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>
25 #include <linux/cpumask.h>
28 * Targeted preemption latency for CPU-bound tasks:
29 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
31 * NOTE: this latency value is not the same as the concept of
32 * 'timeslice length' - timeslices in CFS are of variable length
33 * and have no persistent notion like in traditional, time-slice
34 * based scheduling concepts.
36 * (to see the precise effective timeslice length of your workload,
37 * run vmstat and monitor the context-switches (cs) field)
39 unsigned int sysctl_sched_latency = 6000000ULL;
40 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
43 * The initial- and re-scaling of tunables is configurable
44 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
47 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
48 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
49 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 enum sched_tunable_scaling sysctl_sched_tunable_scaling
52 = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
56 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
58 unsigned int sysctl_sched_min_granularity = 750000ULL;
59 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
62 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
64 static unsigned int sched_nr_latency = 8;
67 * After fork, child runs first. If set to 0 (default) then
68 * parent will (try to) run first.
70 unsigned int sysctl_sched_child_runs_first __read_mostly;
73 * SCHED_OTHER wake-up granularity.
74 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
81 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
83 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
86 * The exponential sliding window over which load is averaged for shares
90 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
92 static const struct sched_class fair_sched_class;
94 /**************************************************************
95 * CFS operations on generic schedulable entities:
98 #ifdef CONFIG_FAIR_GROUP_SCHED
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
106 /* An entity is a task if it doesn't "own" a runqueue */
107 #define entity_is_task(se) (!se->my_q)
109 static inline struct task_struct *task_of(struct sched_entity *se)
111 #ifdef CONFIG_SCHED_DEBUG
112 WARN_ON_ONCE(!entity_is_task(se));
114 return container_of(se, struct task_struct, se);
117 /* Walk up scheduling entities hierarchy */
118 #define for_each_sched_entity(se) \
119 for (; se; se = se->parent)
121 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
126 /* runqueue on which this entity is (to be) queued */
127 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
132 /* runqueue "owned" by this group */
133 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
138 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
139 * another cpu ('this_cpu')
141 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
143 return cfs_rq->tg->cfs_rq[this_cpu];
146 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
148 if (!cfs_rq->on_list) {
150 * Ensure we either appear before our parent (if already
151 * enqueued) or force our parent to appear after us when it is
152 * enqueued. The fact that we always enqueue bottom-up
153 * reduces this to two cases.
155 if (cfs_rq->tg->parent &&
156 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
157 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
158 &rq_of(cfs_rq)->leaf_cfs_rq_list);
160 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
161 &rq_of(cfs_rq)->leaf_cfs_rq_list);
168 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
170 if (cfs_rq->on_list) {
171 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
176 /* Iterate thr' all leaf cfs_rq's on a runqueue */
177 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
178 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
180 /* Do the two (enqueued) entities belong to the same group ? */
182 is_same_group(struct sched_entity *se, struct sched_entity *pse)
184 if (se->cfs_rq == pse->cfs_rq)
190 static inline struct sched_entity *parent_entity(struct sched_entity *se)
195 /* return depth at which a sched entity is present in the hierarchy */
196 static inline int depth_se(struct sched_entity *se)
200 for_each_sched_entity(se)
207 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
209 int se_depth, pse_depth;
212 * preemption test can be made between sibling entities who are in the
213 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
214 * both tasks until we find their ancestors who are siblings of common
218 /* First walk up until both entities are at same depth */
219 se_depth = depth_se(*se);
220 pse_depth = depth_se(*pse);
222 while (se_depth > pse_depth) {
224 *se = parent_entity(*se);
227 while (pse_depth > se_depth) {
229 *pse = parent_entity(*pse);
232 while (!is_same_group(*se, *pse)) {
233 *se = parent_entity(*se);
234 *pse = parent_entity(*pse);
238 #else /* !CONFIG_FAIR_GROUP_SCHED */
240 static inline struct task_struct *task_of(struct sched_entity *se)
242 return container_of(se, struct task_struct, se);
245 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
247 return container_of(cfs_rq, struct rq, cfs);
250 #define entity_is_task(se) 1
252 #define for_each_sched_entity(se) \
253 for (; se; se = NULL)
255 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
257 return &task_rq(p)->cfs;
260 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
262 struct task_struct *p = task_of(se);
263 struct rq *rq = task_rq(p);
268 /* runqueue "owned" by this group */
269 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
274 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
276 return &cpu_rq(this_cpu)->cfs;
279 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
283 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
288 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
291 is_same_group(struct sched_entity *se, struct sched_entity *pse)
296 static inline struct sched_entity *parent_entity(struct sched_entity *se)
302 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
306 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 /**************************************************************
310 * Scheduling class tree data structure manipulation methods:
313 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
315 s64 delta = (s64)(vruntime - min_vruntime);
317 min_vruntime = vruntime;
322 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
324 s64 delta = (s64)(vruntime - min_vruntime);
326 min_vruntime = vruntime;
331 static inline int entity_before(struct sched_entity *a,
332 struct sched_entity *b)
334 return (s64)(a->vruntime - b->vruntime) < 0;
337 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
339 return se->vruntime - cfs_rq->min_vruntime;
342 static void update_min_vruntime(struct cfs_rq *cfs_rq)
344 u64 vruntime = cfs_rq->min_vruntime;
347 vruntime = cfs_rq->curr->vruntime;
349 if (cfs_rq->rb_leftmost) {
350 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
355 vruntime = se->vruntime;
357 vruntime = min_vruntime(vruntime, se->vruntime);
360 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
364 * Enqueue an entity into the rb-tree:
366 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
368 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
369 struct rb_node *parent = NULL;
370 struct sched_entity *entry;
371 s64 key = entity_key(cfs_rq, se);
375 * Find the right place in the rbtree:
379 entry = rb_entry(parent, struct sched_entity, run_node);
381 * We dont care about collisions. Nodes with
382 * the same key stay together.
384 if (key < entity_key(cfs_rq, entry)) {
385 link = &parent->rb_left;
387 link = &parent->rb_right;
393 * Maintain a cache of leftmost tree entries (it is frequently
397 cfs_rq->rb_leftmost = &se->run_node;
399 rb_link_node(&se->run_node, parent, link);
400 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
403 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
405 if (cfs_rq->rb_leftmost == &se->run_node) {
406 struct rb_node *next_node;
408 next_node = rb_next(&se->run_node);
409 cfs_rq->rb_leftmost = next_node;
412 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
415 static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
417 struct rb_node *left = cfs_rq->rb_leftmost;
422 return rb_entry(left, struct sched_entity, run_node);
425 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
427 struct rb_node *next = rb_next(&se->run_node);
432 return rb_entry(next, struct sched_entity, run_node);
435 #ifdef CONFIG_SCHED_DEBUG
436 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
438 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
443 return rb_entry(last, struct sched_entity, run_node);
446 /**************************************************************
447 * Scheduling class statistics methods:
450 int sched_proc_update_handler(struct ctl_table *table, int write,
451 void __user *buffer, size_t *lenp,
454 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
455 int factor = get_update_sysctl_factor();
460 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
461 sysctl_sched_min_granularity);
463 #define WRT_SYSCTL(name) \
464 (normalized_sysctl_##name = sysctl_##name / (factor))
465 WRT_SYSCTL(sched_min_granularity);
466 WRT_SYSCTL(sched_latency);
467 WRT_SYSCTL(sched_wakeup_granularity);
477 static inline unsigned long
478 calc_delta_fair(unsigned long delta, struct sched_entity *se)
480 if (unlikely(se->load.weight != NICE_0_LOAD))
481 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
487 * The idea is to set a period in which each task runs once.
489 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
490 * this period because otherwise the slices get too small.
492 * p = (nr <= nl) ? l : l*nr/nl
494 static u64 __sched_period(unsigned long nr_running)
496 u64 period = sysctl_sched_latency;
497 unsigned long nr_latency = sched_nr_latency;
499 if (unlikely(nr_running > nr_latency)) {
500 period = sysctl_sched_min_granularity;
501 period *= nr_running;
508 * We calculate the wall-time slice from the period by taking a part
509 * proportional to the weight.
513 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
515 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
517 for_each_sched_entity(se) {
518 struct load_weight *load;
519 struct load_weight lw;
521 cfs_rq = cfs_rq_of(se);
522 load = &cfs_rq->load;
524 if (unlikely(!se->on_rq)) {
527 update_load_add(&lw, se->load.weight);
530 slice = calc_delta_mine(slice, se->load.weight, load);
536 * We calculate the vruntime slice of a to be inserted task
540 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
542 return calc_delta_fair(sched_slice(cfs_rq, se), se);
545 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
546 static void update_cfs_shares(struct cfs_rq *cfs_rq);
549 * Update the current task's runtime statistics. Skip current tasks that
550 * are not in our scheduling class.
553 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
554 unsigned long delta_exec)
556 unsigned long delta_exec_weighted;
558 schedstat_set(curr->statistics.exec_max,
559 max((u64)delta_exec, curr->statistics.exec_max));
561 curr->sum_exec_runtime += delta_exec;
562 schedstat_add(cfs_rq, exec_clock, delta_exec);
563 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
565 curr->vruntime += delta_exec_weighted;
566 update_min_vruntime(cfs_rq);
568 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
569 cfs_rq->load_unacc_exec_time += delta_exec;
573 static void update_curr(struct cfs_rq *cfs_rq)
575 struct sched_entity *curr = cfs_rq->curr;
576 u64 now = rq_of(cfs_rq)->clock_task;
577 unsigned long delta_exec;
583 * Get the amount of time the current task was running
584 * since the last time we changed load (this cannot
585 * overflow on 32 bits):
587 delta_exec = (unsigned long)(now - curr->exec_start);
591 __update_curr(cfs_rq, curr, delta_exec);
592 curr->exec_start = now;
594 if (entity_is_task(curr)) {
595 struct task_struct *curtask = task_of(curr);
597 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
598 cpuacct_charge(curtask, delta_exec);
599 account_group_exec_runtime(curtask, delta_exec);
604 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
606 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
610 * Task is being enqueued - update stats:
612 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
615 * Are we enqueueing a waiting task? (for current tasks
616 * a dequeue/enqueue event is a NOP)
618 if (se != cfs_rq->curr)
619 update_stats_wait_start(cfs_rq, se);
623 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
625 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
626 rq_of(cfs_rq)->clock - se->statistics.wait_start));
627 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
628 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
629 rq_of(cfs_rq)->clock - se->statistics.wait_start);
630 #ifdef CONFIG_SCHEDSTATS
631 if (entity_is_task(se)) {
632 trace_sched_stat_wait(task_of(se),
633 rq_of(cfs_rq)->clock - se->statistics.wait_start);
636 schedstat_set(se->statistics.wait_start, 0);
640 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
643 * Mark the end of the wait period if dequeueing a
646 if (se != cfs_rq->curr)
647 update_stats_wait_end(cfs_rq, se);
651 * We are picking a new current task - update its stats:
654 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
657 * We are starting a new run period:
659 se->exec_start = rq_of(cfs_rq)->clock_task;
662 /**************************************************
663 * Scheduling class queueing methods:
666 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
668 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
670 cfs_rq->task_weight += weight;
674 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
680 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
682 update_load_add(&cfs_rq->load, se->load.weight);
683 if (!parent_entity(se))
684 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
685 if (entity_is_task(se)) {
686 add_cfs_task_weight(cfs_rq, se->load.weight);
687 list_add(&se->group_node, &cfs_rq->tasks);
689 cfs_rq->nr_running++;
693 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
695 update_load_sub(&cfs_rq->load, se->load.weight);
696 if (!parent_entity(se))
697 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
698 if (entity_is_task(se)) {
699 add_cfs_task_weight(cfs_rq, -se->load.weight);
700 list_del_init(&se->group_node);
702 cfs_rq->nr_running--;
705 #ifdef CONFIG_FAIR_GROUP_SCHED
707 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
710 struct task_group *tg = cfs_rq->tg;
713 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
714 load_avg -= cfs_rq->load_contribution;
716 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
717 atomic_add(load_avg, &tg->load_weight);
718 cfs_rq->load_contribution += load_avg;
722 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
724 u64 period = sysctl_sched_shares_window;
726 unsigned long load = cfs_rq->load.weight;
728 if (cfs_rq->tg == &root_task_group)
731 now = rq_of(cfs_rq)->clock_task;
732 delta = now - cfs_rq->load_stamp;
734 /* truncate load history at 4 idle periods */
735 if (cfs_rq->load_stamp > cfs_rq->load_last &&
736 now - cfs_rq->load_last > 4 * period) {
737 cfs_rq->load_period = 0;
738 cfs_rq->load_avg = 0;
742 cfs_rq->load_stamp = now;
743 cfs_rq->load_unacc_exec_time = 0;
744 cfs_rq->load_period += delta;
746 cfs_rq->load_last = now;
747 cfs_rq->load_avg += delta * load;
750 /* consider updating load contribution on each fold or truncate */
751 if (global_update || cfs_rq->load_period > period
752 || !cfs_rq->load_period)
753 update_cfs_rq_load_contribution(cfs_rq, global_update);
755 while (cfs_rq->load_period > period) {
757 * Inline assembly required to prevent the compiler
758 * optimising this loop into a divmod call.
