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>
26 #include <linux/slab.h>
27 #include <linux/profile.h>
28 #include <linux/interrupt.h>
30 #include <trace/events/sched.h>
35 * Targeted preemption latency for CPU-bound tasks:
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 * NOTE: this latency value is not the same as the concept of
39 * 'timeslice length' - timeslices in CFS are of variable length
40 * and have no persistent notion like in traditional, time-slice
41 * based scheduling concepts.
43 * (to see the precise effective timeslice length of your workload,
44 * run vmstat and monitor the context-switches (cs) field)
46 unsigned int sysctl_sched_latency = 6000000ULL;
47 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
50 * The initial- and re-scaling of tunables is configurable
51 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
55 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
56 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
58 enum sched_tunable_scaling sysctl_sched_tunable_scaling
59 = SCHED_TUNABLESCALING_LOG;
62 * Minimal preemption granularity for CPU-bound tasks:
63 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
65 unsigned int sysctl_sched_min_granularity = 750000ULL;
66 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
69 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
71 static unsigned int sched_nr_latency = 8;
74 * After fork, child runs first. If set to 0 (default) then
75 * parent will (try to) run first.
77 unsigned int sysctl_sched_child_runs_first __read_mostly;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
99 #ifdef CONFIG_CFS_BANDWIDTH
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
108 * default: 5 msec, units: microseconds
110 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114 * Increase the granularity value when there are more CPUs,
115 * because with more CPUs the 'effective latency' as visible
116 * to users decreases. But the relationship is not linear,
117 * so pick a second-best guess by going with the log2 of the
120 * This idea comes from the SD scheduler of Con Kolivas:
122 static int get_update_sysctl_factor(void)
124 unsigned int cpus = min_t(int, num_online_cpus(), 8);
127 switch (sysctl_sched_tunable_scaling) {
128 case SCHED_TUNABLESCALING_NONE:
131 case SCHED_TUNABLESCALING_LINEAR:
134 case SCHED_TUNABLESCALING_LOG:
136 factor = 1 + ilog2(cpus);
143 static void update_sysctl(void)
145 unsigned int factor = get_update_sysctl_factor();
147 #define SET_SYSCTL(name) \
148 (sysctl_##name = (factor) * normalized_sysctl_##name)
149 SET_SYSCTL(sched_min_granularity);
150 SET_SYSCTL(sched_latency);
151 SET_SYSCTL(sched_wakeup_granularity);
155 void sched_init_granularity(void)
160 #if BITS_PER_LONG == 32
161 # define WMULT_CONST (~0UL)
163 # define WMULT_CONST (1UL << 32)
166 #define WMULT_SHIFT 32
169 * Shift right and round:
171 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
174 * delta *= weight / lw
177 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
178 struct load_weight *lw)
183 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
184 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
185 * 2^SCHED_LOAD_RESOLUTION.
187 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
188 tmp = (u64)delta_exec * scale_load_down(weight);
190 tmp = (u64)delta_exec;
192 if (!lw->inv_weight) {
193 unsigned long w = scale_load_down(lw->weight);
195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
197 else if (unlikely(!w))
198 lw->inv_weight = WMULT_CONST;
200 lw->inv_weight = WMULT_CONST / w;
204 * Check whether we'd overflow the 64-bit multiplication:
206 if (unlikely(tmp > WMULT_CONST))
207 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
210 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
212 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
216 const struct sched_class fair_sched_class;
218 /**************************************************************
219 * CFS operations on generic schedulable entities:
222 #ifdef CONFIG_FAIR_GROUP_SCHED
224 /* cpu runqueue to which this cfs_rq is attached */
225 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
230 /* An entity is a task if it doesn't "own" a runqueue */
231 #define entity_is_task(se) (!se->my_q)
233 static inline struct task_struct *task_of(struct sched_entity *se)
235 #ifdef CONFIG_SCHED_DEBUG
236 WARN_ON_ONCE(!entity_is_task(se));
238 return container_of(se, struct task_struct, se);
241 /* Walk up scheduling entities hierarchy */
242 #define for_each_sched_entity(se) \
243 for (; se; se = se->parent)
245 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
250 /* runqueue on which this entity is (to be) queued */
251 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
256 /* runqueue "owned" by this group */
257 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
262 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
265 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
267 if (!cfs_rq->on_list) {
269 * Ensure we either appear before our parent (if already
270 * enqueued) or force our parent to appear after us when it is
271 * enqueued. The fact that we always enqueue bottom-up
272 * reduces this to two cases.
274 if (cfs_rq->tg->parent &&
275 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
276 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
277 &rq_of(cfs_rq)->leaf_cfs_rq_list);
279 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
280 &rq_of(cfs_rq)->leaf_cfs_rq_list);
284 /* We should have no load, but we need to update last_decay. */
285 update_cfs_rq_blocked_load(cfs_rq, 0);
289 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
291 if (cfs_rq->on_list) {
292 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
297 /* Iterate thr' all leaf cfs_rq's on a runqueue */
298 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
299 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
301 /* Do the two (enqueued) entities belong to the same group ? */
303 is_same_group(struct sched_entity *se, struct sched_entity *pse)
305 if (se->cfs_rq == pse->cfs_rq)
311 static inline struct sched_entity *parent_entity(struct sched_entity *se)
316 /* return depth at which a sched entity is present in the hierarchy */
317 static inline int depth_se(struct sched_entity *se)
321 for_each_sched_entity(se)
328 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
330 int se_depth, pse_depth;
333 * preemption test can be made between sibling entities who are in the
334 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
335 * both tasks until we find their ancestors who are siblings of common
339 /* First walk up until both entities are at same depth */
340 se_depth = depth_se(*se);
341 pse_depth = depth_se(*pse);
343 while (se_depth > pse_depth) {
345 *se = parent_entity(*se);
348 while (pse_depth > se_depth) {
350 *pse = parent_entity(*pse);
353 while (!is_same_group(*se, *pse)) {
354 *se = parent_entity(*se);
355 *pse = parent_entity(*pse);
359 #else /* !CONFIG_FAIR_GROUP_SCHED */
361 static inline struct task_struct *task_of(struct sched_entity *se)
363 return container_of(se, struct task_struct, se);
366 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
368 return container_of(cfs_rq, struct rq, cfs);
371 #define entity_is_task(se) 1
373 #define for_each_sched_entity(se) \
374 for (; se; se = NULL)
376 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
378 return &task_rq(p)->cfs;
381 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
383 struct task_struct *p = task_of(se);
384 struct rq *rq = task_rq(p);
389 /* runqueue "owned" by this group */
390 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
395 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
399 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
403 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
404 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
407 is_same_group(struct sched_entity *se, struct sched_entity *pse)
412 static inline struct sched_entity *parent_entity(struct sched_entity *se)
418 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
422 #endif /* CONFIG_FAIR_GROUP_SCHED */
424 static __always_inline
425 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec);
427 /**************************************************************
428 * Scheduling class tree data structure manipulation methods:
431 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
433 s64 delta = (s64)(vruntime - min_vruntime);
435 min_vruntime = vruntime;
440 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
442 s64 delta = (s64)(vruntime - min_vruntime);
444 min_vruntime = vruntime;
449 static inline int entity_before(struct sched_entity *a,
450 struct sched_entity *b)
452 return (s64)(a->vruntime - b->vruntime) < 0;
455 static void update_min_vruntime(struct cfs_rq *cfs_rq)
457 u64 vruntime = cfs_rq->min_vruntime;
460 vruntime = cfs_rq->curr->vruntime;
462 if (cfs_rq->rb_leftmost) {
463 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
468 vruntime = se->vruntime;
470 vruntime = min_vruntime(vruntime, se->vruntime);
473 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
476 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
481 * Enqueue an entity into the rb-tree:
483 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
485 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
486 struct rb_node *parent = NULL;
487 struct sched_entity *entry;
491 * Find the right place in the rbtree:
495 entry = rb_entry(parent, struct sched_entity, run_node);
497 * We dont care about collisions. Nodes with
498 * the same key stay together.
