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;
973 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
976 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
979 * Update the normalized vruntime before updating min_vruntime
980 * through callig update_curr().
982 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
983 se->vruntime += cfs_rq->min_vruntime;
986 * Update run-time statistics of the 'current'.
989 update_cfs_load(cfs_rq, 0);
990 account_entity_enqueue(cfs_rq, se);
991 update_cfs_shares(cfs_rq);
993 if (flags & ENQUEUE_WAKEUP) {
994 place_entity(cfs_rq, se, 0);
995 enqueue_sleeper(cfs_rq, se);
998 update_stats_enqueue(cfs_rq, se);
999 check_spread(cfs_rq, se);
1000 if (se != cfs_rq->curr)
1001 __enqueue_entity(cfs_rq, se);
1004 if (cfs_rq->nr_running == 1) {
1005 list_add_leaf_cfs_rq(cfs_rq);
1006 check_enqueue_throttle(cfs_rq);
1010 static void __clear_buddies_last(struct sched_entity *se)
1012 for_each_sched_entity(se) {
1013 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1014 if (cfs_rq->last == se)
1015 cfs_rq->last = NULL;
1021 static void __clear_buddies_next(struct sched_entity *se)
1023 for_each_sched_entity(se) {
1024 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1025 if (cfs_rq->next == se)
1026 cfs_rq->next = NULL;
1032 static void __clear_buddies_skip(struct sched_entity *se)
1034 for_each_sched_entity(se) {
1035 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1036 if (cfs_rq->skip == se)
1037 cfs_rq->skip = NULL;
1043 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1045 if (cfs_rq->last == se)
1046 __clear_buddies_last(se);
1048 if (cfs_rq->next == se)
1049 __clear_buddies_next(se);
1051 if (cfs_rq->skip == se)
1052 __clear_buddies_skip(se);
1056 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1059 * Update run-time statistics of the 'current'.
1061 update_curr(cfs_rq);
1063 update_stats_dequeue(cfs_rq, se);
1064 if (flags & DEQUEUE_SLEEP) {
1065 #ifdef CONFIG_SCHEDSTATS
1066 if (entity_is_task(se)) {
1067 struct task_struct *tsk = task_of(se);
1069 if (tsk->state & TASK_INTERRUPTIBLE)
1070 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1071 if (tsk->state & TASK_UNINTERRUPTIBLE)
1072 se->statistics.block_start = rq_of(cfs_rq)->clock;
1077 clear_buddies(cfs_rq, se);
1079 if (se != cfs_rq->curr)
1080 __dequeue_entity(cfs_rq, se);
1082 update_cfs_load(cfs_rq, 0);
1083 account_entity_dequeue(cfs_rq, se);
1086 * Normalize the entity after updating the min_vruntime because the
1087 * update can refer to the ->curr item and we need to reflect this
1088 * movement in our normalized position.
1090 if (!(flags & DEQUEUE_SLEEP))
1091 se->vruntime -= cfs_rq->min_vruntime;
1093 update_min_vruntime(cfs_rq);
1094 update_cfs_shares(cfs_rq);
1098 * Preempt the current task with a newly woken task if needed:
1101 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1103 unsigned long ideal_runtime, delta_exec;
1105 ideal_runtime = sched_slice(cfs_rq, curr);
1106 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1107 if (delta_exec > ideal_runtime) {
1108 resched_task(rq_of(cfs_rq)->curr);
1110 * The current task ran long enough, ensure it doesn't get
1111 * re-elected due to buddy favours.
1113 clear_buddies(cfs_rq, curr);
1118 * Ensure that a task that missed wakeup preemption by a
1119 * narrow margin doesn't have to wait for a full slice.
1120 * This also mitigates buddy induced latencies under load.
1122 if (delta_exec < sysctl_sched_min_granularity)
1125 if (cfs_rq->nr_running > 1) {
1126 struct sched_entity *se = __pick_first_entity(cfs_rq);
1127 s64 delta = curr->vruntime - se->vruntime;
1132 if (delta > ideal_runtime)
1133 resched_task(rq_of(cfs_rq)->curr);
1138 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1140 /* 'current' is not kept within the tree. */
1143 * Any task has to be enqueued before it get to execute on
1144 * a CPU. So account for the time it spent waiting on the
1147 update_stats_wait_end(cfs_rq, se);
1148 __dequeue_entity(cfs_rq, se);
1151 update_stats_curr_start(cfs_rq, se);
1153 #ifdef CONFIG_SCHEDSTATS
1155 * Track our maximum slice length, if the CPU's load is at
1156 * least twice that of our own weight (i.e. dont track it
1157 * when there are only lesser-weight tasks around):
1159 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1160 se->statistics.slice_max = max(se->statistics.slice_max,
1161 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1164 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1168 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1171 * Pick the next process, keeping these things in mind, in this order:
1172 * 1) keep things fair between processes/task groups
1173 * 2) pick the "next" process, since someone really wants that to run
1174 * 3) pick the "last" process, for cache locality
1175 * 4) do not run the "skip" process, if something else is available
1177 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1179 struct sched_entity *se = __pick_first_entity(cfs_rq);
1180 struct sched_entity *left = se;
1183 * Avoid running the skip buddy, if running something else can
1184 * be done without getting too unfair.
1186 if (cfs_rq->skip == se) {
1187 struct sched_entity *second = __pick_next_entity(se);
1188 if (second && wakeup_preempt_entity(second, left) < 1)
1193 * Prefer last buddy, try to return the CPU to a preempted task.
1195 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1199 * Someone really wants this to run. If it's not unfair, run it.
1201 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1204 clear_buddies(cfs_rq, se);
1209 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1211 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1214 * If still on the runqueue then deactivate_task()
1215 * was not called and update_curr() has to be done:
1218 update_curr(cfs_rq);
1220 /* throttle cfs_rqs exceeding runtime */
1221 check_cfs_rq_runtime(cfs_rq);
1223 check_spread(cfs_rq, prev);
1225 update_stats_wait_start(cfs_rq, prev);
1226 /* Put 'current' back into the tree. */
1227 __enqueue_entity(cfs_rq, prev);
1229 cfs_rq->curr = NULL;
1233 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1236 * Update run-time statistics of the 'current'.
1238 update_curr(cfs_rq);
1241 * Update share accounting for long-running entities.
1243 update_entity_shares_tick(cfs_rq);
1245 #ifdef CONFIG_SCHED_HRTICK
1247 * queued ticks are scheduled to match the slice, so don't bother
1248 * validating it and just reschedule.
1251 resched_task(rq_of(cfs_rq)->curr);
1255 * don't let the period tick interfere with the hrtick preemption
1257 if (!sched_feat(DOUBLE_TICK) &&
1258 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1262 if (cfs_rq->nr_running > 1)
1263 check_preempt_tick(cfs_rq, curr);
1267 /**************************************************
1268 * CFS bandwidth control machinery
1271 #ifdef CONFIG_CFS_BANDWIDTH
1273 * default period for cfs group bandwidth.
1274 * default: 0.1s, units: nanoseconds
1276 static inline u64 default_cfs_period(void)
1278 return 100000000ULL;
1281 static inline u64 sched_cfs_bandwidth_slice(void)
1283 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1287 * Replenish runtime according to assigned quota and update expiration time.
1288 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1289 * additional synchronization around rq->lock.
1291 * requires cfs_b->lock
1293 static void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1297 if (cfs_b->quota == RUNTIME_INF)
1300 now = sched_clock_cpu(smp_processor_id());
1301 cfs_b->runtime = cfs_b->quota;
1302 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1305 /* returns 0 on failure to allocate runtime */
1306 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1308 struct task_group *tg = cfs_rq->tg;
1309 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1310 u64 amount = 0, min_amount, expires;
1312 /* note: this is a positive sum as runtime_remaining <= 0 */
1313 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1315 raw_spin_lock(&cfs_b->lock);
1316 if (cfs_b->quota == RUNTIME_INF)
1317 amount = min_amount;
1320 * If the bandwidth pool has become inactive, then at least one
1321 * period must have elapsed since the last consumption.
1322 * Refresh the global state and ensure bandwidth timer becomes
1325 if (!cfs_b->timer_active) {
1326 __refill_cfs_bandwidth_runtime(cfs_b);
1327 __start_cfs_bandwidth(cfs_b);
1330 if (cfs_b->runtime > 0) {
1331 amount = min(cfs_b->runtime, min_amount);
1332 cfs_b->runtime -= amount;
1336 expires = cfs_b->runtime_expires;
1337 raw_spin_unlock(&cfs_b->lock);
1339 cfs_rq->runtime_remaining += amount;
1341 * we may have advanced our local expiration to account for allowed
1342 * spread between our sched_clock and the one on which runtime was
1345 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1346 cfs_rq->runtime_expires = expires;
1348 return cfs_rq->runtime_remaining > 0;
1352 * Note: This depends on the synchronization provided by sched_clock and the
1353 * fact that rq->clock snapshots this value.
1355 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1357 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1358 struct rq *rq = rq_of(cfs_rq);
1360 /* if the deadline is ahead of our clock, nothing to do */
1361 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1364 if (cfs_rq->runtime_remaining < 0)
1368 * If the local deadline has passed we have to consider the
1369 * possibility that our sched_clock is 'fast' and the global deadline
1370 * has not truly expired.
1372 * Fortunately we can check determine whether this the case by checking
1373 * whether the global deadline has advanced.
1376 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1377 /* extend local deadline, drift is bounded above by 2 ticks */
1378 cfs_rq->runtime_expires += TICK_NSEC;
1380 /* global deadline is ahead, expiration has passed */
1381 cfs_rq->runtime_remaining = 0;
1385 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1386 unsigned long delta_exec)
1388 /* dock delta_exec before expiring quota (as it could span periods) */
1389 cfs_rq->runtime_remaining -= delta_exec;
1390 expire_cfs_rq_runtime(cfs_rq);
1392 if (likely(cfs_rq->runtime_remaining > 0))
1396 * if we're unable to extend our runtime we resched so that the active
1397 * hierarchy can be throttled
1399 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1400 resched_task(rq_of(cfs_rq)->curr);
1403 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1404 unsigned long delta_exec)
1406 if (!cfs_rq->runtime_enabled)
1409 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1412 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1414 return cfs_rq->throttled;
1417 /* check whether cfs_rq, or any parent, is throttled */
1418 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1420 return cfs_rq->throttle_count;
1424 * Ensure that neither of the group entities corresponding to src_cpu or
1425 * dest_cpu are members of a throttled hierarchy when performing group
1426 * load-balance operations.