759 * See __iter_div_u64_rem() for another example of this.
761 asm("" : "+rm" (cfs_rq->load_period));
762 cfs_rq->load_period /= 2;
763 cfs_rq->load_avg /= 2;
766 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
767 list_del_leaf_cfs_rq(cfs_rq);
770 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
772 long load_weight, load, shares;
774 load = cfs_rq->load.weight;
776 load_weight = atomic_read(&tg->load_weight);
778 load_weight -= cfs_rq->load_contribution;
780 shares = (tg->shares * load);
782 shares /= load_weight;
784 if (shares < MIN_SHARES)
786 if (shares > tg->shares)
792 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
794 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
795 update_cfs_load(cfs_rq, 0);
796 update_cfs_shares(cfs_rq);
799 # else /* CONFIG_SMP */
800 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
804 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
809 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
812 # endif /* CONFIG_SMP */
813 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
814 unsigned long weight)
817 /* commit outstanding execution time */
818 if (cfs_rq->curr == se)
820 account_entity_dequeue(cfs_rq, se);
823 update_load_set(&se->load, weight);
826 account_entity_enqueue(cfs_rq, se);
829 static void update_cfs_shares(struct cfs_rq *cfs_rq)
831 struct task_group *tg;
832 struct sched_entity *se;
836 se = tg->se[cpu_of(rq_of(cfs_rq))];
840 if (likely(se->load.weight == tg->shares))
843 shares = calc_cfs_shares(cfs_rq, tg);
845 reweight_entity(cfs_rq_of(se), se, shares);
847 #else /* CONFIG_FAIR_GROUP_SCHED */
848 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
852 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
856 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
859 #endif /* CONFIG_FAIR_GROUP_SCHED */
861 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
863 #ifdef CONFIG_SCHEDSTATS
864 struct task_struct *tsk = NULL;
866 if (entity_is_task(se))
869 if (se->statistics.sleep_start) {
870 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
875 if (unlikely(delta > se->statistics.sleep_max))
876 se->statistics.sleep_max = delta;
878 se->statistics.sleep_start = 0;
879 se->statistics.sum_sleep_runtime += delta;
882 account_scheduler_latency(tsk, delta >> 10, 1);
883 trace_sched_stat_sleep(tsk, delta);
886 if (se->statistics.block_start) {
887 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
892 if (unlikely(delta > se->statistics.block_max))
893 se->statistics.block_max = delta;
895 se->statistics.block_start = 0;
896 se->statistics.sum_sleep_runtime += delta;
899 if (tsk->in_iowait) {
900 se->statistics.iowait_sum += delta;
901 se->statistics.iowait_count++;
902 trace_sched_stat_iowait(tsk, delta);
906 * Blocking time is in units of nanosecs, so shift by
907 * 20 to get a milliseconds-range estimation of the
908 * amount of time that the task spent sleeping:
910 if (unlikely(prof_on == SLEEP_PROFILING)) {
911 profile_hits(SLEEP_PROFILING,
912 (void *)get_wchan(tsk),
915 account_scheduler_latency(tsk, delta >> 10, 0);
921 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
923 #ifdef CONFIG_SCHED_DEBUG
924 s64 d = se->vruntime - cfs_rq->min_vruntime;
929 if (d > 3*sysctl_sched_latency)
930 schedstat_inc(cfs_rq, nr_spread_over);
935 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
937 u64 vruntime = cfs_rq->min_vruntime;
940 * The 'current' period is already promised to the current tasks,
941 * however the extra weight of the new task will slow them down a
942 * little, place the new task so that it fits in the slot that
943 * stays open at the end.
945 if (initial && sched_feat(START_DEBIT))
946 vruntime += sched_vslice(cfs_rq, se);
948 /* sleeps up to a single latency don't count. */
950 unsigned long thresh = sysctl_sched_latency;
953 * Halve their sleep time's effect, to allow
954 * for a gentler effect of sleepers:
956 if (sched_feat(GENTLE_FAIR_SLEEPERS))
962 /* ensure we never gain time by being placed backwards. */
963 vruntime = max_vruntime(se->vruntime, vruntime);
965 se->vruntime = vruntime;
969 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
972 * Update the normalized vruntime before updating min_vruntime
973 * through callig update_curr().
975 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
976 se->vruntime += cfs_rq->min_vruntime;
979 * Update run-time statistics of the 'current'.
982 update_cfs_load(cfs_rq, 0);
983 account_entity_enqueue(cfs_rq, se);
984 update_cfs_shares(cfs_rq);
986 if (flags & ENQUEUE_WAKEUP) {
987 place_entity(cfs_rq, se, 0);
988 enqueue_sleeper(cfs_rq, se);
991 update_stats_enqueue(cfs_rq, se);
992 check_spread(cfs_rq, se);
993 if (se != cfs_rq->curr)
994 __enqueue_entity(cfs_rq, se);
997 if (cfs_rq->nr_running == 1)
998 list_add_leaf_cfs_rq(cfs_rq);
1001 static void __clear_buddies_last(struct sched_entity *se)
1003 for_each_sched_entity(se) {
1004 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1005 if (cfs_rq->last == se)
1006 cfs_rq->last = NULL;
1012 static void __clear_buddies_next(struct sched_entity *se)
1014 for_each_sched_entity(se) {
1015 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1016 if (cfs_rq->next == se)
1017 cfs_rq->next = NULL;
1023 static void __clear_buddies_skip(struct sched_entity *se)
1025 for_each_sched_entity(se) {
1026 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1027 if (cfs_rq->skip == se)
1028 cfs_rq->skip = NULL;
1034 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1036 if (cfs_rq->last == se)
1037 __clear_buddies_last(se);
1039 if (cfs_rq->next == se)
1040 __clear_buddies_next(se);
1042 if (cfs_rq->skip == se)
1043 __clear_buddies_skip(se);
1047 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1050 * Update run-time statistics of the 'current'.
1052 update_curr(cfs_rq);
1054 update_stats_dequeue(cfs_rq, se);
1055 if (flags & DEQUEUE_SLEEP) {
1056 #ifdef CONFIG_SCHEDSTATS
1057 if (entity_is_task(se)) {
1058 struct task_struct *tsk = task_of(se);
1060 if (tsk->state & TASK_INTERRUPTIBLE)
1061 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1062 if (tsk->state & TASK_UNINTERRUPTIBLE)
1063 se->statistics.block_start = rq_of(cfs_rq)->clock;
1068 clear_buddies(cfs_rq, se);
1070 if (se != cfs_rq->curr)
1071 __dequeue_entity(cfs_rq, se);
1073 update_cfs_load(cfs_rq, 0);
1074 account_entity_dequeue(cfs_rq, se);
1075 update_min_vruntime(cfs_rq);
1076 update_cfs_shares(cfs_rq);
1079 * Normalize the entity after updating the min_vruntime because the
1080 * update can refer to the ->curr item and we need to reflect this
1081 * movement in our normalized position.
1083 if (!(flags & DEQUEUE_SLEEP))
1084 se->vruntime -= cfs_rq->min_vruntime;
1088 * Preempt the current task with a newly woken task if needed:
1091 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1093 unsigned long ideal_runtime, delta_exec;
1095 ideal_runtime = sched_slice(cfs_rq, curr);
1096 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1097 if (delta_exec > ideal_runtime) {
1098 resched_task(rq_of(cfs_rq)->curr);
1100 * The current task ran long enough, ensure it doesn't get
1101 * re-elected due to buddy favours.
1103 clear_buddies(cfs_rq, curr);
1108 * Ensure that a task that missed wakeup preemption by a
1109 * narrow margin doesn't have to wait for a full slice.
1110 * This also mitigates buddy induced latencies under load.
1112 if (!sched_feat(WAKEUP_PREEMPT))
1115 if (delta_exec < sysctl_sched_min_granularity)
1118 if (cfs_rq->nr_running > 1) {
1119 struct sched_entity *se = __pick_first_entity(cfs_rq);
1120 s64 delta = curr->vruntime - se->vruntime;
1125 if (delta > ideal_runtime)
1126 resched_task(rq_of(cfs_rq)->curr);
1131 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1133 /* 'current' is not kept within the tree. */
1136 * Any task has to be enqueued before it get to execute on
1137 * a CPU. So account for the time it spent waiting on the
1140 update_stats_wait_end(cfs_rq, se);
1141 __dequeue_entity(cfs_rq, se);
1144 update_stats_curr_start(cfs_rq, se);
1146 #ifdef CONFIG_SCHEDSTATS
1148 * Track our maximum slice length, if the CPU's load is at
1149 * least twice that of our own weight (i.e. dont track it
1150 * when there are only lesser-weight tasks around):
1152 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1153 se->statistics.slice_max = max(se->statistics.slice_max,
1154 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1157 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1161 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1164 * Pick the next process, keeping these things in mind, in this order:
1165 * 1) keep things fair between processes/task groups
1166 * 2) pick the "next" process, since someone really wants that to run
1167 * 3) pick the "last" process, for cache locality
1168 * 4) do not run the "skip" process, if something else is available
1170 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1172 struct sched_entity *se = __pick_first_entity(cfs_rq);
1173 struct sched_entity *left = se;
1176 * Avoid running the skip buddy, if running something else can
1177 * be done without getting too unfair.
1179 if (cfs_rq->skip == se) {
1180 struct sched_entity *second = __pick_next_entity(se);
1181 if (second && wakeup_preempt_entity(second, left) < 1)
1186 * Prefer last buddy, try to return the CPU to a preempted task.
1188 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1192 * Someone really wants this to run. If it's not unfair, run it.
