
2182 for_each_leaf_cfs_rq(rq, cfs_rq)
2183 update_shares_cpu(cfs_rq->tg, cpu);
2187 static unsigned long
2188 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2189 unsigned long max_load_move,
2190 struct sched_domain *sd, enum cpu_idle_type idle,
2191 int *all_pinned, int *this_best_prio)
2193 long rem_load_move = max_load_move;
2194 int busiest_cpu = cpu_of(busiest);
2195 struct task_group *tg;
2198 update_h_load(busiest_cpu);
2200 list_for_each_entry_rcu(tg, &task_groups, list) {
2201 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2202 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2203 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2204 u64 rem_load, moved_load;
2209 if (!busiest_cfs_rq->task_weight)
2212 rem_load = (u64)rem_load_move * busiest_weight;
2213 rem_load = div_u64(rem_load, busiest_h_load + 1);
2215 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2216 rem_load, sd, idle, all_pinned, this_best_prio,
2222 moved_load *= busiest_h_load;
2223 moved_load = div_u64(moved_load, busiest_weight + 1);
2225 rem_load_move -= moved_load;
2226 if (rem_load_move < 0)
2231 return max_load_move - rem_load_move;
2234 static inline void update_shares(int cpu)
2238 static unsigned long
2239 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2240 unsigned long max_load_move,
2241 struct sched_domain *sd, enum cpu_idle_type idle,
2242 int *all_pinned, int *this_best_prio)
2244 return balance_tasks(this_rq, this_cpu, busiest,
2245 max_load_move, sd, idle, all_pinned,
2246 this_best_prio, &busiest->cfs);
2251 * move_tasks tries to move up to max_load_move weighted load from busiest to
2252 * this_rq, as part of a balancing operation within domain "sd".
2253 * Returns 1 if successful and 0 otherwise.
2255 * Called with both runqueues locked.
2257 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2258 unsigned long max_load_move,
2259 struct sched_domain *sd, enum cpu_idle_type idle,
2262 unsigned long total_load_moved = 0, load_moved;
2263 int this_best_prio = this_rq->curr->prio;
2266 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2267 max_load_move - total_load_moved,
2268 sd, idle, all_pinned, &this_best_prio);
2270 total_load_moved += load_moved;
2272 #ifdef CONFIG_PREEMPT
2274 * NEWIDLE balancing is a source of latency, so preemptible
2275 * kernels will stop after the first task is pulled to minimize
2276 * the critical section.
2278 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2281 if (raw_spin_is_contended(&this_rq->lock) ||
2282 raw_spin_is_contended(&busiest->lock))
2285 } while (load_moved && max_load_move > total_load_moved);
2287 return total_load_moved > 0;
2290 /********** Helpers for find_busiest_group ************************/
2292 * sd_lb_stats - Structure to store the statistics of a sched_domain
2293 * during load balancing.
2295 struct sd_lb_stats {
2296 struct sched_group *busiest; /* Busiest group in this sd */
2297 struct sched_group *this; /* Local group in this sd */
2298 unsigned long total_load; /* Total load of all groups in sd */
2299 unsigned long total_pwr; /* Total power of all groups in sd */
2300 unsigned long avg_load; /* Average load across all groups in sd */
2302 /** Statistics of this group */
2303 unsigned long this_load;
2304 unsigned long this_load_per_task;
2305 unsigned long this_nr_running;
2306 unsigned long this_has_capacity;
2307 unsigned int this_idle_cpus;
2309 /* Statistics of the busiest group */
2310 unsigned int busiest_idle_cpus;
2311 unsigned long max_load;
2312 unsigned long busiest_load_per_task;
2313 unsigned long busiest_nr_running;
2314 unsigned long busiest_group_capacity;
2315 unsigned long busiest_has_capacity;
2316 unsigned int busiest_group_weight;
2318 int group_imb; /* Is there imbalance in this sd */
2319 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2320 int power_savings_balance; /* Is powersave balance needed for this sd */
2321 struct sched_group *group_min; /* Least loaded group in sd */
2322 struct sched_group *group_leader; /* Group which relieves group_min */
2323 unsigned long min_load_per_task; /* load_per_task in group_min */
2324 unsigned long leader_nr_running; /* Nr running of group_leader */
2325 unsigned long min_nr_running; /* Nr running of group_min */
2330 * sg_lb_stats - stats of a sched_group required for load_balancing
2332 struct sg_lb_stats {
2333 unsigned long avg_load; /*Avg load across the CPUs of the group */
2334 unsigned long group_load; /* Total load over the CPUs of the group */
2335 unsigned long sum_nr_running; /* Nr tasks running in the group */
2336 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2337 unsigned long group_capacity;
2338 unsigned long idle_cpus;
2339 unsigned long group_weight;
2340 int group_imb; /* Is there an imbalance in the group ? */
2341 int group_has_capacity; /* Is there extra capacity in the group? */
2345 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2346 * @group: The group whose first cpu is to be returned.
2348 static inline unsigned int group_first_cpu(struct sched_group *group)
2350 return cpumask_first(sched_group_cpus(group));
2354 * get_sd_load_idx - Obtain the load index for a given sched domain.
2355 * @sd: The sched_domain whose load_idx is to be obtained.
2356 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2358 static inline int get_sd_load_idx(struct sched_domain *sd,
2359 enum cpu_idle_type idle)
2365 load_idx = sd->busy_idx;
2368 case CPU_NEWLY_IDLE:
2369 load_idx = sd->newidle_idx;
2372 load_idx = sd->idle_idx;
2380 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2382 * init_sd_power_savings_stats - Initialize power savings statistics for
2383 * the given sched_domain, during load balancing.
2385 * @sd: Sched domain whose power-savings statistics are to be initialized.
2386 * @sds: Variable containing the statistics for sd.
2387 * @idle: Idle status of the CPU at which we're performing load-balancing.
2389 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2390 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2393 * Busy processors will not participate in power savings
2396 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2397 sds->power_savings_balance = 0;
2399 sds->power_savings_balance = 1;
2400 sds->min_nr_running = ULONG_MAX;
2401 sds->leader_nr_running = 0;
2406 * update_sd_power_savings_stats - Update the power saving stats for a
2407 * sched_domain while performing load balancing.
2409 * @group: sched_group belonging to the sched_domain under consideration.
2410 * @sds: Variable containing the statistics of the sched_domain
2411 * @local_group: Does group contain the CPU for which we're performing
2413 * @sgs: Variable containing the statistics of the group.