1428 static inline int throttled_lb_pair(struct task_group *tg,
1429 int src_cpu, int dest_cpu)
1431 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1433 src_cfs_rq = tg->cfs_rq[src_cpu];
1434 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1436 return throttled_hierarchy(src_cfs_rq) ||
1437 throttled_hierarchy(dest_cfs_rq);
1440 /* updated child weight may affect parent so we have to do this bottom up */
1441 static int tg_unthrottle_up(struct task_group *tg, void *data)
1443 struct rq *rq = data;
1444 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1446 cfs_rq->throttle_count--;
1448 if (!cfs_rq->throttle_count) {
1449 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1451 /* leaving throttled state, advance shares averaging windows */
1452 cfs_rq->load_stamp += delta;
1453 cfs_rq->load_last += delta;
1455 /* update entity weight now that we are on_rq again */
1456 update_cfs_shares(cfs_rq);
1463 static int tg_throttle_down(struct task_group *tg, void *data)
1465 struct rq *rq = data;
1466 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1468 /* group is entering throttled state, record last load */
1469 if (!cfs_rq->throttle_count)
1470 update_cfs_load(cfs_rq, 0);
1471 cfs_rq->throttle_count++;
1476 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1478 struct rq *rq = rq_of(cfs_rq);
1479 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1480 struct sched_entity *se;
1481 long task_delta, dequeue = 1;
1483 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1485 /* account load preceding throttle */
1487 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1490 task_delta = cfs_rq->h_nr_running;
1491 for_each_sched_entity(se) {
1492 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1493 /* throttled entity or throttle-on-deactivate */
1498 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1499 qcfs_rq->h_nr_running -= task_delta;
1501 if (qcfs_rq->load.weight)
1506 rq->nr_running -= task_delta;
1508 cfs_rq->throttled = 1;
1509 cfs_rq->throttled_timestamp = rq->clock;
1510 raw_spin_lock(&cfs_b->lock);
1511 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1512 raw_spin_unlock(&cfs_b->lock);
1515 static void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1517 struct rq *rq = rq_of(cfs_rq);
1518 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1519 struct sched_entity *se;
1523 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1525 cfs_rq->throttled = 0;
1526 raw_spin_lock(&cfs_b->lock);
1527 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1528 list_del_rcu(&cfs_rq->throttled_list);
1529 raw_spin_unlock(&cfs_b->lock);
1530 cfs_rq->throttled_timestamp = 0;
1532 update_rq_clock(rq);
1533 /* update hierarchical throttle state */
1534 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1536 if (!cfs_rq->load.weight)
1539 task_delta = cfs_rq->h_nr_running;
1540 for_each_sched_entity(se) {
1544 cfs_rq = cfs_rq_of(se);
1546 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1547 cfs_rq->h_nr_running += task_delta;
1549 if (cfs_rq_throttled(cfs_rq))
1554 rq->nr_running += task_delta;
1556 /* determine whether we need to wake up potentially idle cpu */
1557 if (rq->curr == rq->idle && rq->cfs.nr_running)
1558 resched_task(rq->curr);
1561 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1562 u64 remaining, u64 expires)
1564 struct cfs_rq *cfs_rq;
1565 u64 runtime = remaining;
1568 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1570 struct rq *rq = rq_of(cfs_rq);
1572 raw_spin_lock(&rq->lock);
1573 if (!cfs_rq_throttled(cfs_rq))
1576 runtime = -cfs_rq->runtime_remaining + 1;
1577 if (runtime > remaining)
1578 runtime = remaining;
1579 remaining -= runtime;
1581 cfs_rq->runtime_remaining += runtime;
1582 cfs_rq->runtime_expires = expires;
1584 /* we check whether we're throttled above */
1585 if (cfs_rq->runtime_remaining > 0)
1586 unthrottle_cfs_rq(cfs_rq);
1589 raw_spin_unlock(&rq->lock);
1600 * Responsible for refilling a task_group's bandwidth and unthrottling its
1601 * cfs_rqs as appropriate. If there has been no activity within the last
1602 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1603 * used to track this state.
1605 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1607 u64 runtime, runtime_expires;
1608 int idle = 1, throttled;
1610 raw_spin_lock(&cfs_b->lock);
1611 /* no need to continue the timer with no bandwidth constraint */
1612 if (cfs_b->quota == RUNTIME_INF)
1615 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1616 /* idle depends on !throttled (for the case of a large deficit) */
1617 idle = cfs_b->idle && !throttled;
1618 cfs_b->nr_periods += overrun;
1620 /* if we're going inactive then everything else can be deferred */
1624 __refill_cfs_bandwidth_runtime(cfs_b);
1627 /* mark as potentially idle for the upcoming period */
1632 /* account preceding periods in which throttling occurred */
1633 cfs_b->nr_throttled += overrun;
1636 * There are throttled entities so we must first use the new bandwidth
1637 * to unthrottle them before making it generally available. This
1638 * ensures that all existing debts will be paid before a new cfs_rq is
1641 runtime = cfs_b->runtime;
1642 runtime_expires = cfs_b->runtime_expires;
1646 * This check is repeated as we are holding onto the new bandwidth
1647 * while we unthrottle. This can potentially race with an unthrottled
1648 * group trying to acquire new bandwidth from the global pool.
1650 while (throttled && runtime > 0) {
1651 raw_spin_unlock(&cfs_b->lock);
1652 /* we can't nest cfs_b->lock while distributing bandwidth */
1653 runtime = distribute_cfs_runtime(cfs_b, runtime,
1655 raw_spin_lock(&cfs_b->lock);
1657 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1660 /* return (any) remaining runtime */
1661 cfs_b->runtime = runtime;
1663 * While we are ensured activity in the period following an
1664 * unthrottle, this also covers the case in which the new bandwidth is
1665 * insufficient to cover the existing bandwidth deficit. (Forcing the
1666 * timer to remain active while there are any throttled entities.)
1671 cfs_b->timer_active = 0;
1672 raw_spin_unlock(&cfs_b->lock);
1678 * When a group wakes up we want to make sure that its quota is not already
1679 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1680 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1682 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
1684 /* an active group must be handled by the update_curr()->put() path */
1685 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
1688 /* ensure the group is not already throttled */
1689 if (cfs_rq_throttled(cfs_rq))
1692 /* update runtime allocation */
1693 account_cfs_rq_runtime(cfs_rq, 0);
1694 if (cfs_rq->runtime_remaining <= 0)
1695 throttle_cfs_rq(cfs_rq);
1698 /* conditionally throttle active cfs_rq's from put_prev_entity() */
1699 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1701 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
1705 * it's possible for a throttled entity to be forced into a running
1706 * state (e.g. set_curr_task), in this case we're finished.
1708 if (cfs_rq_throttled(cfs_rq))
1711 throttle_cfs_rq(cfs_rq);
1714 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1715 unsigned long delta_exec) {}
1716 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
1717 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
1719 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1724 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1729 static inline int throttled_lb_pair(struct task_group *tg,
1730 int src_cpu, int dest_cpu)
1736 /**************************************************
1737 * CFS operations on tasks:
1740 #ifdef CONFIG_SCHED_HRTICK
1741 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1743 struct sched_entity *se = &p->se;
1744 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1746 WARN_ON(task_rq(p) != rq);
1748 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1749 u64 slice = sched_slice(cfs_rq, se);
1750 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1751 s64 delta = slice - ran;
1760 * Don't schedule slices shorter than 10000ns, that just
1761 * doesn't make sense. Rely on vruntime for fairness.
1764 delta = max_t(s64, 10000LL, delta);
1766 hrtick_start(rq, delta);
1771 * called from enqueue/dequeue and updates the hrtick when the
1772 * current task is from our class and nr_running is low enough
1775 static void hrtick_update(struct rq *rq)
1777 struct task_struct *curr = rq->curr;
1779 if (curr->sched_class != &fair_sched_class)
1782 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1783 hrtick_start_fair(rq, curr);
1785 #else /* !CONFIG_SCHED_HRTICK */
1787 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1791 static inline void hrtick_update(struct rq *rq)
1797 * The enqueue_task method is called before nr_running is
1798 * increased. Here we update the fair scheduling stats and
1799 * then put the task into the rbtree:
1802 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1804 struct cfs_rq *cfs_rq;
1805 struct sched_entity *se = &p->se;
1807 for_each_sched_entity(se) {
1810 cfs_rq = cfs_rq_of(se);
1811 enqueue_entity(cfs_rq, se, flags);
1814 * end evaluation on encountering a throttled cfs_rq
1816 * note: in the case of encountering a throttled cfs_rq we will
1817 * post the final h_nr_running increment below.
1819 if (cfs_rq_throttled(cfs_rq))
1821 cfs_rq->h_nr_running++;
1823 flags = ENQUEUE_WAKEUP;
1826 for_each_sched_entity(se) {
1827 cfs_rq = cfs_rq_of(se);
1828 cfs_rq->h_nr_running++;
1830 if (cfs_rq_throttled(cfs_rq))
1833 update_cfs_load(cfs_rq, 0);
1834 update_cfs_shares(cfs_rq);
1842 static void set_next_buddy(struct sched_entity *se);
1845 * The dequeue_task method is called before nr_running is
1846 * decreased. We remove the task from the rbtree and
1847 * update the fair scheduling stats:
1849 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1851 struct cfs_rq *cfs_rq;
1852 struct sched_entity *se = &p->se;
1853 int task_sleep = flags & DEQUEUE_SLEEP;
1855 for_each_sched_entity(se) {
1856 cfs_rq = cfs_rq_of(se);
1857 dequeue_entity(cfs_rq, se, flags);
1860 * end evaluation on encountering a throttled cfs_rq
1862 * note: in the case of encountering a throttled cfs_rq we will
1863 * post the final h_nr_running decrement below.
1865 if (cfs_rq_throttled(cfs_rq))
1867 cfs_rq->h_nr_running--;
1869 /* Don't dequeue parent if it has other entities besides us */
1870 if (cfs_rq->load.weight) {
1872 * Bias pick_next to pick a task from this cfs_rq, as
1873 * p is sleeping when it is within its sched_slice.
1875 if (task_sleep && parent_entity(se))
1876 set_next_buddy(parent_entity(se));
1878 /* avoid re-evaluating load for this entity */
1879 se = parent_entity(se);
1882 flags |= DEQUEUE_SLEEP;
1885 for_each_sched_entity(se) {
1886 cfs_rq = cfs_rq_of(se);
1887 cfs_rq->h_nr_running--;
1889 if (cfs_rq_throttled(cfs_rq))
1892 update_cfs_load(cfs_rq, 0);
1893 update_cfs_shares(cfs_rq);
1903 static void task_waking_fair(struct task_struct *p)
1905 struct sched_entity *se = &p->se;
1906 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1909 #ifndef CONFIG_64BIT
1910 u64 min_vruntime_copy;
1913 min_vruntime_copy = cfs_rq->min_vruntime_copy;
1915 min_vruntime = cfs_rq->min_vruntime;
1916 } while (min_vruntime != min_vruntime_copy);
1918 min_vruntime = cfs_rq->min_vruntime;
1921 se->vruntime -= min_vruntime;
1924 #ifdef CONFIG_FAIR_GROUP_SCHED
1926 * effective_load() calculates the load change as seen from the root_task_group
1928 * Adding load to a group doesn't make a group heavier, but can cause movement
1929 * of group shares between cpus. Assuming the shares were perfectly aligned one
1930 * can calculate the shift in shares.