500 if (entity_before(se, entry)) {
501 link = &parent->rb_left;
503 link = &parent->rb_right;
509 * Maintain a cache of leftmost tree entries (it is frequently
513 cfs_rq->rb_leftmost = &se->run_node;
515 rb_link_node(&se->run_node, parent, link);
516 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
519 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
521 if (cfs_rq->rb_leftmost == &se->run_node) {
522 struct rb_node *next_node;
524 next_node = rb_next(&se->run_node);
525 cfs_rq->rb_leftmost = next_node;
528 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
531 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
533 struct rb_node *left = cfs_rq->rb_leftmost;
538 return rb_entry(left, struct sched_entity, run_node);
541 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
543 struct rb_node *next = rb_next(&se->run_node);
548 return rb_entry(next, struct sched_entity, run_node);
551 #ifdef CONFIG_SCHED_DEBUG
552 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
554 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
559 return rb_entry(last, struct sched_entity, run_node);
562 /**************************************************************
563 * Scheduling class statistics methods:
566 int sched_proc_update_handler(struct ctl_table *table, int write,
567 void __user *buffer, size_t *lenp,
570 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
571 int factor = get_update_sysctl_factor();
576 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
577 sysctl_sched_min_granularity);
579 #define WRT_SYSCTL(name) \
580 (normalized_sysctl_##name = sysctl_##name / (factor))
581 WRT_SYSCTL(sched_min_granularity);
582 WRT_SYSCTL(sched_latency);
583 WRT_SYSCTL(sched_wakeup_granularity);
593 static inline unsigned long
594 calc_delta_fair(unsigned long delta, struct sched_entity *se)
596 if (unlikely(se->load.weight != NICE_0_LOAD))
597 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
603 * The idea is to set a period in which each task runs once.
605 * When there are too many tasks (sched_nr_latency) we have to stretch
606 * this period because otherwise the slices get too small.
608 * p = (nr <= nl) ? l : l*nr/nl
610 static u64 __sched_period(unsigned long nr_running)
612 u64 period = sysctl_sched_latency;
613 unsigned long nr_latency = sched_nr_latency;
615 if (unlikely(nr_running > nr_latency)) {
616 period = sysctl_sched_min_granularity;
617 period *= nr_running;
624 * We calculate the wall-time slice from the period by taking a part
625 * proportional to the weight.
629 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
631 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
633 for_each_sched_entity(se) {
634 struct load_weight *load;
635 struct load_weight lw;
637 cfs_rq = cfs_rq_of(se);
638 load = &cfs_rq->load;
640 if (unlikely(!se->on_rq)) {
643 update_load_add(&lw, se->load.weight);
646 slice = calc_delta_mine(slice, se->load.weight, load);
652 * We calculate the vruntime slice of a to be inserted task
656 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
658 return calc_delta_fair(sched_slice(cfs_rq, se), se);
662 * Update the current task's runtime statistics. Skip current tasks that
663 * are not in our scheduling class.
666 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
667 unsigned long delta_exec)
669 unsigned long delta_exec_weighted;
671 schedstat_set(curr->statistics.exec_max,
672 max((u64)delta_exec, curr->statistics.exec_max));
674 curr->sum_exec_runtime += delta_exec;
675 schedstat_add(cfs_rq, exec_clock, delta_exec);
676 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
678 curr->vruntime += delta_exec_weighted;
679 update_min_vruntime(cfs_rq);
682 static void update_curr(struct cfs_rq *cfs_rq)
684 struct sched_entity *curr = cfs_rq->curr;
685 u64 now = rq_of(cfs_rq)->clock_task;
686 unsigned long delta_exec;
692 * Get the amount of time the current task was running
693 * since the last time we changed load (this cannot
694 * overflow on 32 bits):
696 delta_exec = (unsigned long)(now - curr->exec_start);
700 __update_curr(cfs_rq, curr, delta_exec);
701 curr->exec_start = now;
703 if (entity_is_task(curr)) {
704 struct task_struct *curtask = task_of(curr);
706 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
707 cpuacct_charge(curtask, delta_exec);
708 account_group_exec_runtime(curtask, delta_exec);
711 account_cfs_rq_runtime(cfs_rq, delta_exec);
715 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
717 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
721 * Task is being enqueued - update stats:
723 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
726 * Are we enqueueing a waiting task? (for current tasks
727 * a dequeue/enqueue event is a NOP)
729 if (se != cfs_rq->curr)
730 update_stats_wait_start(cfs_rq, se);
734 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
736 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
737 rq_of(cfs_rq)->clock - se->statistics.wait_start));
738 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
739 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
740 rq_of(cfs_rq)->clock - se->statistics.wait_start);
741 #ifdef CONFIG_SCHEDSTATS
742 if (entity_is_task(se)) {
743 trace_sched_stat_wait(task_of(se),
744 rq_of(cfs_rq)->clock - se->statistics.wait_start);
747 schedstat_set(se->statistics.wait_start, 0);
751 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
754 * Mark the end of the wait period if dequeueing a
757 if (se != cfs_rq->curr)
758 update_stats_wait_end(cfs_rq, se);
762 * We are picking a new current task - update its stats:
765 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
768 * We are starting a new run period:
770 se->exec_start = rq_of(cfs_rq)->clock_task;
773 /**************************************************
774 * Scheduling class queueing methods:
778 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
780 update_load_add(&cfs_rq->load, se->load.weight);
781 if (!parent_entity(se))
782 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
784 if (entity_is_task(se))
785 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
787 cfs_rq->nr_running++;
791 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
793 update_load_sub(&cfs_rq->load, se->load.weight);
794 if (!parent_entity(se))
795 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
796 if (entity_is_task(se))
797 list_del_init(&se->group_node);
798 cfs_rq->nr_running--;
801 #ifdef CONFIG_FAIR_GROUP_SCHED
803 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
808 * Use this CPU's actual weight instead of the last load_contribution
809 * to gain a more accurate current total weight. See
810 * update_cfs_rq_load_contribution().
812 tg_weight = atomic64_read(&tg->load_avg);
813 tg_weight -= cfs_rq->tg_load_contrib;
814 tg_weight += cfs_rq->load.weight;
819 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
821 long tg_weight, load, shares;
823 tg_weight = calc_tg_weight(tg, cfs_rq);
824 load = cfs_rq->load.weight;
826 shares = (tg->shares * load);
830 if (shares < MIN_SHARES)
832 if (shares > tg->shares)
837 # else /* CONFIG_SMP */
838 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
842 # endif /* CONFIG_SMP */
843 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
844 unsigned long weight)
847 /* commit outstanding execution time */
848 if (cfs_rq->curr == se)
850 account_entity_dequeue(cfs_rq, se);
853 update_load_set(&se->load, weight);
856 account_entity_enqueue(cfs_rq, se);
859 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
861 static void update_cfs_shares(struct cfs_rq *cfs_rq)
863 struct task_group *tg;
864 struct sched_entity *se;
868 se = tg->se[cpu_of(rq_of(cfs_rq))];
869 if (!se || throttled_hierarchy(cfs_rq))
872 if (likely(se->load.weight == tg->shares))
875 shares = calc_cfs_shares(cfs_rq, tg);
877 reweight_entity(cfs_rq_of(se), se, shares);
879 #else /* CONFIG_FAIR_GROUP_SCHED */
880 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
883 #endif /* CONFIG_FAIR_GROUP_SCHED */
885 /* Only depends on SMP, FAIR_GROUP_SCHED may be removed when useful in lb */
886 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
888 * We choose a half-life close to 1 scheduling period.