2415 static inline void update_sd_power_savings_stats(struct sched_group *group,
2416 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2419 if (!sds->power_savings_balance)
2423 * If the local group is idle or completely loaded
2424 * no need to do power savings balance at this domain
2426 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2427 !sds->this_nr_running))
2428 sds->power_savings_balance = 0;
2431 * If a group is already running at full capacity or idle,
2432 * don't include that group in power savings calculations
2434 if (!sds->power_savings_balance ||
2435 sgs->sum_nr_running >= sgs->group_capacity ||
2436 !sgs->sum_nr_running)
2440 * Calculate the group which has the least non-idle load.
2441 * This is the group from where we need to pick up the load
2444 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2445 (sgs->sum_nr_running == sds->min_nr_running &&
2446 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2447 sds->group_min = group;
2448 sds->min_nr_running = sgs->sum_nr_running;
2449 sds->min_load_per_task = sgs->sum_weighted_load /
2450 sgs->sum_nr_running;
2454 * Calculate the group which is almost near its
2455 * capacity but still has some space to pick up some load
2456 * from other group and save more power
2458 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2461 if (sgs->sum_nr_running > sds->leader_nr_running ||
2462 (sgs->sum_nr_running == sds->leader_nr_running &&
2463 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2464 sds->group_leader = group;
2465 sds->leader_nr_running = sgs->sum_nr_running;
2470 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2471 * @sds: Variable containing the statistics of the sched_domain
2472 * under consideration.
2473 * @this_cpu: Cpu at which we're currently performing load-balancing.
2474 * @imbalance: Variable to store the imbalance.
2477 * Check if we have potential to perform some power-savings balance.
2478 * If yes, set the busiest group to be the least loaded group in the
2479 * sched_domain, so that it's CPUs can be put to idle.
2481 * Returns 1 if there is potential to perform power-savings balance.
2484 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2485 int this_cpu, unsigned long *imbalance)
2487 if (!sds->power_savings_balance)
2490 if (sds->this != sds->group_leader ||
2491 sds->group_leader == sds->group_min)
2494 *imbalance = sds->min_load_per_task;
2495 sds->busiest = sds->group_min;
2500 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2501 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2502 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2507 static inline void update_sd_power_savings_stats(struct sched_group *group,
2508 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2513 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2514 int this_cpu, unsigned long *imbalance)
2518 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2521 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2523 return SCHED_LOAD_SCALE;
2526 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2528 return default_scale_freq_power(sd, cpu);
2531 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2533 unsigned long weight = sd->span_weight;
2534 unsigned long smt_gain = sd->smt_gain;
2541 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2543 return default_scale_smt_power(sd, cpu);
2546 unsigned long scale_rt_power(int cpu)
2548 struct rq *rq = cpu_rq(cpu);
2549 u64 total, available;
2551 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2553 if (unlikely(total < rq->rt_avg)) {
2554 /* Ensures that power won't end up being negative */
2557 available = total - rq->rt_avg;
2560 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2561 total = SCHED_LOAD_SCALE;
2563 total >>= SCHED_LOAD_SHIFT;
2565 return div_u64(available, total);
2568 static void update_cpu_power(struct sched_domain *sd, int cpu)
2570 unsigned long weight = sd->span_weight;
2571 unsigned long power = SCHED_LOAD_SCALE;
2572 struct sched_group *sdg = sd->groups;
2574 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2575 if (sched_feat(ARCH_POWER))
2576 power *= arch_scale_smt_power(sd, cpu);
2578 power *= default_scale_smt_power(sd, cpu);
2580 power >>= SCHED_LOAD_SHIFT;
2583 sdg->cpu_power_orig = power;
2585 if (sched_feat(ARCH_POWER))
2586 power *= arch_scale_freq_power(sd, cpu);
2588 power *= default_scale_freq_power(sd, cpu);
2590 power >>= SCHED_LOAD_SHIFT;
2592 power *= scale_rt_power(cpu);
2593 power >>= SCHED_LOAD_SHIFT;
2598 cpu_rq(cpu)->cpu_power = power;
2599 sdg->cpu_power = power;
2602 static void update_group_power(struct sched_domain *sd, int cpu)
2604 struct sched_domain *child = sd->child;
2605 struct sched_group *group, *sdg = sd->groups;
2606 unsigned long power;
2609 update_cpu_power(sd, cpu);
2615 group = child->groups;
2617 power += group->cpu_power;
2618 group = group->next;
2619 } while (group != child->groups);
2621 sdg->cpu_power = power;
2625 * Try and fix up capacity for tiny siblings, this is needed when
2626 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2627 * which on its own isn't powerful enough.
2629 * See update_sd_pick_busiest() and check_asym_packing().
2632 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2635 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2637 if (sd->level != SD_LV_SIBLING)
2641 * If ~90% of the cpu_power is still there, we're good.
2643 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2650 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2651 * @sd: The sched_domain whose statistics are to be updated.
2652 * @group: sched_group whose statistics are to be updated.
2653 * @this_cpu: Cpu for which load balance is currently performed.
2654 * @idle: Idle status of this_cpu
2655 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2656 * @sd_idle: Idle status of the sched_domain containing group.
2657 * @local_group: Does group contain this_cpu.
2658 * @cpus: Set of cpus considered for load balancing.
2659 * @balance: Should we balance.
2660 * @sgs: variable to hold the statistics for this group.
2662 static inline void update_sg_lb_stats(struct sched_domain *sd,
2663 struct sched_group *group, int this_cpu,
2664 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2665 int local_group, const struct cpumask *cpus,
2666 int *balance, struct sg_lb_stats *sgs)
2668 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2670 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2671 unsigned long avg_load_per_task = 0;
2674 balance_cpu = group_first_cpu(group);
2676 /* Tally up the load of all CPUs in the group */
2678 min_cpu_load = ~0UL;
2681 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2682 struct rq *rq = cpu_rq(i);
2684 if (*sd_idle && rq->nr_running)
2687 /* Bias balancing toward cpus of our domain */
2689 if (idle_cpu(i) && !first_idle_cpu) {
2694 load = target_load(i, load_idx);
2696 load = source_load(i, load_idx);
2697 if (load > max_cpu_load) {
2698 max_cpu_load = load;
2699 max_nr_running = rq->nr_running;
2701 if (min_cpu_load > load)
2702 min_cpu_load = load;
2705 sgs->group_load += load;
2706 sgs->sum_nr_running += rq->nr_running;
2707 sgs->sum_weighted_load += weighted_cpuload(i);
2713 * First idle cpu or the first cpu(busiest) in this sched group
2714 * is eligible for doing load balancing at this and above
2715 * domains. In the newly idle case, we will allow all the cpu's
2716 * to do the newly idle load balance.