1932 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1934 struct sched_entity *se = tg->se[cpu];
1939 for_each_sched_entity(se) {
1943 w = se->my_q->load.weight;
1945 /* use this cpu's instantaneous contribution */
1946 lw = atomic_read(&tg->load_weight);
1947 lw -= se->my_q->load_contribution;
1952 if (lw > 0 && wl < lw)
1953 wl = (wl * tg->shares) / lw;
1957 /* zero point is MIN_SHARES */
1958 if (wl < MIN_SHARES)
1960 wl -= se->load.weight;
1968 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1969 unsigned long wl, unsigned long wg)
1976 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1978 s64 this_load, load;
1979 int idx, this_cpu, prev_cpu;
1980 unsigned long tl_per_task;
1981 struct task_group *tg;
1982 unsigned long weight;
1986 this_cpu = smp_processor_id();
1987 prev_cpu = task_cpu(p);
1988 load = source_load(prev_cpu, idx);
1989 this_load = target_load(this_cpu, idx);
1992 * If sync wakeup then subtract the (maximum possible)
1993 * effect of the currently running task from the load
1994 * of the current CPU:
1997 tg = task_group(current);
1998 weight = current->se.load.weight;
2000 this_load += effective_load(tg, this_cpu, -weight, -weight);
2001 load += effective_load(tg, prev_cpu, 0, -weight);
2005 weight = p->se.load.weight;
2008 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2009 * due to the sync cause above having dropped this_load to 0, we'll
2010 * always have an imbalance, but there's really nothing you can do
2011 * about that, so that's good too.
2013 * Otherwise check if either cpus are near enough in load to allow this
2014 * task to be woken on this_cpu.
2016 if (this_load > 0) {
2017 s64 this_eff_load, prev_eff_load;
2019 this_eff_load = 100;
2020 this_eff_load *= power_of(prev_cpu);
2021 this_eff_load *= this_load +
2022 effective_load(tg, this_cpu, weight, weight);
2024 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2025 prev_eff_load *= power_of(this_cpu);
2026 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2028 balanced = this_eff_load <= prev_eff_load;
2033 * If the currently running task will sleep within
2034 * a reasonable amount of time then attract this newly
2037 if (sync && balanced)
2040 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2041 tl_per_task = cpu_avg_load_per_task(this_cpu);
2044 (this_load <= load &&
2045 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2047 * This domain has SD_WAKE_AFFINE and
2048 * p is cache cold in this domain, and
2049 * there is no bad imbalance.
2051 schedstat_inc(sd, ttwu_move_affine);
2052 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2060 * find_idlest_group finds and returns the least busy CPU group within the
2063 static struct sched_group *
2064 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2065 int this_cpu, int load_idx)
2067 struct sched_group *idlest = NULL, *group = sd->groups;
2068 unsigned long min_load = ULONG_MAX, this_load = 0;
2069 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2072 unsigned long load, avg_load;
2076 /* Skip over this group if it has no CPUs allowed */
2077 if (!cpumask_intersects(sched_group_cpus(group),
2081 local_group = cpumask_test_cpu(this_cpu,
2082 sched_group_cpus(group));
2084 /* Tally up the load of all CPUs in the group */
2087 for_each_cpu(i, sched_group_cpus(group)) {
2088 /* Bias balancing toward cpus of our domain */
2090 load = source_load(i, load_idx);
2092 load = target_load(i, load_idx);
2097 /* Adjust by relative CPU power of the group */
2098 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2101 this_load = avg_load;
2102 } else if (avg_load < min_load) {
2103 min_load = avg_load;
2106 } while (group = group->next, group != sd->groups);
2108 if (!idlest || 100*this_load < imbalance*min_load)
2114 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2117 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2119 unsigned long load, min_load = ULONG_MAX;
2123 /* Traverse only the allowed CPUs */
2124 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2125 load = weighted_cpuload(i);
2127 if (load < min_load || (load == min_load && i == this_cpu)) {
2137 * Try and locate an idle CPU in the sched_domain.
2139 static int select_idle_sibling(struct task_struct *p, int target)
2141 int cpu = smp_processor_id();
2142 int prev_cpu = task_cpu(p);
2143 struct sched_domain *sd;
2147 * If the task is going to be woken-up on this cpu and if it is
2148 * already idle, then it is the right target.
2150 if (target == cpu && idle_cpu(cpu))
2154 * If the task is going to be woken-up on the cpu where it previously
2155 * ran and if it is currently idle, then it the right target.
2157 if (target == prev_cpu && idle_cpu(prev_cpu))
2161 * Otherwise, iterate the domains and find an elegible idle cpu.
2164 for_each_domain(target, sd) {
2165 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
2168 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
2176 * Lets stop looking for an idle sibling when we reached
2177 * the domain that spans the current cpu and prev_cpu.
2179 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
2180 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
2189 * sched_balance_self: balance the current task (running on cpu) in domains
2190 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2193 * Balance, ie. select the least loaded group.
2195 * Returns the target CPU number, or the same CPU if no balancing is needed.
2197 * preempt must be disabled.
2200 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2202 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2203 int cpu = smp_processor_id();
2204 int prev_cpu = task_cpu(p);
2206 int want_affine = 0;
2208 int sync = wake_flags & WF_SYNC;
2210 if (sd_flag & SD_BALANCE_WAKE) {
2211 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
2217 for_each_domain(cpu, tmp) {
2218 if (!(tmp->flags & SD_LOAD_BALANCE))
2222 * If power savings logic is enabled for a domain, see if we
2223 * are not overloaded, if so, don't balance wider.
2225 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2226 unsigned long power = 0;
2227 unsigned long nr_running = 0;
2228 unsigned long capacity;
2231 for_each_cpu(i, sched_domain_span(tmp)) {
2232 power += power_of(i);
2233 nr_running += cpu_rq(i)->cfs.nr_running;
2236 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2238 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2241 if (nr_running < capacity)
2246 * If both cpu and prev_cpu are part of this domain,
2247 * cpu is a valid SD_WAKE_AFFINE target.
2249 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2250 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2255 if (!want_sd && !want_affine)
2258 if (!(tmp->flags & sd_flag))
2266 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2269 new_cpu = select_idle_sibling(p, prev_cpu);
2274 int load_idx = sd->forkexec_idx;
2275 struct sched_group *group;
2278 if (!(sd->flags & sd_flag)) {
2283 if (sd_flag & SD_BALANCE_WAKE)
2284 load_idx = sd->wake_idx;
2286 group = find_idlest_group(sd, p, cpu, load_idx);
2292 new_cpu = find_idlest_cpu(group, p, cpu);
2293 if (new_cpu == -1 || new_cpu == cpu) {
2294 /* Now try balancing at a lower domain level of cpu */
2299 /* Now try balancing at a lower domain level of new_cpu */
2301 weight = sd->span_weight;
2303 for_each_domain(cpu, tmp) {
2304 if (weight <= tmp->span_weight)
2306 if (tmp->flags & sd_flag)
2309 /* while loop will break here if sd == NULL */
2316 #endif /* CONFIG_SMP */
2318 static unsigned long
2319 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2321 unsigned long gran = sysctl_sched_wakeup_granularity;
2324 * Since its curr running now, convert the gran from real-time
2325 * to virtual-time in his units.
2327 * By using 'se' instead of 'curr' we penalize light tasks, so
2328 * they get preempted easier. That is, if 'se' < 'curr' then
2329 * the resulting gran will be larger, therefore penalizing the
2330 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2331 * be smaller, again penalizing the lighter task.
2333 * This is especially important for buddies when the leftmost
2334 * task is higher priority than the buddy.
2336 return calc_delta_fair(gran, se);
2340 * Should 'se' preempt 'curr'.
2354 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2356 s64 gran, vdiff = curr->vruntime - se->vruntime;
2361 gran = wakeup_gran(curr, se);
2368 static void set_last_buddy(struct sched_entity *se)
2370 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2373 for_each_sched_entity(se)
2374 cfs_rq_of(se)->last = se;
2377 static void set_next_buddy(struct sched_entity *se)
2379 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2382 for_each_sched_entity(se)
2383 cfs_rq_of(se)->next = se;
2386 static void set_skip_buddy(struct sched_entity *se)
2388 for_each_sched_entity(se)
2389 cfs_rq_of(se)->skip = se;
2393 * Preempt the current task with a newly woken task if needed:
2395 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2397 struct task_struct *curr = rq->curr;
2398 struct sched_entity *se = &curr->se, *pse = &p->se;
2399 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2400 int scale = cfs_rq->nr_running >= sched_nr_latency;
2401 int next_buddy_marked = 0;
2403 if (unlikely(se == pse))
2407 * This is possible from callers such as pull_task(), in which we
2408 * unconditionally check_prempt_curr() after an enqueue (which may have
2409 * lead to a throttle). This both saves work and prevents false
2410 * next-buddy nomination below.
2412 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
2415 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2416 set_next_buddy(pse);
2417 next_buddy_marked = 1;
2421 * We can come here with TIF_NEED_RESCHED already set from new task
2424 * Note: this also catches the edge-case of curr being in a throttled
2425 * group (e.g. via set_curr_task), since update_curr() (in the
2426 * enqueue of curr) will have resulted in resched being set. This
2427 * prevents us from potentially nominating it as a false LAST_BUDDY
2430 if (test_tsk_need_resched(curr))
2433 /* Idle tasks are by definition preempted by non-idle tasks. */
2434 if (unlikely(curr->policy == SCHED_IDLE) &&
2435 likely(p->policy != SCHED_IDLE))
2439 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2440 * is driven by the tick):
2442 if (unlikely(p->policy != SCHED_NORMAL))
2445 find_matching_se(&se, &pse);
2446 update_curr(cfs_rq_of(se));
2448 if (wakeup_preempt_entity(se, pse) == 1) {
2450 * Bias pick_next to pick the sched entity that is
2451 * triggering this preemption.
2453 if (!next_buddy_marked)
2454 set_next_buddy(pse);
2463 * Only set the backward buddy when the current task is still
2464 * on the rq. This can happen when a wakeup gets interleaved
2465 * with schedule on the ->pre_schedule() or idle_balance()
2466 * point, either of which can * drop the rq lock.
2468 * Also, during early boot the idle thread is in the fair class,
2469 * for obvious reasons its a bad idea to schedule back to it.
2471 if (unlikely(!se->on_rq || curr == rq->idle))
2474 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2478 static struct task_struct *pick_next_task_fair(struct rq *rq)
2480 struct task_struct *p;
2481 struct cfs_rq *cfs_rq = &rq->cfs;
2482 struct sched_entity *se;
2484 if (!cfs_rq->nr_running)
2488 se = pick_next_entity(cfs_rq);
2489 set_next_entity(cfs_rq, se);
2490 cfs_rq = group_cfs_rq(se);
2494 hrtick_start_fair(rq, p);
2500 * Account for a descheduled task:
2502 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
2504 struct sched_entity *se = &prev->se;
2505 struct cfs_rq *cfs_rq;
2507 for_each_sched_entity(se) {
2508 cfs_rq = cfs_rq_of(se);
2509 put_prev_entity(cfs_rq, se);
2514 * sched_yield() is very simple
2516 * The magic of dealing with the ->skip buddy is in pick_next_entity.
2518 static void yield_task_fair(struct rq *rq)
2520 struct task_struct *curr = rq->curr;
2521 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2522 struct sched_entity *se = &curr->se;
2525 * Are we the only task in the tree?
2527 if (unlikely(rq->nr_running == 1))
2530 clear_buddies(cfs_rq, se);
2532 if (curr->policy != SCHED_BATCH) {
2533 update_rq_clock(rq);
2535 * Update run-time statistics of the 'current'.