889 * Note: The tables below are dependent on this value.
891 #define LOAD_AVG_PERIOD 32
892 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
893 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_MAX_AVG */
895 /* Precomputed fixed inverse multiplies for multiplication by y^n */
896 static const u32 runnable_avg_yN_inv[] = {
897 0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
898 0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
899 0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
900 0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
901 0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
902 0x85aac367, 0x82cd8698,
906 * Precomputed \Sum y^k { 1<=k<=n }. These are floor(true_value) to prevent
907 * over-estimates when re-combining.
909 static const u32 runnable_avg_yN_sum[] = {
910 0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
911 9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
912 17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
917 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
919 static __always_inline u64 decay_load(u64 val, u64 n)
921 unsigned int local_n;
925 else if (unlikely(n > LOAD_AVG_PERIOD * 63))
928 /* after bounds checking we can collapse to 32-bit */
932 * As y^PERIOD = 1/2, we can combine
933 * y^n = 1/2^(n/PERIOD) * k^(n%PERIOD)
934 * With a look-up table which covers k^n (n<PERIOD)
936 * To achieve constant time decay_load.
938 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
939 val >>= local_n / LOAD_AVG_PERIOD;
940 local_n %= LOAD_AVG_PERIOD;
943 val *= runnable_avg_yN_inv[local_n];
944 /* We don't use SRR here since we always want to round down. */
949 * For updates fully spanning n periods, the contribution to runnable
950 * average will be: \Sum 1024*y^n
952 * We can compute this reasonably efficiently by combining:
953 * y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for n <PERIOD}
955 static u32 __compute_runnable_contrib(u64 n)
959 if (likely(n <= LOAD_AVG_PERIOD))
960 return runnable_avg_yN_sum[n];
961 else if (unlikely(n >= LOAD_AVG_MAX_N))
964 /* Compute \Sum k^n combining precomputed values for k^i, \Sum k^j */
966 contrib /= 2; /* y^LOAD_AVG_PERIOD = 1/2 */
967 contrib += runnable_avg_yN_sum[LOAD_AVG_PERIOD];
969 n -= LOAD_AVG_PERIOD;
970 } while (n > LOAD_AVG_PERIOD);
972 contrib = decay_load(contrib, n);
973 return contrib + runnable_avg_yN_sum[n];
977 * We can represent the historical contribution to runnable average as the
978 * coefficients of a geometric series. To do this we sub-divide our runnable
979 * history into segments of approximately 1ms (1024us); label the segment that
980 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
982 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
984 * (now) (~1ms ago) (~2ms ago)
986 * Let u_i denote the fraction of p_i that the entity was runnable.
988 * We then designate the fractions u_i as our co-efficients, yielding the
989 * following representation of historical load:
990 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
992 * We choose y based on the with of a reasonably scheduling period, fixing:
995 * This means that the contribution to load ~32ms ago (u_32) will be weighted
996 * approximately half as much as the contribution to load within the last ms
999 * When a period "rolls over" and we have new u_0`, multiplying the previous
1000 * sum again by y is sufficient to update:
1001 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
1002 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
1004 static __always_inline int __update_entity_runnable_avg(u64 now,
1005 struct sched_avg *sa,
1009 u32 runnable_contrib;
1010 int delta_w, decayed = 0;
1012 delta = now - sa->last_runnable_update;
1014 * This should only happen when time goes backwards, which it
1015 * unfortunately does during sched clock init when we swap over to TSC.
1017 if ((s64)delta < 0) {
1018 sa->last_runnable_update = now;
1023 * Use 1024ns as the unit of measurement since it's a reasonable
1024 * approximation of 1us and fast to compute.
1029 sa->last_runnable_update = now;
1031 /* delta_w is the amount already accumulated against our next period */
1032 delta_w = sa->runnable_avg_period % 1024;
1033 if (delta + delta_w >= 1024) {
1034 /* period roll-over */
1038 * Now that we know we're crossing a period boundary, figure
1039 * out how much from delta we need to complete the current
1040 * period and accrue it.
1042 delta_w = 1024 - delta_w;
1044 sa->runnable_avg_sum += delta_w;
1045 sa->runnable_avg_period += delta_w;
1049 /* Figure out how many additional periods this update spans */
1050 periods = delta / 1024;
1053 sa->runnable_avg_sum = decay_load(sa->runnable_avg_sum,
1055 sa->runnable_avg_period = decay_load(sa->runnable_avg_period,
1058 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
1059 runnable_contrib = __compute_runnable_contrib(periods);
1061 sa->runnable_avg_sum += runnable_contrib;
1062 sa->runnable_avg_period += runnable_contrib;
1065 /* Remainder of delta accrued against u_0` */
1067 sa->runnable_avg_sum += delta;
1068 sa->runnable_avg_period += delta;
1073 /* Synchronize an entity's decay with its parenting cfs_rq.*/
1074 static inline u64 __synchronize_entity_decay(struct sched_entity *se)
1076 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1077 u64 decays = atomic64_read(&cfs_rq->decay_counter);
1079 decays -= se->avg.decay_count;
1083 se->avg.load_avg_contrib = decay_load(se->avg.load_avg_contrib, decays);
1084 se->avg.decay_count = 0;
1089 #ifdef CONFIG_FAIR_GROUP_SCHED
1090 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1093 struct task_group *tg = cfs_rq->tg;
1096 tg_contrib = cfs_rq->runnable_load_avg + cfs_rq->blocked_load_avg;
1097 tg_contrib -= cfs_rq->tg_load_contrib;
1099 if (force_update || abs64(tg_contrib) > cfs_rq->tg_load_contrib / 8) {
1100 atomic64_add(tg_contrib, &tg->load_avg);
1101 cfs_rq->tg_load_contrib += tg_contrib;
1106 * Aggregate cfs_rq runnable averages into an equivalent task_group
1107 * representation for computing load contributions.
1109 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1110 struct cfs_rq *cfs_rq)
1112 struct task_group *tg = cfs_rq->tg;
1115 /* The fraction of a cpu used by this cfs_rq */
1116 contrib = div_u64(sa->runnable_avg_sum << NICE_0_SHIFT,
1117 sa->runnable_avg_period + 1);
1118 contrib -= cfs_rq->tg_runnable_contrib;
1120 if (abs(contrib) > cfs_rq->tg_runnable_contrib / 64) {
1121 atomic_add(contrib, &tg->runnable_avg);
1122 cfs_rq->tg_runnable_contrib += contrib;
1126 static inline void __update_group_entity_contrib(struct sched_entity *se)
1128 struct cfs_rq *cfs_rq = group_cfs_rq(se);
1129 struct task_group *tg = cfs_rq->tg;
1134 contrib = cfs_rq->tg_load_contrib * tg->shares;
1135 se->avg.load_avg_contrib = div64_u64(contrib,
1136 atomic64_read(&tg->load_avg) + 1);
1139 * For group entities we need to compute a correction term in the case
1140 * that they are consuming <1 cpu so that we would contribute the same
1141 * load as a task of equal weight.
1143 * Explicitly co-ordinating this measurement would be expensive, but
1144 * fortunately the sum of each cpus contribution forms a usable
1145 * lower-bound on the true value.
1147 * Consider the aggregate of 2 contributions. Either they are disjoint
1148 * (and the sum represents true value) or they are disjoint and we are
1149 * understating by the aggregate of their overlap.
