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 #ifdef CONFIG_CFS_BANDWIDTH
94 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
95 * each time a cfs_rq requests quota.
97 * Note: in the case that the slice exceeds the runtime remaining (either due
98 * to consumption or the quota being specified to be smaller than the slice)
99 * we will always only issue the remaining available time.
101 * default: 5 msec, units: microseconds
103 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
106 static const struct sched_class fair_sched_class;
108 /**************************************************************
109 * CFS operations on generic schedulable entities:
112 #ifdef CONFIG_FAIR_GROUP_SCHED
114 /* cpu runqueue to which this cfs_rq is attached */
115 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
120 /* An entity is a task if it doesn't "own" a runqueue */
121 #define entity_is_task(se) (!se->my_q)
123 static inline struct task_struct *task_of(struct sched_entity *se)
125 #ifdef CONFIG_SCHED_DEBUG
126 WARN_ON_ONCE(!entity_is_task(se));
128 return container_of(se, struct task_struct, se);
131 /* Walk up scheduling entities hierarchy */
132 #define for_each_sched_entity(se) \
133 for (; se; se = se->parent)
135 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
140 /* runqueue on which this entity is (to be) queued */
141 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
146 /* runqueue "owned" by this group */
147 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
152 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
154 if (!cfs_rq->on_list) {
156 * Ensure we either appear before our parent (if already
157 * enqueued) or force our parent to appear after us when it is
158 * enqueued. The fact that we always enqueue bottom-up
159 * reduces this to two cases.
161 if (cfs_rq->tg->parent &&
162 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
163 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
164 &rq_of(cfs_rq)->leaf_cfs_rq_list);
166 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
167 &rq_of(cfs_rq)->leaf_cfs_rq_list);
174 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
176 if (cfs_rq->on_list) {
177 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
182 /* Iterate thr' all leaf cfs_rq's on a runqueue */
183 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
184 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
186 /* Do the two (enqueued) entities belong to the same group ? */
188 is_same_group(struct sched_entity *se, struct sched_entity *pse)
190 if (se->cfs_rq == pse->cfs_rq)
196 static inline struct sched_entity *parent_entity(struct sched_entity *se)
201 /* return depth at which a sched entity is present in the hierarchy */
202 static inline int depth_se(struct sched_entity *se)
206 for_each_sched_entity(se)
213 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
215 int se_depth, pse_depth;
218 * preemption test can be made between sibling entities who are in the
219 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
220 * both tasks until we find their ancestors who are siblings of common
224 /* First walk up until both entities are at same depth */
225 se_depth = depth_se(*se);
226 pse_depth = depth_se(*pse);
228 while (se_depth > pse_depth) {
230 *se = parent_entity(*se);
233 while (pse_depth > se_depth) {
235 *pse = parent_entity(*pse);
238 while (!is_same_group(*se, *pse)) {
239 *se = parent_entity(*se);
240 *pse = parent_entity(*pse);
244 #else /* !CONFIG_FAIR_GROUP_SCHED */
246 static inline struct task_struct *task_of(struct sched_entity *se)
248 return container_of(se, struct task_struct, se);
251 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 return container_of(cfs_rq, struct rq, cfs);
256 #define entity_is_task(se) 1
258 #define for_each_sched_entity(se) \
259 for (; se; se = NULL)
261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
263 return &task_rq(p)->cfs;
266 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268 struct task_struct *p = task_of(se);
269 struct rq *rq = task_rq(p);
274 /* runqueue "owned" by this group */
275 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
280 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
284 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
289 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
292 is_same_group(struct sched_entity *se, struct sched_entity *pse)
297 static inline struct sched_entity *parent_entity(struct sched_entity *se)
303 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
310 unsigned long delta_exec);
312 /**************************************************************
313 * Scheduling class tree data structure manipulation methods:
316 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
318 s64 delta = (s64)(vruntime - min_vruntime);
320 min_vruntime = vruntime;
325 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
327 s64 delta = (s64)(vruntime - min_vruntime);
329 min_vruntime = vruntime;
334 static inline int entity_before(struct sched_entity *a,
335 struct sched_entity *b)
337 return (s64)(a->vruntime - b->vruntime) < 0;
340 static void update_min_vruntime(struct cfs_rq *cfs_rq)
342 u64 vruntime = cfs_rq->min_vruntime;
345 vruntime = cfs_rq->curr->vruntime;
347 if (cfs_rq->rb_leftmost) {
348 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
353 vruntime = se->vruntime;
355 vruntime = min_vruntime(vruntime, se->vruntime);
358 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
361 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
366 * Enqueue an entity into the rb-tree:
368 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
370 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
371 struct rb_node *parent = NULL;
372 struct sched_entity *entry;
376 * Find the right place in the rbtree:
380 entry = rb_entry(parent, struct sched_entity, run_node);
382 * We dont care about collisions. Nodes with
383 * the same key stay together.
385 if (entity_before(se, entry)) {
386 link = &parent->rb_left;
388 link = &parent->rb_right;
394 * Maintain a cache of leftmost tree entries (it is frequently
398 cfs_rq->rb_leftmost = &se->run_node;
400 rb_link_node(&se->run_node, parent, link);
401 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
404 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
406 if (cfs_rq->rb_leftmost == &se->run_node) {
407 struct rb_node *next_node;
409 next_node = rb_next(&se->run_node);
410 cfs_rq->rb_leftmost = next_node;
413 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
416 static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
418 struct rb_node *left = cfs_rq->rb_leftmost;
423 return rb_entry(left, struct sched_entity, run_node);
426 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
428 struct rb_node *next = rb_next(&se->run_node);
433 return rb_entry(next, struct sched_entity, run_node);
436 #ifdef CONFIG_SCHED_DEBUG
437 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
439 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
444 return rb_entry(last, struct sched_entity, run_node);
447 /**************************************************************
448 * Scheduling class statistics methods:
451 int sched_proc_update_handler(struct ctl_table *table, int write,
452 void __user *buffer, size_t *lenp,
455 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
456 int factor = get_update_sysctl_factor();
461 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
462 sysctl_sched_min_granularity);
464 #define WRT_SYSCTL(name) \
465 (normalized_sysctl_##name = sysctl_##name / (factor))
466 WRT_SYSCTL(sched_min_granularity);
467 WRT_SYSCTL(sched_latency);
468 WRT_SYSCTL(sched_wakeup_granularity);
478 static inline unsigned long
479 calc_delta_fair(unsigned long delta, struct sched_entity *se)
481 if (unlikely(se->load.weight != NICE_0_LOAD))
482 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
488 * The idea is to set a period in which each task runs once.
490 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
491 * this period because otherwise the slices get too small.
493 * p = (nr <= nl) ? l : l*nr/nl
495 static u64 __sched_period(unsigned long nr_running)
497 u64 period = sysctl_sched_latency;
498 unsigned long nr_latency = sched_nr_latency;
500 if (unlikely(nr_running > nr_latency)) {
501 period = sysctl_sched_min_granularity;
502 period *= nr_running;
509 * We calculate the wall-time slice from the period by taking a part
510 * proportional to the weight.
514 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
516 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
518 for_each_sched_entity(se) {
519 struct load_weight *load;
520 struct load_weight lw;
522 cfs_rq = cfs_rq_of(se);
523 load = &cfs_rq->load;
525 if (unlikely(!se->on_rq)) {
528 update_load_add(&lw, se->load.weight);
531 slice = calc_delta_mine(slice, se->load.weight, load);
537 * We calculate the vruntime slice of a to be inserted task
541 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 return calc_delta_fair(sched_slice(cfs_rq, se), se);
546 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
547 static void update_cfs_shares(struct cfs_rq *cfs_rq);
550 * Update the current task's runtime statistics. Skip current tasks that
551 * are not in our scheduling class.
554 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
555 unsigned long delta_exec)
557 unsigned long delta_exec_weighted;
559 schedstat_set(curr->statistics.exec_max,
560 max((u64)delta_exec, curr->statistics.exec_max));
562 curr->sum_exec_runtime += delta_exec;
563 schedstat_add(cfs_rq, exec_clock, delta_exec);
564 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
566 curr->vruntime += delta_exec_weighted;
567 update_min_vruntime(cfs_rq);
569 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
570 cfs_rq->load_unacc_exec_time += delta_exec;
574 static void update_curr(struct cfs_rq *cfs_rq)
576 struct sched_entity *curr = cfs_rq->curr;
577 u64 now = rq_of(cfs_rq)->clock_task;
578 unsigned long delta_exec;
584 * Get the amount of time the current task was running
585 * since the last time we changed load (this cannot
586 * overflow on 32 bits):
588 delta_exec = (unsigned long)(now - curr->exec_start);
592 __update_curr(cfs_rq, curr, delta_exec);
593 curr->exec_start = now;
595 if (entity_is_task(curr)) {
596 struct task_struct *curtask = task_of(curr);
598 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
599 cpuacct_charge(curtask, delta_exec);
600 account_group_exec_runtime(curtask, delta_exec);
603 account_cfs_rq_runtime(cfs_rq, delta_exec);
607 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
609 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
613 * Task is being enqueued - update stats:
615 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
618 * Are we enqueueing a waiting task? (for current tasks
619 * a dequeue/enqueue event is a NOP)
621 if (se != cfs_rq->curr)
622 update_stats_wait_start(cfs_rq, se);
626 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
629 rq_of(cfs_rq)->clock - se->statistics.wait_start));
630 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
631 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
632 rq_of(cfs_rq)->clock - se->statistics.wait_start);
633 #ifdef CONFIG_SCHEDSTATS
634 if (entity_is_task(se)) {
635 trace_sched_stat_wait(task_of(se),
636 rq_of(cfs_rq)->clock - se->statistics.wait_start);
639 schedstat_set(se->statistics.wait_start, 0);
643 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
646 * Mark the end of the wait period if dequeueing a
649 if (se != cfs_rq->curr)
650 update_stats_wait_end(cfs_rq, se);
654 * We are picking a new current task - update its stats:
657 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
660 * We are starting a new run period:
662 se->exec_start = rq_of(cfs_rq)->clock_task;
665 /**************************************************
666 * Scheduling class queueing methods:
669 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
671 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
673 cfs_rq->task_weight += weight;
677 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
683 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
685 update_load_add(&cfs_rq->load, se->load.weight);
686 if (!parent_entity(se))
687 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
688 if (entity_is_task(se)) {
689 add_cfs_task_weight(cfs_rq, se->load.weight);
690 list_add(&se->group_node, &cfs_rq->tasks);
692 cfs_rq->nr_running++;
696 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
698 update_load_sub(&cfs_rq->load, se->load.weight);
699 if (!parent_entity(se))
700 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
701 if (entity_is_task(se)) {
702 add_cfs_task_weight(cfs_rq, -se->load.weight);
703 list_del_init(&se->group_node);
705 cfs_rq->nr_running--;
708 #ifdef CONFIG_FAIR_GROUP_SCHED
709 /* we need this in update_cfs_load and load-balance functions below */
710 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
712 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
715 struct task_group *tg = cfs_rq->tg;
718 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
719 load_avg -= cfs_rq->load_contribution;
721 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
722 atomic_add(load_avg, &tg->load_weight);
723 cfs_rq->load_contribution += load_avg;
727 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
729 u64 period = sysctl_sched_shares_window;
731 unsigned long load = cfs_rq->load.weight;
733 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
736 now = rq_of(cfs_rq)->clock_task;
737 delta = now - cfs_rq->load_stamp;
739 /* truncate load history at 4 idle periods */
740 if (cfs_rq->load_stamp > cfs_rq->load_last &&
741 now - cfs_rq->load_last > 4 * period) {
742 cfs_rq->load_period = 0;
743 cfs_rq->load_avg = 0;
747 cfs_rq->load_stamp = now;
748 cfs_rq->load_unacc_exec_time = 0;
749 cfs_rq->load_period += delta;
751 cfs_rq->load_last = now;
752 cfs_rq->load_avg += delta * load;
755 /* consider updating load contribution on each fold or truncate */
756 if (global_update || cfs_rq->load_period > period
757 || !cfs_rq->load_period)
758 update_cfs_rq_load_contribution(cfs_rq, global_update);
760 while (cfs_rq->load_period > period) {
762 * Inline assembly required to prevent the compiler
763 * optimising this loop into a divmod call.
764 * See __iter_div_u64_rem() for another example of this.