2537 update_curr(cfs_rq);
2543 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2545 struct sched_entity *se = &p->se;
2547 /* throttled hierarchies are not runnable */
2548 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
2551 /* Tell the scheduler that we'd really like pse to run next. */
2554 yield_task_fair(rq);
2560 /**************************************************
2561 * Fair scheduling class load-balancing methods:
2565 * pull_task - move a task from a remote runqueue to the local runqueue.
2566 * Both runqueues must be locked.
2568 static void pull_task(struct rq *src_rq, struct task_struct *p,
2569 struct rq *this_rq, int this_cpu)
2571 deactivate_task(src_rq, p, 0);
2572 set_task_cpu(p, this_cpu);
2573 activate_task(this_rq, p, 0);
2574 check_preempt_curr(this_rq, p, 0);
2578 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2581 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2582 struct sched_domain *sd, enum cpu_idle_type idle,
2585 int tsk_cache_hot = 0;
2587 * We do not migrate tasks that are:
2588 * 1) running (obviously), or
2589 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2590 * 3) are cache-hot on their current CPU.
2592 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2593 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2598 if (task_running(rq, p)) {
2599 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2604 * Aggressive migration if:
2605 * 1) task is cache cold, or
2606 * 2) too many balance attempts have failed.
2609 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2610 if (!tsk_cache_hot ||
2611 sd->nr_balance_failed > sd->cache_nice_tries) {
2612 #ifdef CONFIG_SCHEDSTATS
2613 if (tsk_cache_hot) {
2614 schedstat_inc(sd, lb_hot_gained[idle]);
2615 schedstat_inc(p, se.statistics.nr_forced_migrations);
2621 if (tsk_cache_hot) {
2622 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2629 * move_one_task tries to move exactly one task from busiest to this_rq, as
2630 * part of active balancing operations within "domain".
2631 * Returns 1 if successful and 0 otherwise.
2633 * Called with both runqueues locked.
2636 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2637 struct sched_domain *sd, enum cpu_idle_type idle)
2639 struct task_struct *p, *n;
2640 struct cfs_rq *cfs_rq;
2643 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2644 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2645 if (throttled_lb_pair(task_group(p),
2646 busiest->cpu, this_cpu))
2649 if (!can_migrate_task(p, busiest, this_cpu,
2653 pull_task(busiest, p, this_rq, this_cpu);
2655 * Right now, this is only the second place pull_task()
2656 * is called, so we can safely collect pull_task()
2657 * stats here rather than inside pull_task().
2659 schedstat_inc(sd, lb_gained[idle]);
2667 static unsigned long
2668 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2669 unsigned long max_load_move, struct sched_domain *sd,
2670 enum cpu_idle_type idle, int *all_pinned,
2671 struct cfs_rq *busiest_cfs_rq)
2673 int loops = 0, pulled = 0;
2674 long rem_load_move = max_load_move;
2675 struct task_struct *p, *n;
2677 if (max_load_move == 0)
2680 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2681 if (loops++ > sysctl_sched_nr_migrate)
2684 if ((p->se.load.weight >> 1) > rem_load_move ||
2685 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2689 pull_task(busiest, p, this_rq, this_cpu);
2691 rem_load_move -= p->se.load.weight;
2693 #ifdef CONFIG_PREEMPT
2695 * NEWIDLE balancing is a source of latency, so preemptible
2696 * kernels will stop after the first task is pulled to minimize
2697 * the critical section.
2699 if (idle == CPU_NEWLY_IDLE)
2704 * We only want to steal up to the prescribed amount of
2707 if (rem_load_move <= 0)
2712 * Right now, this is one of only two places pull_task() is called,
2713 * so we can safely collect pull_task() stats here rather than
2714 * inside pull_task().
2716 schedstat_add(sd, lb_gained[idle], pulled);
2718 return max_load_move - rem_load_move;
2721 #ifdef CONFIG_FAIR_GROUP_SCHED
2723 * update tg->load_weight by folding this cpu's load_avg
2725 static int update_shares_cpu(struct task_group *tg, int cpu)
2727 struct cfs_rq *cfs_rq;
2728 unsigned long flags;
2735 cfs_rq = tg->cfs_rq[cpu];
2737 raw_spin_lock_irqsave(&rq->lock, flags);
2739 update_rq_clock(rq);
2740 update_cfs_load(cfs_rq, 1);
2743 * We need to update shares after updating tg->load_weight in
2744 * order to adjust the weight of groups with long running tasks.
2746 update_cfs_shares(cfs_rq);
2748 raw_spin_unlock_irqrestore(&rq->lock, flags);
2753 static void update_shares(int cpu)
2755 struct cfs_rq *cfs_rq;
2756 struct rq *rq = cpu_rq(cpu);
2760 * Iterates the task_group tree in a bottom up fashion, see
2761 * list_add_leaf_cfs_rq() for details.
2763 for_each_leaf_cfs_rq(rq, cfs_rq) {
2764 /* throttled entities do not contribute to load */
2765 if (throttled_hierarchy(cfs_rq))
2768 update_shares_cpu(cfs_rq->tg, cpu);
2774 * Compute the cpu's hierarchical load factor for each task group.
2775 * This needs to be done in a top-down fashion because the load of a child
2776 * group is a fraction of its parents load.
2778 static int tg_load_down(struct task_group *tg, void *data)
2781 long cpu = (long)data;
2784 load = cpu_rq(cpu)->load.weight;
2786 load = tg->parent->cfs_rq[cpu]->h_load;
2787 load *= tg->se[cpu]->load.weight;
2788 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2791 tg->cfs_rq[cpu]->h_load = load;
2796 static void update_h_load(long cpu)
2798 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2801 static unsigned long
2802 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2803 unsigned long max_load_move,
2804 struct sched_domain *sd, enum cpu_idle_type idle,
2807 long rem_load_move = max_load_move;
2808 struct cfs_rq *busiest_cfs_rq;
2811 update_h_load(cpu_of(busiest));
2813 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
2814 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2815 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2816 u64 rem_load, moved_load;
2819 * empty group or part of a throttled hierarchy
2821 if (!busiest_cfs_rq->task_weight ||
2822 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu))
2825 rem_load = (u64)rem_load_move * busiest_weight;
2826 rem_load = div_u64(rem_load, busiest_h_load + 1);
2828 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2829 rem_load, sd, idle, all_pinned,
2835 moved_load *= busiest_h_load;
2836 moved_load = div_u64(moved_load, busiest_weight + 1);
2838 rem_load_move -= moved_load;
2839 if (rem_load_move < 0)
2844 return max_load_move - rem_load_move;
2847 static inline void update_shares(int cpu)
2851 static unsigned long
2852 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2853 unsigned long max_load_move,
2854 struct sched_domain *sd, enum cpu_idle_type idle,
2857 return balance_tasks(this_rq, this_cpu, busiest,
2858 max_load_move, sd, idle, all_pinned,
2864 * move_tasks tries to move up to max_load_move weighted load from busiest to
2865 * this_rq, as part of a balancing operation within domain "sd".
2866 * Returns 1 if successful and 0 otherwise.
2868 * Called with both runqueues locked.
2870 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2871 unsigned long max_load_move,
2872 struct sched_domain *sd, enum cpu_idle_type idle,
2875 unsigned long total_load_moved = 0, load_moved;
2878 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2879 max_load_move - total_load_moved,
2880 sd, idle, all_pinned);
2882 total_load_moved += load_moved;
2884 #ifdef CONFIG_PREEMPT
2886 * NEWIDLE balancing is a source of latency, so preemptible
2887 * kernels will stop after the first task is pulled to minimize
2888 * the critical section.
2890 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2893 if (raw_spin_is_contended(&this_rq->lock) ||
2894 raw_spin_is_contended(&busiest->lock))
2897 } while (load_moved && max_load_move > total_load_moved);
2899 return total_load_moved > 0;
2902 /********** Helpers for find_busiest_group ************************/
2904 * sd_lb_stats - Structure to store the statistics of a sched_domain
2905 * during load balancing.
2907 struct sd_lb_stats {
2908 struct sched_group *busiest; /* Busiest group in this sd */
2909 struct sched_group *this; /* Local group in this sd */
2910 unsigned long total_load; /* Total load of all groups in sd */
2911 unsigned long total_pwr; /* Total power of all groups in sd */
2912 unsigned long avg_load; /* Average load across all groups in sd */
2914 /** Statistics of this group */
2915 unsigned long this_load;
2916 unsigned long this_load_per_task;
2917 unsigned long this_nr_running;
2918 unsigned long this_has_capacity;
2919 unsigned int this_idle_cpus;
2921 /* Statistics of the busiest group */
2922 unsigned int busiest_idle_cpus;
2923 unsigned long max_load;
2924 unsigned long busiest_load_per_task;
2925 unsigned long busiest_nr_running;
2926 unsigned long busiest_group_capacity;
2927 unsigned long busiest_has_capacity;
2928 unsigned int busiest_group_weight;
2930 int group_imb; /* Is there imbalance in this sd */
2931 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2932 int power_savings_balance; /* Is powersave balance needed for this sd */
2933 struct sched_group *group_min; /* Least loaded group in sd */
2934 struct sched_group *group_leader; /* Group which relieves group_min */
2935 unsigned long min_load_per_task; /* load_per_task in group_min */
2936 unsigned long leader_nr_running; /* Nr running of group_leader */
2937 unsigned long min_nr_running; /* Nr running of group_min */
2942 * sg_lb_stats - stats of a sched_group required for load_balancing
2944 struct sg_lb_stats {
2945 unsigned long avg_load; /*Avg load across the CPUs of the group */
2946 unsigned long group_load; /* Total load over the CPUs of the group */
2947 unsigned long sum_nr_running; /* Nr tasks running in the group */
2948 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2949 unsigned long group_capacity;
2950 unsigned long idle_cpus;
2951 unsigned long group_weight;
2952 int group_imb; /* Is there an imbalance in the group ? */
2953 int group_has_capacity; /* Is there extra capacity in the group? */
2957 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2958 * @group: The group whose first cpu is to be returned.
2960 static inline unsigned int group_first_cpu(struct sched_group *group)
2962 return cpumask_first(sched_group_cpus(group));
2966 * get_sd_load_idx - Obtain the load index for a given sched domain.
2967 * @sd: The sched_domain whose load_idx is to be obtained.
2968 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2970 static inline int get_sd_load_idx(struct sched_domain *sd,
2971 enum cpu_idle_type idle)
2977 load_idx = sd->busy_idx;
2980 case CPU_NEWLY_IDLE:
2981 load_idx = sd->newidle_idx;
2984 load_idx = sd->idle_idx;
2992 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2994 * init_sd_power_savings_stats - Initialize power savings statistics for
2995 * the given sched_domain, during load balancing.
2997 * @sd: Sched domain whose power-savings statistics are to be initialized.
2998 * @sds: Variable containing the statistics for sd.
2999 * @idle: Idle status of the CPU at which we're performing load-balancing.
3001 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3002 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3005 * Busy processors will not participate in power savings
3008 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3009 sds->power_savings_balance = 0;
3011 sds->power_savings_balance = 1;
3012 sds->min_nr_running = ULONG_MAX;
3013 sds->leader_nr_running = 0;
3018 * update_sd_power_savings_stats - Update the power saving stats for a
3019 * sched_domain while performing load balancing.
3021 * @group: sched_group belonging to the sched_domain under consideration.
3022 * @sds: Variable containing the statistics of the sched_domain
3023 * @local_group: Does group contain the CPU for which we're performing
3025 * @sgs: Variable containing the statistics of the group.