2718 if (idle != CPU_NEWLY_IDLE && local_group) {
2719 if (balance_cpu != this_cpu) {
2723 update_group_power(sd, this_cpu);
2726 /* Adjust by relative CPU power of the group */
2727 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2730 * Consider the group unbalanced when the imbalance is larger
2731 * than the average weight of two tasks.
2733 * APZ: with cgroup the avg task weight can vary wildly and
2734 * might not be a suitable number - should we keep a
2735 * normalized nr_running number somewhere that negates
2738 if (sgs->sum_nr_running)
2739 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2741 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2744 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2745 if (!sgs->group_capacity)
2746 sgs->group_capacity = fix_small_capacity(sd, group);
2747 sgs->group_weight = group->group_weight;
2749 if (sgs->group_capacity > sgs->sum_nr_running)
2750 sgs->group_has_capacity = 1;
2754 * update_sd_pick_busiest - return 1 on busiest group
2755 * @sd: sched_domain whose statistics are to be checked
2756 * @sds: sched_domain statistics
2757 * @sg: sched_group candidate to be checked for being the busiest
2758 * @sgs: sched_group statistics
2759 * @this_cpu: the current cpu
2761 * Determine if @sg is a busier group than the previously selected
2764 static bool update_sd_pick_busiest(struct sched_domain *sd,
2765 struct sd_lb_stats *sds,
2766 struct sched_group *sg,
2767 struct sg_lb_stats *sgs,
2770 if (sgs->avg_load <= sds->max_load)
2773 if (sgs->sum_nr_running > sgs->group_capacity)
2780 * ASYM_PACKING needs to move all the work to the lowest
2781 * numbered CPUs in the group, therefore mark all groups
2782 * higher than ourself as busy.
2784 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2785 this_cpu < group_first_cpu(sg)) {
2789 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2797 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2798 * @sd: sched_domain whose statistics are to be updated.
2799 * @this_cpu: Cpu for which load balance is currently performed.
2800 * @idle: Idle status of this_cpu
2801 * @sd_idle: Idle status of the sched_domain containing sg.
2802 * @cpus: Set of cpus considered for load balancing.
2803 * @balance: Should we balance.
2804 * @sds: variable to hold the statistics for this sched_domain.
2806 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2807 enum cpu_idle_type idle, int *sd_idle,
2808 const struct cpumask *cpus, int *balance,
2809 struct sd_lb_stats *sds)
2811 struct sched_domain *child = sd->child;
2812 struct sched_group *sg = sd->groups;
2813 struct sg_lb_stats sgs;
2814 int load_idx, prefer_sibling = 0;
2816 if (child && child->flags & SD_PREFER_SIBLING)
2819 init_sd_power_savings_stats(sd, sds, idle);
2820 load_idx = get_sd_load_idx(sd, idle);
2825 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2826 memset(&sgs, 0, sizeof(sgs));
2827 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2828 local_group, cpus, balance, &sgs);
2830 if (local_group && !(*balance))
2833 sds->total_load += sgs.group_load;
2834 sds->total_pwr += sg->cpu_power;
2837 * In case the child domain prefers tasks go to siblings
2838 * first, lower the sg capacity to one so that we'll try
2839 * and move all the excess tasks away. We lower the capacity
2840 * of a group only if the local group has the capacity to fit
2841 * these excess tasks, i.e. nr_running < group_capacity. The
2842 * extra check prevents the case where you always pull from the
2843 * heaviest group when it is already under-utilized (possible
2844 * with a large weight task outweighs the tasks on the system).
2846 if (prefer_sibling && !local_group && sds->this_has_capacity)
2847 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2850 sds->this_load = sgs.avg_load;
2852 sds->this_nr_running = sgs.sum_nr_running;
2853 sds->this_load_per_task = sgs.sum_weighted_load;
2854 sds->this_has_capacity = sgs.group_has_capacity;
2855 sds->this_idle_cpus = sgs.idle_cpus;
2856 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2857 sds->max_load = sgs.avg_load;
2859 sds->busiest_nr_running = sgs.sum_nr_running;
2860 sds->busiest_idle_cpus = sgs.idle_cpus;
2861 sds->busiest_group_capacity = sgs.group_capacity;
2862 sds->busiest_load_per_task = sgs.sum_weighted_load;
2863 sds->busiest_has_capacity = sgs.group_has_capacity;
2864 sds->busiest_group_weight = sgs.group_weight;
2865 sds->group_imb = sgs.group_imb;
2868 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2870 } while (sg != sd->groups);
2873 int __weak arch_sd_sibling_asym_packing(void)
2875 return 0*SD_ASYM_PACKING;
2879 * check_asym_packing - Check to see if the group is packed into the
2882 * This is primarily intended to used at the sibling level. Some
2883 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2884 * case of POWER7, it can move to lower SMT modes only when higher
2885 * threads are idle. When in lower SMT modes, the threads will
2886 * perform better since they share less core resources. Hence when we
2887 * have idle threads, we want them to be the higher ones.
2889 * This packing function is run on idle threads. It checks to see if
2890 * the busiest CPU in this domain (core in the P7 case) has a higher
2891 * CPU number than the packing function is being run on. Here we are
2892 * assuming lower CPU number will be equivalent to lower a SMT thread
2895 * Returns 1 when packing is required and a task should be moved to
2896 * this CPU. The amount of the imbalance is returned in *imbalance.
2898 * @sd: The sched_domain whose packing is to be checked.
2899 * @sds: Statistics of the sched_domain which is to be packed
2900 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2901 * @imbalance: returns amount of imbalanced due to packing.
2903 static int check_asym_packing(struct sched_domain *sd,
2904 struct sd_lb_stats *sds,
2905 int this_cpu, unsigned long *imbalance)
2909 if (!(sd->flags & SD_ASYM_PACKING))
2915 busiest_cpu = group_first_cpu(sds->busiest);
2916 if (this_cpu > busiest_cpu)
2919 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2925 * fix_small_imbalance - Calculate the minor imbalance that exists
2926 * amongst the groups of a sched_domain, during
2928 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2929 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2930 * @imbalance: Variable to store the imbalance.