1194 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1197 clear_buddies(cfs_rq, se);
1202 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1205 * If still on the runqueue then deactivate_task()
1206 * was not called and update_curr() has to be done:
1209 update_curr(cfs_rq);
1211 check_spread(cfs_rq, prev);
1213 update_stats_wait_start(cfs_rq, prev);
1214 /* Put 'current' back into the tree. */
1215 __enqueue_entity(cfs_rq, prev);
1217 cfs_rq->curr = NULL;
1221 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1224 * Update run-time statistics of the 'current'.
1226 update_curr(cfs_rq);
1229 * Update share accounting for long-running entities.
1231 update_entity_shares_tick(cfs_rq);
1233 #ifdef CONFIG_SCHED_HRTICK
1235 * queued ticks are scheduled to match the slice, so don't bother
1236 * validating it and just reschedule.
1239 resched_task(rq_of(cfs_rq)->curr);
1243 * don't let the period tick interfere with the hrtick preemption
1245 if (!sched_feat(DOUBLE_TICK) &&
1246 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1250 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1251 check_preempt_tick(cfs_rq, curr);
1254 /**************************************************
1255 * CFS operations on tasks:
1258 #ifdef CONFIG_SCHED_HRTICK
1259 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1261 struct sched_entity *se = &p->se;
1262 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1264 WARN_ON(task_rq(p) != rq);
1266 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1267 u64 slice = sched_slice(cfs_rq, se);
1268 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1269 s64 delta = slice - ran;
1278 * Don't schedule slices shorter than 10000ns, that just
1279 * doesn't make sense. Rely on vruntime for fairness.
1282 delta = max_t(s64, 10000LL, delta);
1284 hrtick_start(rq, delta);
1289 * called from enqueue/dequeue and updates the hrtick when the
1290 * current task is from our class and nr_running is low enough
1293 static void hrtick_update(struct rq *rq)
1295 struct task_struct *curr = rq->curr;
1297 if (curr->sched_class != &fair_sched_class)
1300 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1301 hrtick_start_fair(rq, curr);
1303 #else /* !CONFIG_SCHED_HRTICK */
1305 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1309 static inline void hrtick_update(struct rq *rq)
1315 * The enqueue_task method is called before nr_running is
1316 * increased. Here we update the fair scheduling stats and
1317 * then put the task into the rbtree:
1320 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1322 struct cfs_rq *cfs_rq;
1323 struct sched_entity *se = &p->se;
1325 for_each_sched_entity(se) {
1328 cfs_rq = cfs_rq_of(se);
1329 enqueue_entity(cfs_rq, se, flags);
1330 flags = ENQUEUE_WAKEUP;
1333 for_each_sched_entity(se) {
1334 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1336 update_cfs_load(cfs_rq, 0);
1337 update_cfs_shares(cfs_rq);
1344 * The dequeue_task method is called before nr_running is
1345 * decreased. We remove the task from the rbtree and
1346 * update the fair scheduling stats:
1348 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1350 struct cfs_rq *cfs_rq;
1351 struct sched_entity *se = &p->se;
1353 for_each_sched_entity(se) {
1354 cfs_rq = cfs_rq_of(se);
1355 dequeue_entity(cfs_rq, se, flags);
1357 /* Don't dequeue parent if it has other entities besides us */
1358 if (cfs_rq->load.weight)
1360 flags |= DEQUEUE_SLEEP;
1363 for_each_sched_entity(se) {
1364 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1366 update_cfs_load(cfs_rq, 0);
1367 update_cfs_shares(cfs_rq);
1375 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1377 struct sched_entity *se = &p->se;
1378 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1380 se->vruntime -= cfs_rq->min_vruntime;
1383 #ifdef CONFIG_FAIR_GROUP_SCHED
1385 * effective_load() calculates the load change as seen from the root_task_group
1387 * Adding load to a group doesn't make a group heavier, but can cause movement
1388 * of group shares between cpus. Assuming the shares were perfectly aligned one
1389 * can calculate the shift in shares.
1391 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1393 struct sched_entity *se = tg->se[cpu];
1398 for_each_sched_entity(se) {
1402 w = se->my_q->load.weight;
1404 /* use this cpu's instantaneous contribution */
1405 lw = atomic_read(&tg->load_weight);
1406 lw -= se->my_q->load_contribution;
1411 if (lw > 0 && wl < lw)
1412 wl = (wl * tg->shares) / lw;
1416 /* zero point is MIN_SHARES */
1417 if (wl < MIN_SHARES)
1419 wl -= se->load.weight;
1428 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1429 unsigned long wl, unsigned long wg)
1436 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1438 s64 this_load, load;
1439 int idx, this_cpu, prev_cpu;
1440 unsigned long tl_per_task;
1441 struct task_group *tg;
1442 unsigned long weight;
1446 this_cpu = smp_processor_id();
1447 prev_cpu = task_cpu(p);
1448 load = source_load(prev_cpu, idx);
1449 this_load = target_load(this_cpu, idx);
1452 * If sync wakeup then subtract the (maximum possible)
1453 * effect of the currently running task from the load
1454 * of the current CPU:
1458 tg = task_group(current);
1459 weight = current->se.load.weight;
1461 this_load += effective_load(tg, this_cpu, -weight, -weight);
1462 load += effective_load(tg, prev_cpu, 0, -weight);
1466 weight = p->se.load.weight;
1469 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1470 * due to the sync cause above having dropped this_load to 0, we'll
1471 * always have an imbalance, but there's really nothing you can do
1472 * about that, so that's good too.
1474 * Otherwise check if either cpus are near enough in load to allow this
1475 * task to be woken on this_cpu.
1477 if (this_load > 0) {
1478 s64 this_eff_load, prev_eff_load;
1480 this_eff_load = 100;
1481 this_eff_load *= power_of(prev_cpu);
1482 this_eff_load *= this_load +
1483 effective_load(tg, this_cpu, weight, weight);
1485 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1486 prev_eff_load *= power_of(this_cpu);
1487 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1489 balanced = this_eff_load <= prev_eff_load;
1495 * If the currently running task will sleep within
1496 * a reasonable amount of time then attract this newly
1499 if (sync && balanced)
1502 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1503 tl_per_task = cpu_avg_load_per_task(this_cpu);
1506 (this_load <= load &&
1507 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1509 * This domain has SD_WAKE_AFFINE and
1510 * p is cache cold in this domain, and
1511 * there is no bad imbalance.
1513 schedstat_inc(sd, ttwu_move_affine);
1514 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1522 * find_idlest_group finds and returns the least busy CPU group within the
1525 static struct sched_group *
1526 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1527 int this_cpu, int load_idx)
1529 struct sched_group *idlest = NULL, *group = sd->groups;
1530 unsigned long min_load = ULONG_MAX, this_load = 0;
1531 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1534 unsigned long load, avg_load;
1538 /* Skip over this group if it has no CPUs allowed */
1539 if (!cpumask_intersects(sched_group_cpus(group),
1543 local_group = cpumask_test_cpu(this_cpu,
1544 sched_group_cpus(group));
1546 /* Tally up the load of all CPUs in the group */
1549 for_each_cpu(i, sched_group_cpus(group)) {
1550 /* Bias balancing toward cpus of our domain */
1552 load = source_load(i, load_idx);
1554 load = target_load(i, load_idx);
1559 /* Adjust by relative CPU power of the group */
1560 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1563 this_load = avg_load;
1564 } else if (avg_load < min_load) {
1565 min_load = avg_load;
1568 } while (group = group->next, group != sd->groups);
1570 if (!idlest || 100*this_load < imbalance*min_load)
1576 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1579 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1581 unsigned long load, min_load = ULONG_MAX;
1585 /* Traverse only the allowed CPUs */
1586 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1587 load = weighted_cpuload(i);
1589 if (load < min_load || (load == min_load && i == this_cpu)) {
1599 * Try and locate an idle CPU in the sched_domain.
1601 static int select_idle_sibling(struct task_struct *p, int target)
1603 int cpu = smp_processor_id();
1604 int prev_cpu = task_cpu(p);
1605 struct sched_domain *sd;
1609 * If the task is going to be woken-up on this cpu and if it is
1610 * already idle, then it is the right target.
1612 if (target == cpu && idle_cpu(cpu))
1616 * If the task is going to be woken-up on the cpu where it previously
1617 * ran and if it is currently idle, then it the right target.
1619 if (target == prev_cpu && idle_cpu(prev_cpu))
1623 * Otherwise, iterate the domains and find an elegible idle cpu.
1626 for_each_domain(target, sd) {
1627 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1630 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1638 * Lets stop looking for an idle sibling when we reached
1639 * the domain that spans the current cpu and prev_cpu.
1641 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1642 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1651 * sched_balance_self: balance the current task (running on cpu) in domains
1652 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1655 * Balance, ie. select the least loaded group.
1657 * Returns the target CPU number, or the same CPU if no balancing is needed.
1659 * preempt must be disabled.
1662 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1664 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1665 int cpu = smp_processor_id();
1666 int prev_cpu = task_cpu(p);
1668 int want_affine = 0;
1670 int sync = wake_flags & WF_SYNC;
1672 if (sd_flag & SD_BALANCE_WAKE) {
1673 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1679 for_each_domain(cpu, tmp) {
1680 if (!(tmp->flags & SD_LOAD_BALANCE))
1684 * If power savings logic is enabled for a domain, see if we
1685 * are not overloaded, if so, don't balance wider.
1687 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1688 unsigned long power = 0;
1689 unsigned long nr_running = 0;
1690 unsigned long capacity;
1693 for_each_cpu(i, sched_domain_span(tmp)) {
1694 power += power_of(i);
1695 nr_running += cpu_rq(i)->cfs.nr_running;
1698 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1700 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1703 if (nr_running < capacity)
1708 * If both cpu and prev_cpu are part of this domain,
1709 * cpu is a valid SD_WAKE_AFFINE target.
1711 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1712 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1717 if (!want_sd && !want_affine)
1720 if (!(tmp->flags & sd_flag))
1728 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1731 new_cpu = select_idle_sibling(p, prev_cpu);
1736 int load_idx = sd->forkexec_idx;
1737 struct sched_group *group;
1740 if (!(sd->flags & sd_flag)) {
1745 if (sd_flag & SD_BALANCE_WAKE)
1746 load_idx = sd->wake_idx;
1748 group = find_idlest_group(sd, p, cpu, load_idx);
1754 new_cpu = find_idlest_cpu(group, p, cpu);
1755 if (new_cpu == -1 || new_cpu == cpu) {
1756 /* Now try balancing at a lower domain level of cpu */
1761 /* Now try balancing at a lower domain level of new_cpu */
1763 weight = sd->span_weight;
1765 for_each_domain(cpu, tmp) {
1766 if (weight <= tmp->span_weight)
1768 if (tmp->flags & sd_flag)
1771 /* while loop will break here if sd == NULL */
1778 #endif /* CONFIG_SMP */
1780 static unsigned long
1781 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1783 unsigned long gran = sysctl_sched_wakeup_granularity;
1786 * Since its curr running now, convert the gran from real-time
1787 * to virtual-time in his units.
1789 * By using 'se' instead of 'curr' we penalize light tasks, so
1790 * they get preempted easier. That is, if 'se' < 'curr' then
1791 * the resulting gran will be larger, therefore penalizing the
1792 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1793 * be smaller, again penalizing the lighter task.
1795 * This is especially important for buddies when the leftmost
1796 * task is higher priority than the buddy.