1151 * Extending this to N cpus, for a given overlap, the maximum amount we
1152 * understand is then n_i(n_i+1)/2 * w_i where n_i is the number of
1153 * cpus that overlap for this interval and w_i is the interval width.
1155 * On a small machine; the first term is well-bounded which bounds the
1156 * total error since w_i is a subset of the period. Whereas on a
1157 * larger machine, while this first term can be larger, if w_i is the
1158 * of consequential size guaranteed to see n_i*w_i quickly converge to
1159 * our upper bound of 1-cpu.
1161 runnable_avg = atomic_read(&tg->runnable_avg);
1162 if (runnable_avg < NICE_0_LOAD) {
1163 se->avg.load_avg_contrib *= runnable_avg;
1164 se->avg.load_avg_contrib >>= NICE_0_SHIFT;
1168 static inline void __update_cfs_rq_tg_load_contrib(struct cfs_rq *cfs_rq,
1169 int force_update) {}
1170 static inline void __update_tg_runnable_avg(struct sched_avg *sa,
1171 struct cfs_rq *cfs_rq) {}
1172 static inline void __update_group_entity_contrib(struct sched_entity *se) {}
1175 static inline void __update_task_entity_contrib(struct sched_entity *se)
1179 /* avoid overflowing a 32-bit type w/ SCHED_LOAD_SCALE */
1180 contrib = se->avg.runnable_avg_sum * scale_load_down(se->load.weight);
1181 contrib /= (se->avg.runnable_avg_period + 1);
1182 se->avg.load_avg_contrib = scale_load(contrib);
1185 /* Compute the current contribution to load_avg by se, return any delta */
1186 static long __update_entity_load_avg_contrib(struct sched_entity *se)
1188 long old_contrib = se->avg.load_avg_contrib;
1190 if (entity_is_task(se)) {
1191 __update_task_entity_contrib(se);
1193 __update_tg_runnable_avg(&se->avg, group_cfs_rq(se));
1194 __update_group_entity_contrib(se);
1197 return se->avg.load_avg_contrib - old_contrib;
1200 static inline void subtract_blocked_load_contrib(struct cfs_rq *cfs_rq,
1203 if (likely(load_contrib < cfs_rq->blocked_load_avg))
1204 cfs_rq->blocked_load_avg -= load_contrib;
1206 cfs_rq->blocked_load_avg = 0;
1209 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
1211 /* Update a sched_entity's runnable average */
1212 static inline void update_entity_load_avg(struct sched_entity *se,
1215 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1220 * For a group entity we need to use their owned cfs_rq_clock_task() in
1221 * case they are the parent of a throttled hierarchy.
1223 if (entity_is_task(se))
1224 now = cfs_rq_clock_task(cfs_rq);
1226 now = cfs_rq_clock_task(group_cfs_rq(se));
1228 if (!__update_entity_runnable_avg(now, &se->avg, se->on_rq))
1231 contrib_delta = __update_entity_load_avg_contrib(se);
1237 cfs_rq->runnable_load_avg += contrib_delta;
1239 subtract_blocked_load_contrib(cfs_rq, -contrib_delta);
1243 * Decay the load contributed by all blocked children and account this so that
1244 * their contribution may appropriately discounted when they wake up.
1246 static void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq, int force_update)
1248 u64 now = cfs_rq_clock_task(cfs_rq) >> 20;
1251 decays = now - cfs_rq->last_decay;
1252 if (!decays && !force_update)
1255 if (atomic64_read(&cfs_rq->removed_load)) {
1256 u64 removed_load = atomic64_xchg(&cfs_rq->removed_load, 0);
1257 subtract_blocked_load_contrib(cfs_rq, removed_load);
1261 cfs_rq->blocked_load_avg = decay_load(cfs_rq->blocked_load_avg,
1263 atomic64_add(decays, &cfs_rq->decay_counter);
1264 cfs_rq->last_decay = now;
1267 __update_cfs_rq_tg_load_contrib(cfs_rq, force_update);
1268 update_cfs_shares(cfs_rq);
1271 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1273 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable);
1274 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1277 /* Add the load generated by se into cfs_rq's child load-average */
1278 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1279 struct sched_entity *se,
1283 * We track migrations using entity decay_count <= 0, on a wake-up
1284 * migration we use a negative decay count to track the remote decays
1285 * accumulated while sleeping.
1287 if (unlikely(se->avg.decay_count <= 0)) {
1288 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1289 if (se->avg.decay_count) {
1291 * In a wake-up migration we have to approximate the
1292 * time sleeping. This is because we can't synchronize
1293 * clock_task between the two cpus, and it is not
1294 * guaranteed to be read-safe. Instead, we can
1295 * approximate this using our carried decays, which are
1296 * explicitly atomically readable.
1298 se->avg.last_runnable_update -= (-se->avg.decay_count)
1300 update_entity_load_avg(se, 0);
1301 /* Indicate that we're now synchronized and on-rq */
1302 se->avg.decay_count = 0;
1306 __synchronize_entity_decay(se);
1309 /* migrated tasks did not contribute to our blocked load */
1311 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1312 update_entity_load_avg(se, 0);
1315 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1316 /* we force update consideration on load-balancer moves */
1317 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1321 * Remove se's load from this cfs_rq child load-average, if the entity is
1322 * transitioning to a blocked state we track its projected decay using
1325 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1326 struct sched_entity *se,
1329 update_entity_load_avg(se, 1);
1330 /* we force update consideration on load-balancer moves */
1331 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1333 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1335 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1336 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1337 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1340 static inline void update_entity_load_avg(struct sched_entity *se,
1341 int update_cfs_rq) {}
1342 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1343 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1344 struct sched_entity *se,
1346 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1347 struct sched_entity *se,
1349 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1350 int force_update) {}
1353 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1355 #ifdef CONFIG_SCHEDSTATS
1356 struct task_struct *tsk = NULL;
1358 if (entity_is_task(se))
1361 if (se->statistics.sleep_start) {
1362 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1367 if (unlikely(delta > se->statistics.sleep_max))
1368 se->statistics.sleep_max = delta;
1370 se->statistics.sleep_start = 0;
1371 se->statistics.sum_sleep_runtime += delta;
1374 account_scheduler_latency(tsk, delta >> 10, 1);
1375 trace_sched_stat_sleep(tsk, delta);
1378 if (se->statistics.block_start) {
1379 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1384 if (unlikely(delta > se->statistics.block_max))
1385 se->statistics.block_max = delta;
1387 se->statistics.block_start = 0;
1388 se->statistics.sum_sleep_runtime += delta;
1391 if (tsk->in_iowait) {
1392 se->statistics.iowait_sum += delta;
1393 se->statistics.iowait_count++;
1394 trace_sched_stat_iowait(tsk, delta);
1397 trace_sched_stat_blocked(tsk, delta);
1400 * Blocking time is in units of nanosecs, so shift by
1401 * 20 to get a milliseconds-range estimation of the
1402 * amount of time that the task spent sleeping:
1404 if (unlikely(prof_on == SLEEP_PROFILING)) {
1405 profile_hits(SLEEP_PROFILING,
1406 (void *)get_wchan(tsk),
1409 account_scheduler_latency(tsk, delta >> 10, 0);
1415 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1417 #ifdef CONFIG_SCHED_DEBUG
1418 s64 d = se->vruntime - cfs_rq->min_vruntime;
1423 if (d > 3*sysctl_sched_latency)
1424 schedstat_inc(cfs_rq, nr_spread_over);
1429 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1431 u64 vruntime = cfs_rq->min_vruntime;
1434 * The 'current' period is already promised to the current tasks,
1435 * however the extra weight of the new task will slow them down a
1436 * little, place the new task so that it fits in the slot that
1437 * stays open at the end.