766 asm("" : "+rm" (cfs_rq->load_period));
767 cfs_rq->load_period /= 2;
768 cfs_rq->load_avg /= 2;
771 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
772 list_del_leaf_cfs_rq(cfs_rq);
775 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
777 long load_weight, load, shares;
779 load = cfs_rq->load.weight;
781 load_weight = atomic_read(&tg->load_weight);
783 load_weight -= cfs_rq->load_contribution;
785 shares = (tg->shares * load);
787 shares /= load_weight;
789 if (shares < MIN_SHARES)
791 if (shares > tg->shares)
797 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
799 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
800 update_cfs_load(cfs_rq, 0);
801 update_cfs_shares(cfs_rq);
804 # else /* CONFIG_SMP */
805 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
809 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
814 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
817 # endif /* CONFIG_SMP */
818 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
819 unsigned long weight)
822 /* commit outstanding execution time */
823 if (cfs_rq->curr == se)
825 account_entity_dequeue(cfs_rq, se);
828 update_load_set(&se->load, weight);
831 account_entity_enqueue(cfs_rq, se);
834 static void update_cfs_shares(struct cfs_rq *cfs_rq)
836 struct task_group *tg;
837 struct sched_entity *se;
841 se = tg->se[cpu_of(rq_of(cfs_rq))];
842 if (!se || throttled_hierarchy(cfs_rq))
845 if (likely(se->load.weight == tg->shares))
848 shares = calc_cfs_shares(cfs_rq, tg);
850 reweight_entity(cfs_rq_of(se), se, shares);
852 #else /* CONFIG_FAIR_GROUP_SCHED */
853 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
857 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
861 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
864 #endif /* CONFIG_FAIR_GROUP_SCHED */
866 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
868 #ifdef CONFIG_SCHEDSTATS
869 struct task_struct *tsk = NULL;
871 if (entity_is_task(se))
874 if (se->statistics.sleep_start) {
875 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
880 if (unlikely(delta > se->statistics.sleep_max))
881 se->statistics.sleep_max = delta;
883 se->statistics.sleep_start = 0;
884 se->statistics.sum_sleep_runtime += delta;
887 account_scheduler_latency(tsk, delta >> 10, 1);
888 trace_sched_stat_sleep(tsk, delta);
891 if (se->statistics.block_start) {
892 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
897 if (unlikely(delta > se->statistics.block_max))
898 se->statistics.block_max = delta;
900 se->statistics.block_start = 0;
901 se->statistics.sum_sleep_runtime += delta;
904 if (tsk->in_iowait) {
905 se->statistics.iowait_sum += delta;
906 se->statistics.iowait_count++;
907 trace_sched_stat_iowait(tsk, delta);
911 * Blocking time is in units of nanosecs, so shift by
912 * 20 to get a milliseconds-range estimation of the
913 * amount of time that the task spent sleeping:
915 if (unlikely(prof_on == SLEEP_PROFILING)) {
916 profile_hits(SLEEP_PROFILING,
917 (void *)get_wchan(tsk),
920 account_scheduler_latency(tsk, delta >> 10, 0);
926 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
928 #ifdef CONFIG_SCHED_DEBUG
929 s64 d = se->vruntime - cfs_rq->min_vruntime;
934 if (d > 3*sysctl_sched_latency)
935 schedstat_inc(cfs_rq, nr_spread_over);
940 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
942 u64 vruntime = cfs_rq->min_vruntime;
945 * The 'current' period is already promised to the current tasks,
946 * however the extra weight of the new task will slow them down a
947 * little, place the new task so that it fits in the slot that
948 * stays open at the end.
950 if (initial && sched_feat(START_DEBIT))
951 vruntime += sched_vslice(cfs_rq, se);
953 /* sleeps up to a single latency don't count. */
955 unsigned long thresh = sysctl_sched_latency;
958 * Halve their sleep time's effect, to allow
959 * for a gentler effect of sleepers:
961 if (sched_feat(GENTLE_FAIR_SLEEPERS))
967 /* ensure we never gain time by being placed backwards. */
968 vruntime = max_vruntime(se->vruntime, vruntime);
970 se->vruntime = vruntime;
974 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
977 * Update the normalized vruntime before updating min_vruntime
978 * through callig update_curr().
980 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
981 se->vruntime += cfs_rq->min_vruntime;
984 * Update run-time statistics of the 'current'.
987 update_cfs_load(cfs_rq, 0);
988 account_entity_enqueue(cfs_rq, se);
989 update_cfs_shares(cfs_rq);
991 if (flags & ENQUEUE_WAKEUP) {
992 place_entity(cfs_rq, se, 0);
993 enqueue_sleeper(cfs_rq, se);
996 update_stats_enqueue(cfs_rq, se);
997 check_spread(cfs_rq, se);
998 if (se != cfs_rq->curr)
999 __enqueue_entity(cfs_rq, se);
1002 if (cfs_rq->nr_running == 1)
1003 list_add_leaf_cfs_rq(cfs_rq);
1006 static void __clear_buddies_last(struct sched_entity *se)
1008 for_each_sched_entity(se) {
1009 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1010 if (cfs_rq->last == se)
1011 cfs_rq->last = NULL;
1017 static void __clear_buddies_next(struct sched_entity *se)
1019 for_each_sched_entity(se) {
1020 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1021 if (cfs_rq->next == se)
1022 cfs_rq->next = NULL;
1028 static void __clear_buddies_skip(struct sched_entity *se)
1030 for_each_sched_entity(se) {
1031 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1032 if (cfs_rq->skip == se)
1033 cfs_rq->skip = NULL;
1039 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1041 if (cfs_rq->last == se)
1042 __clear_buddies_last(se);
1044 if (cfs_rq->next == se)
1045 __clear_buddies_next(se);
1047 if (cfs_rq->skip == se)
1048 __clear_buddies_skip(se);
1052 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1055 * Update run-time statistics of the 'current'.
1057 update_curr(cfs_rq);
1059 update_stats_dequeue(cfs_rq, se);
1060 if (flags & DEQUEUE_SLEEP) {
1061 #ifdef CONFIG_SCHEDSTATS
1062 if (entity_is_task(se)) {
1063 struct task_struct *tsk = task_of(se);
1065 if (tsk->state & TASK_INTERRUPTIBLE)
1066 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1067 if (tsk->state & TASK_UNINTERRUPTIBLE)
1068 se->statistics.block_start = rq_of(cfs_rq)->clock;
1073 clear_buddies(cfs_rq, se);
1075 if (se != cfs_rq->curr)
1076 __dequeue_entity(cfs_rq, se);
1078 update_cfs_load(cfs_rq, 0);
1079 account_entity_dequeue(cfs_rq, se);
1082 * Normalize the entity after updating the min_vruntime because the
1083 * update can refer to the ->curr item and we need to reflect this
1084 * movement in our normalized position.
1086 if (!(flags & DEQUEUE_SLEEP))
1087 se->vruntime -= cfs_rq->min_vruntime;
1089 update_min_vruntime(cfs_rq);
1090 update_cfs_shares(cfs_rq);
1094 * Preempt the current task with a newly woken task if needed:
1097 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1099 unsigned long ideal_runtime, delta_exec;
1101 ideal_runtime = sched_slice(cfs_rq, curr);
1102 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1103 if (delta_exec > ideal_runtime) {
1104 resched_task(rq_of(cfs_rq)->curr);
1106 * The current task ran long enough, ensure it doesn't get
1107 * re-elected due to buddy favours.
1109 clear_buddies(cfs_rq, curr);
1114 * Ensure that a task that missed wakeup preemption by a
1115 * narrow margin doesn't have to wait for a full slice.
1116 * This also mitigates buddy induced latencies under load.
1118 if (delta_exec < sysctl_sched_min_granularity)
1121 if (cfs_rq->nr_running > 1) {
1122 struct sched_entity *se = __pick_first_entity(cfs_rq);
1123 s64 delta = curr->vruntime - se->vruntime;
1128 if (delta > ideal_runtime)
1129 resched_task(rq_of(cfs_rq)->curr);
1134 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1136 /* 'current' is not kept within the tree. */
1139 * Any task has to be enqueued before it get to execute on
1140 * a CPU. So account for the time it spent waiting on the
1143 update_stats_wait_end(cfs_rq, se);
1144 __dequeue_entity(cfs_rq, se);
1147 update_stats_curr_start(cfs_rq, se);
1149 #ifdef CONFIG_SCHEDSTATS
1151 * Track our maximum slice length, if the CPU's load is at
1152 * least twice that of our own weight (i.e. dont track it
1153 * when there are only lesser-weight tasks around):
1155 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1156 se->statistics.slice_max = max(se->statistics.slice_max,
1157 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1160 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1164 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1167 * Pick the next process, keeping these things in mind, in this order:
1168 * 1) keep things fair between processes/task groups
1169 * 2) pick the "next" process, since someone really wants that to run
1170 * 3) pick the "last" process, for cache locality
1171 * 4) do not run the "skip" process, if something else is available
1173 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1175 struct sched_entity *se = __pick_first_entity(cfs_rq);
1176 struct sched_entity *left = se;
1179 * Avoid running the skip buddy, if running something else can
1180 * be done without getting too unfair.
1182 if (cfs_rq->skip == se) {
1183 struct sched_entity *second = __pick_next_entity(se);
1184 if (second && wakeup_preempt_entity(second, left) < 1)
1189 * Prefer last buddy, try to return the CPU to a preempted task.
1191 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1195 * Someone really wants this to run. If it's not unfair, run it.
1197 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1200 clear_buddies(cfs_rq, se);
1205 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1208 * If still on the runqueue then deactivate_task()
1209 * was not called and update_curr() has to be done:
1212 update_curr(cfs_rq);
1214 check_spread(cfs_rq, prev);
1216 update_stats_wait_start(cfs_rq, prev);
1217 /* Put 'current' back into the tree. */
1218 __enqueue_entity(cfs_rq, prev);
1220 cfs_rq->curr = NULL;
1224 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1227 * Update run-time statistics of the 'current'.
1229 update_curr(cfs_rq);
1232 * Update share accounting for long-running entities.
1234 update_entity_shares_tick(cfs_rq);
1236 #ifdef CONFIG_SCHED_HRTICK
1238 * queued ticks are scheduled to match the slice, so don't bother
1239 * validating it and just reschedule.
1242 resched_task(rq_of(cfs_rq)->curr);
1246 * don't let the period tick interfere with the hrtick preemption
1248 if (!sched_feat(DOUBLE_TICK) &&
1249 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1253 if (cfs_rq->nr_running > 1)
1254 check_preempt_tick(cfs_rq, curr);
1258 /**************************************************
1259 * CFS bandwidth control machinery
1262 #ifdef CONFIG_CFS_BANDWIDTH
1264 * default period for cfs group bandwidth.
1265 * default: 0.1s, units: nanoseconds
1267 static inline u64 default_cfs_period(void)
1269 return 100000000ULL;
1272 static inline u64 sched_cfs_bandwidth_slice(void)
1274 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1278 * Replenish runtime according to assigned quota and update expiration time.
1279 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1280 * additional synchronization around rq->lock.
1282 * requires cfs_b->lock
1284 static void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1288 if (cfs_b->quota == RUNTIME_INF)
1291 now = sched_clock_cpu(smp_processor_id());
1292 cfs_b->runtime = cfs_b->quota;
1293 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1296 /* returns 0 on failure to allocate runtime */
1297 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1299 struct task_group *tg = cfs_rq->tg;
1300 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1301 u64 amount = 0, min_amount, expires;
1303 /* note: this is a positive sum as runtime_remaining <= 0 */
1304 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1306 raw_spin_lock(&cfs_b->lock);
1307 if (cfs_b->quota == RUNTIME_INF)
1308 amount = min_amount;
1311 * If the bandwidth pool has become inactive, then at least one
1312 * period must have elapsed since the last consumption.
1313 * Refresh the global state and ensure bandwidth timer becomes
1316 if (!cfs_b->timer_active) {
1317 __refill_cfs_bandwidth_runtime(cfs_b);
1318 __start_cfs_bandwidth(cfs_b);
1321 if (cfs_b->runtime > 0) {
1322 amount = min(cfs_b->runtime, min_amount);
1323 cfs_b->runtime -= amount;
1327 expires = cfs_b->runtime_expires;
1328 raw_spin_unlock(&cfs_b->lock);
1330 cfs_rq->runtime_remaining += amount;
1332 * we may have advanced our local expiration to account for allowed
1333 * spread between our sched_clock and the one on which runtime was
1336 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1337 cfs_rq->runtime_expires = expires;
1339 return cfs_rq->runtime_remaining > 0;
1343 * Note: This depends on the synchronization provided by sched_clock and the
1344 * fact that rq->clock snapshots this value.
1346 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1348 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1349 struct rq *rq = rq_of(cfs_rq);
1351 /* if the deadline is ahead of our clock, nothing to do */
1352 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1355 if (cfs_rq->runtime_remaining < 0)
1359 * If the local deadline has passed we have to consider the
1360 * possibility that our sched_clock is 'fast' and the global deadline
1361 * has not truly expired.
1363 * Fortunately we can check determine whether this the case by checking
1364 * whether the global deadline has advanced.
1367 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1368 /* extend local deadline, drift is bounded above by 2 ticks */
1369 cfs_rq->runtime_expires += TICK_NSEC;
1371 /* global deadline is ahead, expiration has passed */
1372 cfs_rq->runtime_remaining = 0;
1376 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1377 unsigned long delta_exec)
1379 /* dock delta_exec before expiring quota (as it could span periods) */
1380 cfs_rq->runtime_remaining -= delta_exec;
1381 expire_cfs_rq_runtime(cfs_rq);
1383 if (likely(cfs_rq->runtime_remaining > 0))
1387 * if we're unable to extend our runtime we resched so that the active
1388 * hierarchy can be throttled
1390 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1391 resched_task(rq_of(cfs_rq)->curr);
1394 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1395 unsigned long delta_exec)
1397 if (!cfs_rq->runtime_enabled)
1400 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1403 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1405 return cfs_rq->throttled;
1408 /* check whether cfs_rq, or any parent, is throttled */
1409 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1411 return cfs_rq->throttle_count;
1415 * Ensure that neither of the group entities corresponding to src_cpu or
1416 * dest_cpu are members of a throttled hierarchy when performing group
1417 * load-balance operations.