1798 return calc_delta_fair(gran, se);
1802 * Should 'se' preempt 'curr'.
1816 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1818 s64 gran, vdiff = curr->vruntime - se->vruntime;
1823 gran = wakeup_gran(curr, se);
1830 static void set_last_buddy(struct sched_entity *se)
1832 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1833 for_each_sched_entity(se)
1834 cfs_rq_of(se)->last = se;
1838 static void set_next_buddy(struct sched_entity *se)
1840 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1841 for_each_sched_entity(se)
1842 cfs_rq_of(se)->next = se;
1846 static void set_skip_buddy(struct sched_entity *se)
1848 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1849 for_each_sched_entity(se)
1850 cfs_rq_of(se)->skip = se;
1855 * Preempt the current task with a newly woken task if needed:
1857 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1859 struct task_struct *curr = rq->curr;
1860 struct sched_entity *se = &curr->se, *pse = &p->se;
1861 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1862 int scale = cfs_rq->nr_running >= sched_nr_latency;
1864 if (unlikely(se == pse))
1867 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1868 set_next_buddy(pse);
1871 * We can come here with TIF_NEED_RESCHED already set from new task
1874 if (test_tsk_need_resched(curr))
1877 /* Idle tasks are by definition preempted by non-idle tasks. */
1878 if (unlikely(curr->policy == SCHED_IDLE) &&
1879 likely(p->policy != SCHED_IDLE))
1883 * Batch and idle tasks do not preempt non-idle tasks (their preemption
1884 * is driven by the tick):
1886 if (unlikely(p->policy != SCHED_NORMAL))
1890 if (!sched_feat(WAKEUP_PREEMPT))
1893 update_curr(cfs_rq);
1894 find_matching_se(&se, &pse);
1896 if (wakeup_preempt_entity(se, pse) == 1)
1904 * Only set the backward buddy when the current task is still
1905 * on the rq. This can happen when a wakeup gets interleaved
1906 * with schedule on the ->pre_schedule() or idle_balance()
1907 * point, either of which can * drop the rq lock.
1909 * Also, during early boot the idle thread is in the fair class,
1910 * for obvious reasons its a bad idea to schedule back to it.
1912 if (unlikely(!se->on_rq || curr == rq->idle))
1915 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1919 static struct task_struct *pick_next_task_fair(struct rq *rq)
1921 struct task_struct *p;
1922 struct cfs_rq *cfs_rq = &rq->cfs;
1923 struct sched_entity *se;
1925 if (!cfs_rq->nr_running)
1929 se = pick_next_entity(cfs_rq);
1930 set_next_entity(cfs_rq, se);
1931 cfs_rq = group_cfs_rq(se);
1935 hrtick_start_fair(rq, p);
1941 * Account for a descheduled task:
1943 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1945 struct sched_entity *se = &prev->se;
1946 struct cfs_rq *cfs_rq;
1948 for_each_sched_entity(se) {
1949 cfs_rq = cfs_rq_of(se);
1950 put_prev_entity(cfs_rq, se);
1955 * sched_yield() is very simple
1957 * The magic of dealing with the ->skip buddy is in pick_next_entity.
1959 static void yield_task_fair(struct rq *rq)
1961 struct task_struct *curr = rq->curr;
1962 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1963 struct sched_entity *se = &curr->se;
1966 * Are we the only task in the tree?
1968 if (unlikely(rq->nr_running == 1))
1971 clear_buddies(cfs_rq, se);
1973 if (curr->policy != SCHED_BATCH) {
1974 update_rq_clock(rq);
1976 * Update run-time statistics of the 'current'.
1978 update_curr(cfs_rq);
1984 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
1986 struct sched_entity *se = &p->se;
1991 /* Tell the scheduler that we'd really like pse to run next. */
1994 yield_task_fair(rq);
2000 /**************************************************
2001 * Fair scheduling class load-balancing methods:
2005 * pull_task - move a task from a remote runqueue to the local runqueue.
2006 * Both runqueues must be locked.
2008 static void pull_task(struct rq *src_rq, struct task_struct *p,
2009 struct rq *this_rq, int this_cpu)
2011 deactivate_task(src_rq, p, 0);
2012 set_task_cpu(p, this_cpu);
2013 activate_task(this_rq, p, 0);
2014 check_preempt_curr(this_rq, p, 0);
2018 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2021 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2022 struct sched_domain *sd, enum cpu_idle_type idle,
2025 int tsk_cache_hot = 0;
2027 * We do not migrate tasks that are:
2028 * 1) running (obviously), or
2029 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2030 * 3) are cache-hot on their current CPU.
2032 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2033 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2038 if (task_running(rq, p)) {
2039 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2044 * Aggressive migration if:
2045 * 1) task is cache cold, or
2046 * 2) too many balance attempts have failed.
2049 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2050 if (!tsk_cache_hot ||
2051 sd->nr_balance_failed > sd->cache_nice_tries) {
2052 #ifdef CONFIG_SCHEDSTATS
2053 if (tsk_cache_hot) {
2054 schedstat_inc(sd, lb_hot_gained[idle]);
2055 schedstat_inc(p, se.statistics.nr_forced_migrations);
2061 if (tsk_cache_hot) {
2062 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2069 * move_one_task tries to move exactly one task from busiest to this_rq, as
2070 * part of active balancing operations within "domain".
2071 * Returns 1 if successful and 0 otherwise.
2073 * Called with both runqueues locked.
2076 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2077 struct sched_domain *sd, enum cpu_idle_type idle)
2079 struct task_struct *p, *n;
2080 struct cfs_rq *cfs_rq;
2083 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2084 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2086 if (!can_migrate_task(p, busiest, this_cpu,
2090 pull_task(busiest, p, this_rq, this_cpu);
2092 * Right now, this is only the second place pull_task()
2093 * is called, so we can safely collect pull_task()
2094 * stats here rather than inside pull_task().
2096 schedstat_inc(sd, lb_gained[idle]);
2104 static unsigned long
2105 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2106 unsigned long max_load_move, struct sched_domain *sd,
2107 enum cpu_idle_type idle, int *all_pinned,
2108 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
2110 int loops = 0, pulled = 0;
2111 long rem_load_move = max_load_move;
2112 struct task_struct *p, *n;
2114 if (max_load_move == 0)
2117 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2118 if (loops++ > sysctl_sched_nr_migrate)
2121 if ((p->se.load.weight >> 1) > rem_load_move ||
2122 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2126 pull_task(busiest, p, this_rq, this_cpu);
2128 rem_load_move -= p->se.load.weight;
2130 #ifdef CONFIG_PREEMPT
2132 * NEWIDLE balancing is a source of latency, so preemptible
2133 * kernels will stop after the first task is pulled to minimize
2134 * the critical section.
2136 if (idle == CPU_NEWLY_IDLE)
2141 * We only want to steal up to the prescribed amount of
2144 if (rem_load_move <= 0)
2147 if (p->prio < *this_best_prio)
2148 *this_best_prio = p->prio;
2152 * Right now, this is one of only two places pull_task() is called,
2153 * so we can safely collect pull_task() stats here rather than
2154 * inside pull_task().
2156 schedstat_add(sd, lb_gained[idle], pulled);
2158 return max_load_move - rem_load_move;
2161 #ifdef CONFIG_FAIR_GROUP_SCHED
2163 * update tg->load_weight by folding this cpu's load_avg
2165 static int update_shares_cpu(struct task_group *tg, int cpu)
2167 struct cfs_rq *cfs_rq;
2168 unsigned long flags;
2175 cfs_rq = tg->cfs_rq[cpu];
2177 raw_spin_lock_irqsave(&rq->lock, flags);
2179 update_rq_clock(rq);
2180 update_cfs_load(cfs_rq, 1);
2183 * We need to update shares after updating tg->load_weight in
2184 * order to adjust the weight of groups with long running tasks.
2186 update_cfs_shares(cfs_rq);
2188 raw_spin_unlock_irqrestore(&rq->lock, flags);
2193 static void update_shares(int cpu)
2195 struct cfs_rq *cfs_rq;
2196 struct rq *rq = cpu_rq(cpu);
2199 for_each_leaf_cfs_rq(rq, cfs_rq)
2200 update_shares_cpu(cfs_rq->tg, cpu);
2204 static unsigned long
2205 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2206 unsigned long max_load_move,
2207 struct sched_domain *sd, enum cpu_idle_type idle,
2208 int *all_pinned, int *this_best_prio)
2210 long rem_load_move = max_load_move;
2211 int busiest_cpu = cpu_of(busiest);
2212 struct task_group *tg;
2215 update_h_load(busiest_cpu);
2217 list_for_each_entry_rcu(tg, &task_groups, list) {
2218 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2219 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2220 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2221 u64 rem_load, moved_load;
2226 if (!busiest_cfs_rq->task_weight)
2229 rem_load = (u64)rem_load_move * busiest_weight;
2230 rem_load = div_u64(rem_load, busiest_h_load + 1);
2232 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2233 rem_load, sd, idle, all_pinned, this_best_prio,
2239 moved_load *= busiest_h_load;
2240 moved_load = div_u64(moved_load, busiest_weight + 1);
2242 rem_load_move -= moved_load;
2243 if (rem_load_move < 0)
2248 return max_load_move - rem_load_move;
2251 static inline void update_shares(int cpu)
2255 static unsigned long
2256 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2257 unsigned long max_load_move,
2258 struct sched_domain *sd, enum cpu_idle_type idle,
2259 int *all_pinned, int *this_best_prio)
2261 return balance_tasks(this_rq, this_cpu, busiest,
2262 max_load_move, sd, idle, all_pinned,
2263 this_best_prio, &busiest->cfs);
2268 * move_tasks tries to move up to max_load_move weighted load from busiest to
2269 * this_rq, as part of a balancing operation within domain "sd".
2270 * Returns 1 if successful and 0 otherwise.
2272 * Called with both runqueues locked.
2274 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2275 unsigned long max_load_move,
2276 struct sched_domain *sd, enum cpu_idle_type idle,
2279 unsigned long total_load_moved = 0, load_moved;
2280 int this_best_prio = this_rq->curr->prio;
2283 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2284 max_load_move - total_load_moved,
2285 sd, idle, all_pinned, &this_best_prio);
2287 total_load_moved += load_moved;
2289 #ifdef CONFIG_PREEMPT
2291 * NEWIDLE balancing is a source of latency, so preemptible
2292 * kernels will stop after the first task is pulled to minimize
2293 * the critical section.
2295 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2298 if (raw_spin_is_contended(&this_rq->lock) ||
2299 raw_spin_is_contended(&busiest->lock))
2302 } while (load_moved && max_load_move > total_load_moved);
2304 return total_load_moved > 0;
2307 /********** Helpers for find_busiest_group ************************/
2309 * sd_lb_stats - Structure to store the statistics of a sched_domain
2310 * during load balancing.