3027 static inline void update_sd_power_savings_stats(struct sched_group *group,
3028 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3031 if (!sds->power_savings_balance)
3035 * If the local group is idle or completely loaded
3036 * no need to do power savings balance at this domain
3038 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3039 !sds->this_nr_running))
3040 sds->power_savings_balance = 0;
3043 * If a group is already running at full capacity or idle,
3044 * don't include that group in power savings calculations
3046 if (!sds->power_savings_balance ||
3047 sgs->sum_nr_running >= sgs->group_capacity ||
3048 !sgs->sum_nr_running)
3052 * Calculate the group which has the least non-idle load.
3053 * This is the group from where we need to pick up the load
3056 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3057 (sgs->sum_nr_running == sds->min_nr_running &&
3058 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3059 sds->group_min = group;
3060 sds->min_nr_running = sgs->sum_nr_running;
3061 sds->min_load_per_task = sgs->sum_weighted_load /
3062 sgs->sum_nr_running;
3066 * Calculate the group which is almost near its
3067 * capacity but still has some space to pick up some load
3068 * from other group and save more power
3070 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3073 if (sgs->sum_nr_running > sds->leader_nr_running ||
3074 (sgs->sum_nr_running == sds->leader_nr_running &&
3075 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3076 sds->group_leader = group;
3077 sds->leader_nr_running = sgs->sum_nr_running;
3082 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3083 * @sds: Variable containing the statistics of the sched_domain
3084 * under consideration.
3085 * @this_cpu: Cpu at which we're currently performing load-balancing.
3086 * @imbalance: Variable to store the imbalance.
3089 * Check if we have potential to perform some power-savings balance.
3090 * If yes, set the busiest group to be the least loaded group in the
3091 * sched_domain, so that it's CPUs can be put to idle.
3093 * Returns 1 if there is potential to perform power-savings balance.
3096 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3097 int this_cpu, unsigned long *imbalance)
3099 if (!sds->power_savings_balance)
3102 if (sds->this != sds->group_leader ||
3103 sds->group_leader == sds->group_min)
3106 *imbalance = sds->min_load_per_task;
3107 sds->busiest = sds->group_min;
3112 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3113 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3114 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3119 static inline void update_sd_power_savings_stats(struct sched_group *group,
3120 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3125 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3126 int this_cpu, unsigned long *imbalance)
3130 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3133 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3135 return SCHED_POWER_SCALE;
3138 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3140 return default_scale_freq_power(sd, cpu);
3143 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3145 unsigned long weight = sd->span_weight;
3146 unsigned long smt_gain = sd->smt_gain;
3153 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3155 return default_scale_smt_power(sd, cpu);
3158 unsigned long scale_rt_power(int cpu)
3160 struct rq *rq = cpu_rq(cpu);
3161 u64 total, available;
3163 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3165 if (unlikely(total < rq->rt_avg)) {
3166 /* Ensures that power won't end up being negative */
3169 available = total - rq->rt_avg;
3172 if (unlikely((s64)total < SCHED_POWER_SCALE))
3173 total = SCHED_POWER_SCALE;
3175 total >>= SCHED_POWER_SHIFT;
3177 return div_u64(available, total);
3180 static void update_cpu_power(struct sched_domain *sd, int cpu)
3182 unsigned long weight = sd->span_weight;
3183 unsigned long power = SCHED_POWER_SCALE;
3184 struct sched_group *sdg = sd->groups;
3186 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3187 if (sched_feat(ARCH_POWER))
3188 power *= arch_scale_smt_power(sd, cpu);
3190 power *= default_scale_smt_power(sd, cpu);
3192 power >>= SCHED_POWER_SHIFT;
3195 sdg->sgp->power_orig = power;
3197 if (sched_feat(ARCH_POWER))
3198 power *= arch_scale_freq_power(sd, cpu);
3200 power *= default_scale_freq_power(sd, cpu);
3202 power >>= SCHED_POWER_SHIFT;
3204 power *= scale_rt_power(cpu);
3205 power >>= SCHED_POWER_SHIFT;
3210 cpu_rq(cpu)->cpu_power = power;
3211 sdg->sgp->power = power;
3214 static void update_group_power(struct sched_domain *sd, int cpu)
3216 struct sched_domain *child = sd->child;
3217 struct sched_group *group, *sdg = sd->groups;
3218 unsigned long power;
3221 update_cpu_power(sd, cpu);
3227 group = child->groups;
3229 power += group->sgp->power;
3230 group = group->next;
3231 } while (group != child->groups);
3233 sdg->sgp->power = power;
3237 * Try and fix up capacity for tiny siblings, this is needed when
3238 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3239 * which on its own isn't powerful enough.
3241 * See update_sd_pick_busiest() and check_asym_packing().
3244 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3247 * Only siblings can have significantly less than SCHED_POWER_SCALE
3249 if (!(sd->flags & SD_SHARE_CPUPOWER))
3253 * If ~90% of the cpu_power is still there, we're good.
3255 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3262 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3263 * @sd: The sched_domain whose statistics are to be updated.
3264 * @group: sched_group whose statistics are to be updated.
3265 * @this_cpu: Cpu for which load balance is currently performed.
3266 * @idle: Idle status of this_cpu
3267 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3268 * @local_group: Does group contain this_cpu.
3269 * @cpus: Set of cpus considered for load balancing.
3270 * @balance: Should we balance.
3271 * @sgs: variable to hold the statistics for this group.
3273 static inline void update_sg_lb_stats(struct sched_domain *sd,
3274 struct sched_group *group, int this_cpu,
3275 enum cpu_idle_type idle, int load_idx,
3276 int local_group, const struct cpumask *cpus,
3277 int *balance, struct sg_lb_stats *sgs)
3279 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3281 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3282 unsigned long avg_load_per_task = 0;
3285 balance_cpu = group_first_cpu(group);
3287 /* Tally up the load of all CPUs in the group */
3289 min_cpu_load = ~0UL;
3292 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3293 struct rq *rq = cpu_rq(i);
3295 /* Bias balancing toward cpus of our domain */
3297 if (idle_cpu(i) && !first_idle_cpu) {
3302 load = target_load(i, load_idx);
3304 load = source_load(i, load_idx);
3305 if (load > max_cpu_load) {
3306 max_cpu_load = load;
3307 max_nr_running = rq->nr_running;
3309 if (min_cpu_load > load)
3310 min_cpu_load = load;
3313 sgs->group_load += load;
3314 sgs->sum_nr_running += rq->nr_running;
3315 sgs->sum_weighted_load += weighted_cpuload(i);
3321 * First idle cpu or the first cpu(busiest) in this sched group
3322 * is eligible for doing load balancing at this and above
3323 * domains. In the newly idle case, we will allow all the cpu's
3324 * to do the newly idle load balance.
3326 if (idle != CPU_NEWLY_IDLE && local_group) {
3327 if (balance_cpu != this_cpu) {
3331 update_group_power(sd, this_cpu);
3334 /* Adjust by relative CPU power of the group */
3335 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3338 * Consider the group unbalanced when the imbalance is larger
3339 * than the average weight of a task.
3341 * APZ: with cgroup the avg task weight can vary wildly and
3342 * might not be a suitable number - should we keep a
3343 * normalized nr_running number somewhere that negates
3346 if (sgs->sum_nr_running)
3347 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3349 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3352 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3354 if (!sgs->group_capacity)
3355 sgs->group_capacity = fix_small_capacity(sd, group);
3356 sgs->group_weight = group->group_weight;
3358 if (sgs->group_capacity > sgs->sum_nr_running)
3359 sgs->group_has_capacity = 1;
3363 * update_sd_pick_busiest - return 1 on busiest group
3364 * @sd: sched_domain whose statistics are to be checked
3365 * @sds: sched_domain statistics
3366 * @sg: sched_group candidate to be checked for being the busiest
3367 * @sgs: sched_group statistics
3368 * @this_cpu: the current cpu
3370 * Determine if @sg is a busier group than the previously selected
3373 static bool update_sd_pick_busiest(struct sched_domain *sd,
3374 struct sd_lb_stats *sds,
3375 struct sched_group *sg,
3376 struct sg_lb_stats *sgs,
3379 if (sgs->avg_load <= sds->max_load)
3382 if (sgs->sum_nr_running > sgs->group_capacity)
3389 * ASYM_PACKING needs to move all the work to the lowest
3390 * numbered CPUs in the group, therefore mark all groups
3391 * higher than ourself as busy.
3393 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3394 this_cpu < group_first_cpu(sg)) {
3398 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3406 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3407 * @sd: sched_domain whose statistics are to be updated.
3408 * @this_cpu: Cpu for which load balance is currently performed.
3409 * @idle: Idle status of this_cpu
3410 * @cpus: Set of cpus considered for load balancing.
3411 * @balance: Should we balance.
3412 * @sds: variable to hold the statistics for this sched_domain.
3414 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3415 enum cpu_idle_type idle, const struct cpumask *cpus,
3416 int *balance, struct sd_lb_stats *sds)
3418 struct sched_domain *child = sd->child;
3419 struct sched_group *sg = sd->groups;
3420 struct sg_lb_stats sgs;
3421 int load_idx, prefer_sibling = 0;
3423 if (child && child->flags & SD_PREFER_SIBLING)
3426 init_sd_power_savings_stats(sd, sds, idle);
3427 load_idx = get_sd_load_idx(sd, idle);
3432 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3433 memset(&sgs, 0, sizeof(sgs));
3434 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3435 local_group, cpus, balance, &sgs);
3437 if (local_group && !(*balance))
3440 sds->total_load += sgs.group_load;
3441 sds->total_pwr += sg->sgp->power;
3444 * In case the child domain prefers tasks go to siblings
3445 * first, lower the sg capacity to one so that we'll try
3446 * and move all the excess tasks away. We lower the capacity
3447 * of a group only if the local group has the capacity to fit
3448 * these excess tasks, i.e. nr_running < group_capacity. The
3449 * extra check prevents the case where you always pull from the
3450 * heaviest group when it is already under-utilized (possible
3451 * with a large weight task outweighs the tasks on the system).
3453 if (prefer_sibling && !local_group && sds->this_has_capacity)
3454 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3457 sds->this_load = sgs.avg_load;
3459 sds->this_nr_running = sgs.sum_nr_running;
3460 sds->this_load_per_task = sgs.sum_weighted_load;
3461 sds->this_has_capacity = sgs.group_has_capacity;
3462 sds->this_idle_cpus = sgs.idle_cpus;
3463 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
3464 sds->max_load = sgs.avg_load;
3466 sds->busiest_nr_running = sgs.sum_nr_running;
3467 sds->busiest_idle_cpus = sgs.idle_cpus;
3468 sds->busiest_group_capacity = sgs.group_capacity;
3469 sds->busiest_load_per_task = sgs.sum_weighted_load;
3470 sds->busiest_has_capacity = sgs.group_has_capacity;
3471 sds->busiest_group_weight = sgs.group_weight;
3472 sds->group_imb = sgs.group_imb;
3475 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
3477 } while (sg != sd->groups);
3480 int __weak arch_sd_sibling_asym_packing(void)
3482 return 0*SD_ASYM_PACKING;
3486 * check_asym_packing - Check to see if the group is packed into the
3489 * This is primarily intended to used at the sibling level. Some
3490 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3491 * case of POWER7, it can move to lower SMT modes only when higher
3492 * threads are idle. When in lower SMT modes, the threads will
3493 * perform better since they share less core resources. Hence when we
3494 * have idle threads, we want them to be the higher ones.