2932 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2933 int this_cpu, unsigned long *imbalance)
2935 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2936 unsigned int imbn = 2;
2937 unsigned long scaled_busy_load_per_task;
2939 if (sds->this_nr_running) {
2940 sds->this_load_per_task /= sds->this_nr_running;
2941 if (sds->busiest_load_per_task >
2942 sds->this_load_per_task)
2945 sds->this_load_per_task =
2946 cpu_avg_load_per_task(this_cpu);
2948 scaled_busy_load_per_task = sds->busiest_load_per_task
2950 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2952 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2953 (scaled_busy_load_per_task * imbn)) {
2954 *imbalance = sds->busiest_load_per_task;
2959 * OK, we don't have enough imbalance to justify moving tasks,
2960 * however we may be able to increase total CPU power used by
2964 pwr_now += sds->busiest->cpu_power *
2965 min(sds->busiest_load_per_task, sds->max_load);
2966 pwr_now += sds->this->cpu_power *
2967 min(sds->this_load_per_task, sds->this_load);
2968 pwr_now /= SCHED_LOAD_SCALE;
2970 /* Amount of load we'd subtract */
2971 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2972 sds->busiest->cpu_power;
2973 if (sds->max_load > tmp)
2974 pwr_move += sds->busiest->cpu_power *
2975 min(sds->busiest_load_per_task, sds->max_load - tmp);
2977 /* Amount of load we'd add */
2978 if (sds->max_load * sds->busiest->cpu_power <
2979 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2980 tmp = (sds->max_load * sds->busiest->cpu_power) /
2981 sds->this->cpu_power;
2983 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2984 sds->this->cpu_power;
2985 pwr_move += sds->this->cpu_power *
2986 min(sds->this_load_per_task, sds->this_load + tmp);
2987 pwr_move /= SCHED_LOAD_SCALE;
2989 /* Move if we gain throughput */
2990 if (pwr_move > pwr_now)
2991 *imbalance = sds->busiest_load_per_task;
2995 * calculate_imbalance - Calculate the amount of imbalance present within the
2996 * groups of a given sched_domain during load balance.
2997 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2998 * @this_cpu: Cpu for which currently load balance is being performed.
2999 * @imbalance: The variable to store the imbalance.
3001 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3002 unsigned long *imbalance)
3004 unsigned long max_pull, load_above_capacity = ~0UL;
3006 sds->busiest_load_per_task /= sds->busiest_nr_running;
3007 if (sds->group_imb) {
3008 sds->busiest_load_per_task =
3009 min(sds->busiest_load_per_task, sds->avg_load);
3013 * In the presence of smp nice balancing, certain scenarios can have
3014 * max load less than avg load(as we skip the groups at or below
3015 * its cpu_power, while calculating max_load..)
3017 if (sds->max_load < sds->avg_load) {
3019 return fix_small_imbalance(sds, this_cpu, imbalance);
3022 if (!sds->group_imb) {
3024 * Don't want to pull so many tasks that a group would go idle.
3026 load_above_capacity = (sds->busiest_nr_running -
3027 sds->busiest_group_capacity);
3029 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
3031 load_above_capacity /= sds->busiest->cpu_power;
3035 * We're trying to get all the cpus to the average_load, so we don't
3036 * want to push ourselves above the average load, nor do we wish to
3037 * reduce the max loaded cpu below the average load. At the same time,
3038 * we also don't want to reduce the group load below the group capacity
3039 * (so that we can implement power-savings policies etc). Thus we look
3040 * for the minimum possible imbalance.
3041 * Be careful of negative numbers as they'll appear as very large values
3042 * with unsigned longs.
3044 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3046 /* How much load to actually move to equalise the imbalance */
3047 *imbalance = min(max_pull * sds->busiest->cpu_power,
3048 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3052 * if *imbalance is less than the average load per runnable task
3053 * there is no gaurantee that any tasks will be moved so we'll have
3054 * a think about bumping its value to force at least one task to be
3057 if (*imbalance < sds->busiest_load_per_task)
3058 return fix_small_imbalance(sds, this_cpu, imbalance);
3062 /******* find_busiest_group() helpers end here *********************/
3065 * find_busiest_group - Returns the busiest group within the sched_domain
3066 * if there is an imbalance. If there isn't an imbalance, and
3067 * the user has opted for power-savings, it returns a group whose
3068 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3069 * such a group exists.
3071 * Also calculates the amount of weighted load which should be moved
3072 * to restore balance.
3074 * @sd: The sched_domain whose busiest group is to be returned.
3075 * @this_cpu: The cpu for which load balancing is currently being performed.
3076 * @imbalance: Variable which stores amount of weighted load which should
3077 * be moved to restore balance/put a group to idle.
3078 * @idle: The idle status of this_cpu.
3079 * @sd_idle: The idleness of sd
3080 * @cpus: The set of CPUs under consideration for load-balancing.
3081 * @balance: Pointer to a variable indicating if this_cpu
3082 * is the appropriate cpu to perform load balancing at this_level.
3084 * Returns: - the busiest group if imbalance exists.
3085 * - If no imbalance and user has opted for power-savings balance,
3086 * return the least loaded group whose CPUs can be
3087 * put to idle by rebalancing its tasks onto our group.
3089 static struct sched_group *
3090 find_busiest_group(struct sched_domain *sd, int this_cpu,
3091 unsigned long *imbalance, enum cpu_idle_type idle,
3092 int *sd_idle, const struct cpumask *cpus, int *balance)
3094 struct sd_lb_stats sds;
3096 memset(&sds, 0, sizeof(sds));
3099 * Compute the various statistics relavent for load balancing at
3102 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3105 /* Cases where imbalance does not exist from POV of this_cpu */
3106 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3108 * 2) There is no busy sibling group to pull from.
3109 * 3) This group is the busiest group.
3110 * 4) This group is more busy than the avg busieness at this
3112 * 5) The imbalance is within the specified limit.
3114 * Note: when doing newidle balance, if the local group has excess
3115 * capacity (i.e. nr_running < group_capacity) and the busiest group
3116 * does not have any capacity, we force a load balance to pull tasks
3117 * to the local group. In this case, we skip past checks 3, 4 and 5.
3122 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3123 check_asym_packing(sd, &sds, this_cpu, imbalance))
3126 if (!sds.busiest || sds.busiest_nr_running == 0)
3129 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3130 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3131 !sds.busiest_has_capacity)
3134 if (sds.this_load >= sds.max_load)
3137 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3139 if (sds.this_load >= sds.avg_load)
3143 * In the CPU_NEWLY_IDLE, use imbalance_pct to be conservative.
3144 * And to check for busy balance use !idle_cpu instead of
3145 * CPU_NOT_IDLE. This is because HT siblings will use CPU_NOT_IDLE
3146 * even when they are idle.