2312 struct sd_lb_stats {
2313 struct sched_group *busiest; /* Busiest group in this sd */
2314 struct sched_group *this; /* Local group in this sd */
2315 unsigned long total_load; /* Total load of all groups in sd */
2316 unsigned long total_pwr; /* Total power of all groups in sd */
2317 unsigned long avg_load; /* Average load across all groups in sd */
2319 /** Statistics of this group */
2320 unsigned long this_load;
2321 unsigned long this_load_per_task;
2322 unsigned long this_nr_running;
2323 unsigned long this_has_capacity;
2324 unsigned int this_idle_cpus;
2326 /* Statistics of the busiest group */
2327 unsigned int busiest_idle_cpus;
2328 unsigned long max_load;
2329 unsigned long busiest_load_per_task;
2330 unsigned long busiest_nr_running;
2331 unsigned long busiest_group_capacity;
2332 unsigned long busiest_has_capacity;
2333 unsigned int busiest_group_weight;
2335 int group_imb; /* Is there imbalance in this sd */
2336 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2337 int power_savings_balance; /* Is powersave balance needed for this sd */
2338 struct sched_group *group_min; /* Least loaded group in sd */
2339 struct sched_group *group_leader; /* Group which relieves group_min */
2340 unsigned long min_load_per_task; /* load_per_task in group_min */
2341 unsigned long leader_nr_running; /* Nr running of group_leader */
2342 unsigned long min_nr_running; /* Nr running of group_min */
2347 * sg_lb_stats - stats of a sched_group required for load_balancing
2349 struct sg_lb_stats {
2350 unsigned long avg_load; /*Avg load across the CPUs of the group */
2351 unsigned long group_load; /* Total load over the CPUs of the group */
2352 unsigned long sum_nr_running; /* Nr tasks running in the group */
2353 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2354 unsigned long group_capacity;
2355 unsigned long idle_cpus;
2356 unsigned long group_weight;
2357 int group_imb; /* Is there an imbalance in the group ? */
2358 int group_has_capacity; /* Is there extra capacity in the group? */
2362 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2363 * @group: The group whose first cpu is to be returned.
2365 static inline unsigned int group_first_cpu(struct sched_group *group)
2367 return cpumask_first(sched_group_cpus(group));
2371 * get_sd_load_idx - Obtain the load index for a given sched domain.
2372 * @sd: The sched_domain whose load_idx is to be obtained.
2373 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2375 static inline int get_sd_load_idx(struct sched_domain *sd,
2376 enum cpu_idle_type idle)
2382 load_idx = sd->busy_idx;
2385 case CPU_NEWLY_IDLE:
2386 load_idx = sd->newidle_idx;
2389 load_idx = sd->idle_idx;
2397 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2399 * init_sd_power_savings_stats - Initialize power savings statistics for
2400 * the given sched_domain, during load balancing.
2402 * @sd: Sched domain whose power-savings statistics are to be initialized.
2403 * @sds: Variable containing the statistics for sd.
2404 * @idle: Idle status of the CPU at which we're performing load-balancing.
2406 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2407 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2410 * Busy processors will not participate in power savings
2413 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2414 sds->power_savings_balance = 0;
2416 sds->power_savings_balance = 1;
2417 sds->min_nr_running = ULONG_MAX;
2418 sds->leader_nr_running = 0;
2423 * update_sd_power_savings_stats - Update the power saving stats for a
2424 * sched_domain while performing load balancing.
2426 * @group: sched_group belonging to the sched_domain under consideration.
2427 * @sds: Variable containing the statistics of the sched_domain
2428 * @local_group: Does group contain the CPU for which we're performing
2430 * @sgs: Variable containing the statistics of the group.
2432 static inline void update_sd_power_savings_stats(struct sched_group *group,
2433 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2436 if (!sds->power_savings_balance)
2440 * If the local group is idle or completely loaded
2441 * no need to do power savings balance at this domain
2443 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2444 !sds->this_nr_running))
2445 sds->power_savings_balance = 0;
2448 * If a group is already running at full capacity or idle,
2449 * don't include that group in power savings calculations
2451 if (!sds->power_savings_balance ||
2452 sgs->sum_nr_running >= sgs->group_capacity ||
2453 !sgs->sum_nr_running)
2457 * Calculate the group which has the least non-idle load.
2458 * This is the group from where we need to pick up the load
2461 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2462 (sgs->sum_nr_running == sds->min_nr_running &&
2463 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2464 sds->group_min = group;
2465 sds->min_nr_running = sgs->sum_nr_running;
2466 sds->min_load_per_task = sgs->sum_weighted_load /
2467 sgs->sum_nr_running;
2471 * Calculate the group which is almost near its
2472 * capacity but still has some space to pick up some load
2473 * from other group and save more power
2475 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2478 if (sgs->sum_nr_running > sds->leader_nr_running ||
2479 (sgs->sum_nr_running == sds->leader_nr_running &&
2480 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2481 sds->group_leader = group;
2482 sds->leader_nr_running = sgs->sum_nr_running;
2487 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2488 * @sds: Variable containing the statistics of the sched_domain
2489 * under consideration.
2490 * @this_cpu: Cpu at which we're currently performing load-balancing.
2491 * @imbalance: Variable to store the imbalance.
2494 * Check if we have potential to perform some power-savings balance.
2495 * If yes, set the busiest group to be the least loaded group in the
2496 * sched_domain, so that it's CPUs can be put to idle.
2498 * Returns 1 if there is potential to perform power-savings balance.
2501 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2502 int this_cpu, unsigned long *imbalance)
2504 if (!sds->power_savings_balance)
2507 if (sds->this != sds->group_leader ||
2508 sds->group_leader == sds->group_min)
2511 *imbalance = sds->min_load_per_task;
2512 sds->busiest = sds->group_min;
2517 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2518 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2519 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2524 static inline void update_sd_power_savings_stats(struct sched_group *group,
2525 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2530 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2531 int this_cpu, unsigned long *imbalance)
2535 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2538 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2540 return SCHED_LOAD_SCALE;
2543 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2545 return default_scale_freq_power(sd, cpu);
2548 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2550 unsigned long weight = sd->span_weight;
2551 unsigned long smt_gain = sd->smt_gain;
2558 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2560 return default_scale_smt_power(sd, cpu);
2563 unsigned long scale_rt_power(int cpu)
2565 struct rq *rq = cpu_rq(cpu);
2566 u64 total, available;
2568 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2570 if (unlikely(total < rq->rt_avg)) {
2571 /* Ensures that power won't end up being negative */
2574 available = total - rq->rt_avg;
2577 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2578 total = SCHED_LOAD_SCALE;
2580 total >>= SCHED_LOAD_SHIFT;
2582 return div_u64(available, total);
2585 static void update_cpu_power(struct sched_domain *sd, int cpu)
2587 unsigned long weight = sd->span_weight;
2588 unsigned long power = SCHED_LOAD_SCALE;
2589 struct sched_group *sdg = sd->groups;
2591 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2592 if (sched_feat(ARCH_POWER))
2593 power *= arch_scale_smt_power(sd, cpu);
2595 power *= default_scale_smt_power(sd, cpu);
2597 power >>= SCHED_LOAD_SHIFT;
2600 sdg->cpu_power_orig = power;
2602 if (sched_feat(ARCH_POWER))
2603 power *= arch_scale_freq_power(sd, cpu);
2605 power *= default_scale_freq_power(sd, cpu);
2607 power >>= SCHED_LOAD_SHIFT;
2609 power *= scale_rt_power(cpu);
2610 power >>= SCHED_LOAD_SHIFT;
2615 cpu_rq(cpu)->cpu_power = power;
2616 sdg->cpu_power = power;
2619 static void update_group_power(struct sched_domain *sd, int cpu)
2621 struct sched_domain *child = sd->child;
2622 struct sched_group *group, *sdg = sd->groups;
2623 unsigned long power;
2626 update_cpu_power(sd, cpu);
2632 group = child->groups;
2634 power += group->cpu_power;
2635 group = group->next;
2636 } while (group != child->groups);
2638 sdg->cpu_power = power;
2642 * Try and fix up capacity for tiny siblings, this is needed when
2643 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2644 * which on its own isn't powerful enough.
2646 * See update_sd_pick_busiest() and check_asym_packing().
2649 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2652 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2654 if (!(sd->flags & SD_SHARE_CPUPOWER))
2658 * If ~90% of the cpu_power is still there, we're good.
2660 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2667 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2668 * @sd: The sched_domain whose statistics are to be updated.
2669 * @group: sched_group whose statistics are to be updated.
2670 * @this_cpu: Cpu for which load balance is currently performed.
2671 * @idle: Idle status of this_cpu
2672 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2673 * @local_group: Does group contain this_cpu.
2674 * @cpus: Set of cpus considered for load balancing.
2675 * @balance: Should we balance.
2676 * @sgs: variable to hold the statistics for this group.
2678 static inline void update_sg_lb_stats(struct sched_domain *sd,
2679 struct sched_group *group, int this_cpu,
2680 enum cpu_idle_type idle, int load_idx,
2681 int local_group, const struct cpumask *cpus,
2682 int *balance, struct sg_lb_stats *sgs)
2684 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2686 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2687 unsigned long avg_load_per_task = 0;
2690 balance_cpu = group_first_cpu(group);
2692 /* Tally up the load of all CPUs in the group */
2694 min_cpu_load = ~0UL;
2697 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2698 struct rq *rq = cpu_rq(i);
2700 /* Bias balancing toward cpus of our domain */
2702 if (idle_cpu(i) && !first_idle_cpu) {
2707 load = target_load(i, load_idx);
2709 load = source_load(i, load_idx);
2710 if (load > max_cpu_load) {
2711 max_cpu_load = load;
2712 max_nr_running = rq->nr_running;
2714 if (min_cpu_load > load)
2715 min_cpu_load = load;
2718 sgs->group_load += load;
2719 sgs->sum_nr_running += rq->nr_running;
2720 sgs->sum_weighted_load += weighted_cpuload(i);
2726 * First idle cpu or the first cpu(busiest) in this sched group
2727 * is eligible for doing load balancing at this and above
2728 * domains. In the newly idle case, we will allow all the cpu's
2729 * to do the newly idle load balance.
2731 if (idle != CPU_NEWLY_IDLE && local_group) {
2732 if (balance_cpu != this_cpu) {
2736 update_group_power(sd, this_cpu);
2739 /* Adjust by relative CPU power of the group */
2740 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2743 * Consider the group unbalanced when the imbalance is larger
2744 * than the average weight of a task.
2746 * APZ: with cgroup the avg task weight can vary wildly and
2747 * might not be a suitable number - should we keep a
2748 * normalized nr_running number somewhere that negates
2751 if (sgs->sum_nr_running)
2752 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2754 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
2757 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2758 if (!sgs->group_capacity)
2759 sgs->group_capacity = fix_small_capacity(sd, group);
2760 sgs->group_weight = group->group_weight;
2762 if (sgs->group_capacity > sgs->sum_nr_running)
2763 sgs->group_has_capacity = 1;
2767 * update_sd_pick_busiest - return 1 on busiest group
2768 * @sd: sched_domain whose statistics are to be checked
2769 * @sds: sched_domain statistics
2770 * @sg: sched_group candidate to be checked for being the busiest
2771 * @sgs: sched_group statistics
2772 * @this_cpu: the current cpu
2774 * Determine if @sg is a busier group than the previously selected
2777 static bool update_sd_pick_busiest(struct sched_domain *sd,
2778 struct sd_lb_stats *sds,
2779 struct sched_group *sg,
2780 struct sg_lb_stats *sgs,
2783 if (sgs->avg_load <= sds->max_load)
2786 if (sgs->sum_nr_running > sgs->group_capacity)
2793 * ASYM_PACKING needs to move all the work to the lowest
2794 * numbered CPUs in the group, therefore mark all groups
2795 * higher than ourself as busy.
2797 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2798 this_cpu < group_first_cpu(sg)) {
2802 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2810 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2811 * @sd: sched_domain whose statistics are to be updated.