3496 * This packing function is run on idle threads. It checks to see if
3497 * the busiest CPU in this domain (core in the P7 case) has a higher
3498 * CPU number than the packing function is being run on. Here we are
3499 * assuming lower CPU number will be equivalent to lower a SMT thread
3502 * Returns 1 when packing is required and a task should be moved to
3503 * this CPU. The amount of the imbalance is returned in *imbalance.
3505 * @sd: The sched_domain whose packing is to be checked.
3506 * @sds: Statistics of the sched_domain which is to be packed
3507 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3508 * @imbalance: returns amount of imbalanced due to packing.
3510 static int check_asym_packing(struct sched_domain *sd,
3511 struct sd_lb_stats *sds,
3512 int this_cpu, unsigned long *imbalance)
3516 if (!(sd->flags & SD_ASYM_PACKING))
3522 busiest_cpu = group_first_cpu(sds->busiest);
3523 if (this_cpu > busiest_cpu)
3526 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
3532 * fix_small_imbalance - Calculate the minor imbalance that exists
3533 * amongst the groups of a sched_domain, during
3535 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3536 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3537 * @imbalance: Variable to store the imbalance.
3539 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3540 int this_cpu, unsigned long *imbalance)
3542 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3543 unsigned int imbn = 2;
3544 unsigned long scaled_busy_load_per_task;
3546 if (sds->this_nr_running) {
3547 sds->this_load_per_task /= sds->this_nr_running;
3548 if (sds->busiest_load_per_task >
3549 sds->this_load_per_task)
3552 sds->this_load_per_task =
3553 cpu_avg_load_per_task(this_cpu);
3555 scaled_busy_load_per_task = sds->busiest_load_per_task
3556 * SCHED_POWER_SCALE;
3557 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3559 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3560 (scaled_busy_load_per_task * imbn)) {
3561 *imbalance = sds->busiest_load_per_task;
3566 * OK, we don't have enough imbalance to justify moving tasks,
3567 * however we may be able to increase total CPU power used by
3571 pwr_now += sds->busiest->sgp->power *
3572 min(sds->busiest_load_per_task, sds->max_load);
3573 pwr_now += sds->this->sgp->power *
3574 min(sds->this_load_per_task, sds->this_load);
3575 pwr_now /= SCHED_POWER_SCALE;
3577 /* Amount of load we'd subtract */
3578 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3579 sds->busiest->sgp->power;
3580 if (sds->max_load > tmp)
3581 pwr_move += sds->busiest->sgp->power *
3582 min(sds->busiest_load_per_task, sds->max_load - tmp);
3584 /* Amount of load we'd add */
3585 if (sds->max_load * sds->busiest->sgp->power <
3586 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3587 tmp = (sds->max_load * sds->busiest->sgp->power) /
3588 sds->this->sgp->power;
3590 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3591 sds->this->sgp->power;
3592 pwr_move += sds->this->sgp->power *
3593 min(sds->this_load_per_task, sds->this_load + tmp);
3594 pwr_move /= SCHED_POWER_SCALE;
3596 /* Move if we gain throughput */
3597 if (pwr_move > pwr_now)
3598 *imbalance = sds->busiest_load_per_task;
3602 * calculate_imbalance - Calculate the amount of imbalance present within the
3603 * groups of a given sched_domain during load balance.
3604 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3605 * @this_cpu: Cpu for which currently load balance is being performed.
3606 * @imbalance: The variable to store the imbalance.
3608 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3609 unsigned long *imbalance)
3611 unsigned long max_pull, load_above_capacity = ~0UL;
3613 sds->busiest_load_per_task /= sds->busiest_nr_running;
3614 if (sds->group_imb) {
3615 sds->busiest_load_per_task =
3616 min(sds->busiest_load_per_task, sds->avg_load);
3620 * In the presence of smp nice balancing, certain scenarios can have
3621 * max load less than avg load(as we skip the groups at or below
3622 * its cpu_power, while calculating max_load..)
3624 if (sds->max_load < sds->avg_load) {
3626 return fix_small_imbalance(sds, this_cpu, imbalance);
3629 if (!sds->group_imb) {
3631 * Don't want to pull so many tasks that a group would go idle.
3633 load_above_capacity = (sds->busiest_nr_running -
3634 sds->busiest_group_capacity);
3636 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3638 load_above_capacity /= sds->busiest->sgp->power;
3642 * We're trying to get all the cpus to the average_load, so we don't
3643 * want to push ourselves above the average load, nor do we wish to
3644 * reduce the max loaded cpu below the average load. At the same time,
3645 * we also don't want to reduce the group load below the group capacity
3646 * (so that we can implement power-savings policies etc). Thus we look
3647 * for the minimum possible imbalance.
3648 * Be careful of negative numbers as they'll appear as very large values
3649 * with unsigned longs.
3651 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3653 /* How much load to actually move to equalise the imbalance */
3654 *imbalance = min(max_pull * sds->busiest->sgp->power,
3655 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3656 / SCHED_POWER_SCALE;
3659 * if *imbalance is less than the average load per runnable task
3660 * there is no guarantee that any tasks will be moved so we'll have
3661 * a think about bumping its value to force at least one task to be
3664 if (*imbalance < sds->busiest_load_per_task)
3665 return fix_small_imbalance(sds, this_cpu, imbalance);
3669 /******* find_busiest_group() helpers end here *********************/
3672 * find_busiest_group - Returns the busiest group within the sched_domain
3673 * if there is an imbalance. If there isn't an imbalance, and
3674 * the user has opted for power-savings, it returns a group whose
3675 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3676 * such a group exists.
3678 * Also calculates the amount of weighted load which should be moved
3679 * to restore balance.
3681 * @sd: The sched_domain whose busiest group is to be returned.
3682 * @this_cpu: The cpu for which load balancing is currently being performed.
3683 * @imbalance: Variable which stores amount of weighted load which should
3684 * be moved to restore balance/put a group to idle.
3685 * @idle: The idle status of this_cpu.
3686 * @cpus: The set of CPUs under consideration for load-balancing.
3687 * @balance: Pointer to a variable indicating if this_cpu
3688 * is the appropriate cpu to perform load balancing at this_level.
3690 * Returns: - the busiest group if imbalance exists.
3691 * - If no imbalance and user has opted for power-savings balance,
3692 * return the least loaded group whose CPUs can be
3693 * put to idle by rebalancing its tasks onto our group.
3695 static struct sched_group *
3696 find_busiest_group(struct sched_domain *sd, int this_cpu,
3697 unsigned long *imbalance, enum cpu_idle_type idle,
3698 const struct cpumask *cpus, int *balance)
3700 struct sd_lb_stats sds;
3702 memset(&sds, 0, sizeof(sds));
3705 * Compute the various statistics relavent for load balancing at
3708 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3711 * this_cpu is not the appropriate cpu to perform load balancing at
3717 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3718 check_asym_packing(sd, &sds, this_cpu, imbalance))
3721 /* There is no busy sibling group to pull tasks from */
3722 if (!sds.busiest || sds.busiest_nr_running == 0)
3725 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3728 * If the busiest group is imbalanced the below checks don't
3729 * work because they assumes all things are equal, which typically
3730 * isn't true due to cpus_allowed constraints and the like.
3735 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3736 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3737 !sds.busiest_has_capacity)
3741 * If the local group is more busy than the selected busiest group
3742 * don't try and pull any tasks.
3744 if (sds.this_load >= sds.max_load)
3748 * Don't pull any tasks if this group is already above the domain
3751 if (sds.this_load >= sds.avg_load)
3754 if (idle == CPU_IDLE) {
3756 * This cpu is idle. If the busiest group load doesn't
3757 * have more tasks than the number of available cpu's and
3758 * there is no imbalance between this and busiest group
3759 * wrt to idle cpu's, it is balanced.
3761 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3762 sds.busiest_nr_running <= sds.busiest_group_weight)
3766 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3767 * imbalance_pct to be conservative.
3769 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3774 /* Looks like there is an imbalance. Compute it */
3775 calculate_imbalance(&sds, this_cpu, imbalance);
3780 * There is no obvious imbalance. But check if we can do some balancing
3783 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3791 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3794 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3795 enum cpu_idle_type idle, unsigned long imbalance,
3796 const struct cpumask *cpus)
3798 struct rq *busiest = NULL, *rq;
3799 unsigned long max_load = 0;
3802 for_each_cpu(i, sched_group_cpus(group)) {
3803 unsigned long power = power_of(i);
3804 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3809 capacity = fix_small_capacity(sd, group);
3811 if (!cpumask_test_cpu(i, cpus))
3815 wl = weighted_cpuload(i);
3818 * When comparing with imbalance, use weighted_cpuload()
3819 * which is not scaled with the cpu power.
3821 if (capacity && rq->nr_running == 1 && wl > imbalance)
3825 * For the load comparisons with the other cpu's, consider
3826 * the weighted_cpuload() scaled with the cpu power, so that
3827 * the load can be moved away from the cpu that is potentially
3828 * running at a lower capacity.
3830 wl = (wl * SCHED_POWER_SCALE) / power;
3832 if (wl > max_load) {
3842 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3843 * so long as it is large enough.
3845 #define MAX_PINNED_INTERVAL 512
3847 /* Working cpumask for load_balance and load_balance_newidle. */
3848 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3850 static int need_active_balance(struct sched_domain *sd, int idle,
3851 int busiest_cpu, int this_cpu)
3853 if (idle == CPU_NEWLY_IDLE) {
3856 * ASYM_PACKING needs to force migrate tasks from busy but
3857 * higher numbered CPUs in order to pack all tasks in the
3858 * lowest numbered CPUs.
3860 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3864 * The only task running in a non-idle cpu can be moved to this
3865 * cpu in an attempt to completely freeup the other CPU
3868 * The package power saving logic comes from
3869 * find_busiest_group(). If there are no imbalance, then
3870 * f_b_g() will return NULL. However when sched_mc={1,2} then
3871 * f_b_g() will select a group from which a running task may be
3872 * pulled to this cpu in order to make the other package idle.
3873 * If there is no opportunity to make a package idle and if
3874 * there are no imbalance, then f_b_g() will return NULL and no
3875 * action will be taken in load_balance_newidle().
3877 * Under normal task pull operation due to imbalance, there
3878 * will be more than one task in the source run queue and
3879 * move_tasks() will succeed. ld_moved will be true and this
3880 * active balance code will not be triggered.
3882 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3886 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3889 static int active_load_balance_cpu_stop(void *data);
3892 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3893 * tasks if there is an imbalance.