1439 if (initial && sched_feat(START_DEBIT))
1440 vruntime += sched_vslice(cfs_rq, se);
1442 /* sleeps up to a single latency don't count. */
1444 unsigned long thresh = sysctl_sched_latency;
1447 * Halve their sleep time's effect, to allow
1448 * for a gentler effect of sleepers:
1450 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1456 /* ensure we never gain time by being placed backwards. */
1457 vruntime = max_vruntime(se->vruntime, vruntime);
1459 se->vruntime = vruntime;
1462 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1465 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1468 * Update the normalized vruntime before updating min_vruntime
1469 * through callig update_curr().
1471 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1472 se->vruntime += cfs_rq->min_vruntime;
1475 * Update run-time statistics of the 'current'.
1477 update_curr(cfs_rq);
1478 account_entity_enqueue(cfs_rq, se);
1479 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1481 if (flags & ENQUEUE_WAKEUP) {
1482 place_entity(cfs_rq, se, 0);
1483 enqueue_sleeper(cfs_rq, se);
1486 update_stats_enqueue(cfs_rq, se);
1487 check_spread(cfs_rq, se);
1488 if (se != cfs_rq->curr)
1489 __enqueue_entity(cfs_rq, se);
1492 if (cfs_rq->nr_running == 1) {
1493 list_add_leaf_cfs_rq(cfs_rq);
1494 check_enqueue_throttle(cfs_rq);
1498 static void __clear_buddies_last(struct sched_entity *se)
1500 for_each_sched_entity(se) {
1501 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1502 if (cfs_rq->last == se)
1503 cfs_rq->last = NULL;
1509 static void __clear_buddies_next(struct sched_entity *se)
1511 for_each_sched_entity(se) {
1512 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1513 if (cfs_rq->next == se)
1514 cfs_rq->next = NULL;
1520 static void __clear_buddies_skip(struct sched_entity *se)
1522 for_each_sched_entity(se) {
1523 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1524 if (cfs_rq->skip == se)
1525 cfs_rq->skip = NULL;
1531 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1533 if (cfs_rq->last == se)
1534 __clear_buddies_last(se);
1536 if (cfs_rq->next == se)
1537 __clear_buddies_next(se);
1539 if (cfs_rq->skip == se)
1540 __clear_buddies_skip(se);
1543 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1546 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1549 * Update run-time statistics of the 'current'.
1551 update_curr(cfs_rq);
1553 update_stats_dequeue(cfs_rq, se);
1554 if (flags & DEQUEUE_SLEEP) {
1555 #ifdef CONFIG_SCHEDSTATS
1556 if (entity_is_task(se)) {
1557 struct task_struct *tsk = task_of(se);
1559 if (tsk->state & TASK_INTERRUPTIBLE)
1560 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1561 if (tsk->state & TASK_UNINTERRUPTIBLE)
1562 se->statistics.block_start = rq_of(cfs_rq)->clock;
1567 clear_buddies(cfs_rq, se);
1569 if (se != cfs_rq->curr)
1570 __dequeue_entity(cfs_rq, se);
1571 account_entity_dequeue(cfs_rq, se);
1572 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1575 * Normalize the entity after updating the min_vruntime because the
1576 * update can refer to the ->curr item and we need to reflect this
1577 * movement in our normalized position.
1579 if (!(flags & DEQUEUE_SLEEP))
1580 se->vruntime -= cfs_rq->min_vruntime;
1582 /* return excess runtime on last dequeue */
1583 return_cfs_rq_runtime(cfs_rq);
1585 update_min_vruntime(cfs_rq);
1590 * Preempt the current task with a newly woken task if needed:
1593 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1595 unsigned long ideal_runtime, delta_exec;
1596 struct sched_entity *se;
1599 ideal_runtime = sched_slice(cfs_rq, curr);
1600 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1601 if (delta_exec > ideal_runtime) {
1602 resched_task(rq_of(cfs_rq)->curr);
1604 * The current task ran long enough, ensure it doesn't get
1605 * re-elected due to buddy favours.
1607 clear_buddies(cfs_rq, curr);
1612 * Ensure that a task that missed wakeup preemption by a
1613 * narrow margin doesn't have to wait for a full slice.
1614 * This also mitigates buddy induced latencies under load.
1616 if (delta_exec < sysctl_sched_min_granularity)
1619 se = __pick_first_entity(cfs_rq);
1620 delta = curr->vruntime - se->vruntime;
1625 if (delta > ideal_runtime)
1626 resched_task(rq_of(cfs_rq)->curr);
1630 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1632 /* 'current' is not kept within the tree. */
1635 * Any task has to be enqueued before it get to execute on
1636 * a CPU. So account for the time it spent waiting on the
1639 update_stats_wait_end(cfs_rq, se);
1640 __dequeue_entity(cfs_rq, se);
1643 update_stats_curr_start(cfs_rq, se);
1645 #ifdef CONFIG_SCHEDSTATS
1647 * Track our maximum slice length, if the CPU's load is at
1648 * least twice that of our own weight (i.e. dont track it
1649 * when there are only lesser-weight tasks around):
1651 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1652 se->statistics.slice_max = max(se->statistics.slice_max,
1653 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1656 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1660 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1663 * Pick the next process, keeping these things in mind, in this order:
1664 * 1) keep things fair between processes/task groups
1665 * 2) pick the "next" process, since someone really wants that to run
1666 * 3) pick the "last" process, for cache locality
1667 * 4) do not run the "skip" process, if something else is available
1669 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1671 struct sched_entity *se = __pick_first_entity(cfs_rq);
1672 struct sched_entity *left = se;
1675 * Avoid running the skip buddy, if running something else can
1676 * be done without getting too unfair.
1678 if (cfs_rq->skip == se) {
1679 struct sched_entity *second = __pick_next_entity(se);
1680 if (second && wakeup_preempt_entity(second, left) < 1)
1685 * Prefer last buddy, try to return the CPU to a preempted task.
1687 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1691 * Someone really wants this to run. If it's not unfair, run it.
1693 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1696 clear_buddies(cfs_rq, se);
1701 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1703 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1706 * If still on the runqueue then deactivate_task()
1707 * was not called and update_curr() has to be done:
1710 update_curr(cfs_rq);
1712 /* throttle cfs_rqs exceeding runtime */
1713 check_cfs_rq_runtime(cfs_rq);
1715 check_spread(cfs_rq, prev);
1717 update_stats_wait_start(cfs_rq, prev);
1718 /* Put 'current' back into the tree. */
1719 __enqueue_entity(cfs_rq, prev);
1720 /* in !on_rq case, update occurred at dequeue */
1721 update_entity_load_avg(prev, 1);
1723 cfs_rq->curr = NULL;
1727 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1730 * Update run-time statistics of the 'current'.
1732 update_curr(cfs_rq);
1735 * Ensure that runnable average is periodically updated.
1737 update_entity_load_avg(curr, 1);
1738 update_cfs_rq_blocked_load(cfs_rq, 1);
1740 #ifdef CONFIG_SCHED_HRTICK
1742 * queued ticks are scheduled to match the slice, so don't bother
1743 * validating it and just reschedule.