2812 * @this_cpu: Cpu for which load balance is currently performed.
2813 * @idle: Idle status of this_cpu
2814 * @cpus: Set of cpus considered for load balancing.
2815 * @balance: Should we balance.
2816 * @sds: variable to hold the statistics for this sched_domain.
2818 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2819 enum cpu_idle_type idle, const struct cpumask *cpus,
2820 int *balance, struct sd_lb_stats *sds)
2822 struct sched_domain *child = sd->child;
2823 struct sched_group *sg = sd->groups;
2824 struct sg_lb_stats sgs;
2825 int load_idx, prefer_sibling = 0;
2827 if (child && child->flags & SD_PREFER_SIBLING)
2830 init_sd_power_savings_stats(sd, sds, idle);
2831 load_idx = get_sd_load_idx(sd, idle);
2836 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2837 memset(&sgs, 0, sizeof(sgs));
2838 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
2839 local_group, cpus, balance, &sgs);
2841 if (local_group && !(*balance))
2844 sds->total_load += sgs.group_load;
2845 sds->total_pwr += sg->cpu_power;
2848 * In case the child domain prefers tasks go to siblings
2849 * first, lower the sg capacity to one so that we'll try
2850 * and move all the excess tasks away. We lower the capacity
2851 * of a group only if the local group has the capacity to fit
2852 * these excess tasks, i.e. nr_running < group_capacity. The
2853 * extra check prevents the case where you always pull from the
2854 * heaviest group when it is already under-utilized (possible
2855 * with a large weight task outweighs the tasks on the system).
2857 if (prefer_sibling && !local_group && sds->this_has_capacity)
2858 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2861 sds->this_load = sgs.avg_load;
2863 sds->this_nr_running = sgs.sum_nr_running;
2864 sds->this_load_per_task = sgs.sum_weighted_load;
2865 sds->this_has_capacity = sgs.group_has_capacity;
2866 sds->this_idle_cpus = sgs.idle_cpus;
2867 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2868 sds->max_load = sgs.avg_load;
2870 sds->busiest_nr_running = sgs.sum_nr_running;
2871 sds->busiest_idle_cpus = sgs.idle_cpus;
2872 sds->busiest_group_capacity = sgs.group_capacity;
2873 sds->busiest_load_per_task = sgs.sum_weighted_load;
2874 sds->busiest_has_capacity = sgs.group_has_capacity;
2875 sds->busiest_group_weight = sgs.group_weight;
2876 sds->group_imb = sgs.group_imb;
2879 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2881 } while (sg != sd->groups);
2884 int __weak arch_sd_sibling_asym_packing(void)
2886 return 0*SD_ASYM_PACKING;
2890 * check_asym_packing - Check to see if the group is packed into the
2893 * This is primarily intended to used at the sibling level. Some
2894 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2895 * case of POWER7, it can move to lower SMT modes only when higher
2896 * threads are idle. When in lower SMT modes, the threads will
2897 * perform better since they share less core resources. Hence when we
2898 * have idle threads, we want them to be the higher ones.
2900 * This packing function is run on idle threads. It checks to see if
2901 * the busiest CPU in this domain (core in the P7 case) has a higher
2902 * CPU number than the packing function is being run on. Here we are
2903 * assuming lower CPU number will be equivalent to lower a SMT thread
2906 * Returns 1 when packing is required and a task should be moved to
2907 * this CPU. The amount of the imbalance is returned in *imbalance.
2909 * @sd: The sched_domain whose packing is to be checked.
2910 * @sds: Statistics of the sched_domain which is to be packed
2911 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2912 * @imbalance: returns amount of imbalanced due to packing.
2914 static int check_asym_packing(struct sched_domain *sd,
2915 struct sd_lb_stats *sds,
2916 int this_cpu, unsigned long *imbalance)
2920 if (!(sd->flags & SD_ASYM_PACKING))
2926 busiest_cpu = group_first_cpu(sds->busiest);
2927 if (this_cpu > busiest_cpu)
2930 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2936 * fix_small_imbalance - Calculate the minor imbalance that exists
2937 * amongst the groups of a sched_domain, during
2939 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2940 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2941 * @imbalance: Variable to store the imbalance.
2943 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2944 int this_cpu, unsigned long *imbalance)
2946 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2947 unsigned int imbn = 2;
2948 unsigned long scaled_busy_load_per_task;
2950 if (sds->this_nr_running) {
2951 sds->this_load_per_task /= sds->this_nr_running;
2952 if (sds->busiest_load_per_task >
2953 sds->this_load_per_task)
2956 sds->this_load_per_task =
2957 cpu_avg_load_per_task(this_cpu);
2959 scaled_busy_load_per_task = sds->busiest_load_per_task
2961 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2963 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2964 (scaled_busy_load_per_task * imbn)) {
2965 *imbalance = sds->busiest_load_per_task;
2970 * OK, we don't have enough imbalance to justify moving tasks,
2971 * however we may be able to increase total CPU power used by
2975 pwr_now += sds->busiest->cpu_power *
2976 min(sds->busiest_load_per_task, sds->max_load);
2977 pwr_now += sds->this->cpu_power *
2978 min(sds->this_load_per_task, sds->this_load);
2979 pwr_now /= SCHED_LOAD_SCALE;
2981 /* Amount of load we'd subtract */
2982 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2983 sds->busiest->cpu_power;
2984 if (sds->max_load > tmp)
2985 pwr_move += sds->busiest->cpu_power *
2986 min(sds->busiest_load_per_task, sds->max_load - tmp);
2988 /* Amount of load we'd add */
2989 if (sds->max_load * sds->busiest->cpu_power <
2990 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2991 tmp = (sds->max_load * sds->busiest->cpu_power) /
2992 sds->this->cpu_power;
2994 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2995 sds->this->cpu_power;
2996 pwr_move += sds->this->cpu_power *
2997 min(sds->this_load_per_task, sds->this_load + tmp);
2998 pwr_move /= SCHED_LOAD_SCALE;
3000 /* Move if we gain throughput */
3001 if (pwr_move > pwr_now)
3002 *imbalance = sds->busiest_load_per_task;
3006 * calculate_imbalance - Calculate the amount of imbalance present within the
3007 * groups of a given sched_domain during load balance.
3008 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3009 * @this_cpu: Cpu for which currently load balance is being performed.
3010 * @imbalance: The variable to store the imbalance.
3012 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3013 unsigned long *imbalance)
3015 unsigned long max_pull, load_above_capacity = ~0UL;
3017 sds->busiest_load_per_task /= sds->busiest_nr_running;
3018 if (sds->group_imb) {
3019 sds->busiest_load_per_task =
3020 min(sds->busiest_load_per_task, sds->avg_load);
3024 * In the presence of smp nice balancing, certain scenarios can have
3025 * max load less than avg load(as we skip the groups at or below
3026 * its cpu_power, while calculating max_load..)
3028 if (sds->max_load < sds->avg_load) {
3030 return fix_small_imbalance(sds, this_cpu, imbalance);
3033 if (!sds->group_imb) {
3035 * Don't want to pull so many tasks that a group would go idle.
3037 load_above_capacity = (sds->busiest_nr_running -
3038 sds->busiest_group_capacity);
3040 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
3042 load_above_capacity /= sds->busiest->cpu_power;
3046 * We're trying to get all the cpus to the average_load, so we don't
3047 * want to push ourselves above the average load, nor do we wish to
3048 * reduce the max loaded cpu below the average load. At the same time,
3049 * we also don't want to reduce the group load below the group capacity
3050 * (so that we can implement power-savings policies etc). Thus we look
3051 * for the minimum possible imbalance.
3052 * Be careful of negative numbers as they'll appear as very large values
3053 * with unsigned longs.
3055 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3057 /* How much load to actually move to equalise the imbalance */
3058 *imbalance = min(max_pull * sds->busiest->cpu_power,
3059 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3063 * if *imbalance is less than the average load per runnable task
3064 * there is no guarantee that any tasks will be moved so we'll have
3065 * a think about bumping its value to force at least one task to be
3068 if (*imbalance < sds->busiest_load_per_task)
3069 return fix_small_imbalance(sds, this_cpu, imbalance);
3073 /******* find_busiest_group() helpers end here *********************/
3076 * find_busiest_group - Returns the busiest group within the sched_domain
3077 * if there is an imbalance. If there isn't an imbalance, and
3078 * the user has opted for power-savings, it returns a group whose
3079 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3080 * such a group exists.
3082 * Also calculates the amount of weighted load which should be moved
3083 * to restore balance.
3085 * @sd: The sched_domain whose busiest group is to be returned.
3086 * @this_cpu: The cpu for which load balancing is currently being performed.
3087 * @imbalance: Variable which stores amount of weighted load which should
3088 * be moved to restore balance/put a group to idle.
3089 * @idle: The idle status of this_cpu.
3090 * @cpus: The set of CPUs under consideration for load-balancing.
3091 * @balance: Pointer to a variable indicating if this_cpu
3092 * is the appropriate cpu to perform load balancing at this_level.
3094 * Returns: - the busiest group if imbalance exists.
3095 * - If no imbalance and user has opted for power-savings balance,
3096 * return the least loaded group whose CPUs can be
3097 * put to idle by rebalancing its tasks onto our group.
3099 static struct sched_group *
3100 find_busiest_group(struct sched_domain *sd, int this_cpu,
3101 unsigned long *imbalance, enum cpu_idle_type idle,
3102 const struct cpumask *cpus, int *balance)
3104 struct sd_lb_stats sds;
3106 memset(&sds, 0, sizeof(sds));
3109 * Compute the various statistics relavent for load balancing at
3112 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3115 * this_cpu is not the appropriate cpu to perform load balancing at
3121 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3122 check_asym_packing(sd, &sds, this_cpu, imbalance))
3125 /* There is no busy sibling group to pull tasks from */
3126 if (!sds.busiest || sds.busiest_nr_running == 0)
3129 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3132 * If the busiest group is imbalanced the below checks don't
3133 * work because they assumes all things are equal, which typically
3134 * isn't true due to cpus_allowed constraints and the like.
3139 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3140 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3141 !sds.busiest_has_capacity)
3145 * If the local group is more busy than the selected busiest group
3146 * don't try and pull any tasks.
3148 if (sds.this_load >= sds.max_load)
3152 * Don't pull any tasks if this group is already above the domain
3155 if (sds.this_load >= sds.avg_load)
3158 if (idle == CPU_IDLE) {
3160 * This cpu is idle. If the busiest group load doesn't
3161 * have more tasks than the number of available cpu's and
3162 * there is no imbalance between this and busiest group
3163 * wrt to idle cpu's, it is balanced.
3165 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3166 sds.busiest_nr_running <= sds.busiest_group_weight)
3170 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3171 * imbalance_pct to be conservative.
3173 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3178 /* Looks like there is an imbalance. Compute it */
3179 calculate_imbalance(&sds, this_cpu, imbalance);
3184 * There is no obvious imbalance. But check if we can do some balancing
3187 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3195 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3198 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3199 enum cpu_idle_type idle, unsigned long imbalance,
3200 const struct cpumask *cpus)
3202 struct rq *busiest = NULL, *rq;
3203 unsigned long max_load = 0;
3206 for_each_cpu(i, sched_group_cpus(group)) {
3207 unsigned long power = power_of(i);
3208 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3212 capacity = fix_small_capacity(sd, group);
3214 if (!cpumask_test_cpu(i, cpus))
3218 wl = weighted_cpuload(i);
3221 * When comparing with imbalance, use weighted_cpuload()
3222 * which is not scaled with the cpu power.