1419 static inline int throttled_lb_pair(struct task_group *tg,
1420 int src_cpu, int dest_cpu)
1422 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1424 src_cfs_rq = tg->cfs_rq[src_cpu];
1425 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1427 return throttled_hierarchy(src_cfs_rq) ||
1428 throttled_hierarchy(dest_cfs_rq);
1431 /* updated child weight may affect parent so we have to do this bottom up */
1432 static int tg_unthrottle_up(struct task_group *tg, void *data)
1434 struct rq *rq = data;
1435 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1437 cfs_rq->throttle_count--;
1439 if (!cfs_rq->throttle_count) {
1440 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1442 /* leaving throttled state, advance shares averaging windows */
1443 cfs_rq->load_stamp += delta;
1444 cfs_rq->load_last += delta;
1446 /* update entity weight now that we are on_rq again */
1447 update_cfs_shares(cfs_rq);
1454 static int tg_throttle_down(struct task_group *tg, void *data)
1456 struct rq *rq = data;
1457 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1459 /* group is entering throttled state, record last load */
1460 if (!cfs_rq->throttle_count)
1461 update_cfs_load(cfs_rq, 0);
1462 cfs_rq->throttle_count++;
1467 static __used void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1469 struct rq *rq = rq_of(cfs_rq);
1470 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1471 struct sched_entity *se;
1472 long task_delta, dequeue = 1;
1474 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1476 /* account load preceding throttle */
1478 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1481 task_delta = cfs_rq->h_nr_running;
1482 for_each_sched_entity(se) {
1483 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1484 /* throttled entity or throttle-on-deactivate */
1489 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1490 qcfs_rq->h_nr_running -= task_delta;
1492 if (qcfs_rq->load.weight)
1497 rq->nr_running -= task_delta;
1499 cfs_rq->throttled = 1;
1500 raw_spin_lock(&cfs_b->lock);
1501 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1502 raw_spin_unlock(&cfs_b->lock);
1505 static void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1507 struct rq *rq = rq_of(cfs_rq);
1508 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1509 struct sched_entity *se;
1513 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1515 cfs_rq->throttled = 0;
1516 raw_spin_lock(&cfs_b->lock);
1517 list_del_rcu(&cfs_rq->throttled_list);
1518 raw_spin_unlock(&cfs_b->lock);
1520 update_rq_clock(rq);
1521 /* update hierarchical throttle state */
1522 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1524 if (!cfs_rq->load.weight)
1527 task_delta = cfs_rq->h_nr_running;
1528 for_each_sched_entity(se) {
1532 cfs_rq = cfs_rq_of(se);
1534 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1535 cfs_rq->h_nr_running += task_delta;
1537 if (cfs_rq_throttled(cfs_rq))
1542 rq->nr_running += task_delta;
1544 /* determine whether we need to wake up potentially idle cpu */
1545 if (rq->curr == rq->idle && rq->cfs.nr_running)
1546 resched_task(rq->curr);
1549 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1550 u64 remaining, u64 expires)
1552 struct cfs_rq *cfs_rq;
1553 u64 runtime = remaining;
1556 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1558 struct rq *rq = rq_of(cfs_rq);
1560 raw_spin_lock(&rq->lock);
1561 if (!cfs_rq_throttled(cfs_rq))
1564 runtime = -cfs_rq->runtime_remaining + 1;
1565 if (runtime > remaining)
1566 runtime = remaining;
1567 remaining -= runtime;
1569 cfs_rq->runtime_remaining += runtime;
1570 cfs_rq->runtime_expires = expires;
1572 /* we check whether we're throttled above */
1573 if (cfs_rq->runtime_remaining > 0)
1574 unthrottle_cfs_rq(cfs_rq);
1577 raw_spin_unlock(&rq->lock);
1588 * Responsible for refilling a task_group's bandwidth and unthrottling its
1589 * cfs_rqs as appropriate. If there has been no activity within the last
1590 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1591 * used to track this state.
1593 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1595 u64 runtime, runtime_expires;
1596 int idle = 1, throttled;
1598 raw_spin_lock(&cfs_b->lock);
1599 /* no need to continue the timer with no bandwidth constraint */
1600 if (cfs_b->quota == RUNTIME_INF)
1603 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1604 /* idle depends on !throttled (for the case of a large deficit) */
1605 idle = cfs_b->idle && !throttled;
1607 /* if we're going inactive then everything else can be deferred */
1611 __refill_cfs_bandwidth_runtime(cfs_b);
1614 /* mark as potentially idle for the upcoming period */
1620 * There are throttled entities so we must first use the new bandwidth
1621 * to unthrottle them before making it generally available. This
1622 * ensures that all existing debts will be paid before a new cfs_rq is
1625 runtime = cfs_b->runtime;
1626 runtime_expires = cfs_b->runtime_expires;
1630 * This check is repeated as we are holding onto the new bandwidth
1631 * while we unthrottle. This can potentially race with an unthrottled
1632 * group trying to acquire new bandwidth from the global pool.
1634 while (throttled && runtime > 0) {
1635 raw_spin_unlock(&cfs_b->lock);
1636 /* we can't nest cfs_b->lock while distributing bandwidth */
1637 runtime = distribute_cfs_runtime(cfs_b, runtime,
1639 raw_spin_lock(&cfs_b->lock);
1641 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1644 /* return (any) remaining runtime */
1645 cfs_b->runtime = runtime;
1647 * While we are ensured activity in the period following an
1648 * unthrottle, this also covers the case in which the new bandwidth is
1649 * insufficient to cover the existing bandwidth deficit. (Forcing the
1650 * timer to remain active while there are any throttled entities.)
1655 cfs_b->timer_active = 0;
1656 raw_spin_unlock(&cfs_b->lock);
1661 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1662 unsigned long delta_exec) {}
1664 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1669 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1674 static inline int throttled_lb_pair(struct task_group *tg,
1675 int src_cpu, int dest_cpu)
1681 /**************************************************
1682 * CFS operations on tasks:
1685 #ifdef CONFIG_SCHED_HRTICK
1686 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1688 struct sched_entity *se = &p->se;
1689 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1691 WARN_ON(task_rq(p) != rq);
1693 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1694 u64 slice = sched_slice(cfs_rq, se);
1695 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1696 s64 delta = slice - ran;
1705 * Don't schedule slices shorter than 10000ns, that just
1706 * doesn't make sense. Rely on vruntime for fairness.
1709 delta = max_t(s64, 10000LL, delta);
1711 hrtick_start(rq, delta);
1716 * called from enqueue/dequeue and updates the hrtick when the
1717 * current task is from our class and nr_running is low enough
1720 static void hrtick_update(struct rq *rq)
1722 struct task_struct *curr = rq->curr;
1724 if (curr->sched_class != &fair_sched_class)
1727 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1728 hrtick_start_fair(rq, curr);
1730 #else /* !CONFIG_SCHED_HRTICK */
1732 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1736 static inline void hrtick_update(struct rq *rq)
1742 * The enqueue_task method is called before nr_running is
1743 * increased. Here we update the fair scheduling stats and
1744 * then put the task into the rbtree:
1747 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1749 struct cfs_rq *cfs_rq;
1750 struct sched_entity *se = &p->se;
1752 for_each_sched_entity(se) {
1755 cfs_rq = cfs_rq_of(se);
1756 enqueue_entity(cfs_rq, se, flags);
1759 * end evaluation on encountering a throttled cfs_rq
1761 * note: in the case of encountering a throttled cfs_rq we will
1762 * post the final h_nr_running increment below.
1764 if (cfs_rq_throttled(cfs_rq))
1766 cfs_rq->h_nr_running++;
1768 flags = ENQUEUE_WAKEUP;
1771 for_each_sched_entity(se) {
1772 cfs_rq = cfs_rq_of(se);
1773 cfs_rq->h_nr_running++;
1775 if (cfs_rq_throttled(cfs_rq))
1778 update_cfs_load(cfs_rq, 0);
1779 update_cfs_shares(cfs_rq);
1787 static void set_next_buddy(struct sched_entity *se);
1790 * The dequeue_task method is called before nr_running is
1791 * decreased. We remove the task from the rbtree and
1792 * update the fair scheduling stats:
1794 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1796 struct cfs_rq *cfs_rq;
1797 struct sched_entity *se = &p->se;
1798 int task_sleep = flags & DEQUEUE_SLEEP;
1800 for_each_sched_entity(se) {
1801 cfs_rq = cfs_rq_of(se);
1802 dequeue_entity(cfs_rq, se, flags);
1805 * end evaluation on encountering a throttled cfs_rq
1807 * note: in the case of encountering a throttled cfs_rq we will
1808 * post the final h_nr_running decrement below.
1810 if (cfs_rq_throttled(cfs_rq))
1812 cfs_rq->h_nr_running--;
1814 /* Don't dequeue parent if it has other entities besides us */
1815 if (cfs_rq->load.weight) {
1817 * Bias pick_next to pick a task from this cfs_rq, as
1818 * p is sleeping when it is within its sched_slice.
1820 if (task_sleep && parent_entity(se))
1821 set_next_buddy(parent_entity(se));
1823 /* avoid re-evaluating load for this entity */
1824 se = parent_entity(se);
1827 flags |= DEQUEUE_SLEEP;
1830 for_each_sched_entity(se) {
1831 cfs_rq = cfs_rq_of(se);
1832 cfs_rq->h_nr_running--;
1834 if (cfs_rq_throttled(cfs_rq))
1837 update_cfs_load(cfs_rq, 0);
1838 update_cfs_shares(cfs_rq);
1848 static void task_waking_fair(struct task_struct *p)
1850 struct sched_entity *se = &p->se;
1851 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1854 #ifndef CONFIG_64BIT
1855 u64 min_vruntime_copy;
1858 min_vruntime_copy = cfs_rq->min_vruntime_copy;
1860 min_vruntime = cfs_rq->min_vruntime;
1861 } while (min_vruntime != min_vruntime_copy);
1863 min_vruntime = cfs_rq->min_vruntime;
1866 se->vruntime -= min_vruntime;
1869 #ifdef CONFIG_FAIR_GROUP_SCHED
1871 * effective_load() calculates the load change as seen from the root_task_group
1873 * Adding load to a group doesn't make a group heavier, but can cause movement
1874 * of group shares between cpus. Assuming the shares were perfectly aligned one
1875 * can calculate the shift in shares.
1877 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1879 struct sched_entity *se = tg->se[cpu];
1884 for_each_sched_entity(se) {
1888 w = se->my_q->load.weight;
1890 /* use this cpu's instantaneous contribution */
1891 lw = atomic_read(&tg->load_weight);
1892 lw -= se->my_q->load_contribution;
1897 if (lw > 0 && wl < lw)
1898 wl = (wl * tg->shares) / lw;
1902 /* zero point is MIN_SHARES */
1903 if (wl < MIN_SHARES)
1905 wl -= se->load.weight;
1913 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1914 unsigned long wl, unsigned long wg)
1921 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1923 s64 this_load, load;
1924 int idx, this_cpu, prev_cpu;
1925 unsigned long tl_per_task;
1926 struct task_group *tg;
1927 unsigned long weight;
1931 this_cpu = smp_processor_id();
1932 prev_cpu = task_cpu(p);
1933 load = source_load(prev_cpu, idx);
1934 this_load = target_load(this_cpu, idx);
1937 * If sync wakeup then subtract the (maximum possible)
1938 * effect of the currently running task from the load
1939 * of the current CPU:
1942 tg = task_group(current);
1943 weight = current->se.load.weight;
1945 this_load += effective_load(tg, this_cpu, -weight, -weight);
1946 load += effective_load(tg, prev_cpu, 0, -weight);
1950 weight = p->se.load.weight;
1953 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1954 * due to the sync cause above having dropped this_load to 0, we'll
1955 * always have an imbalance, but there's really nothing you can do
1956 * about that, so that's good too.
1958 * Otherwise check if either cpus are near enough in load to allow this
1959 * task to be woken on this_cpu.
1961 if (this_load > 0) {
1962 s64 this_eff_load, prev_eff_load;
1964 this_eff_load = 100;
1965 this_eff_load *= power_of(prev_cpu);
1966 this_eff_load *= this_load +
1967 effective_load(tg, this_cpu, weight, weight);
1969 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1970 prev_eff_load *= power_of(this_cpu);
1971 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1973 balanced = this_eff_load <= prev_eff_load;
1978 * If the currently running task will sleep within
1979 * a reasonable amount of time then attract this newly
1982 if (sync && balanced)
1985 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1986 tl_per_task = cpu_avg_load_per_task(this_cpu);
1989 (this_load <= load &&
1990 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1992 * This domain has SD_WAKE_AFFINE and
1993 * p is cache cold in this domain, and
1994 * there is no bad imbalance.