1746 resched_task(rq_of(cfs_rq)->curr);
1750 * don't let the period tick interfere with the hrtick preemption
1752 if (!sched_feat(DOUBLE_TICK) &&
1753 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1757 if (cfs_rq->nr_running > 1)
1758 check_preempt_tick(cfs_rq, curr);
1762 /**************************************************
1763 * CFS bandwidth control machinery
1766 #ifdef CONFIG_CFS_BANDWIDTH
1768 #ifdef HAVE_JUMP_LABEL
1769 static struct static_key __cfs_bandwidth_used;
1771 static inline bool cfs_bandwidth_used(void)
1773 return static_key_false(&__cfs_bandwidth_used);
1776 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1778 /* only need to count groups transitioning between enabled/!enabled */
1779 if (enabled && !was_enabled)
1780 static_key_slow_inc(&__cfs_bandwidth_used);
1781 else if (!enabled && was_enabled)
1782 static_key_slow_dec(&__cfs_bandwidth_used);
1784 #else /* HAVE_JUMP_LABEL */
1785 static bool cfs_bandwidth_used(void)
1790 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1791 #endif /* HAVE_JUMP_LABEL */
1794 * default period for cfs group bandwidth.
1795 * default: 0.1s, units: nanoseconds
1797 static inline u64 default_cfs_period(void)
1799 return 100000000ULL;
1802 static inline u64 sched_cfs_bandwidth_slice(void)
1804 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1808 * Replenish runtime according to assigned quota and update expiration time.
1809 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1810 * additional synchronization around rq->lock.
1812 * requires cfs_b->lock
1814 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1818 if (cfs_b->quota == RUNTIME_INF)
1821 now = sched_clock_cpu(smp_processor_id());
1822 cfs_b->runtime = cfs_b->quota;
1823 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1826 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1828 return &tg->cfs_bandwidth;
1831 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
1832 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
1834 if (unlikely(cfs_rq->throttle_count))
1835 return cfs_rq->throttled_clock_task;
1837 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
1840 /* returns 0 on failure to allocate runtime */
1841 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1843 struct task_group *tg = cfs_rq->tg;
1844 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1845 u64 amount = 0, min_amount, expires;
1847 /* note: this is a positive sum as runtime_remaining <= 0 */
1848 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1850 raw_spin_lock(&cfs_b->lock);
1851 if (cfs_b->quota == RUNTIME_INF)
1852 amount = min_amount;
1855 * If the bandwidth pool has become inactive, then at least one
1856 * period must have elapsed since the last consumption.
1857 * Refresh the global state and ensure bandwidth timer becomes
1860 if (!cfs_b->timer_active) {
1861 __refill_cfs_bandwidth_runtime(cfs_b);
1862 __start_cfs_bandwidth(cfs_b);
1865 if (cfs_b->runtime > 0) {
1866 amount = min(cfs_b->runtime, min_amount);
1867 cfs_b->runtime -= amount;
1871 expires = cfs_b->runtime_expires;
1872 raw_spin_unlock(&cfs_b->lock);
1874 cfs_rq->runtime_remaining += amount;
1876 * we may have advanced our local expiration to account for allowed
1877 * spread between our sched_clock and the one on which runtime was
1880 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1881 cfs_rq->runtime_expires = expires;
1883 return cfs_rq->runtime_remaining > 0;
1887 * Note: This depends on the synchronization provided by sched_clock and the
1888 * fact that rq->clock snapshots this value.
1890 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1892 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1893 struct rq *rq = rq_of(cfs_rq);
1895 /* if the deadline is ahead of our clock, nothing to do */
1896 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1899 if (cfs_rq->runtime_remaining < 0)
1903 * If the local deadline has passed we have to consider the
1904 * possibility that our sched_clock is 'fast' and the global deadline
1905 * has not truly expired.
1907 * Fortunately we can check determine whether this the case by checking
1908 * whether the global deadline has advanced.
1911 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1912 /* extend local deadline, drift is bounded above by 2 ticks */
1913 cfs_rq->runtime_expires += TICK_NSEC;
1915 /* global deadline is ahead, expiration has passed */
1916 cfs_rq->runtime_remaining = 0;
1920 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1921 unsigned long delta_exec)
1923 /* dock delta_exec before expiring quota (as it could span periods) */
1924 cfs_rq->runtime_remaining -= delta_exec;
1925 expire_cfs_rq_runtime(cfs_rq);
1927 if (likely(cfs_rq->runtime_remaining > 0))
1931 * if we're unable to extend our runtime we resched so that the active
1932 * hierarchy can be throttled
1934 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1935 resched_task(rq_of(cfs_rq)->curr);
1938 static __always_inline
1939 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1941 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1944 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1947 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1949 return cfs_bandwidth_used() && cfs_rq->throttled;
1952 /* check whether cfs_rq, or any parent, is throttled */
1953 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1955 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1959 * Ensure that neither of the group entities corresponding to src_cpu or
1960 * dest_cpu are members of a throttled hierarchy when performing group
1961 * load-balance operations.
1963 static inline int throttled_lb_pair(struct task_group *tg,
1964 int src_cpu, int dest_cpu)
1966 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1968 src_cfs_rq = tg->cfs_rq[src_cpu];
1969 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1971 return throttled_hierarchy(src_cfs_rq) ||
1972 throttled_hierarchy(dest_cfs_rq);
1975 /* updated child weight may affect parent so we have to do this bottom up */
1976 static int tg_unthrottle_up(struct task_group *tg, void *data)
1978 struct rq *rq = data;
1979 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1981 cfs_rq->throttle_count--;
1983 if (!cfs_rq->throttle_count) {
1984 /* adjust cfs_rq_clock_task() */
1985 cfs_rq->throttled_clock_task_time += rq->clock_task -
1986 cfs_rq->throttled_clock_task;
1993 static int tg_throttle_down(struct task_group *tg, void *data)
1995 struct rq *rq = data;
1996 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1998 /* group is entering throttled state, stop time */
1999 if (!cfs_rq->throttle_count)
2000 cfs_rq->throttled_clock_task = rq->clock_task;
2001 cfs_rq->throttle_count++;
2006 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2008 struct rq *rq = rq_of(cfs_rq);
2009 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2010 struct sched_entity *se;
2011 long task_delta, dequeue = 1;
2013 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2015 /* freeze hierarchy runnable averages while throttled */
2017 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2020 task_delta = cfs_rq->h_nr_running;
2021 for_each_sched_entity(se) {
2022 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2023 /* throttled entity or throttle-on-deactivate */
2028 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2029 qcfs_rq->h_nr_running -= task_delta;
2031 if (qcfs_rq->load.weight)
2036 rq->nr_running -= task_delta;
2038 cfs_rq->throttled = 1;
2039 cfs_rq->throttled_clock = rq->clock;
2040 raw_spin_lock(&cfs_b->lock);
2041 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2042 raw_spin_unlock(&cfs_b->lock);
2045 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2047 struct rq *rq = rq_of(cfs_rq);
2048 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2049 struct sched_entity *se;
2053 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2055 cfs_rq->throttled = 0;
2056 raw_spin_lock(&cfs_b->lock);
2057 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2058 list_del_rcu(&cfs_rq->throttled_list);
2059 raw_spin_unlock(&cfs_b->lock);
2061 update_rq_clock(rq);
2062 /* update hierarchical throttle state */
2063 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2065 if (!cfs_rq->load.weight)
2068 task_delta = cfs_rq->h_nr_running;
2069 for_each_sched_entity(se) {
2073 cfs_rq = cfs_rq_of(se);
2075 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2076 cfs_rq->h_nr_running += task_delta;
2078 if (cfs_rq_throttled(cfs_rq))
2083 rq->nr_running += task_delta;
2085 /* determine whether we need to wake up potentially idle cpu */
2086 if (rq->curr == rq->idle && rq->cfs.nr_running)
2087 resched_task(rq->curr);
2090 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2091 u64 remaining, u64 expires)
2093 struct cfs_rq *cfs_rq;
2094 u64 runtime = remaining;
2097 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2099 struct rq *rq = rq_of(cfs_rq);
2101 raw_spin_lock(&rq->lock);
2102 if (!cfs_rq_throttled(cfs_rq))
2105 runtime = -cfs_rq->runtime_remaining + 1;
2106 if (runtime > remaining)
2107 runtime = remaining;
2108 remaining -= runtime;
2110 cfs_rq->runtime_remaining += runtime;
2111 cfs_rq->runtime_expires = expires;
2113 /* we check whether we're throttled above */
2114 if (cfs_rq->runtime_remaining > 0)
2115 unthrottle_cfs_rq(cfs_rq);
2118 raw_spin_unlock(&rq->lock);
2129 * Responsible for refilling a task_group's bandwidth and unthrottling its
2130 * cfs_rqs as appropriate. If there has been no activity within the last
2131 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2132 * used to track this state.