3895 static int load_balance(int this_cpu, struct rq *this_rq,
3896 struct sched_domain *sd, enum cpu_idle_type idle,
3899 int ld_moved, all_pinned = 0, active_balance = 0;
3900 struct sched_group *group;
3901 unsigned long imbalance;
3903 unsigned long flags;
3904 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3906 cpumask_copy(cpus, cpu_active_mask);
3908 schedstat_inc(sd, lb_count[idle]);
3911 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
3918 schedstat_inc(sd, lb_nobusyg[idle]);
3922 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3924 schedstat_inc(sd, lb_nobusyq[idle]);
3928 BUG_ON(busiest == this_rq);
3930 schedstat_add(sd, lb_imbalance[idle], imbalance);
3933 if (busiest->nr_running > 1) {
3935 * Attempt to move tasks. If find_busiest_group has found
3936 * an imbalance but busiest->nr_running <= 1, the group is
3937 * still unbalanced. ld_moved simply stays zero, so it is
3938 * correctly treated as an imbalance.
3941 local_irq_save(flags);
3942 double_rq_lock(this_rq, busiest);
3943 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3944 imbalance, sd, idle, &all_pinned);
3945 double_rq_unlock(this_rq, busiest);
3946 local_irq_restore(flags);
3949 * some other cpu did the load balance for us.
3951 if (ld_moved && this_cpu != smp_processor_id())
3952 resched_cpu(this_cpu);
3954 /* All tasks on this runqueue were pinned by CPU affinity */
3955 if (unlikely(all_pinned)) {
3956 cpumask_clear_cpu(cpu_of(busiest), cpus);
3957 if (!cpumask_empty(cpus))
3964 schedstat_inc(sd, lb_failed[idle]);
3966 * Increment the failure counter only on periodic balance.
3967 * We do not want newidle balance, which can be very
3968 * frequent, pollute the failure counter causing
3969 * excessive cache_hot migrations and active balances.
3971 if (idle != CPU_NEWLY_IDLE)
3972 sd->nr_balance_failed++;
3974 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
3975 raw_spin_lock_irqsave(&busiest->lock, flags);
3977 /* don't kick the active_load_balance_cpu_stop,
3978 * if the curr task on busiest cpu can't be
3981 if (!cpumask_test_cpu(this_cpu,
3982 &busiest->curr->cpus_allowed)) {
3983 raw_spin_unlock_irqrestore(&busiest->lock,
3986 goto out_one_pinned;
3990 * ->active_balance synchronizes accesses to
3991 * ->active_balance_work. Once set, it's cleared
3992 * only after active load balance is finished.
3994 if (!busiest->active_balance) {
3995 busiest->active_balance = 1;
3996 busiest->push_cpu = this_cpu;
3999 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4002 stop_one_cpu_nowait(cpu_of(busiest),
4003 active_load_balance_cpu_stop, busiest,
4004 &busiest->active_balance_work);
4007 * We've kicked active balancing, reset the failure
4010 sd->nr_balance_failed = sd->cache_nice_tries+1;
4013 sd->nr_balance_failed = 0;
4015 if (likely(!active_balance)) {
4016 /* We were unbalanced, so reset the balancing interval */
4017 sd->balance_interval = sd->min_interval;
4020 * If we've begun active balancing, start to back off. This
4021 * case may not be covered by the all_pinned logic if there
4022 * is only 1 task on the busy runqueue (because we don't call
4025 if (sd->balance_interval < sd->max_interval)
4026 sd->balance_interval *= 2;
4032 schedstat_inc(sd, lb_balanced[idle]);
4034 sd->nr_balance_failed = 0;
4037 /* tune up the balancing interval */
4038 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4039 (sd->balance_interval < sd->max_interval))
4040 sd->balance_interval *= 2;
4048 * idle_balance is called by schedule() if this_cpu is about to become
4049 * idle. Attempts to pull tasks from other CPUs.
4051 static void idle_balance(int this_cpu, struct rq *this_rq)
4053 struct sched_domain *sd;
4054 int pulled_task = 0;
4055 unsigned long next_balance = jiffies + HZ;
4057 this_rq->idle_stamp = this_rq->clock;
4059 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4063 * Drop the rq->lock, but keep IRQ/preempt disabled.
4065 raw_spin_unlock(&this_rq->lock);
4067 update_shares(this_cpu);
4069 for_each_domain(this_cpu, sd) {
4070 unsigned long interval;
4073 if (!(sd->flags & SD_LOAD_BALANCE))
4076 if (sd->flags & SD_BALANCE_NEWIDLE) {
4077 /* If we've pulled tasks over stop searching: */
4078 pulled_task = load_balance(this_cpu, this_rq,
4079 sd, CPU_NEWLY_IDLE, &balance);
4082 interval = msecs_to_jiffies(sd->balance_interval);
4083 if (time_after(next_balance, sd->last_balance + interval))
4084 next_balance = sd->last_balance + interval;
4086 this_rq->idle_stamp = 0;
4092 raw_spin_lock(&this_rq->lock);
4094 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4096 * We are going idle. next_balance may be set based on
4097 * a busy processor. So reset next_balance.
4099 this_rq->next_balance = next_balance;
4104 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4105 * running tasks off the busiest CPU onto idle CPUs. It requires at
4106 * least 1 task to be running on each physical CPU where possible, and
4107 * avoids physical / logical imbalances.
4109 static int active_load_balance_cpu_stop(void *data)
4111 struct rq *busiest_rq = data;
4112 int busiest_cpu = cpu_of(busiest_rq);
4113 int target_cpu = busiest_rq->push_cpu;
4114 struct rq *target_rq = cpu_rq(target_cpu);
4115 struct sched_domain *sd;
4117 raw_spin_lock_irq(&busiest_rq->lock);
4119 /* make sure the requested cpu hasn't gone down in the meantime */
4120 if (unlikely(busiest_cpu != smp_processor_id() ||
4121 !busiest_rq->active_balance))
4124 /* Is there any task to move? */
4125 if (busiest_rq->nr_running <= 1)
4129 * This condition is "impossible", if it occurs
4130 * we need to fix it. Originally reported by
4131 * Bjorn Helgaas on a 128-cpu setup.
4133 BUG_ON(busiest_rq == target_rq);
4135 /* move a task from busiest_rq to target_rq */
4136 double_lock_balance(busiest_rq, target_rq);
4138 /* Search for an sd spanning us and the target CPU. */
4140 for_each_domain(target_cpu, sd) {
4141 if ((sd->flags & SD_LOAD_BALANCE) &&
4142 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4147 schedstat_inc(sd, alb_count);
4149 if (move_one_task(target_rq, target_cpu, busiest_rq,
4151 schedstat_inc(sd, alb_pushed);
4153 schedstat_inc(sd, alb_failed);
4156 double_unlock_balance(busiest_rq, target_rq);
4158 busiest_rq->active_balance = 0;
4159 raw_spin_unlock_irq(&busiest_rq->lock);
4165 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
4167 static void trigger_sched_softirq(void *data)
4169 raise_softirq_irqoff(SCHED_SOFTIRQ);
4172 static inline void init_sched_softirq_csd(struct call_single_data *csd)
4174 csd->func = trigger_sched_softirq;
4181 * idle load balancing details
4182 * - One of the idle CPUs nominates itself as idle load_balancer, while
4184 * - This idle load balancer CPU will also go into tickless mode when
4185 * it is idle, just like all other idle CPUs
4186 * - When one of the busy CPUs notice that there may be an idle rebalancing
4187 * needed, they will kick the idle load balancer, which then does idle
4188 * load balancing for all the idle CPUs.
4191 atomic_t load_balancer;
4192 atomic_t first_pick_cpu;
4193 atomic_t second_pick_cpu;
4194 cpumask_var_t idle_cpus_mask;
4195 cpumask_var_t grp_idle_mask;
4196 unsigned long next_balance; /* in jiffy units */
4197 } nohz ____cacheline_aligned;
4199 int get_nohz_load_balancer(void)
4201 return atomic_read(&nohz.load_balancer);
4204 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4206 * lowest_flag_domain - Return lowest sched_domain containing flag.
4207 * @cpu: The cpu whose lowest level of sched domain is to
4209 * @flag: The flag to check for the lowest sched_domain
4210 * for the given cpu.
4212 * Returns the lowest sched_domain of a cpu which contains the given flag.
4214 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4216 struct sched_domain *sd;
4218 for_each_domain(cpu, sd)
4219 if (sd->flags & flag)
4226 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4227 * @cpu: The cpu whose domains we're iterating over.
4228 * @sd: variable holding the value of the power_savings_sd
4230 * @flag: The flag to filter the sched_domains to be iterated.
4232 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4233 * set, starting from the lowest sched_domain to the highest.
4235 #define for_each_flag_domain(cpu, sd, flag) \
4236 for (sd = lowest_flag_domain(cpu, flag); \
4237 (sd && (sd->flags & flag)); sd = sd->parent)
4240 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4241 * @ilb_group: group to be checked for semi-idleness
4243 * Returns: 1 if the group is semi-idle. 0 otherwise.
4245 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4246 * and atleast one non-idle CPU. This helper function checks if the given
4247 * sched_group is semi-idle or not.
4249 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4251 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
4252 sched_group_cpus(ilb_group));
4255 * A sched_group is semi-idle when it has atleast one busy cpu
4256 * and atleast one idle cpu.
4258 if (cpumask_empty(nohz.grp_idle_mask))
4261 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
4267 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4268 * @cpu: The cpu which is nominating a new idle_load_balancer.
4270 * Returns: Returns the id of the idle load balancer if it exists,
4271 * Else, returns >= nr_cpu_ids.
4273 * This algorithm picks the idle load balancer such that it belongs to a
4274 * semi-idle powersavings sched_domain. The idea is to try and avoid
4275 * completely idle packages/cores just for the purpose of idle load balancing
4276 * when there are other idle cpu's which are better suited for that job.
4278 static int find_new_ilb(int cpu)
4280 struct sched_domain *sd;
4281 struct sched_group *ilb_group;
4282 int ilb = nr_cpu_ids;
4285 * Have idle load balancer selection from semi-idle packages only
4286 * when power-aware load balancing is enabled
4288 if (!(sched_smt_power_savings || sched_mc_power_savings))
4292 * Optimize for the case when we have no idle CPUs or only one
4293 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4295 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4299 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4300 ilb_group = sd->groups;
4303 if (is_semi_idle_group(ilb_group)) {
4304 ilb = cpumask_first(nohz.grp_idle_mask);
4308 ilb_group = ilb_group->next;
4310 } while (ilb_group != sd->groups);
4318 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4319 static inline int find_new_ilb(int call_cpu)
4326 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4327 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4328 * CPU (if there is one).
4330 static void nohz_balancer_kick(int cpu)
4334 nohz.next_balance++;
4336 ilb_cpu = get_nohz_load_balancer();
4338 if (ilb_cpu >= nr_cpu_ids) {
4339 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4340 if (ilb_cpu >= nr_cpu_ids)
4344 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4345 struct call_single_data *cp;
4347 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4348 cp = &per_cpu(remote_sched_softirq_cb, cpu);
4349 __smp_call_function_single(ilb_cpu, cp, 0);
4355 * This routine will try to nominate the ilb (idle load balancing)
4356 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4357 * load balancing on behalf of all those cpus.
4359 * When the ilb owner becomes busy, we will not have new ilb owner until some
4360 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4361 * idle load balancing by kicking one of the idle CPUs.
4363 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4364 * ilb owner CPU in future (when there is a need for idle load balancing on
4365 * behalf of all idle CPUs).