1996 schedstat_inc(sd, ttwu_move_affine);
1997 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2005 * find_idlest_group finds and returns the least busy CPU group within the
2008 static struct sched_group *
2009 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2010 int this_cpu, int load_idx)
2012 struct sched_group *idlest = NULL, *group = sd->groups;
2013 unsigned long min_load = ULONG_MAX, this_load = 0;
2014 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2017 unsigned long load, avg_load;
2021 /* Skip over this group if it has no CPUs allowed */
2022 if (!cpumask_intersects(sched_group_cpus(group),
2026 local_group = cpumask_test_cpu(this_cpu,
2027 sched_group_cpus(group));
2029 /* Tally up the load of all CPUs in the group */
2032 for_each_cpu(i, sched_group_cpus(group)) {
2033 /* Bias balancing toward cpus of our domain */
2035 load = source_load(i, load_idx);
2037 load = target_load(i, load_idx);
2042 /* Adjust by relative CPU power of the group */
2043 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2046 this_load = avg_load;
2047 } else if (avg_load < min_load) {
2048 min_load = avg_load;
2051 } while (group = group->next, group != sd->groups);
2053 if (!idlest || 100*this_load < imbalance*min_load)
2059 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2062 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2064 unsigned long load, min_load = ULONG_MAX;
2068 /* Traverse only the allowed CPUs */
2069 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2070 load = weighted_cpuload(i);
2072 if (load < min_load || (load == min_load && i == this_cpu)) {
2082 * Try and locate an idle CPU in the sched_domain.
2084 static int select_idle_sibling(struct task_struct *p, int target)
2086 int cpu = smp_processor_id();
2087 int prev_cpu = task_cpu(p);
2088 struct sched_domain *sd;
2092 * If the task is going to be woken-up on this cpu and if it is
2093 * already idle, then it is the right target.
2095 if (target == cpu && idle_cpu(cpu))
2099 * If the task is going to be woken-up on the cpu where it previously
2100 * ran and if it is currently idle, then it the right target.
2102 if (target == prev_cpu && idle_cpu(prev_cpu))
2106 * Otherwise, iterate the domains and find an elegible idle cpu.
2109 for_each_domain(target, sd) {
2110 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
2113 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
2121 * Lets stop looking for an idle sibling when we reached
2122 * the domain that spans the current cpu and prev_cpu.
2124 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
2125 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
2134 * sched_balance_self: balance the current task (running on cpu) in domains
2135 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2138 * Balance, ie. select the least loaded group.
2140 * Returns the target CPU number, or the same CPU if no balancing is needed.
2142 * preempt must be disabled.
2145 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2147 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2148 int cpu = smp_processor_id();
2149 int prev_cpu = task_cpu(p);
2151 int want_affine = 0;
2153 int sync = wake_flags & WF_SYNC;
2155 if (sd_flag & SD_BALANCE_WAKE) {
2156 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
2162 for_each_domain(cpu, tmp) {
2163 if (!(tmp->flags & SD_LOAD_BALANCE))
2167 * If power savings logic is enabled for a domain, see if we
2168 * are not overloaded, if so, don't balance wider.
2170 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2171 unsigned long power = 0;
2172 unsigned long nr_running = 0;
2173 unsigned long capacity;
2176 for_each_cpu(i, sched_domain_span(tmp)) {
2177 power += power_of(i);
2178 nr_running += cpu_rq(i)->cfs.nr_running;
2181 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2183 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2186 if (nr_running < capacity)
2191 * If both cpu and prev_cpu are part of this domain,
2192 * cpu is a valid SD_WAKE_AFFINE target.
2194 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2195 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2200 if (!want_sd && !want_affine)
2203 if (!(tmp->flags & sd_flag))
2211 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2214 new_cpu = select_idle_sibling(p, prev_cpu);
2219 int load_idx = sd->forkexec_idx;
2220 struct sched_group *group;
2223 if (!(sd->flags & sd_flag)) {
2228 if (sd_flag & SD_BALANCE_WAKE)
2229 load_idx = sd->wake_idx;
2231 group = find_idlest_group(sd, p, cpu, load_idx);
2237 new_cpu = find_idlest_cpu(group, p, cpu);
2238 if (new_cpu == -1 || new_cpu == cpu) {
2239 /* Now try balancing at a lower domain level of cpu */
2244 /* Now try balancing at a lower domain level of new_cpu */
2246 weight = sd->span_weight;
2248 for_each_domain(cpu, tmp) {
2249 if (weight <= tmp->span_weight)
2251 if (tmp->flags & sd_flag)
2254 /* while loop will break here if sd == NULL */
2261 #endif /* CONFIG_SMP */
2263 static unsigned long
2264 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2266 unsigned long gran = sysctl_sched_wakeup_granularity;
2269 * Since its curr running now, convert the gran from real-time
2270 * to virtual-time in his units.
2272 * By using 'se' instead of 'curr' we penalize light tasks, so
2273 * they get preempted easier. That is, if 'se' < 'curr' then
2274 * the resulting gran will be larger, therefore penalizing the
2275 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2276 * be smaller, again penalizing the lighter task.
2278 * This is especially important for buddies when the leftmost
2279 * task is higher priority than the buddy.
2281 return calc_delta_fair(gran, se);
2285 * Should 'se' preempt 'curr'.
2299 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2301 s64 gran, vdiff = curr->vruntime - se->vruntime;
2306 gran = wakeup_gran(curr, se);
2313 static void set_last_buddy(struct sched_entity *se)
2315 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2318 for_each_sched_entity(se)
2319 cfs_rq_of(se)->last = se;
2322 static void set_next_buddy(struct sched_entity *se)
2324 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2327 for_each_sched_entity(se)
2328 cfs_rq_of(se)->next = se;
2331 static void set_skip_buddy(struct sched_entity *se)
2333 for_each_sched_entity(se)
2334 cfs_rq_of(se)->skip = se;
2338 * Preempt the current task with a newly woken task if needed:
2340 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2342 struct task_struct *curr = rq->curr;
2343 struct sched_entity *se = &curr->se, *pse = &p->se;
2344 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2345 int scale = cfs_rq->nr_running >= sched_nr_latency;
2346 int next_buddy_marked = 0;
2348 if (unlikely(se == pse))
2351 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2352 set_next_buddy(pse);
2353 next_buddy_marked = 1;
2357 * We can come here with TIF_NEED_RESCHED already set from new task
2360 if (test_tsk_need_resched(curr))
2363 /* Idle tasks are by definition preempted by non-idle tasks. */
2364 if (unlikely(curr->policy == SCHED_IDLE) &&
2365 likely(p->policy != SCHED_IDLE))
2369 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2370 * is driven by the tick):
2372 if (unlikely(p->policy != SCHED_NORMAL))
2375 find_matching_se(&se, &pse);
2376 update_curr(cfs_rq_of(se));
2378 if (wakeup_preempt_entity(se, pse) == 1) {
2380 * Bias pick_next to pick the sched entity that is
2381 * triggering this preemption.
2383 if (!next_buddy_marked)
2384 set_next_buddy(pse);
2393 * Only set the backward buddy when the current task is still
2394 * on the rq. This can happen when a wakeup gets interleaved
2395 * with schedule on the ->pre_schedule() or idle_balance()
2396 * point, either of which can * drop the rq lock.
2398 * Also, during early boot the idle thread is in the fair class,
2399 * for obvious reasons its a bad idea to schedule back to it.
2401 if (unlikely(!se->on_rq || curr == rq->idle))
2404 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2408 static struct task_struct *pick_next_task_fair(struct rq *rq)
2410 struct task_struct *p;
2411 struct cfs_rq *cfs_rq = &rq->cfs;
2412 struct sched_entity *se;
2414 if (!cfs_rq->nr_running)
2418 se = pick_next_entity(cfs_rq);
2419 set_next_entity(cfs_rq, se);
2420 cfs_rq = group_cfs_rq(se);
2424 hrtick_start_fair(rq, p);
2430 * Account for a descheduled task:
2432 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
2434 struct sched_entity *se = &prev->se;
2435 struct cfs_rq *cfs_rq;
2437 for_each_sched_entity(se) {
2438 cfs_rq = cfs_rq_of(se);
2439 put_prev_entity(cfs_rq, se);
2444 * sched_yield() is very simple
2446 * The magic of dealing with the ->skip buddy is in pick_next_entity.
2448 static void yield_task_fair(struct rq *rq)
2450 struct task_struct *curr = rq->curr;
2451 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2452 struct sched_entity *se = &curr->se;
2455 * Are we the only task in the tree?
2457 if (unlikely(rq->nr_running == 1))
2460 clear_buddies(cfs_rq, se);
2462 if (curr->policy != SCHED_BATCH) {
2463 update_rq_clock(rq);
2465 * Update run-time statistics of the 'current'.
2467 update_curr(cfs_rq);
2473 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2475 struct sched_entity *se = &p->se;
2480 /* Tell the scheduler that we'd really like pse to run next. */
2483 yield_task_fair(rq);
2489 /**************************************************
2490 * Fair scheduling class load-balancing methods:
2494 * pull_task - move a task from a remote runqueue to the local runqueue.
2495 * Both runqueues must be locked.
2497 static void pull_task(struct rq *src_rq, struct task_struct *p,
2498 struct rq *this_rq, int this_cpu)
2500 deactivate_task(src_rq, p, 0);
2501 set_task_cpu(p, this_cpu);
2502 activate_task(this_rq, p, 0);
2503 check_preempt_curr(this_rq, p, 0);
2507 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2510 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2511 struct sched_domain *sd, enum cpu_idle_type idle,
2514 int tsk_cache_hot = 0;
2516 * We do not migrate tasks that are:
2517 * 1) running (obviously), or
2518 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2519 * 3) are cache-hot on their current CPU.
2521 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2522 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2527 if (task_running(rq, p)) {
2528 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2533 * Aggressive migration if:
2534 * 1) task is cache cold, or
2535 * 2) too many balance attempts have failed.
2538 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2539 if (!tsk_cache_hot ||
2540 sd->nr_balance_failed > sd->cache_nice_tries) {
2541 #ifdef CONFIG_SCHEDSTATS
2542 if (tsk_cache_hot) {
2543 schedstat_inc(sd, lb_hot_gained[idle]);
2544 schedstat_inc(p, se.statistics.nr_forced_migrations);
2550 if (tsk_cache_hot) {
2551 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2558 * move_one_task tries to move exactly one task from busiest to this_rq, as
2559 * part of active balancing operations within "domain".
2560 * Returns 1 if successful and 0 otherwise.
2562 * Called with both runqueues locked.
2565 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2566 struct sched_domain *sd, enum cpu_idle_type idle)
2568 struct task_struct *p, *n;
2569 struct cfs_rq *cfs_rq;
2572 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2573 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2574 if (throttled_lb_pair(task_group(p),
2575 busiest->cpu, this_cpu))
2578 if (!can_migrate_task(p, busiest, this_cpu,
2582 pull_task(busiest, p, this_rq, this_cpu);
2584 * Right now, this is only the second place pull_task()
2585 * is called, so we can safely collect pull_task()
2586 * stats here rather than inside pull_task().
2588 schedstat_inc(sd, lb_gained[idle]);
2596 static unsigned long
2597 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2598 unsigned long max_load_move, struct sched_domain *sd,
2599 enum cpu_idle_type idle, int *all_pinned,
2600 struct cfs_rq *busiest_cfs_rq)
2602 int loops = 0, pulled = 0;
2603 long rem_load_move = max_load_move;
2604 struct task_struct *p, *n;
2606 if (max_load_move == 0)
2609 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2610 if (loops++ > sysctl_sched_nr_migrate)
2613 if ((p->se.load.weight >> 1) > rem_load_move ||
2614 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2618 pull_task(busiest, p, this_rq, this_cpu);
2620 rem_load_move -= p->se.load.weight;
2622 #ifdef CONFIG_PREEMPT
2624 * NEWIDLE balancing is a source of latency, so preemptible
2625 * kernels will stop after the first task is pulled to minimize
2626 * the critical section.
2628 if (idle == CPU_NEWLY_IDLE)
2633 * We only want to steal up to the prescribed amount of
2636 if (rem_load_move <= 0)
2641 * Right now, this is one of only two places pull_task() is called,
2642 * so we can safely collect pull_task() stats here rather than
2643 * inside pull_task().
2645 schedstat_add(sd, lb_gained[idle], pulled);
2647 return max_load_move - rem_load_move;
2650 #ifdef CONFIG_FAIR_GROUP_SCHED
2652 * update tg->load_weight by folding this cpu's load_avg
2654 static int update_shares_cpu(struct task_group *tg, int cpu)
2656 struct cfs_rq *cfs_rq;
2657 unsigned long flags;
2664 cfs_rq = tg->cfs_rq[cpu];
2666 raw_spin_lock_irqsave(&rq->lock, flags);
2668 update_rq_clock(rq);
2669 update_cfs_load(cfs_rq, 1);
2672 * We need to update shares after updating tg->load_weight in
2673 * order to adjust the weight of groups with long running tasks.