3224 if (capacity && rq->nr_running == 1 && wl > imbalance)
3228 * For the load comparisons with the other cpu's, consider
3229 * the weighted_cpuload() scaled with the cpu power, so that
3230 * the load can be moved away from the cpu that is potentially
3231 * running at a lower capacity.
3233 wl = (wl * SCHED_LOAD_SCALE) / power;
3235 if (wl > max_load) {
3245 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3246 * so long as it is large enough.
3248 #define MAX_PINNED_INTERVAL 512
3250 /* Working cpumask for load_balance and load_balance_newidle. */
3251 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3253 static int need_active_balance(struct sched_domain *sd, int idle,
3254 int busiest_cpu, int this_cpu)
3256 if (idle == CPU_NEWLY_IDLE) {
3259 * ASYM_PACKING needs to force migrate tasks from busy but
3260 * higher numbered CPUs in order to pack all tasks in the
3261 * lowest numbered CPUs.
3263 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3267 * The only task running in a non-idle cpu can be moved to this
3268 * cpu in an attempt to completely freeup the other CPU
3271 * The package power saving logic comes from
3272 * find_busiest_group(). If there are no imbalance, then
3273 * f_b_g() will return NULL. However when sched_mc={1,2} then
3274 * f_b_g() will select a group from which a running task may be
3275 * pulled to this cpu in order to make the other package idle.
3276 * If there is no opportunity to make a package idle and if
3277 * there are no imbalance, then f_b_g() will return NULL and no
3278 * action will be taken in load_balance_newidle().
3280 * Under normal task pull operation due to imbalance, there
3281 * will be more than one task in the source run queue and
3282 * move_tasks() will succeed. ld_moved will be true and this
3283 * active balance code will not be triggered.
3285 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3289 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3292 static int active_load_balance_cpu_stop(void *data);
3295 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3296 * tasks if there is an imbalance.
3298 static int load_balance(int this_cpu, struct rq *this_rq,
3299 struct sched_domain *sd, enum cpu_idle_type idle,
3302 int ld_moved, all_pinned = 0, active_balance = 0;
3303 struct sched_group *group;
3304 unsigned long imbalance;
3306 unsigned long flags;
3307 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3309 cpumask_copy(cpus, cpu_active_mask);
3311 schedstat_inc(sd, lb_count[idle]);
3314 group = find_busiest_group(sd, this_cpu, &imbalance, 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.
3344 local_irq_save(flags);
3345 double_rq_lock(this_rq, busiest);
3346 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3347 imbalance, sd, idle, &all_pinned);
3348 double_rq_unlock(this_rq, busiest);
3349 local_irq_restore(flags);
3352 * some other cpu did the load balance for us.
3354 if (ld_moved && this_cpu != smp_processor_id())
3355 resched_cpu(this_cpu);
3357 /* All tasks on this runqueue were pinned by CPU affinity */
3358 if (unlikely(all_pinned)) {
3359 cpumask_clear_cpu(cpu_of(busiest), cpus);
3360 if (!cpumask_empty(cpus))
3367 schedstat_inc(sd, lb_failed[idle]);
3369 * Increment the failure counter only on periodic balance.
3370 * We do not want newidle balance, which can be very
3371 * frequent, pollute the failure counter causing
3372 * excessive cache_hot migrations and active balances.
3374 if (idle != CPU_NEWLY_IDLE)
3375 sd->nr_balance_failed++;
3377 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
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;
3435 schedstat_inc(sd, lb_balanced[idle]);
3437 sd->nr_balance_failed = 0;
3440 /* tune up the balancing interval */
3441 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3442 (sd->balance_interval < sd->max_interval))
3443 sd->balance_interval *= 2;
3451 * idle_balance is called by schedule() if this_cpu is about to become
3452 * idle. Attempts to pull tasks from other CPUs.
3454 static void idle_balance(int this_cpu, struct rq *this_rq)
3456 struct sched_domain *sd;
3457 int pulled_task = 0;
3458 unsigned long next_balance = jiffies + HZ;
3460 this_rq->idle_stamp = this_rq->clock;
3462 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3466 * Drop the rq->lock, but keep IRQ/preempt disabled.
3468 raw_spin_unlock(&this_rq->lock);
3470 update_shares(this_cpu);
3472 for_each_domain(this_cpu, sd) {
3473 unsigned long interval;
3476 if (!(sd->flags & SD_LOAD_BALANCE))
3479 if (sd->flags & SD_BALANCE_NEWIDLE) {
3480 /* If we've pulled tasks over stop searching: */
3481 pulled_task = load_balance(this_cpu, this_rq,
3482 sd, CPU_NEWLY_IDLE, &balance);
3485 interval = msecs_to_jiffies(sd->balance_interval);
3486 if (time_after(next_balance, sd->last_balance + interval))
3487 next_balance = sd->last_balance + interval;
3489 this_rq->idle_stamp = 0;
3495 raw_spin_lock(&this_rq->lock);
3497 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3499 * We are going idle. next_balance may be set based on
3500 * a busy processor. So reset next_balance.
3502 this_rq->next_balance = next_balance;
3507 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3508 * running tasks off the busiest CPU onto idle CPUs. It requires at
3509 * least 1 task to be running on each physical CPU where possible, and
3510 * avoids physical / logical imbalances.
3512 static int active_load_balance_cpu_stop(void *data)
3514 struct rq *busiest_rq = data;
3515 int busiest_cpu = cpu_of(busiest_rq);
3516 int target_cpu = busiest_rq->push_cpu;
3517 struct rq *target_rq = cpu_rq(target_cpu);
3518 struct sched_domain *sd;
3520 raw_spin_lock_irq(&busiest_rq->lock);
3522 /* make sure the requested cpu hasn't gone down in the meantime */
3523 if (unlikely(busiest_cpu != smp_processor_id() ||
3524 !busiest_rq->active_balance))
3527 /* Is there any task to move? */
3528 if (busiest_rq->nr_running <= 1)
3532 * This condition is "impossible", if it occurs
3533 * we need to fix it. Originally reported by
3534 * Bjorn Helgaas on a 128-cpu setup.
3536 BUG_ON(busiest_rq == target_rq);
3538 /* move a task from busiest_rq to target_rq */
3539 double_lock_balance(busiest_rq, target_rq);
3541 /* Search for an sd spanning us and the target CPU. */
3543 for_each_domain(target_cpu, sd) {
3544 if ((sd->flags & SD_LOAD_BALANCE) &&
3545 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3550 schedstat_inc(sd, alb_count);
3552 if (move_one_task(target_rq, target_cpu, busiest_rq,
3554 schedstat_inc(sd, alb_pushed);
3556 schedstat_inc(sd, alb_failed);
3559 double_unlock_balance(busiest_rq, target_rq);
3561 busiest_rq->active_balance = 0;
3562 raw_spin_unlock_irq(&busiest_rq->lock);
3568 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3570 static void trigger_sched_softirq(void *data)
3572 raise_softirq_irqoff(SCHED_SOFTIRQ);
3575 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3577 csd->func = trigger_sched_softirq;
3584 * idle load balancing details
3585 * - One of the idle CPUs nominates itself as idle load_balancer, while
3587 * - This idle load balancer CPU will also go into tickless mode when
3588 * it is idle, just like all other idle CPUs
3589 * - When one of the busy CPUs notice that there may be an idle rebalancing
3590 * needed, they will kick the idle load balancer, which then does idle
3591 * load balancing for all the idle CPUs.
3594 atomic_t load_balancer;
3595 atomic_t first_pick_cpu;
3596 atomic_t second_pick_cpu;
3597 cpumask_var_t idle_cpus_mask;
3598 cpumask_var_t grp_idle_mask;
3599 unsigned long next_balance; /* in jiffy units */
3600 } nohz ____cacheline_aligned;
3602 int get_nohz_load_balancer(void)
3604 return atomic_read(&nohz.load_balancer);
3607 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3609 * lowest_flag_domain - Return lowest sched_domain containing flag.
3610 * @cpu: The cpu whose lowest level of sched domain is to
3612 * @flag: The flag to check for the lowest sched_domain
3613 * for the given cpu.
3615 * Returns the lowest sched_domain of a cpu which contains the given flag.
3617 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3619 struct sched_domain *sd;
3621 for_each_domain(cpu, sd)
3622 if (sd && (sd->flags & flag))
3629 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3630 * @cpu: The cpu whose domains we're iterating over.
3631 * @sd: variable holding the value of the power_savings_sd
3633 * @flag: The flag to filter the sched_domains to be iterated.
3635 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3636 * set, starting from the lowest sched_domain to the highest.
3638 #define for_each_flag_domain(cpu, sd, flag) \
3639 for (sd = lowest_flag_domain(cpu, flag); \
3640 (sd && (sd->flags & flag)); sd = sd->parent)
3643 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3644 * @ilb_group: group to be checked for semi-idleness
3646 * Returns: 1 if the group is semi-idle. 0 otherwise.
3648 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3649 * and atleast one non-idle CPU. This helper function checks if the given
3650 * sched_group is semi-idle or not.
3652 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3654 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3655 sched_group_cpus(ilb_group));
3658 * A sched_group is semi-idle when it has atleast one busy cpu
3659 * and atleast one idle cpu.
3661 if (cpumask_empty(nohz.grp_idle_mask))
3664 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3670 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3671 * @cpu: The cpu which is nominating a new idle_load_balancer.
3673 * Returns: Returns the id of the idle load balancer if it exists,
3674 * Else, returns >= nr_cpu_ids.
3676 * This algorithm picks the idle load balancer such that it belongs to a
3677 * semi-idle powersavings sched_domain. The idea is to try and avoid
3678 * completely idle packages/cores just for the purpose of idle load balancing
3679 * when there are other idle cpu's which are better suited for that job.
3681 static int find_new_ilb(int cpu)
3683 struct sched_domain *sd;
3684 struct sched_group *ilb_group;
3685 int ilb = nr_cpu_ids;
3688 * Have idle load balancer selection from semi-idle packages only
3689 * when power-aware load balancing is enabled
3691 if (!(sched_smt_power_savings || sched_mc_power_savings))
3695 * Optimize for the case when we have no idle CPUs or only one
3696 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3698 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3702 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3703 ilb_group = sd->groups;
3706 if (is_semi_idle_group(ilb_group)) {
3707 ilb = cpumask_first(nohz.grp_idle_mask);
3711 ilb_group = ilb_group->next;
3713 } while (ilb_group != sd->groups);
3721 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3722 static inline int find_new_ilb(int call_cpu)
3729 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3730 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3731 * CPU (if there is one).
3733 static void nohz_balancer_kick(int cpu)
3737 nohz.next_balance++;
3739 ilb_cpu = get_nohz_load_balancer();
3741 if (ilb_cpu >= nr_cpu_ids) {
3742 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3743 if (ilb_cpu >= nr_cpu_ids)
3747 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3748 struct call_single_data *cp;
3750 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3751 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3752 __smp_call_function_single(ilb_cpu, cp, 0);
3758 * This routine will try to nominate the ilb (idle load balancing)
3759 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3760 * load balancing on behalf of all those cpus.
3762 * When the ilb owner becomes busy, we will not have new ilb owner until some
3763 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3764 * idle load balancing by kicking one of the idle CPUs.