3148 if (idle == CPU_NEWLY_IDLE || !idle_cpu(this_cpu)) {
3149 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3153 * This cpu is idle. If the busiest group load doesn't
3154 * have more tasks than the number of available cpu's and
3155 * there is no imbalance between this and busiest group
3156 * wrt to idle cpu's, it is balanced.
3158 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3159 sds.busiest_nr_running <= sds.busiest_group_weight)
3164 /* Looks like there is an imbalance. Compute it */
3165 calculate_imbalance(&sds, this_cpu, imbalance);
3170 * There is no obvious imbalance. But check if we can do some balancing
3173 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3181 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3184 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3185 enum cpu_idle_type idle, unsigned long imbalance,
3186 const struct cpumask *cpus)
3188 struct rq *busiest = NULL, *rq;
3189 unsigned long max_load = 0;
3192 for_each_cpu(i, sched_group_cpus(group)) {
3193 unsigned long power = power_of(i);
3194 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3198 capacity = fix_small_capacity(sd, group);
3200 if (!cpumask_test_cpu(i, cpus))
3204 wl = weighted_cpuload(i);
3207 * When comparing with imbalance, use weighted_cpuload()
3208 * which is not scaled with the cpu power.
3210 if (capacity && rq->nr_running == 1 && wl > imbalance)
3214 * For the load comparisons with the other cpu's, consider
3215 * the weighted_cpuload() scaled with the cpu power, so that
3216 * the load can be moved away from the cpu that is potentially
3217 * running at a lower capacity.
3219 wl = (wl * SCHED_LOAD_SCALE) / power;
3221 if (wl > max_load) {
3231 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3232 * so long as it is large enough.
3234 #define MAX_PINNED_INTERVAL 512
3236 /* Working cpumask for load_balance and load_balance_newidle. */
3237 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3239 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
3240 int busiest_cpu, int this_cpu)
3242 if (idle == CPU_NEWLY_IDLE) {
3245 * ASYM_PACKING needs to force migrate tasks from busy but
3246 * higher numbered CPUs in order to pack all tasks in the
3247 * lowest numbered CPUs.
3249 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3253 * The only task running in a non-idle cpu can be moved to this
3254 * cpu in an attempt to completely freeup the other CPU
3257 * The package power saving logic comes from
3258 * find_busiest_group(). If there are no imbalance, then
3259 * f_b_g() will return NULL. However when sched_mc={1,2} then
3260 * f_b_g() will select a group from which a running task may be
3261 * pulled to this cpu in order to make the other package idle.
3262 * If there is no opportunity to make a package idle and if
3263 * there are no imbalance, then f_b_g() will return NULL and no
3264 * action will be taken in load_balance_newidle().
3266 * Under normal task pull operation due to imbalance, there
3267 * will be more than one task in the source run queue and
3268 * move_tasks() will succeed. ld_moved will be true and this
3269 * active balance code will not be triggered.
3271 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3272 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3275 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3279 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3282 static int active_load_balance_cpu_stop(void *data);
3285 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3286 * tasks if there is an imbalance.
3288 static int load_balance(int this_cpu, struct rq *this_rq,
3289 struct sched_domain *sd, enum cpu_idle_type idle,
3292 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3293 struct sched_group *group;
3294 unsigned long imbalance;
3296 unsigned long flags;
3297 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3299 cpumask_copy(cpus, cpu_active_mask);
3302 * When power savings policy is enabled for the parent domain, idle
3303 * sibling can pick up load irrespective of busy siblings. In this case,
3304 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3305 * portraying it as CPU_NOT_IDLE.
3307 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3308 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3311 schedstat_inc(sd, lb_count[idle]);
3314 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3321 schedstat_inc(sd, lb_nobusyg[idle]);
3325 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3327 schedstat_inc(sd, lb_nobusyq[idle]);
3331 BUG_ON(busiest == this_rq);
3333 schedstat_add(sd, lb_imbalance[idle], imbalance);
3336 if (busiest->nr_running > 1) {
3338 * Attempt to move tasks. If find_busiest_group has found
3339 * an imbalance but busiest->nr_running <= 1, the group is
3340 * still unbalanced. ld_moved simply stays zero, so it is
3341 * correctly treated as an imbalance.
3343 local_irq_save(flags);
3344 double_rq_lock(this_rq, busiest);
3345 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3346 imbalance, sd, idle, &all_pinned);
3347 double_rq_unlock(this_rq, busiest);
3348 local_irq_restore(flags);
3351 * some other cpu did the load balance for us.
3353 if (ld_moved && this_cpu != smp_processor_id())
3354 resched_cpu(this_cpu);
3356 /* All tasks on this runqueue were pinned by CPU affinity */
3357 if (unlikely(all_pinned)) {
3358 cpumask_clear_cpu(cpu_of(busiest), cpus);
3359 if (!cpumask_empty(cpus))
3366 schedstat_inc(sd, lb_failed[idle]);
3368 * Increment the failure counter only on periodic balance.
3369 * We do not want newidle balance, which can be very
3370 * frequent, pollute the failure counter causing
3371 * excessive cache_hot migrations and active balances.
3373 if (idle != CPU_NEWLY_IDLE)
3374 sd->nr_balance_failed++;
3376 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3378 raw_spin_lock_irqsave(&busiest->lock, flags);
3380 /* don't kick the active_load_balance_cpu_stop,
3381 * if the curr task on busiest cpu can't be
3384 if (!cpumask_test_cpu(this_cpu,
3385 &busiest->curr->cpus_allowed)) {
3386 raw_spin_unlock_irqrestore(&busiest->lock,
3389 goto out_one_pinned;
3393 * ->active_balance synchronizes accesses to
3394 * ->active_balance_work. Once set, it's cleared
3395 * only after active load balance is finished.
3397 if (!busiest->active_balance) {
3398 busiest->active_balance = 1;
3399 busiest->push_cpu = this_cpu;
3402 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3405 stop_one_cpu_nowait(cpu_of(busiest),
3406 active_load_balance_cpu_stop, busiest,
3407 &busiest->active_balance_work);
3410 * We've kicked active balancing, reset the failure
3413 sd->nr_balance_failed = sd->cache_nice_tries+1;
3416 sd->nr_balance_failed = 0;
3418 if (likely(!active_balance)) {
3419 /* We were unbalanced, so reset the balancing interval */
3420 sd->balance_interval = sd->min_interval;
3423 * If we've begun active balancing, start to back off. This
3424 * case may not be covered by the all_pinned logic if there
3425 * is only 1 task on the busy runqueue (because we don't call
3428 if (sd->balance_interval < sd->max_interval)
3429 sd->balance_interval *= 2;
3432 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3433 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3439 schedstat_inc(sd, lb_balanced[idle]);
3441 sd->nr_balance_failed = 0;
3444 /* tune up the balancing interval */
3445 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3446 (sd->balance_interval < sd->max_interval))
3447 sd->balance_interval *= 2;
3449 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3450 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3459 * idle_balance is called by schedule() if this_cpu is about to become
3460 * idle. Attempts to pull tasks from other CPUs.