2675 update_cfs_shares(cfs_rq);
2677 raw_spin_unlock_irqrestore(&rq->lock, flags);
2682 static void update_shares(int cpu)
2684 struct cfs_rq *cfs_rq;
2685 struct rq *rq = cpu_rq(cpu);
2689 * Iterates the task_group tree in a bottom up fashion, see
2690 * list_add_leaf_cfs_rq() for details.
2692 for_each_leaf_cfs_rq(rq, cfs_rq) {
2693 /* throttled entities do not contribute to load */
2694 if (throttled_hierarchy(cfs_rq))
2697 update_shares_cpu(cfs_rq->tg, cpu);
2703 * Compute the cpu's hierarchical load factor for each task group.
2704 * This needs to be done in a top-down fashion because the load of a child
2705 * group is a fraction of its parents load.
2707 static int tg_load_down(struct task_group *tg, void *data)
2710 long cpu = (long)data;
2713 load = cpu_rq(cpu)->load.weight;
2715 load = tg->parent->cfs_rq[cpu]->h_load;
2716 load *= tg->se[cpu]->load.weight;
2717 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2720 tg->cfs_rq[cpu]->h_load = load;
2725 static void update_h_load(long cpu)
2727 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2730 static unsigned long
2731 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2732 unsigned long max_load_move,
2733 struct sched_domain *sd, enum cpu_idle_type idle,
2736 long rem_load_move = max_load_move;
2737 struct cfs_rq *busiest_cfs_rq;
2740 update_h_load(cpu_of(busiest));
2742 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
2743 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2744 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2745 u64 rem_load, moved_load;
2748 * empty group or part of a throttled hierarchy
2750 if (!busiest_cfs_rq->task_weight ||
2751 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu))
2754 rem_load = (u64)rem_load_move * busiest_weight;
2755 rem_load = div_u64(rem_load, busiest_h_load + 1);
2757 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2758 rem_load, sd, idle, all_pinned,
2764 moved_load *= busiest_h_load;
2765 moved_load = div_u64(moved_load, busiest_weight + 1);
2767 rem_load_move -= moved_load;
2768 if (rem_load_move < 0)
2773 return max_load_move - rem_load_move;
2776 static inline void update_shares(int cpu)
2780 static unsigned long
2781 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2782 unsigned long max_load_move,
2783 struct sched_domain *sd, enum cpu_idle_type idle,
2786 return balance_tasks(this_rq, this_cpu, busiest,
2787 max_load_move, sd, idle, all_pinned,
2793 * move_tasks tries to move up to max_load_move weighted load from busiest to
2794 * this_rq, as part of a balancing operation within domain "sd".
2795 * Returns 1 if successful and 0 otherwise.
2797 * Called with both runqueues locked.
2799 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2800 unsigned long max_load_move,
2801 struct sched_domain *sd, enum cpu_idle_type idle,
2804 unsigned long total_load_moved = 0, load_moved;
2807 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2808 max_load_move - total_load_moved,
2809 sd, idle, all_pinned);
2811 total_load_moved += load_moved;
2813 #ifdef CONFIG_PREEMPT
2815 * NEWIDLE balancing is a source of latency, so preemptible
2816 * kernels will stop after the first task is pulled to minimize
2817 * the critical section.
2819 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2822 if (raw_spin_is_contended(&this_rq->lock) ||
2823 raw_spin_is_contended(&busiest->lock))
2826 } while (load_moved && max_load_move > total_load_moved);
2828 return total_load_moved > 0;
2831 /********** Helpers for find_busiest_group ************************/
2833 * sd_lb_stats - Structure to store the statistics of a sched_domain
2834 * during load balancing.
2836 struct sd_lb_stats {
2837 struct sched_group *busiest; /* Busiest group in this sd */
2838 struct sched_group *this; /* Local group in this sd */
2839 unsigned long total_load; /* Total load of all groups in sd */
2840 unsigned long total_pwr; /* Total power of all groups in sd */
2841 unsigned long avg_load; /* Average load across all groups in sd */
2843 /** Statistics of this group */
2844 unsigned long this_load;
2845 unsigned long this_load_per_task;
2846 unsigned long this_nr_running;
2847 unsigned long this_has_capacity;
2848 unsigned int this_idle_cpus;
2850 /* Statistics of the busiest group */
2851 unsigned int busiest_idle_cpus;
2852 unsigned long max_load;
2853 unsigned long busiest_load_per_task;
2854 unsigned long busiest_nr_running;
2855 unsigned long busiest_group_capacity;
2856 unsigned long busiest_has_capacity;
2857 unsigned int busiest_group_weight;
2859 int group_imb; /* Is there imbalance in this sd */
2860 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2861 int power_savings_balance; /* Is powersave balance needed for this sd */
2862 struct sched_group *group_min; /* Least loaded group in sd */
2863 struct sched_group *group_leader; /* Group which relieves group_min */
2864 unsigned long min_load_per_task; /* load_per_task in group_min */
2865 unsigned long leader_nr_running; /* Nr running of group_leader */
2866 unsigned long min_nr_running; /* Nr running of group_min */
2871 * sg_lb_stats - stats of a sched_group required for load_balancing
2873 struct sg_lb_stats {
2874 unsigned long avg_load; /*Avg load across the CPUs of the group */
2875 unsigned long group_load; /* Total load over the CPUs of the group */
2876 unsigned long sum_nr_running; /* Nr tasks running in the group */
2877 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2878 unsigned long group_capacity;
2879 unsigned long idle_cpus;
2880 unsigned long group_weight;
2881 int group_imb; /* Is there an imbalance in the group ? */
2882 int group_has_capacity; /* Is there extra capacity in the group? */
2886 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2887 * @group: The group whose first cpu is to be returned.
2889 static inline unsigned int group_first_cpu(struct sched_group *group)
2891 return cpumask_first(sched_group_cpus(group));
2895 * get_sd_load_idx - Obtain the load index for a given sched domain.
2896 * @sd: The sched_domain whose load_idx is to be obtained.
2897 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2899 static inline int get_sd_load_idx(struct sched_domain *sd,
2900 enum cpu_idle_type idle)
2906 load_idx = sd->busy_idx;
2909 case CPU_NEWLY_IDLE:
2910 load_idx = sd->newidle_idx;
2913 load_idx = sd->idle_idx;
2921 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2923 * init_sd_power_savings_stats - Initialize power savings statistics for
2924 * the given sched_domain, during load balancing.
2926 * @sd: Sched domain whose power-savings statistics are to be initialized.
2927 * @sds: Variable containing the statistics for sd.
2928 * @idle: Idle status of the CPU at which we're performing load-balancing.
2930 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2931 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2934 * Busy processors will not participate in power savings
2937 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2938 sds->power_savings_balance = 0;
2940 sds->power_savings_balance = 1;
2941 sds->min_nr_running = ULONG_MAX;
2942 sds->leader_nr_running = 0;
2947 * update_sd_power_savings_stats - Update the power saving stats for a
2948 * sched_domain while performing load balancing.
2950 * @group: sched_group belonging to the sched_domain under consideration.
2951 * @sds: Variable containing the statistics of the sched_domain
2952 * @local_group: Does group contain the CPU for which we're performing
2954 * @sgs: Variable containing the statistics of the group.
2956 static inline void update_sd_power_savings_stats(struct sched_group *group,
2957 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2960 if (!sds->power_savings_balance)
2964 * If the local group is idle or completely loaded
2965 * no need to do power savings balance at this domain
2967 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2968 !sds->this_nr_running))
2969 sds->power_savings_balance = 0;
2972 * If a group is already running at full capacity or idle,
2973 * don't include that group in power savings calculations
2975 if (!sds->power_savings_balance ||
2976 sgs->sum_nr_running >= sgs->group_capacity ||
2977 !sgs->sum_nr_running)
2981 * Calculate the group which has the least non-idle load.
2982 * This is the group from where we need to pick up the load
2985 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2986 (sgs->sum_nr_running == sds->min_nr_running &&
2987 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2988 sds->group_min = group;
2989 sds->min_nr_running = sgs->sum_nr_running;
2990 sds->min_load_per_task = sgs->sum_weighted_load /
2991 sgs->sum_nr_running;
2995 * Calculate the group which is almost near its
2996 * capacity but still has some space to pick up some load
2997 * from other group and save more power
2999 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3002 if (sgs->sum_nr_running > sds->leader_nr_running ||
3003 (sgs->sum_nr_running == sds->leader_nr_running &&
3004 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3005 sds->group_leader = group;
3006 sds->leader_nr_running = sgs->sum_nr_running;
3011 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3012 * @sds: Variable containing the statistics of the sched_domain
3013 * under consideration.
3014 * @this_cpu: Cpu at which we're currently performing load-balancing.
3015 * @imbalance: Variable to store the imbalance.
3018 * Check if we have potential to perform some power-savings balance.
3019 * If yes, set the busiest group to be the least loaded group in the
3020 * sched_domain, so that it's CPUs can be put to idle.
3022 * Returns 1 if there is potential to perform power-savings balance.
3025 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3026 int this_cpu, unsigned long *imbalance)
3028 if (!sds->power_savings_balance)
3031 if (sds->this != sds->group_leader ||
3032 sds->group_leader == sds->group_min)
3035 *imbalance = sds->min_load_per_task;
3036 sds->busiest = sds->group_min;
3041 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3042 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3043 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3048 static inline void update_sd_power_savings_stats(struct sched_group *group,
3049 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3054 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3055 int this_cpu, unsigned long *imbalance)
3059 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3062 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3064 return SCHED_POWER_SCALE;
3067 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3069 return default_scale_freq_power(sd, cpu);
3072 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3074 unsigned long weight = sd->span_weight;
3075 unsigned long smt_gain = sd->smt_gain;
3082 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3084 return default_scale_smt_power(sd, cpu);
3087 unsigned long scale_rt_power(int cpu)
3089 struct rq *rq = cpu_rq(cpu);
3090 u64 total, available;
3092 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3094 if (unlikely(total < rq->rt_avg)) {
3095 /* Ensures that power won't end up being negative */
3098 available = total - rq->rt_avg;
3101 if (unlikely((s64)total < SCHED_POWER_SCALE))
3102 total = SCHED_POWER_SCALE;
3104 total >>= SCHED_POWER_SHIFT;
3106 return div_u64(available, total);
3109 static void update_cpu_power(struct sched_domain *sd, int cpu)
3111 unsigned long weight = sd->span_weight;
3112 unsigned long power = SCHED_POWER_SCALE;
3113 struct sched_group *sdg = sd->groups;
3115 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3116 if (sched_feat(ARCH_POWER))
3117 power *= arch_scale_smt_power(sd, cpu);
3119 power *= default_scale_smt_power(sd, cpu);
3121 power >>= SCHED_POWER_SHIFT;
3124 sdg->sgp->power_orig = power;
3126 if (sched_feat(ARCH_POWER))
3127 power *= arch_scale_freq_power(sd, cpu);
3129 power *= default_scale_freq_power(sd, cpu);
3131 power >>= SCHED_POWER_SHIFT;
3133 power *= scale_rt_power(cpu);
3134 power >>= SCHED_POWER_SHIFT;
3139 cpu_rq(cpu)->cpu_power = power;
3140 sdg->sgp->power = power;
3143 static void update_group_power(struct sched_domain *sd, int cpu)
3145 struct sched_domain *child = sd->child;
3146 struct sched_group *group, *sdg = sd->groups;
3147 unsigned long power;
3150 update_cpu_power(sd, cpu);
3156 group = child->groups;
3158 power += group->sgp->power;
3159 group = group->next;
3160 } while (group != child->groups);
3162 sdg->sgp->power = power;
3166 * Try and fix up capacity for tiny siblings, this is needed when
3167 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3168 * which on its own isn't powerful enough.
3170 * See update_sd_pick_busiest() and check_asym_packing().
3173 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3176 * Only siblings can have significantly less than SCHED_POWER_SCALE
3178 if (!(sd->flags & SD_SHARE_CPUPOWER))
3182 * If ~90% of the cpu_power is still there, we're good.
3184 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3191 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3192 * @sd: The sched_domain whose statistics are to be updated.
3193 * @group: sched_group whose statistics are to be updated.
3194 * @this_cpu: Cpu for which load balance is currently performed.
3195 * @idle: Idle status of this_cpu
3196 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3197 * @local_group: Does group contain this_cpu.
3198 * @cpus: Set of cpus considered for load balancing.
3199 * @balance: Should we balance.
3200 * @sgs: variable to hold the statistics for this group.