3766 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3767 * ilb owner CPU in future (when there is a need for idle load balancing on
3768 * behalf of all idle CPUs).
3770 void select_nohz_load_balancer(int stop_tick)
3772 int cpu = smp_processor_id();
3775 if (!cpu_active(cpu)) {
3776 if (atomic_read(&nohz.load_balancer) != cpu)
3780 * If we are going offline and still the leader,
3783 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3790 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3792 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3793 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3794 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3795 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3797 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3800 /* make me the ilb owner */
3801 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3806 * Check to see if there is a more power-efficient
3809 new_ilb = find_new_ilb(cpu);
3810 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3811 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3812 resched_cpu(new_ilb);
3818 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3821 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3823 if (atomic_read(&nohz.load_balancer) == cpu)
3824 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3832 static DEFINE_SPINLOCK(balancing);
3834 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3837 * Scale the max load_balance interval with the number of CPUs in the system.
3838 * This trades load-balance latency on larger machines for less cross talk.
3840 static void update_max_interval(void)
3842 max_load_balance_interval = HZ*num_online_cpus()/10;
3846 * It checks each scheduling domain to see if it is due to be balanced,
3847 * and initiates a balancing operation if so.
3849 * Balancing parameters are set up in arch_init_sched_domains.
3851 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3854 struct rq *rq = cpu_rq(cpu);
3855 unsigned long interval;
3856 struct sched_domain *sd;
3857 /* Earliest time when we have to do rebalance again */
3858 unsigned long next_balance = jiffies + 60*HZ;
3859 int update_next_balance = 0;
3865 for_each_domain(cpu, sd) {
3866 if (!(sd->flags & SD_LOAD_BALANCE))
3869 interval = sd->balance_interval;
3870 if (idle != CPU_IDLE)
3871 interval *= sd->busy_factor;
3873 /* scale ms to jiffies */
3874 interval = msecs_to_jiffies(interval);
3875 interval = clamp(interval, 1UL, max_load_balance_interval);
3877 need_serialize = sd->flags & SD_SERIALIZE;
3879 if (need_serialize) {
3880 if (!spin_trylock(&balancing))
3884 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3885 if (load_balance(cpu, rq, sd, idle, &balance)) {
3887 * We've pulled tasks over so either we're no
3890 idle = CPU_NOT_IDLE;
3892 sd->last_balance = jiffies;
3895 spin_unlock(&balancing);
3897 if (time_after(next_balance, sd->last_balance + interval)) {
3898 next_balance = sd->last_balance + interval;
3899 update_next_balance = 1;
3903 * Stop the load balance at this level. There is another
3904 * CPU in our sched group which is doing load balancing more
3913 * next_balance will be updated only when there is a need.
3914 * When the cpu is attached to null domain for ex, it will not be
3917 if (likely(update_next_balance))
3918 rq->next_balance = next_balance;
3923 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3924 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3926 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3928 struct rq *this_rq = cpu_rq(this_cpu);
3932 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3935 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3936 if (balance_cpu == this_cpu)
3940 * If this cpu gets work to do, stop the load balancing
3941 * work being done for other cpus. Next load
3942 * balancing owner will pick it up.
3944 if (need_resched()) {
3945 this_rq->nohz_balance_kick = 0;
3949 raw_spin_lock_irq(&this_rq->lock);
3950 update_rq_clock(this_rq);
3951 update_cpu_load(this_rq);
3952 raw_spin_unlock_irq(&this_rq->lock);
3954 rebalance_domains(balance_cpu, CPU_IDLE);
3956 rq = cpu_rq(balance_cpu);
3957 if (time_after(this_rq->next_balance, rq->next_balance))
3958 this_rq->next_balance = rq->next_balance;
3960 nohz.next_balance = this_rq->next_balance;
3961 this_rq->nohz_balance_kick = 0;
3965 * Current heuristic for kicking the idle load balancer
3966 * - first_pick_cpu is the one of the busy CPUs. It will kick
3967 * idle load balancer when it has more than one process active. This
3968 * eliminates the need for idle load balancing altogether when we have
3969 * only one running process in the system (common case).
3970 * - If there are more than one busy CPU, idle load balancer may have
3971 * to run for active_load_balance to happen (i.e., two busy CPUs are
3972 * SMT or core siblings and can run better if they move to different
3973 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3974 * which will kick idle load balancer as soon as it has any load.
3976 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3978 unsigned long now = jiffies;
3980 int first_pick_cpu, second_pick_cpu;
3982 if (time_before(now, nohz.next_balance))
3985 if (rq->idle_at_tick)
3988 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3989 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3991 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3992 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3995 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3996 if (ret == nr_cpu_ids || ret == cpu) {
3997 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3998 if (rq->nr_running > 1)
4001 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4002 if (ret == nr_cpu_ids || ret == cpu) {
4010 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4014 * run_rebalance_domains is triggered when needed from the scheduler tick.
4015 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4017 static void run_rebalance_domains(struct softirq_action *h)
4019 int this_cpu = smp_processor_id();
4020 struct rq *this_rq = cpu_rq(this_cpu);
4021 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4022 CPU_IDLE : CPU_NOT_IDLE;
4024 rebalance_domains(this_cpu, idle);
4027 * If this cpu has a pending nohz_balance_kick, then do the
4028 * balancing on behalf of the other idle cpus whose ticks are
4031 nohz_idle_balance(this_cpu, idle);
4034 static inline int on_null_domain(int cpu)
4036 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4040 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4042 static inline void trigger_load_balance(struct rq *rq, int cpu)
4044 /* Don't need to rebalance while attached to NULL domain */
4045 if (time_after_eq(jiffies, rq->next_balance) &&
4046 likely(!on_null_domain(cpu)))
4047 raise_softirq(SCHED_SOFTIRQ);
4049 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4050 nohz_balancer_kick(cpu);
4054 static void rq_online_fair(struct rq *rq)
4059 static void rq_offline_fair(struct rq *rq)
4064 #else /* CONFIG_SMP */
4067 * on UP we do not need to balance between CPUs:
4069 static inline void idle_balance(int cpu, struct rq *rq)
4073 #endif /* CONFIG_SMP */
4076 * scheduler tick hitting a task of our scheduling class:
4078 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4080 struct cfs_rq *cfs_rq;
4081 struct sched_entity *se = &curr->se;
4083 for_each_sched_entity(se) {
4084 cfs_rq = cfs_rq_of(se);
4085 entity_tick(cfs_rq, se, queued);
4090 * called on fork with the child task as argument from the parent's context
4091 * - child not yet on the tasklist
4092 * - preemption disabled
4094 static void task_fork_fair(struct task_struct *p)
4096 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4097 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4098 int this_cpu = smp_processor_id();
4099 struct rq *rq = this_rq();
4100 unsigned long flags;
4102 raw_spin_lock_irqsave(&rq->lock, flags);
4104 update_rq_clock(rq);
4106 if (unlikely(task_cpu(p) != this_cpu)) {
4108 __set_task_cpu(p, this_cpu);
4112 update_curr(cfs_rq);
4115 se->vruntime = curr->vruntime;
4116 place_entity(cfs_rq, se, 1);
4118 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4120 * Upon rescheduling, sched_class::put_prev_task() will place
4121 * 'current' within the tree based on its new key value.
4123 swap(curr->vruntime, se->vruntime);
4124 resched_task(rq->curr);
4127 se->vruntime -= cfs_rq->min_vruntime;
4129 raw_spin_unlock_irqrestore(&rq->lock, flags);
4133 * Priority of the task has changed. Check to see if we preempt
4137 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4143 * Reschedule if we are currently running on this runqueue and
4144 * our priority decreased, or if we are not currently running on
4145 * this runqueue and our priority is higher than the current's
4147 if (rq->curr == p) {
4148 if (p->prio > oldprio)
4149 resched_task(rq->curr);
4151 check_preempt_curr(rq, p, 0);
4154 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4156 struct sched_entity *se = &p->se;
4157 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4160 * Ensure the task's vruntime is normalized, so that when its
4161 * switched back to the fair class the enqueue_entity(.flags=0) will
4162 * do the right thing.
4164 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4165 * have normalized the vruntime, if it was !on_rq, then only when
4166 * the task is sleeping will it still have non-normalized vruntime.
4168 if (!se->on_rq && p->state != TASK_RUNNING) {
4170 * Fix up our vruntime so that the current sleep doesn't
4171 * cause 'unlimited' sleep bonus.
4173 place_entity(cfs_rq, se, 0);
4174 se->vruntime -= cfs_rq->min_vruntime;
4179 * We switched to the sched_fair class.
4181 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4187 * We were most likely switched from sched_rt, so
4188 * kick off the schedule if running, otherwise just see
4189 * if we can still preempt the current task.
4192 resched_task(rq->curr);
4194 check_preempt_curr(rq, p, 0);
4197 /* Account for a task changing its policy or group.
4199 * This routine is mostly called to set cfs_rq->curr field when a task
4200 * migrates between groups/classes.
4202 static void set_curr_task_fair(struct rq *rq)
4204 struct sched_entity *se = &rq->curr->se;
4206 for_each_sched_entity(se)
4207 set_next_entity(cfs_rq_of(se), se);
4210 #ifdef CONFIG_FAIR_GROUP_SCHED
4211 static void task_move_group_fair(struct task_struct *p, int on_rq)
4214 * If the task was not on the rq at the time of this cgroup movement
4215 * it must have been asleep, sleeping tasks keep their ->vruntime
4216 * absolute on their old rq until wakeup (needed for the fair sleeper
4217 * bonus in place_entity()).
4219 * If it was on the rq, we've just 'preempted' it, which does convert
4220 * ->vruntime to a relative base.
4222 * Make sure both cases convert their relative position when migrating
4223 * to another cgroup's rq. This does somewhat interfere with the
4224 * fair sleeper stuff for the first placement, but who cares.
4227 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4228 set_task_rq(p, task_cpu(p));
4230 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4234 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4236 struct sched_entity *se = &task->se;
4237 unsigned int rr_interval = 0;
4240 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4243 if (rq->cfs.load.weight)
4244 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4250 * All the scheduling class methods:
4252 static const struct sched_class fair_sched_class = {
4253 .next = &idle_sched_class,
4254 .enqueue_task = enqueue_task_fair,
4255 .dequeue_task = dequeue_task_fair,
4256 .yield_task = yield_task_fair,
4257 .yield_to_task = yield_to_task_fair,
4259 .check_preempt_curr = check_preempt_wakeup,
4261 .pick_next_task = pick_next_task_fair,
4262 .put_prev_task = put_prev_task_fair,
4265 .select_task_rq = select_task_rq_fair,
4267 .rq_online = rq_online_fair,
4268 .rq_offline = rq_offline_fair,
4270 .task_waking = task_waking_fair,
4273 .set_curr_task = set_curr_task_fair,
4274 .task_tick = task_tick_fair,
4275 .task_fork = task_fork_fair,
4277 .prio_changed = prio_changed_fair,
4278 .switched_from = switched_from_fair,
4279 .switched_to = switched_to_fair,
4281 .get_rr_interval = get_rr_interval_fair,
4283 #ifdef CONFIG_FAIR_GROUP_SCHED
4284 .task_move_group = task_move_group_fair,
4288 #ifdef CONFIG_SCHED_DEBUG
4289 static void print_cfs_stats(struct seq_file *m, int cpu)
4291 struct cfs_rq *cfs_rq;
4294 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4295 print_cfs_rq(m, cpu, cfs_rq);