3462 static void idle_balance(int this_cpu, struct rq *this_rq)
3464 struct sched_domain *sd;
3465 int pulled_task = 0;
3466 unsigned long next_balance = jiffies + HZ;
3468 this_rq->idle_stamp = this_rq->clock;
3470 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3474 * Drop the rq->lock, but keep IRQ/preempt disabled.
3476 raw_spin_unlock(&this_rq->lock);
3478 update_shares(this_cpu);
3479 for_each_domain(this_cpu, sd) {
3480 unsigned long interval;
3483 if (!(sd->flags & SD_LOAD_BALANCE))
3486 if (sd->flags & SD_BALANCE_NEWIDLE) {
3487 /* If we've pulled tasks over stop searching: */
3488 pulled_task = load_balance(this_cpu, this_rq,
3489 sd, CPU_NEWLY_IDLE, &balance);
3492 interval = msecs_to_jiffies(sd->balance_interval);
3493 if (time_after(next_balance, sd->last_balance + interval))
3494 next_balance = sd->last_balance + interval;
3496 this_rq->idle_stamp = 0;
3501 raw_spin_lock(&this_rq->lock);
3503 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3505 * We are going idle. next_balance may be set based on
3506 * a busy processor. So reset next_balance.
3508 this_rq->next_balance = next_balance;
3513 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3514 * running tasks off the busiest CPU onto idle CPUs. It requires at
3515 * least 1 task to be running on each physical CPU where possible, and
3516 * avoids physical / logical imbalances.
3518 static int active_load_balance_cpu_stop(void *data)
3520 struct rq *busiest_rq = data;
3521 int busiest_cpu = cpu_of(busiest_rq);
3522 int target_cpu = busiest_rq->push_cpu;
3523 struct rq *target_rq = cpu_rq(target_cpu);
3524 struct sched_domain *sd;
3526 raw_spin_lock_irq(&busiest_rq->lock);
3528 /* make sure the requested cpu hasn't gone down in the meantime */
3529 if (unlikely(busiest_cpu != smp_processor_id() ||
3530 !busiest_rq->active_balance))
3533 /* Is there any task to move? */
3534 if (busiest_rq->nr_running <= 1)
3538 * This condition is "impossible", if it occurs
3539 * we need to fix it. Originally reported by
3540 * Bjorn Helgaas on a 128-cpu setup.
3542 BUG_ON(busiest_rq == target_rq);
3544 /* move a task from busiest_rq to target_rq */
3545 double_lock_balance(busiest_rq, target_rq);
3547 /* Search for an sd spanning us and the target CPU. */
3548 for_each_domain(target_cpu, sd) {
3549 if ((sd->flags & SD_LOAD_BALANCE) &&
3550 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3555 schedstat_inc(sd, alb_count);
3557 if (move_one_task(target_rq, target_cpu, busiest_rq,
3559 schedstat_inc(sd, alb_pushed);
3561 schedstat_inc(sd, alb_failed);
3563 double_unlock_balance(busiest_rq, target_rq);
3565 busiest_rq->active_balance = 0;
3566 raw_spin_unlock_irq(&busiest_rq->lock);
3572 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3574 static void trigger_sched_softirq(void *data)
3576 raise_softirq_irqoff(SCHED_SOFTIRQ);
3579 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3581 csd->func = trigger_sched_softirq;
3588 * idle load balancing details
3589 * - One of the idle CPUs nominates itself as idle load_balancer, while
3591 * - This idle load balancer CPU will also go into tickless mode when
3592 * it is idle, just like all other idle CPUs
3593 * - When one of the busy CPUs notice that there may be an idle rebalancing
3594 * needed, they will kick the idle load balancer, which then does idle
3595 * load balancing for all the idle CPUs.
3598 atomic_t load_balancer;
3599 atomic_t first_pick_cpu;
3600 atomic_t second_pick_cpu;
3601 cpumask_var_t idle_cpus_mask;
3602 cpumask_var_t grp_idle_mask;
3603 unsigned long next_balance; /* in jiffy units */
3604 } nohz ____cacheline_aligned;
3606 int get_nohz_load_balancer(void)
3608 return atomic_read(&nohz.load_balancer);
3611 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3613 * lowest_flag_domain - Return lowest sched_domain containing flag.
3614 * @cpu: The cpu whose lowest level of sched domain is to
3616 * @flag: The flag to check for the lowest sched_domain
3617 * for the given cpu.
3619 * Returns the lowest sched_domain of a cpu which contains the given flag.
3621 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3623 struct sched_domain *sd;
3625 for_each_domain(cpu, sd)
3626 if (sd && (sd->flags & flag))
3633 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3634 * @cpu: The cpu whose domains we're iterating over.
3635 * @sd: variable holding the value of the power_savings_sd
3637 * @flag: The flag to filter the sched_domains to be iterated.
3639 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3640 * set, starting from the lowest sched_domain to the highest.
3642 #define for_each_flag_domain(cpu, sd, flag) \
3643 for (sd = lowest_flag_domain(cpu, flag); \
3644 (sd && (sd->flags & flag)); sd = sd->parent)
3647 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3648 * @ilb_group: group to be checked for semi-idleness
3650 * Returns: 1 if the group is semi-idle. 0 otherwise.
3652 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3653 * and atleast one non-idle CPU. This helper function checks if the given
3654 * sched_group is semi-idle or not.
3656 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3658 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3659 sched_group_cpus(ilb_group));
3662 * A sched_group is semi-idle when it has atleast one busy cpu
3663 * and atleast one idle cpu.
3665 if (cpumask_empty(nohz.grp_idle_mask))
3668 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3674 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3675 * @cpu: The cpu which is nominating a new idle_load_balancer.
3677 * Returns: Returns the id of the idle load balancer if it exists,
3678 * Else, returns >= nr_cpu_ids.