3202 static inline void update_sg_lb_stats(struct sched_domain *sd,
3203 struct sched_group *group, int this_cpu,
3204 enum cpu_idle_type idle, int load_idx,
3205 int local_group, const struct cpumask *cpus,
3206 int *balance, struct sg_lb_stats *sgs)
3208 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3210 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3211 unsigned long avg_load_per_task = 0;
3214 balance_cpu = group_first_cpu(group);
3216 /* Tally up the load of all CPUs in the group */
3218 min_cpu_load = ~0UL;
3221 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3222 struct rq *rq = cpu_rq(i);
3224 /* Bias balancing toward cpus of our domain */
3226 if (idle_cpu(i) && !first_idle_cpu) {
3231 load = target_load(i, load_idx);
3233 load = source_load(i, load_idx);
3234 if (load > max_cpu_load) {
3235 max_cpu_load = load;
3236 max_nr_running = rq->nr_running;
3238 if (min_cpu_load > load)
3239 min_cpu_load = load;
3242 sgs->group_load += load;
3243 sgs->sum_nr_running += rq->nr_running;
3244 sgs->sum_weighted_load += weighted_cpuload(i);
3250 * First idle cpu or the first cpu(busiest) in this sched group
3251 * is eligible for doing load balancing at this and above
3252 * domains. In the newly idle case, we will allow all the cpu's
3253 * to do the newly idle load balance.
3255 if (idle != CPU_NEWLY_IDLE && local_group) {
3256 if (balance_cpu != this_cpu) {
3260 update_group_power(sd, this_cpu);
3263 /* Adjust by relative CPU power of the group */
3264 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3267 * Consider the group unbalanced when the imbalance is larger
3268 * than the average weight of a task.
3270 * APZ: with cgroup the avg task weight can vary wildly and
3271 * might not be a suitable number - should we keep a
3272 * normalized nr_running number somewhere that negates
3275 if (sgs->sum_nr_running)
3276 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3278 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3281 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3283 if (!sgs->group_capacity)
3284 sgs->group_capacity = fix_small_capacity(sd, group);
3285 sgs->group_weight = group->group_weight;
3287 if (sgs->group_capacity > sgs->sum_nr_running)
3288 sgs->group_has_capacity = 1;
3292 * update_sd_pick_busiest - return 1 on busiest group
3293 * @sd: sched_domain whose statistics are to be checked
3294 * @sds: sched_domain statistics
3295 * @sg: sched_group candidate to be checked for being the busiest
3296 * @sgs: sched_group statistics
3297 * @this_cpu: the current cpu
3299 * Determine if @sg is a busier group than the previously selected
3302 static bool update_sd_pick_busiest(struct sched_domain *sd,
3303 struct sd_lb_stats *sds,
3304 struct sched_group *sg,
3305 struct sg_lb_stats *sgs,
3308 if (sgs->avg_load <= sds->max_load)
3311 if (sgs->sum_nr_running > sgs->group_capacity)
3318 * ASYM_PACKING needs to move all the work to the lowest
3319 * numbered CPUs in the group, therefore mark all groups
3320 * higher than ourself as busy.
3322 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3323 this_cpu < group_first_cpu(sg)) {
3327 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3335 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3336 * @sd: sched_domain whose statistics are to be updated.
3337 * @this_cpu: Cpu for which load balance is currently performed.
3338 * @idle: Idle status of this_cpu
3339 * @cpus: Set of cpus considered for load balancing.
3340 * @balance: Should we balance.
3341 * @sds: variable to hold the statistics for this sched_domain.
3343 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3344 enum cpu_idle_type idle, const struct cpumask *cpus,
3345 int *balance, struct sd_lb_stats *sds)
3347 struct sched_domain *child = sd->child;
3348 struct sched_group *sg = sd->groups;
3349 struct sg_lb_stats sgs;
3350 int load_idx, prefer_sibling = 0;
3352 if (child && child->flags & SD_PREFER_SIBLING)
3355 init_sd_power_savings_stats(sd, sds, idle);
3356 load_idx = get_sd_load_idx(sd, idle);
3361 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3362 memset(&sgs, 0, sizeof(sgs));
3363 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3364 local_group, cpus, balance, &sgs);
3366 if (local_group && !(*balance))
3369 sds->total_load += sgs.group_load;
3370 sds->total_pwr += sg->sgp->power;
3373 * In case the child domain prefers tasks go to siblings
3374 * first, lower the sg capacity to one so that we'll try
3375 * and move all the excess tasks away. We lower the capacity
3376 * of a group only if the local group has the capacity to fit
3377 * these excess tasks, i.e. nr_running < group_capacity. The
3378 * extra check prevents the case where you always pull from the
3379 * heaviest group when it is already under-utilized (possible
3380 * with a large weight task outweighs the tasks on the system).
3382 if (prefer_sibling && !local_group && sds->this_has_capacity)
3383 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3386 sds->this_load = sgs.avg_load;
3388 sds->this_nr_running = sgs.sum_nr_running;
3389 sds->this_load_per_task = sgs.sum_weighted_load;
3390 sds->this_has_capacity = sgs.group_has_capacity;
3391 sds->this_idle_cpus = sgs.idle_cpus;
3392 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
3393 sds->max_load = sgs.avg_load;
3395 sds->busiest_nr_running = sgs.sum_nr_running;
3396 sds->busiest_idle_cpus = sgs.idle_cpus;
3397 sds->busiest_group_capacity = sgs.group_capacity;
3398 sds->busiest_load_per_task = sgs.sum_weighted_load;
3399 sds->busiest_has_capacity = sgs.group_has_capacity;
3400 sds->busiest_group_weight = sgs.group_weight;
3401 sds->group_imb = sgs.group_imb;
3404 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
3406 } while (sg != sd->groups);
3409 int __weak arch_sd_sibling_asym_packing(void)
3411 return 0*SD_ASYM_PACKING;
3415 * check_asym_packing - Check to see if the group is packed into the
3418 * This is primarily intended to used at the sibling level. Some
3419 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3420 * case of POWER7, it can move to lower SMT modes only when higher
3421 * threads are idle. When in lower SMT modes, the threads will
3422 * perform better since they share less core resources. Hence when we
3423 * have idle threads, we want them to be the higher ones.
3425 * This packing function is run on idle threads. It checks to see if
3426 * the busiest CPU in this domain (core in the P7 case) has a higher
3427 * CPU number than the packing function is being run on. Here we are
3428 * assuming lower CPU number will be equivalent to lower a SMT thread
3431 * Returns 1 when packing is required and a task should be moved to
3432 * this CPU. The amount of the imbalance is returned in *imbalance.
3434 * @sd: The sched_domain whose packing is to be checked.
3435 * @sds: Statistics of the sched_domain which is to be packed
3436 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3437 * @imbalance: returns amount of imbalanced due to packing.
3439 static int check_asym_packing(struct sched_domain *sd,
3440 struct sd_lb_stats *sds,
3441 int this_cpu, unsigned long *imbalance)
3445 if (!(sd->flags & SD_ASYM_PACKING))
3451 busiest_cpu = group_first_cpu(sds->busiest);
3452 if (this_cpu > busiest_cpu)
3455 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
3461 * fix_small_imbalance - Calculate the minor imbalance that exists
3462 * amongst the groups of a sched_domain, during
3464 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3465 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3466 * @imbalance: Variable to store the imbalance.
3468 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3469 int this_cpu, unsigned long *imbalance)
3471 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3472 unsigned int imbn = 2;
3473 unsigned long scaled_busy_load_per_task;
3475 if (sds->this_nr_running) {
3476 sds->this_load_per_task /= sds->this_nr_running;
3477 if (sds->busiest_load_per_task >
3478 sds->this_load_per_task)
3481 sds->this_load_per_task =
3482 cpu_avg_load_per_task(this_cpu);
3484 scaled_busy_load_per_task = sds->busiest_load_per_task
3485 * SCHED_POWER_SCALE;
3486 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3488 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3489 (scaled_busy_load_per_task * imbn)) {
3490 *imbalance = sds->busiest_load_per_task;
3495 * OK, we don't have enough imbalance to justify moving tasks,
3496 * however we may be able to increase total CPU power used by
3500 pwr_now += sds->busiest->sgp->power *
3501 min(sds->busiest_load_per_task, sds->max_load);
3502 pwr_now += sds->this->sgp->power *
3503 min(sds->this_load_per_task, sds->this_load);
3504 pwr_now /= SCHED_POWER_SCALE;
3506 /* Amount of load we'd subtract */
3507 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3508 sds->busiest->sgp->power;
3509 if (sds->max_load > tmp)
3510 pwr_move += sds->busiest->sgp->power *
3511 min(sds->busiest_load_per_task, sds->max_load - tmp);
3513 /* Amount of load we'd add */
3514 if (sds->max_load * sds->busiest->sgp->power <
3515 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3516 tmp = (sds->max_load * sds->busiest->sgp->power) /
3517 sds->this->sgp->power;
3519 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3520 sds->this->sgp->power;
3521 pwr_move += sds->this->sgp->power *
3522 min(sds->this_load_per_task, sds->this_load + tmp);
3523 pwr_move /= SCHED_POWER_SCALE;
3525 /* Move if we gain throughput */
3526 if (pwr_move > pwr_now)
3527 *imbalance = sds->busiest_load_per_task;
3531 * calculate_imbalance - Calculate the amount of imbalance present within the
3532 * groups of a given sched_domain during load balance.
3533 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3534 * @this_cpu: Cpu for which currently load balance is being performed.
3535 * @imbalance: The variable to store the imbalance.
3537 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3538 unsigned long *imbalance)
3540 unsigned long max_pull, load_above_capacity = ~0UL;
3542 sds->busiest_load_per_task /= sds->busiest_nr_running;
3543 if (sds->group_imb) {
3544 sds->busiest_load_per_task =
3545 min(sds->busiest_load_per_task, sds->avg_load);
3549 * In the presence of smp nice balancing, certain scenarios can have
3550 * max load less than avg load(as we skip the groups at or below
3551 * its cpu_power, while calculating max_load..)
3553 if (sds->max_load < sds->avg_load) {
3555 return fix_small_imbalance(sds, this_cpu, imbalance);
3558 if (!sds->group_imb) {
3560 * Don't want to pull so many tasks that a group would go idle.
3562 load_above_capacity = (sds->busiest_nr_running -
3563 sds->busiest_group_capacity);
3565 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3567 load_above_capacity /= sds->busiest->sgp->power;
3571 * We're trying to get all the cpus to the average_load, so we don't
3572 * want to push ourselves above the average load, nor do we wish to
3573 * reduce the max loaded cpu below the average load. At the same time,
3574 * we also don't want to reduce the group load below the group capacity
3575 * (so that we can implement power-savings policies etc). Thus we look
3576 * for the minimum possible imbalance.
3577 * Be careful of negative numbers as they'll appear as very large values
3578 * with unsigned longs.
3580 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3582 /* How much load to actually move to equalise the imbalance */
3583 *imbalance = min(max_pull * sds->busiest->sgp->power,
3584 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3585 / SCHED_POWER_SCALE;
3588 * if *imbalance is less than the average load per runnable task
3589 * there is no guarantee that any tasks will be moved so we'll have
3590 * a think about bumping its value to force at least one task to be
3593 if (*imbalance < sds->busiest_load_per_task)
3594 return fix_small_imbalance(sds, this_cpu, imbalance);
3598 /******* find_busiest_group() helpers end here *********************/
3601 * find_busiest_group - Returns the busiest group within the sched_domain
3602 * if there is an imbalance. If there isn't an imbalance, and
3603 * the user has opted for power-savings, it returns a group whose
3604 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3605 * such a group exists.
3607 * Also calculates the amount of weighted load which should be moved
3608 * to restore balance.
3610 * @sd: The sched_domain whose busiest group is to be returned.
3611 * @this_cpu: The cpu for which load balancing is currently being performed.
3612 * @imbalance: Variable which stores amount of weighted load which should
3613 * be moved to restore balance/put a group to idle.
3614 * @idle: The idle status of this_cpu.
3615 * @cpus: The set of CPUs under consideration for load-balancing.
3616 * @balance: Pointer to a variable indicating if this_cpu
3617 * is the appropriate cpu to perform load balancing at this_level.
3619 * Returns: - the busiest group if imbalance exists.
3620 * - If no imbalance and user has opted for power-savings balance,
3621 * return the least loaded group whose CPUs can be
3622 * put to idle by rebalancing its tasks onto our group.
3624 static struct sched_group *
3625 find_busiest_group(struct sched_domain *sd, int this_cpu,
3626 unsigned long *imbalance, enum cpu_idle_type idle,
3627 const struct cpumask *cpus, int *balance)
3629 struct sd_lb_stats sds;
3631 memset(&sds, 0, sizeof(sds));
3634 * Compute the various statistics relavent for load balancing at
3637 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3640 * this_cpu is not the appropriate cpu to perform load balancing at
3646 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3647 check_asym_packing(sd, &sds, this_cpu, imbalance))
3650 /* There is no busy sibling group to pull tasks from */
3651 if (!sds.busiest || sds.busiest_nr_running == 0)
3654 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3657 * If the busiest group is imbalanced the below checks don't
3658 * work because they assumes all things are equal, which typically
3659 * isn't true due to cpus_allowed constraints and the like.
3664 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3665 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3666 !sds.busiest_has_capacity)
3670 * If the local group is more busy than the selected busiest group
3671 * don't try and pull any tasks.