2134 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2136 u64 runtime, runtime_expires;
2137 int idle = 1, throttled;
2139 raw_spin_lock(&cfs_b->lock);
2140 /* no need to continue the timer with no bandwidth constraint */
2141 if (cfs_b->quota == RUNTIME_INF)
2144 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2145 /* idle depends on !throttled (for the case of a large deficit) */
2146 idle = cfs_b->idle && !throttled;
2147 cfs_b->nr_periods += overrun;
2149 /* if we're going inactive then everything else can be deferred */
2153 __refill_cfs_bandwidth_runtime(cfs_b);
2156 /* mark as potentially idle for the upcoming period */
2161 /* account preceding periods in which throttling occurred */
2162 cfs_b->nr_throttled += overrun;
2165 * There are throttled entities so we must first use the new bandwidth
2166 * to unthrottle them before making it generally available. This
2167 * ensures that all existing debts will be paid before a new cfs_rq is
2170 runtime = cfs_b->runtime;
2171 runtime_expires = cfs_b->runtime_expires;
2175 * This check is repeated as we are holding onto the new bandwidth
2176 * while we unthrottle. This can potentially race with an unthrottled
2177 * group trying to acquire new bandwidth from the global pool.
2179 while (throttled && runtime > 0) {
2180 raw_spin_unlock(&cfs_b->lock);
2181 /* we can't nest cfs_b->lock while distributing bandwidth */
2182 runtime = distribute_cfs_runtime(cfs_b, runtime,
2184 raw_spin_lock(&cfs_b->lock);
2186 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2189 /* return (any) remaining runtime */
2190 cfs_b->runtime = runtime;
2192 * While we are ensured activity in the period following an
2193 * unthrottle, this also covers the case in which the new bandwidth is
2194 * insufficient to cover the existing bandwidth deficit. (Forcing the
2195 * timer to remain active while there are any throttled entities.)
2200 cfs_b->timer_active = 0;
2201 raw_spin_unlock(&cfs_b->lock);
2206 /* a cfs_rq won't donate quota below this amount */
2207 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2208 /* minimum remaining period time to redistribute slack quota */
2209 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2210 /* how long we wait to gather additional slack before distributing */
2211 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2213 /* are we near the end of the current quota period? */
2214 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2216 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2219 /* if the call-back is running a quota refresh is already occurring */
2220 if (hrtimer_callback_running(refresh_timer))
2223 /* is a quota refresh about to occur? */
2224 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2225 if (remaining < min_expire)
2231 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2233 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2235 /* if there's a quota refresh soon don't bother with slack */
2236 if (runtime_refresh_within(cfs_b, min_left))
2239 start_bandwidth_timer(&cfs_b->slack_timer,
2240 ns_to_ktime(cfs_bandwidth_slack_period));
2243 /* we know any runtime found here is valid as update_curr() precedes return */
2244 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2246 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2247 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2249 if (slack_runtime <= 0)
2252 raw_spin_lock(&cfs_b->lock);
2253 if (cfs_b->quota != RUNTIME_INF &&
2254 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2255 cfs_b->runtime += slack_runtime;
2257 /* we are under rq->lock, defer unthrottling using a timer */
2258 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2259 !list_empty(&cfs_b->throttled_cfs_rq))
2260 start_cfs_slack_bandwidth(cfs_b);
2262 raw_spin_unlock(&cfs_b->lock);
2264 /* even if it's not valid for return we don't want to try again */
2265 cfs_rq->runtime_remaining -= slack_runtime;
2268 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2270 if (!cfs_bandwidth_used())
2273 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2276 __return_cfs_rq_runtime(cfs_rq);
2280 * This is done with a timer (instead of inline with bandwidth return) since
2281 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2283 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2285 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2288 /* confirm we're still not at a refresh boundary */
2289 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2292 raw_spin_lock(&cfs_b->lock);
2293 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2294 runtime = cfs_b->runtime;
2297 expires = cfs_b->runtime_expires;
2298 raw_spin_unlock(&cfs_b->lock);
2303 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2305 raw_spin_lock(&cfs_b->lock);
2306 if (expires == cfs_b->runtime_expires)
2307 cfs_b->runtime = runtime;
2308 raw_spin_unlock(&cfs_b->lock);
2312 * When a group wakes up we want to make sure that its quota is not already
2313 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2314 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2316 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2318 if (!cfs_bandwidth_used())
2321 /* an active group must be handled by the update_curr()->put() path */
2322 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2325 /* ensure the group is not already throttled */
2326 if (cfs_rq_throttled(cfs_rq))
2329 /* update runtime allocation */
2330 account_cfs_rq_runtime(cfs_rq, 0);
2331 if (cfs_rq->runtime_remaining <= 0)
2332 throttle_cfs_rq(cfs_rq);
2335 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2336 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2338 if (!cfs_bandwidth_used())
2341 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2345 * it's possible for a throttled entity to be forced into a running
2346 * state (e.g. set_curr_task), in this case we're finished.
2348 if (cfs_rq_throttled(cfs_rq))
2351 throttle_cfs_rq(cfs_rq);
2354 static inline u64 default_cfs_period(void);
2355 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2356 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2358 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2360 struct cfs_bandwidth *cfs_b =
2361 container_of(timer, struct cfs_bandwidth, slack_timer);
2362 do_sched_cfs_slack_timer(cfs_b);
2364 return HRTIMER_NORESTART;
2367 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2369 struct cfs_bandwidth *cfs_b =
2370 container_of(timer, struct cfs_bandwidth, period_timer);
2376 now = hrtimer_cb_get_time(timer);
2377 overrun = hrtimer_forward(timer, now, cfs_b->period);
2382 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2385 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2388 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2390 raw_spin_lock_init(&cfs_b->lock);
2392 cfs_b->quota = RUNTIME_INF;
2393 cfs_b->period = ns_to_ktime(default_cfs_period());
2395 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2396 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2397 cfs_b->period_timer.function = sched_cfs_period_timer;
2398 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2399 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2402 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2404 cfs_rq->runtime_enabled = 0;
2405 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2408 /* requires cfs_b->lock, may release to reprogram timer */
2409 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2412 * The timer may be active because we're trying to set a new bandwidth
2413 * period or because we're racing with the tear-down path
2414 * (timer_active==0 becomes visible before the hrtimer call-back
2415 * terminates). In either case we ensure that it's re-programmed
2417 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2418 raw_spin_unlock(&cfs_b->lock);
2419 /* ensure cfs_b->lock is available while we wait */
2420 hrtimer_cancel(&cfs_b->period_timer);
2422 raw_spin_lock(&cfs_b->lock);
2423 /* if someone else restarted the timer then we're done */
2424 if (cfs_b->timer_active)
2428 cfs_b->timer_active = 1;
2429 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2432 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2434 hrtimer_cancel(&cfs_b->period_timer);
2435 hrtimer_cancel(&cfs_b->slack_timer);
2438 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2440 struct cfs_rq *cfs_rq;
2442 for_each_leaf_cfs_rq(rq, cfs_rq) {
2443 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2445 if (!cfs_rq->runtime_enabled)
2449 * clock_task is not advancing so we just need to make sure
2450 * there's some valid quota amount
2452 cfs_rq->runtime_remaining = cfs_b->quota;
2453 if (cfs_rq_throttled(cfs_rq))
2454 unthrottle_cfs_rq(cfs_rq);
2458 #else /* CONFIG_CFS_BANDWIDTH */
2459 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2461 return rq_of(cfs_rq)->clock_task;
2464 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2465 unsigned long delta_exec) {}
2466 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2467 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2468 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2470 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2475 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2480 static inline int throttled_lb_pair(struct task_group *tg,
2481 int src_cpu, int dest_cpu)
2486 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2488 #ifdef CONFIG_FAIR_GROUP_SCHED
2489 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2492 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2496 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2497 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2499 #endif /* CONFIG_CFS_BANDWIDTH */
2501 /**************************************************
2502 * CFS operations on tasks:
2505 #ifdef CONFIG_SCHED_HRTICK
2506 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2508 struct sched_entity *se = &p->se;
2509 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2511 WARN_ON(task_rq(p) != rq);
2513 if (cfs_rq->nr_running > 1) {
2514 u64 slice = sched_slice(cfs_rq, se);
2515 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2516 s64 delta = slice - ran;
2525 * Don't schedule slices shorter than 10000ns, that just
2526 * doesn't make sense. Rely on vruntime for fairness.