3680 * This algorithm picks the idle load balancer such that it belongs to a
3681 * semi-idle powersavings sched_domain. The idea is to try and avoid
3682 * completely idle packages/cores just for the purpose of idle load balancing
3683 * when there are other idle cpu's which are better suited for that job.
3685 static int find_new_ilb(int cpu)
3687 struct sched_domain *sd;
3688 struct sched_group *ilb_group;
3691 * Have idle load balancer selection from semi-idle packages only
3692 * when power-aware load balancing is enabled
3694 if (!(sched_smt_power_savings || sched_mc_power_savings))
3698 * Optimize for the case when we have no idle CPUs or only one
3699 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3701 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3704 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3705 ilb_group = sd->groups;
3708 if (is_semi_idle_group(ilb_group))
3709 return cpumask_first(nohz.grp_idle_mask);
3711 ilb_group = ilb_group->next;
3713 } while (ilb_group != sd->groups);
3719 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3720 static inline int find_new_ilb(int call_cpu)
3727 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3728 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3729 * CPU (if there is one).
3731 static void nohz_balancer_kick(int cpu)
3735 nohz.next_balance++;
3737 ilb_cpu = get_nohz_load_balancer();
3739 if (ilb_cpu >= nr_cpu_ids) {
3740 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3741 if (ilb_cpu >= nr_cpu_ids)
3745 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3746 struct call_single_data *cp;
3748 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3749 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3750 __smp_call_function_single(ilb_cpu, cp, 0);
3756 * This routine will try to nominate the ilb (idle load balancing)
3757 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3758 * load balancing on behalf of all those cpus.
3760 * When the ilb owner becomes busy, we will not have new ilb owner until some
3761 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3762 * idle load balancing by kicking one of the idle CPUs.
3764 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3765 * ilb owner CPU in future (when there is a need for idle load balancing on
3766 * behalf of all idle CPUs).
3768 void select_nohz_load_balancer(int stop_tick)
3770 int cpu = smp_processor_id();
3773 if (!cpu_active(cpu)) {
3774 if (atomic_read(&nohz.load_balancer) != cpu)
3778 * If we are going offline and still the leader,
3781 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3788 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3790 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3791 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3792 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3793 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3795 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3798 /* make me the ilb owner */
3799 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3804 * Check to see if there is a more power-efficient
3807 new_ilb = find_new_ilb(cpu);
3808 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3809 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3810 resched_cpu(new_ilb);
3816 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3819 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3821 if (atomic_read(&nohz.load_balancer) == cpu)
3822 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3830 static DEFINE_SPINLOCK(balancing);
3833 * It checks each scheduling domain to see if it is due to be balanced,
3834 * and initiates a balancing operation if so.
3836 * Balancing parameters are set up in arch_init_sched_domains.
3838 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3841 struct rq *rq = cpu_rq(cpu);
3842 unsigned long interval;
3843 struct sched_domain *sd;
3844 /* Earliest time when we have to do rebalance again */
3845 unsigned long next_balance = jiffies + 60*HZ;
3846 int update_next_balance = 0;
3851 for_each_domain(cpu, sd) {
3852 if (!(sd->flags & SD_LOAD_BALANCE))
3855 interval = sd->balance_interval;
3856 if (idle != CPU_IDLE)
3857 interval *= sd->busy_factor;
3859 /* scale ms to jiffies */
3860 interval = msecs_to_jiffies(interval);
3861 if (unlikely(!interval))
3863 if (interval > HZ*NR_CPUS/10)
3864 interval = HZ*NR_CPUS/10;
3866 need_serialize = sd->flags & SD_SERIALIZE;
3868 if (need_serialize) {
3869 if (!spin_trylock(&balancing))
3873 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3874 if (load_balance(cpu, rq, sd, idle, &balance)) {
3876 * We've pulled tasks over so either we're no
3877 * longer idle, or one of our SMT siblings is
3880 idle = CPU_NOT_IDLE;
3882 sd->last_balance = jiffies;
3885 spin_unlock(&balancing);
3887 if (time_after(next_balance, sd->last_balance + interval)) {
3888 next_balance = sd->last_balance + interval;
3889 update_next_balance = 1;
3893 * Stop the load balance at this level. There is another
3894 * CPU in our sched group which is doing load balancing more
3902 * next_balance will be updated only when there is a need.
3903 * When the cpu is attached to null domain for ex, it will not be
3906 if (likely(update_next_balance))
3907 rq->next_balance = next_balance;
3912 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3913 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3915 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3917 struct rq *this_rq = cpu_rq(this_cpu);
3921 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3924 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3925 if (balance_cpu == this_cpu)
3929 * If this cpu gets work to do, stop the load balancing
3930 * work being done for other cpus. Next load
3931 * balancing owner will pick it up.
3933 if (need_resched()) {
3934 this_rq->nohz_balance_kick = 0;
3938 raw_spin_lock_irq(&this_rq->lock);
3939 update_rq_clock(this_rq);
3940 update_cpu_load(this_rq);
3941 raw_spin_unlock_irq(&this_rq->lock);
3943 rebalance_domains(balance_cpu, CPU_IDLE);
3945 rq = cpu_rq(balance_cpu);
3946 if (time_after(this_rq->next_balance, rq->next_balance))
3947 this_rq->next_balance = rq->next_balance;
3949 nohz.next_balance = this_rq->next_balance;
3950 this_rq->nohz_balance_kick = 0;
3954 * Current heuristic for kicking the idle load balancer
3955 * - first_pick_cpu is the one of the busy CPUs. It will kick
3956 * idle load balancer when it has more than one process active. This
3957 * eliminates the need for idle load balancing altogether when we have
3958 * only one running process in the system (common case).
3959 * - If there are more than one busy CPU, idle load balancer may have
3960 * to run for active_load_balance to happen (i.e., two busy CPUs are
3961 * SMT or core siblings and can run better if they move to different
3962 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3963 * which will kick idle load balancer as soon as it has any load.
3965 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3967 unsigned long now = jiffies;
3969 int first_pick_cpu, second_pick_cpu;
3971 if (time_before(now, nohz.next_balance))
3974 if (rq->idle_at_tick)
3977 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3978 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3980 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3981 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3984 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3985 if (ret == nr_cpu_ids || ret == cpu) {
3986 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3987 if (rq->nr_running > 1)
3990 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3991 if (ret == nr_cpu_ids || ret == cpu) {
3999 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4003 * run_rebalance_domains is triggered when needed from the scheduler tick.