3673 if (sds.this_load >= sds.max_load)
3677 * Don't pull any tasks if this group is already above the domain
3680 if (sds.this_load >= sds.avg_load)
3683 if (idle == CPU_IDLE) {
3685 * This cpu is idle. If the busiest group load doesn't
3686 * have more tasks than the number of available cpu's and
3687 * there is no imbalance between this and busiest group
3688 * wrt to idle cpu's, it is balanced.
3690 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3691 sds.busiest_nr_running <= sds.busiest_group_weight)
3695 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3696 * imbalance_pct to be conservative.
3698 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3703 /* Looks like there is an imbalance. Compute it */
3704 calculate_imbalance(&sds, this_cpu, imbalance);
3709 * There is no obvious imbalance. But check if we can do some balancing
3712 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3720 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3723 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3724 enum cpu_idle_type idle, unsigned long imbalance,
3725 const struct cpumask *cpus)
3727 struct rq *busiest = NULL, *rq;
3728 unsigned long max_load = 0;
3731 for_each_cpu(i, sched_group_cpus(group)) {
3732 unsigned long power = power_of(i);
3733 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3738 capacity = fix_small_capacity(sd, group);
3740 if (!cpumask_test_cpu(i, cpus))
3744 wl = weighted_cpuload(i);
3747 * When comparing with imbalance, use weighted_cpuload()
3748 * which is not scaled with the cpu power.
3750 if (capacity && rq->nr_running == 1 && wl > imbalance)
3754 * For the load comparisons with the other cpu's, consider
3755 * the weighted_cpuload() scaled with the cpu power, so that
3756 * the load can be moved away from the cpu that is potentially
3757 * running at a lower capacity.
3759 wl = (wl * SCHED_POWER_SCALE) / power;
3761 if (wl > max_load) {
3771 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3772 * so long as it is large enough.
3774 #define MAX_PINNED_INTERVAL 512
3776 /* Working cpumask for load_balance and load_balance_newidle. */
3777 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3779 static int need_active_balance(struct sched_domain *sd, int idle,
3780 int busiest_cpu, int this_cpu)
3782 if (idle == CPU_NEWLY_IDLE) {
3785 * ASYM_PACKING needs to force migrate tasks from busy but
3786 * higher numbered CPUs in order to pack all tasks in the
3787 * lowest numbered CPUs.
3789 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3793 * The only task running in a non-idle cpu can be moved to this
3794 * cpu in an attempt to completely freeup the other CPU
3797 * The package power saving logic comes from
3798 * find_busiest_group(). If there are no imbalance, then
3799 * f_b_g() will return NULL. However when sched_mc={1,2} then
3800 * f_b_g() will select a group from which a running task may be
3801 * pulled to this cpu in order to make the other package idle.
3802 * If there is no opportunity to make a package idle and if
3803 * there are no imbalance, then f_b_g() will return NULL and no
3804 * action will be taken in load_balance_newidle().
3806 * Under normal task pull operation due to imbalance, there
3807 * will be more than one task in the source run queue and
3808 * move_tasks() will succeed. ld_moved will be true and this
3809 * active balance code will not be triggered.
3811 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3815 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3818 static int active_load_balance_cpu_stop(void *data);
3821 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3822 * tasks if there is an imbalance.
3824 static int load_balance(int this_cpu, struct rq *this_rq,
3825 struct sched_domain *sd, enum cpu_idle_type idle,
3828 int ld_moved, all_pinned = 0, active_balance = 0;
3829 struct sched_group *group;
3830 unsigned long imbalance;
3832 unsigned long flags;
3833 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3835 cpumask_copy(cpus, cpu_active_mask);
3837 schedstat_inc(sd, lb_count[idle]);
3840 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
3847 schedstat_inc(sd, lb_nobusyg[idle]);
3851 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3853 schedstat_inc(sd, lb_nobusyq[idle]);
3857 BUG_ON(busiest == this_rq);
3859 schedstat_add(sd, lb_imbalance[idle], imbalance);
3862 if (busiest->nr_running > 1) {
3864 * Attempt to move tasks. If find_busiest_group has found
3865 * an imbalance but busiest->nr_running <= 1, the group is
3866 * still unbalanced. ld_moved simply stays zero, so it is
3867 * correctly treated as an imbalance.
3870 local_irq_save(flags);
3871 double_rq_lock(this_rq, busiest);
3872 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3873 imbalance, sd, idle, &all_pinned);
3874 double_rq_unlock(this_rq, busiest);
3875 local_irq_restore(flags);
3878 * some other cpu did the load balance for us.
3880 if (ld_moved && this_cpu != smp_processor_id())
3881 resched_cpu(this_cpu);
3883 /* All tasks on this runqueue were pinned by CPU affinity */
3884 if (unlikely(all_pinned)) {
3885 cpumask_clear_cpu(cpu_of(busiest), cpus);
3886 if (!cpumask_empty(cpus))
3893 schedstat_inc(sd, lb_failed[idle]);
3895 * Increment the failure counter only on periodic balance.
3896 * We do not want newidle balance, which can be very
3897 * frequent, pollute the failure counter causing
3898 * excessive cache_hot migrations and active balances.
3900 if (idle != CPU_NEWLY_IDLE)
3901 sd->nr_balance_failed++;
3903 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
3904 raw_spin_lock_irqsave(&busiest->lock, flags);
3906 /* don't kick the active_load_balance_cpu_stop,
3907 * if the curr task on busiest cpu can't be
3910 if (!cpumask_test_cpu(this_cpu,
3911 &busiest->curr->cpus_allowed)) {
3912 raw_spin_unlock_irqrestore(&busiest->lock,
3915 goto out_one_pinned;
3919 * ->active_balance synchronizes accesses to
3920 * ->active_balance_work. Once set, it's cleared
3921 * only after active load balance is finished.
3923 if (!busiest->active_balance) {
3924 busiest->active_balance = 1;
3925 busiest->push_cpu = this_cpu;
3928 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3931 stop_one_cpu_nowait(cpu_of(busiest),
3932 active_load_balance_cpu_stop, busiest,
3933 &busiest->active_balance_work);
3936 * We've kicked active balancing, reset the failure
3939 sd->nr_balance_failed = sd->cache_nice_tries+1;
3942 sd->nr_balance_failed = 0;
3944 if (likely(!active_balance)) {
3945 /* We were unbalanced, so reset the balancing interval */
3946 sd->balance_interval = sd->min_interval;
3949 * If we've begun active balancing, start to back off. This
3950 * case may not be covered by the all_pinned logic if there
3951 * is only 1 task on the busy runqueue (because we don't call
3954 if (sd->balance_interval < sd->max_interval)
3955 sd->balance_interval *= 2;
3961 schedstat_inc(sd, lb_balanced[idle]);
3963 sd->nr_balance_failed = 0;
3966 /* tune up the balancing interval */
3967 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3968 (sd->balance_interval < sd->max_interval))
3969 sd->balance_interval *= 2;
3977 * idle_balance is called by schedule() if this_cpu is about to become
3978 * idle. Attempts to pull tasks from other CPUs.
3980 static void idle_balance(int this_cpu, struct rq *this_rq)
3982 struct sched_domain *sd;
3983 int pulled_task = 0;
3984 unsigned long next_balance = jiffies + HZ;
3986 this_rq->idle_stamp = this_rq->clock;
3988 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3992 * Drop the rq->lock, but keep IRQ/preempt disabled.
3994 raw_spin_unlock(&this_rq->lock);
3996 update_shares(this_cpu);
3998 for_each_domain(this_cpu, sd) {
3999 unsigned long interval;
4002 if (!(sd->flags & SD_LOAD_BALANCE))
4005 if (sd->flags & SD_BALANCE_NEWIDLE) {
4006 /* If we've pulled tasks over stop searching: */
4007 pulled_task = load_balance(this_cpu, this_rq,
4008 sd, CPU_NEWLY_IDLE, &balance);
4011 interval = msecs_to_jiffies(sd->balance_interval);
4012 if (time_after(next_balance, sd->last_balance + interval))
4013 next_balance = sd->last_balance + interval;
4015 this_rq->idle_stamp = 0;
4021 raw_spin_lock(&this_rq->lock);
4023 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4025 * We are going idle. next_balance may be set based on
4026 * a busy processor. So reset next_balance.
4028 this_rq->next_balance = next_balance;
4033 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4034 * running tasks off the busiest CPU onto idle CPUs. It requires at
4035 * least 1 task to be running on each physical CPU where possible, and
4036 * avoids physical / logical imbalances.
4038 static int active_load_balance_cpu_stop(void *data)
4040 struct rq *busiest_rq = data;
4041 int busiest_cpu = cpu_of(busiest_rq);
4042 int target_cpu = busiest_rq->push_cpu;
4043 struct rq *target_rq = cpu_rq(target_cpu);
4044 struct sched_domain *sd;
4046 raw_spin_lock_irq(&busiest_rq->lock);
4048 /* make sure the requested cpu hasn't gone down in the meantime */
4049 if (unlikely(busiest_cpu != smp_processor_id() ||
4050 !busiest_rq->active_balance))
4053 /* Is there any task to move? */
4054 if (busiest_rq->nr_running <= 1)
4058 * This condition is "impossible", if it occurs
4059 * we need to fix it. Originally reported by
4060 * Bjorn Helgaas on a 128-cpu setup.
4062 BUG_ON(busiest_rq == target_rq);
4064 /* move a task from busiest_rq to target_rq */
4065 double_lock_balance(busiest_rq, target_rq);
4067 /* Search for an sd spanning us and the target CPU. */
4069 for_each_domain(target_cpu, sd) {
4070 if ((sd->flags & SD_LOAD_BALANCE) &&
4071 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4076 schedstat_inc(sd, alb_count);
4078 if (move_one_task(target_rq, target_cpu, busiest_rq,
4080 schedstat_inc(sd, alb_pushed);
4082 schedstat_inc(sd, alb_failed);
4085 double_unlock_balance(busiest_rq, target_rq);
4087 busiest_rq->active_balance = 0;
4088 raw_spin_unlock_irq(&busiest_rq->lock);
4094 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
4096 static void trigger_sched_softirq(void *data)
4098 raise_softirq_irqoff(SCHED_SOFTIRQ);
4101 static inline void init_sched_softirq_csd(struct call_single_data *csd)
4103 csd->func = trigger_sched_softirq;
4110 * idle load balancing details
4111 * - One of the idle CPUs nominates itself as idle load_balancer, while
4113 * - This idle load balancer CPU will also go into tickless mode when
4114 * it is idle, just like all other idle CPUs
4115 * - When one of the busy CPUs notice that there may be an idle rebalancing
4116 * needed, they will kick the idle load balancer, which then does idle
4117 * load balancing for all the idle CPUs.
4120 atomic_t load_balancer;
4121 atomic_t first_pick_cpu;
4122 atomic_t second_pick_cpu;
4123 cpumask_var_t idle_cpus_mask;
4124 cpumask_var_t grp_idle_mask;
4125 unsigned long next_balance; /* in jiffy units */
4126 } nohz ____cacheline_aligned;
4128 int get_nohz_load_balancer(void)
4130 return atomic_read(&nohz.load_balancer);
4133 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4135 * lowest_flag_domain - Return lowest sched_domain containing flag.
4136 * @cpu: The cpu whose lowest level of sched domain is to
4138 * @flag: The flag to check for the lowest sched_domain
4139 * for the given cpu.
4141 * Returns the lowest sched_domain of a cpu which contains the given flag.
4143 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4145 struct sched_domain *sd;
4147 for_each_domain(cpu, sd)
4148 if (sd->flags & flag)
4155 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4156 * @cpu: The cpu whose domains we're iterating over.
4157 * @sd: variable holding the value of the power_savings_sd
4159 * @flag: The flag to filter the sched_domains to be iterated.
4161 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4162 * set, starting from the lowest sched_domain to the highest.
4164 #define for_each_flag_domain(cpu, sd, flag) \
4165 for (sd = lowest_flag_domain(cpu, flag); \
4166 (sd && (sd->flags & flag)); sd = sd->parent)
4169 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4170 * @ilb_group: group to be checked for semi-idleness
4172 * Returns: 1 if the group is semi-idle. 0 otherwise.
4174 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4175 * and atleast one non-idle CPU. This helper function checks if the given
4176 * sched_group is semi-idle or not.
4178 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4180 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
4181 sched_group_cpus(ilb_group));
4184 * A sched_group is semi-idle when it has atleast one busy cpu
4185 * and atleast one idle cpu.
4187 if (cpumask_empty(nohz.grp_idle_mask))
4190 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
4196 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4197 * @cpu: The cpu which is nominating a new idle_load_balancer.
4199 * Returns: Returns the id of the idle load balancer if it exists,
4200 * Else, returns >= nr_cpu_ids.
4202 * This algorithm picks the idle load balancer such that it belongs to a
4203 * semi-idle powersavings sched_domain. The idea is to try and avoid
4204 * completely idle packages/cores just for the purpose of idle load balancing
4205 * when there are other idle cpu's which are better suited for that job.