4367 void select_nohz_load_balancer(int stop_tick)
4369 int cpu = smp_processor_id();
4372 if (!cpu_active(cpu)) {
4373 if (atomic_read(&nohz.load_balancer) != cpu)
4377 * If we are going offline and still the leader,
4380 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4387 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4389 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4390 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4391 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4392 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4394 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4397 /* make me the ilb owner */
4398 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4403 * Check to see if there is a more power-efficient
4406 new_ilb = find_new_ilb(cpu);
4407 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4408 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4409 resched_cpu(new_ilb);
4415 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4418 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4420 if (atomic_read(&nohz.load_balancer) == cpu)
4421 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4429 static DEFINE_SPINLOCK(balancing);
4431 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4434 * Scale the max load_balance interval with the number of CPUs in the system.
4435 * This trades load-balance latency on larger machines for less cross talk.
4437 static void update_max_interval(void)
4439 max_load_balance_interval = HZ*num_online_cpus()/10;
4443 * It checks each scheduling domain to see if it is due to be balanced,
4444 * and initiates a balancing operation if so.
4446 * Balancing parameters are set up in arch_init_sched_domains.
4448 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4451 struct rq *rq = cpu_rq(cpu);
4452 unsigned long interval;
4453 struct sched_domain *sd;
4454 /* Earliest time when we have to do rebalance again */
4455 unsigned long next_balance = jiffies + 60*HZ;
4456 int update_next_balance = 0;
4462 for_each_domain(cpu, sd) {
4463 if (!(sd->flags & SD_LOAD_BALANCE))
4466 interval = sd->balance_interval;
4467 if (idle != CPU_IDLE)
4468 interval *= sd->busy_factor;
4470 /* scale ms to jiffies */
4471 interval = msecs_to_jiffies(interval);
4472 interval = clamp(interval, 1UL, max_load_balance_interval);
4474 need_serialize = sd->flags & SD_SERIALIZE;
4476 if (need_serialize) {
4477 if (!spin_trylock(&balancing))
4481 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4482 if (load_balance(cpu, rq, sd, idle, &balance)) {
4484 * We've pulled tasks over so either we're no
4487 idle = CPU_NOT_IDLE;
4489 sd->last_balance = jiffies;
4492 spin_unlock(&balancing);
4494 if (time_after(next_balance, sd->last_balance + interval)) {
4495 next_balance = sd->last_balance + interval;
4496 update_next_balance = 1;
4500 * Stop the load balance at this level. There is another
4501 * CPU in our sched group which is doing load balancing more
4510 * next_balance will be updated only when there is a need.
4511 * When the cpu is attached to null domain for ex, it will not be
4514 if (likely(update_next_balance))
4515 rq->next_balance = next_balance;
4520 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4521 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4523 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4525 struct rq *this_rq = cpu_rq(this_cpu);
4529 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
4532 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4533 if (balance_cpu == this_cpu)
4537 * If this cpu gets work to do, stop the load balancing
4538 * work being done for other cpus. Next load
4539 * balancing owner will pick it up.
4541 if (need_resched()) {
4542 this_rq->nohz_balance_kick = 0;
4546 raw_spin_lock_irq(&this_rq->lock);
4547 update_rq_clock(this_rq);
4548 update_cpu_load(this_rq);
4549 raw_spin_unlock_irq(&this_rq->lock);
4551 rebalance_domains(balance_cpu, CPU_IDLE);
4553 rq = cpu_rq(balance_cpu);
4554 if (time_after(this_rq->next_balance, rq->next_balance))
4555 this_rq->next_balance = rq->next_balance;
4557 nohz.next_balance = this_rq->next_balance;
4558 this_rq->nohz_balance_kick = 0;
4562 * Current heuristic for kicking the idle load balancer
4563 * - first_pick_cpu is the one of the busy CPUs. It will kick
4564 * idle load balancer when it has more than one process active. This
4565 * eliminates the need for idle load balancing altogether when we have
4566 * only one running process in the system (common case).
4567 * - If there are more than one busy CPU, idle load balancer may have
4568 * to run for active_load_balance to happen (i.e., two busy CPUs are
4569 * SMT or core siblings and can run better if they move to different
4570 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4571 * which will kick idle load balancer as soon as it has any load.
4573 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4575 unsigned long now = jiffies;
4577 int first_pick_cpu, second_pick_cpu;
4579 if (time_before(now, nohz.next_balance))
4582 if (rq->idle_at_tick)
4585 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4586 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4588 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4589 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4592 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4593 if (ret == nr_cpu_ids || ret == cpu) {
4594 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4595 if (rq->nr_running > 1)
4598 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4599 if (ret == nr_cpu_ids || ret == cpu) {
4607 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4611 * run_rebalance_domains is triggered when needed from the scheduler tick.
4612 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4614 static void run_rebalance_domains(struct softirq_action *h)
4616 int this_cpu = smp_processor_id();
4617 struct rq *this_rq = cpu_rq(this_cpu);
4618 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4619 CPU_IDLE : CPU_NOT_IDLE;
4621 rebalance_domains(this_cpu, idle);
4624 * If this cpu has a pending nohz_balance_kick, then do the
4625 * balancing on behalf of the other idle cpus whose ticks are
4628 nohz_idle_balance(this_cpu, idle);
4631 static inline int on_null_domain(int cpu)
4633 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4637 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4639 static inline void trigger_load_balance(struct rq *rq, int cpu)
4641 /* Don't need to rebalance while attached to NULL domain */
4642 if (time_after_eq(jiffies, rq->next_balance) &&
4643 likely(!on_null_domain(cpu)))
4644 raise_softirq(SCHED_SOFTIRQ);
4646 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4647 nohz_balancer_kick(cpu);
4651 static void rq_online_fair(struct rq *rq)
4656 static void rq_offline_fair(struct rq *rq)
4661 #else /* CONFIG_SMP */
4664 * on UP we do not need to balance between CPUs:
4666 static inline void idle_balance(int cpu, struct rq *rq)
4670 #endif /* CONFIG_SMP */
4673 * scheduler tick hitting a task of our scheduling class:
4675 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4677 struct cfs_rq *cfs_rq;
4678 struct sched_entity *se = &curr->se;
4680 for_each_sched_entity(se) {
4681 cfs_rq = cfs_rq_of(se);
4682 entity_tick(cfs_rq, se, queued);
4687 * called on fork with the child task as argument from the parent's context
4688 * - child not yet on the tasklist
4689 * - preemption disabled
4691 static void task_fork_fair(struct task_struct *p)
4693 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4694 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4695 int this_cpu = smp_processor_id();
4696 struct rq *rq = this_rq();
4697 unsigned long flags;
4699 raw_spin_lock_irqsave(&rq->lock, flags);
4701 update_rq_clock(rq);
4703 if (unlikely(task_cpu(p) != this_cpu)) {
4705 __set_task_cpu(p, this_cpu);
4709 update_curr(cfs_rq);
4712 se->vruntime = curr->vruntime;
4713 place_entity(cfs_rq, se, 1);
4715 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4717 * Upon rescheduling, sched_class::put_prev_task() will place
4718 * 'current' within the tree based on its new key value.
4720 swap(curr->vruntime, se->vruntime);
4721 resched_task(rq->curr);
4724 se->vruntime -= cfs_rq->min_vruntime;
4726 raw_spin_unlock_irqrestore(&rq->lock, flags);
4730 * Priority of the task has changed. Check to see if we preempt
4734 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4740 * Reschedule if we are currently running on this runqueue and
4741 * our priority decreased, or if we are not currently running on
4742 * this runqueue and our priority is higher than the current's
4744 if (rq->curr == p) {
4745 if (p->prio > oldprio)
4746 resched_task(rq->curr);
4748 check_preempt_curr(rq, p, 0);
4751 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4753 struct sched_entity *se = &p->se;
4754 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4757 * Ensure the task's vruntime is normalized, so that when its
4758 * switched back to the fair class the enqueue_entity(.flags=0) will
4759 * do the right thing.
4761 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4762 * have normalized the vruntime, if it was !on_rq, then only when
4763 * the task is sleeping will it still have non-normalized vruntime.
4765 if (!se->on_rq && p->state != TASK_RUNNING) {
4767 * Fix up our vruntime so that the current sleep doesn't
4768 * cause 'unlimited' sleep bonus.
4770 place_entity(cfs_rq, se, 0);
4771 se->vruntime -= cfs_rq->min_vruntime;
4776 * We switched to the sched_fair class.
4778 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4784 * We were most likely switched from sched_rt, so
4785 * kick off the schedule if running, otherwise just see
4786 * if we can still preempt the current task.
4789 resched_task(rq->curr);
4791 check_preempt_curr(rq, p, 0);
4794 /* Account for a task changing its policy or group.
4796 * This routine is mostly called to set cfs_rq->curr field when a task
4797 * migrates between groups/classes.
4799 static void set_curr_task_fair(struct rq *rq)
4801 struct sched_entity *se = &rq->curr->se;
4803 for_each_sched_entity(se) {
4804 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4806 set_next_entity(cfs_rq, se);
4807 /* ensure bandwidth has been allocated on our new cfs_rq */
4808 account_cfs_rq_runtime(cfs_rq, 0);
4812 #ifdef CONFIG_FAIR_GROUP_SCHED
4813 static void task_move_group_fair(struct task_struct *p, int on_rq)
4816 * If the task was not on the rq at the time of this cgroup movement
4817 * it must have been asleep, sleeping tasks keep their ->vruntime
4818 * absolute on their old rq until wakeup (needed for the fair sleeper
4819 * bonus in place_entity()).
4821 * If it was on the rq, we've just 'preempted' it, which does convert
4822 * ->vruntime to a relative base.
4824 * Make sure both cases convert their relative position when migrating
4825 * to another cgroup's rq. This does somewhat interfere with the
4826 * fair sleeper stuff for the first placement, but who cares.
4829 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4830 set_task_rq(p, task_cpu(p));
4832 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4836 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4838 struct sched_entity *se = &task->se;
4839 unsigned int rr_interval = 0;
4842 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4845 if (rq->cfs.load.weight)
4846 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4852 * All the scheduling class methods:
4854 static const struct sched_class fair_sched_class = {
4855 .next = &idle_sched_class,
4856 .enqueue_task = enqueue_task_fair,
4857 .dequeue_task = dequeue_task_fair,
4858 .yield_task = yield_task_fair,
4859 .yield_to_task = yield_to_task_fair,
4861 .check_preempt_curr = check_preempt_wakeup,
4863 .pick_next_task = pick_next_task_fair,
4864 .put_prev_task = put_prev_task_fair,
4867 .select_task_rq = select_task_rq_fair,
4869 .rq_online = rq_online_fair,
4870 .rq_offline = rq_offline_fair,
4872 .task_waking = task_waking_fair,
4875 .set_curr_task = set_curr_task_fair,
4876 .task_tick = task_tick_fair,
4877 .task_fork = task_fork_fair,
4879 .prio_changed = prio_changed_fair,
4880 .switched_from = switched_from_fair,
4881 .switched_to = switched_to_fair,
4883 .get_rr_interval = get_rr_interval_fair,
4885 #ifdef CONFIG_FAIR_GROUP_SCHED
4886 .task_move_group = task_move_group_fair,
4890 #ifdef CONFIG_SCHED_DEBUG
4891 static void print_cfs_stats(struct seq_file *m, int cpu)
4893 struct cfs_rq *cfs_rq;
4896 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4897 print_cfs_rq(m, cpu, cfs_rq);