4004 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4006 static void run_rebalance_domains(struct softirq_action *h)
4008 int this_cpu = smp_processor_id();
4009 struct rq *this_rq = cpu_rq(this_cpu);
4010 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4011 CPU_IDLE : CPU_NOT_IDLE;
4013 rebalance_domains(this_cpu, idle);
4016 * If this cpu has a pending nohz_balance_kick, then do the
4017 * balancing on behalf of the other idle cpus whose ticks are
4020 nohz_idle_balance(this_cpu, idle);
4023 static inline int on_null_domain(int cpu)
4025 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4029 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4031 static inline void trigger_load_balance(struct rq *rq, int cpu)
4033 /* Don't need to rebalance while attached to NULL domain */
4034 if (time_after_eq(jiffies, rq->next_balance) &&
4035 likely(!on_null_domain(cpu)))
4036 raise_softirq(SCHED_SOFTIRQ);
4038 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4039 nohz_balancer_kick(cpu);
4043 static void rq_online_fair(struct rq *rq)
4048 static void rq_offline_fair(struct rq *rq)
4053 #else /* CONFIG_SMP */
4056 * on UP we do not need to balance between CPUs:
4058 static inline void idle_balance(int cpu, struct rq *rq)
4062 #endif /* CONFIG_SMP */
4065 * scheduler tick hitting a task of our scheduling class:
4067 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4069 struct cfs_rq *cfs_rq;
4070 struct sched_entity *se = &curr->se;
4072 for_each_sched_entity(se) {
4073 cfs_rq = cfs_rq_of(se);
4074 entity_tick(cfs_rq, se, queued);
4079 * called on fork with the child task as argument from the parent's context
4080 * - child not yet on the tasklist
4081 * - preemption disabled
4083 static void task_fork_fair(struct task_struct *p)
4085 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4086 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4087 int this_cpu = smp_processor_id();
4088 struct rq *rq = this_rq();
4089 unsigned long flags;
4091 raw_spin_lock_irqsave(&rq->lock, flags);
4093 update_rq_clock(rq);
4095 if (unlikely(task_cpu(p) != this_cpu)) {
4097 __set_task_cpu(p, this_cpu);
4101 update_curr(cfs_rq);
4104 se->vruntime = curr->vruntime;
4105 place_entity(cfs_rq, se, 1);
4107 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4109 * Upon rescheduling, sched_class::put_prev_task() will place
4110 * 'current' within the tree based on its new key value.
4112 swap(curr->vruntime, se->vruntime);
4113 resched_task(rq->curr);
4116 se->vruntime -= cfs_rq->min_vruntime;
4118 raw_spin_unlock_irqrestore(&rq->lock, flags);
4122 * Priority of the task has changed. Check to see if we preempt
4126 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4132 * Reschedule if we are currently running on this runqueue and
4133 * our priority decreased, or if we are not currently running on
4134 * this runqueue and our priority is higher than the current's
4136 if (rq->curr == p) {
4137 if (p->prio > oldprio)
4138 resched_task(rq->curr);
4140 check_preempt_curr(rq, p, 0);
4143 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4145 struct sched_entity *se = &p->se;
4146 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4149 * Ensure the task's vruntime is normalized, so that when its
4150 * switched back to the fair class the enqueue_entity(.flags=0) will
4151 * do the right thing.
4153 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4154 * have normalized the vruntime, if it was !on_rq, then only when
4155 * the task is sleeping will it still have non-normalized vruntime.
4157 if (!se->on_rq && p->state != TASK_RUNNING) {
4159 * Fix up our vruntime so that the current sleep doesn't
4160 * cause 'unlimited' sleep bonus.
4162 place_entity(cfs_rq, se, 0);
4163 se->vruntime -= cfs_rq->min_vruntime;
4168 * We switched to the sched_fair class.
4170 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4176 * We were most likely switched from sched_rt, so
4177 * kick off the schedule if running, otherwise just see
4178 * if we can still preempt the current task.
4181 resched_task(rq->curr);
4183 check_preempt_curr(rq, p, 0);
4186 /* Account for a task changing its policy or group.
4188 * This routine is mostly called to set cfs_rq->curr field when a task
4189 * migrates between groups/classes.
4191 static void set_curr_task_fair(struct rq *rq)
4193 struct sched_entity *se = &rq->curr->se;
4195 for_each_sched_entity(se)
4196 set_next_entity(cfs_rq_of(se), se);
4199 #ifdef CONFIG_FAIR_GROUP_SCHED
4200 static void task_move_group_fair(struct task_struct *p, int on_rq)
4203 * If the task was not on the rq at the time of this cgroup movement
4204 * it must have been asleep, sleeping tasks keep their ->vruntime
4205 * absolute on their old rq until wakeup (needed for the fair sleeper
4206 * bonus in place_entity()).
4208 * If it was on the rq, we've just 'preempted' it, which does convert
4209 * ->vruntime to a relative base.
4211 * Make sure both cases convert their relative position when migrating
4212 * to another cgroup's rq. This does somewhat interfere with the
4213 * fair sleeper stuff for the first placement, but who cares.
4216 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4217 set_task_rq(p, task_cpu(p));
4219 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4223 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4225 struct sched_entity *se = &task->se;
4226 unsigned int rr_interval = 0;
4229 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4232 if (rq->cfs.load.weight)
4233 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4239 * All the scheduling class methods:
4241 static const struct sched_class fair_sched_class = {
4242 .next = &idle_sched_class,
4243 .enqueue_task = enqueue_task_fair,
4244 .dequeue_task = dequeue_task_fair,
4245 .yield_task = yield_task_fair,
4247 .check_preempt_curr = check_preempt_wakeup,
4249 .pick_next_task = pick_next_task_fair,
4250 .put_prev_task = put_prev_task_fair,
4253 .select_task_rq = select_task_rq_fair,
4255 .rq_online = rq_online_fair,
4256 .rq_offline = rq_offline_fair,
4258 .task_waking = task_waking_fair,
4261 .set_curr_task = set_curr_task_fair,
4262 .task_tick = task_tick_fair,
4263 .task_fork = task_fork_fair,
4265 .prio_changed = prio_changed_fair,
4266 .switched_from = switched_from_fair,
4267 .switched_to = switched_to_fair,
4269 .get_rr_interval = get_rr_interval_fair,
4271 #ifdef CONFIG_FAIR_GROUP_SCHED
4272 .task_move_group = task_move_group_fair,
4276 #ifdef CONFIG_SCHED_DEBUG
4277 static void print_cfs_stats(struct seq_file *m, int cpu)
4279 struct cfs_rq *cfs_rq;
4282 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4283 print_cfs_rq(m, cpu, cfs_rq);