4207 static int find_new_ilb(int cpu)
4209 struct sched_domain *sd;
4210 struct sched_group *ilb_group;
4211 int ilb = nr_cpu_ids;
4214 * Have idle load balancer selection from semi-idle packages only
4215 * when power-aware load balancing is enabled
4217 if (!(sched_smt_power_savings || sched_mc_power_savings))
4221 * Optimize for the case when we have no idle CPUs or only one
4222 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4224 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4228 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4229 ilb_group = sd->groups;
4232 if (is_semi_idle_group(ilb_group)) {
4233 ilb = cpumask_first(nohz.grp_idle_mask);
4237 ilb_group = ilb_group->next;
4239 } while (ilb_group != sd->groups);
4247 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4248 static inline int find_new_ilb(int call_cpu)
4255 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4256 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4257 * CPU (if there is one).
4259 static void nohz_balancer_kick(int cpu)
4263 nohz.next_balance++;
4265 ilb_cpu = get_nohz_load_balancer();
4267 if (ilb_cpu >= nr_cpu_ids) {
4268 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4269 if (ilb_cpu >= nr_cpu_ids)
4273 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4274 struct call_single_data *cp;
4276 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4277 cp = &per_cpu(remote_sched_softirq_cb, cpu);
4278 __smp_call_function_single(ilb_cpu, cp, 0);
4284 * This routine will try to nominate the ilb (idle load balancing)
4285 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4286 * load balancing on behalf of all those cpus.
4288 * When the ilb owner becomes busy, we will not have new ilb owner until some
4289 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4290 * idle load balancing by kicking one of the idle CPUs.
4292 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4293 * ilb owner CPU in future (when there is a need for idle load balancing on
4294 * behalf of all idle CPUs).
4296 void select_nohz_load_balancer(int stop_tick)
4298 int cpu = smp_processor_id();
4301 if (!cpu_active(cpu)) {
4302 if (atomic_read(&nohz.load_balancer) != cpu)
4306 * If we are going offline and still the leader,
4309 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4316 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4318 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4319 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4320 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4321 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4323 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4326 /* make me the ilb owner */
4327 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4332 * Check to see if there is a more power-efficient
4335 new_ilb = find_new_ilb(cpu);
4336 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4337 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4338 resched_cpu(new_ilb);
4344 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4347 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4349 if (atomic_read(&nohz.load_balancer) == cpu)
4350 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4358 static DEFINE_SPINLOCK(balancing);
4360 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4363 * Scale the max load_balance interval with the number of CPUs in the system.
4364 * This trades load-balance latency on larger machines for less cross talk.
4366 static void update_max_interval(void)
4368 max_load_balance_interval = HZ*num_online_cpus()/10;
4372 * It checks each scheduling domain to see if it is due to be balanced,
4373 * and initiates a balancing operation if so.
4375 * Balancing parameters are set up in arch_init_sched_domains.
4377 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4380 struct rq *rq = cpu_rq(cpu);
4381 unsigned long interval;
4382 struct sched_domain *sd;
4383 /* Earliest time when we have to do rebalance again */
4384 unsigned long next_balance = jiffies + 60*HZ;
4385 int update_next_balance = 0;
4391 for_each_domain(cpu, sd) {
4392 if (!(sd->flags & SD_LOAD_BALANCE))
4395 interval = sd->balance_interval;
4396 if (idle != CPU_IDLE)
4397 interval *= sd->busy_factor;
4399 /* scale ms to jiffies */
4400 interval = msecs_to_jiffies(interval);
4401 interval = clamp(interval, 1UL, max_load_balance_interval);
4403 need_serialize = sd->flags & SD_SERIALIZE;
4405 if (need_serialize) {
4406 if (!spin_trylock(&balancing))
4410 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4411 if (load_balance(cpu, rq, sd, idle, &balance)) {
4413 * We've pulled tasks over so either we're no
4416 idle = CPU_NOT_IDLE;
4418 sd->last_balance = jiffies;
4421 spin_unlock(&balancing);
4423 if (time_after(next_balance, sd->last_balance + interval)) {
4424 next_balance = sd->last_balance + interval;
4425 update_next_balance = 1;
4429 * Stop the load balance at this level. There is another
4430 * CPU in our sched group which is doing load balancing more
4439 * next_balance will be updated only when there is a need.
4440 * When the cpu is attached to null domain for ex, it will not be
4443 if (likely(update_next_balance))
4444 rq->next_balance = next_balance;
4449 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4450 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4452 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4454 struct rq *this_rq = cpu_rq(this_cpu);
4458 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
4461 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4462 if (balance_cpu == this_cpu)
4466 * If this cpu gets work to do, stop the load balancing
4467 * work being done for other cpus. Next load
4468 * balancing owner will pick it up.
4470 if (need_resched()) {
4471 this_rq->nohz_balance_kick = 0;
4475 raw_spin_lock_irq(&this_rq->lock);
4476 update_rq_clock(this_rq);
4477 update_cpu_load(this_rq);
4478 raw_spin_unlock_irq(&this_rq->lock);
4480 rebalance_domains(balance_cpu, CPU_IDLE);
4482 rq = cpu_rq(balance_cpu);
4483 if (time_after(this_rq->next_balance, rq->next_balance))
4484 this_rq->next_balance = rq->next_balance;
4486 nohz.next_balance = this_rq->next_balance;
4487 this_rq->nohz_balance_kick = 0;
4491 * Current heuristic for kicking the idle load balancer
4492 * - first_pick_cpu is the one of the busy CPUs. It will kick
4493 * idle load balancer when it has more than one process active. This
4494 * eliminates the need for idle load balancing altogether when we have
4495 * only one running process in the system (common case).
4496 * - If there are more than one busy CPU, idle load balancer may have
4497 * to run for active_load_balance to happen (i.e., two busy CPUs are
4498 * SMT or core siblings and can run better if they move to different
4499 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4500 * which will kick idle load balancer as soon as it has any load.
4502 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4504 unsigned long now = jiffies;
4506 int first_pick_cpu, second_pick_cpu;
4508 if (time_before(now, nohz.next_balance))
4511 if (rq->idle_at_tick)
4514 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4515 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4517 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4518 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4521 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4522 if (ret == nr_cpu_ids || ret == cpu) {
4523 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4524 if (rq->nr_running > 1)
4527 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4528 if (ret == nr_cpu_ids || ret == cpu) {
4536 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4540 * run_rebalance_domains is triggered when needed from the scheduler tick.
4541 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4543 static void run_rebalance_domains(struct softirq_action *h)
4545 int this_cpu = smp_processor_id();
4546 struct rq *this_rq = cpu_rq(this_cpu);
4547 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4548 CPU_IDLE : CPU_NOT_IDLE;
4550 rebalance_domains(this_cpu, idle);
4553 * If this cpu has a pending nohz_balance_kick, then do the
4554 * balancing on behalf of the other idle cpus whose ticks are
4557 nohz_idle_balance(this_cpu, idle);
4560 static inline int on_null_domain(int cpu)
4562 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4566 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4568 static inline void trigger_load_balance(struct rq *rq, int cpu)
4570 /* Don't need to rebalance while attached to NULL domain */
4571 if (time_after_eq(jiffies, rq->next_balance) &&
4572 likely(!on_null_domain(cpu)))
4573 raise_softirq(SCHED_SOFTIRQ);
4575 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4576 nohz_balancer_kick(cpu);
4580 static void rq_online_fair(struct rq *rq)
4585 static void rq_offline_fair(struct rq *rq)
4590 #else /* CONFIG_SMP */
4593 * on UP we do not need to balance between CPUs:
4595 static inline void idle_balance(int cpu, struct rq *rq)
4599 #endif /* CONFIG_SMP */
4602 * scheduler tick hitting a task of our scheduling class:
4604 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4606 struct cfs_rq *cfs_rq;
4607 struct sched_entity *se = &curr->se;
4609 for_each_sched_entity(se) {
4610 cfs_rq = cfs_rq_of(se);
4611 entity_tick(cfs_rq, se, queued);
4616 * called on fork with the child task as argument from the parent's context
4617 * - child not yet on the tasklist
4618 * - preemption disabled
4620 static void task_fork_fair(struct task_struct *p)
4622 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4623 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4624 int this_cpu = smp_processor_id();
4625 struct rq *rq = this_rq();
4626 unsigned long flags;
4628 raw_spin_lock_irqsave(&rq->lock, flags);
4630 update_rq_clock(rq);
4632 if (unlikely(task_cpu(p) != this_cpu)) {
4634 __set_task_cpu(p, this_cpu);
4638 update_curr(cfs_rq);
4641 se->vruntime = curr->vruntime;
4642 place_entity(cfs_rq, se, 1);
4644 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4646 * Upon rescheduling, sched_class::put_prev_task() will place
4647 * 'current' within the tree based on its new key value.
4649 swap(curr->vruntime, se->vruntime);
4650 resched_task(rq->curr);
4653 se->vruntime -= cfs_rq->min_vruntime;
4655 raw_spin_unlock_irqrestore(&rq->lock, flags);
4659 * Priority of the task has changed. Check to see if we preempt
4663 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4669 * Reschedule if we are currently running on this runqueue and
4670 * our priority decreased, or if we are not currently running on
4671 * this runqueue and our priority is higher than the current's
4673 if (rq->curr == p) {
4674 if (p->prio > oldprio)
4675 resched_task(rq->curr);
4677 check_preempt_curr(rq, p, 0);
4680 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4682 struct sched_entity *se = &p->se;
4683 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4686 * Ensure the task's vruntime is normalized, so that when its
4687 * switched back to the fair class the enqueue_entity(.flags=0) will
4688 * do the right thing.
4690 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4691 * have normalized the vruntime, if it was !on_rq, then only when
4692 * the task is sleeping will it still have non-normalized vruntime.
4694 if (!se->on_rq && p->state != TASK_RUNNING) {
4696 * Fix up our vruntime so that the current sleep doesn't
4697 * cause 'unlimited' sleep bonus.
4699 place_entity(cfs_rq, se, 0);
4700 se->vruntime -= cfs_rq->min_vruntime;
4705 * We switched to the sched_fair class.
4707 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4713 * We were most likely switched from sched_rt, so
4714 * kick off the schedule if running, otherwise just see
4715 * if we can still preempt the current task.
4718 resched_task(rq->curr);
4720 check_preempt_curr(rq, p, 0);
4723 /* Account for a task changing its policy or group.
4725 * This routine is mostly called to set cfs_rq->curr field when a task
4726 * migrates between groups/classes.
4728 static void set_curr_task_fair(struct rq *rq)
4730 struct sched_entity *se = &rq->curr->se;
4732 for_each_sched_entity(se) {
4733 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4735 set_next_entity(cfs_rq, se);
4736 /* ensure bandwidth has been allocated on our new cfs_rq */
4737 account_cfs_rq_runtime(cfs_rq, 0);
4741 #ifdef CONFIG_FAIR_GROUP_SCHED
4742 static void task_move_group_fair(struct task_struct *p, int on_rq)
4745 * If the task was not on the rq at the time of this cgroup movement
4746 * it must have been asleep, sleeping tasks keep their ->vruntime
4747 * absolute on their old rq until wakeup (needed for the fair sleeper
4748 * bonus in place_entity()).
4750 * If it was on the rq, we've just 'preempted' it, which does convert
4751 * ->vruntime to a relative base.
4753 * Make sure both cases convert their relative position when migrating
4754 * to another cgroup's rq. This does somewhat interfere with the
4755 * fair sleeper stuff for the first placement, but who cares.
4758 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4759 set_task_rq(p, task_cpu(p));
4761 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4765 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4767 struct sched_entity *se = &task->se;
4768 unsigned int rr_interval = 0;
4771 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4774 if (rq->cfs.load.weight)
4775 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4781 * All the scheduling class methods:
4783 static const struct sched_class fair_sched_class = {
4784 .next = &idle_sched_class,
4785 .enqueue_task = enqueue_task_fair,
4786 .dequeue_task = dequeue_task_fair,
4787 .yield_task = yield_task_fair,
4788 .yield_to_task = yield_to_task_fair,
4790 .check_preempt_curr = check_preempt_wakeup,
4792 .pick_next_task = pick_next_task_fair,
4793 .put_prev_task = put_prev_task_fair,
4796 .select_task_rq = select_task_rq_fair,
4798 .rq_online = rq_online_fair,
4799 .rq_offline = rq_offline_fair,
4801 .task_waking = task_waking_fair,
4804 .set_curr_task = set_curr_task_fair,
4805 .task_tick = task_tick_fair,
4806 .task_fork = task_fork_fair,
4808 .prio_changed = prio_changed_fair,
4809 .switched_from = switched_from_fair,
4810 .switched_to = switched_to_fair,
4812 .get_rr_interval = get_rr_interval_fair,
4814 #ifdef CONFIG_FAIR_GROUP_SCHED
4815 .task_move_group = task_move_group_fair,
4819 #ifdef CONFIG_SCHED_DEBUG
4820 static void print_cfs_stats(struct seq_file *m, int cpu)
4822 struct cfs_rq *cfs_rq;
4825 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4826 print_cfs_rq(m, cpu, cfs_rq);