2529 delta = max_t(s64, 10000LL, delta);
2531 hrtick_start(rq, delta);
2536 * called from enqueue/dequeue and updates the hrtick when the
2537 * current task is from our class and nr_running is low enough
2540 static void hrtick_update(struct rq *rq)
2542 struct task_struct *curr = rq->curr;
2544 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2547 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2548 hrtick_start_fair(rq, curr);
2550 #else /* !CONFIG_SCHED_HRTICK */
2552 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2556 static inline void hrtick_update(struct rq *rq)
2562 * The enqueue_task method is called before nr_running is
2563 * increased. Here we update the fair scheduling stats and
2564 * then put the task into the rbtree:
2567 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2569 struct cfs_rq *cfs_rq;
2570 struct sched_entity *se = &p->se;
2572 for_each_sched_entity(se) {
2575 cfs_rq = cfs_rq_of(se);
2576 enqueue_entity(cfs_rq, se, flags);
2579 * end evaluation on encountering a throttled cfs_rq
2581 * note: in the case of encountering a throttled cfs_rq we will
2582 * post the final h_nr_running increment below.
2584 if (cfs_rq_throttled(cfs_rq))
2586 cfs_rq->h_nr_running++;
2588 flags = ENQUEUE_WAKEUP;
2591 for_each_sched_entity(se) {
2592 cfs_rq = cfs_rq_of(se);
2593 cfs_rq->h_nr_running++;
2595 if (cfs_rq_throttled(cfs_rq))
2598 update_entity_load_avg(se, 1);
2599 update_cfs_rq_blocked_load(cfs_rq, 0);
2603 update_rq_runnable_avg(rq, rq->nr_running);
2609 static void set_next_buddy(struct sched_entity *se);
2612 * The dequeue_task method is called before nr_running is
2613 * decreased. We remove the task from the rbtree and
2614 * update the fair scheduling stats:
2616 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2618 struct cfs_rq *cfs_rq;
2619 struct sched_entity *se = &p->se;
2620 int task_sleep = flags & DEQUEUE_SLEEP;
2622 for_each_sched_entity(se) {
2623 cfs_rq = cfs_rq_of(se);
2624 dequeue_entity(cfs_rq, se, flags);
2627 * end evaluation on encountering a throttled cfs_rq
2629 * note: in the case of encountering a throttled cfs_rq we will
2630 * post the final h_nr_running decrement below.
2632 if (cfs_rq_throttled(cfs_rq))
2634 cfs_rq->h_nr_running--;
2636 /* Don't dequeue parent if it has other entities besides us */
2637 if (cfs_rq->load.weight) {
2639 * Bias pick_next to pick a task from this cfs_rq, as
2640 * p is sleeping when it is within its sched_slice.
2642 if (task_sleep && parent_entity(se))
2643 set_next_buddy(parent_entity(se));
2645 /* avoid re-evaluating load for this entity */
2646 se = parent_entity(se);
2649 flags |= DEQUEUE_SLEEP;
2652 for_each_sched_entity(se) {
2653 cfs_rq = cfs_rq_of(se);
2654 cfs_rq->h_nr_running--;
2656 if (cfs_rq_throttled(cfs_rq))
2659 update_entity_load_avg(se, 1);
2660 update_cfs_rq_blocked_load(cfs_rq, 0);
2665 update_rq_runnable_avg(rq, 1);
2671 /* Used instead of source_load when we know the type == 0 */
2672 static unsigned long weighted_cpuload(const int cpu)
2674 return cpu_rq(cpu)->load.weight;
2678 * Return a low guess at the load of a migration-source cpu weighted
2679 * according to the scheduling class and "nice" value.
2681 * We want to under-estimate the load of migration sources, to
2682 * balance conservatively.
2684 static unsigned long source_load(int cpu, int type)
2686 struct rq *rq = cpu_rq(cpu);
2687 unsigned long total = weighted_cpuload(cpu);
2689 if (type == 0 || !sched_feat(LB_BIAS))
2692 return min(rq->cpu_load[type-1], total);
2696 * Return a high guess at the load of a migration-target cpu weighted
2697 * according to the scheduling class and "nice" value.
2699 static unsigned long target_load(int cpu, int type)
2701 struct rq *rq = cpu_rq(cpu);
2702 unsigned long total = weighted_cpuload(cpu);
2704 if (type == 0 || !sched_feat(LB_BIAS))
2707 return max(rq->cpu_load[type-1], total);
2710 static unsigned long power_of(int cpu)
2712 return cpu_rq(cpu)->cpu_power;
2715 static unsigned long cpu_avg_load_per_task(int cpu)
2717 struct rq *rq = cpu_rq(cpu);
2718 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2721 return rq->load.weight / nr_running;
2727 static void task_waking_fair(struct task_struct *p)
2729 struct sched_entity *se = &p->se;
2730 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2733 #ifndef CONFIG_64BIT
2734 u64 min_vruntime_copy;
2737 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2739 min_vruntime = cfs_rq->min_vruntime;
2740 } while (min_vruntime != min_vruntime_copy);
2742 min_vruntime = cfs_rq->min_vruntime;
2745 se->vruntime -= min_vruntime;
2748 #ifdef CONFIG_FAIR_GROUP_SCHED
2750 * effective_load() calculates the load change as seen from the root_task_group
2752 * Adding load to a group doesn't make a group heavier, but can cause movement
2753 * of group shares between cpus. Assuming the shares were perfectly aligned one
2754 * can calculate the shift in shares.
2756 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2757 * on this @cpu and results in a total addition (subtraction) of @wg to the
2758 * total group weight.
2760 * Given a runqueue weight distribution (rw_i) we can compute a shares
2761 * distribution (s_i) using:
2763 * s_i = rw_i / \Sum rw_j (1)
2765 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2766 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2767 * shares distribution (s_i):
2769 * rw_i = { 2, 4, 1, 0 }
2770 * s_i = { 2/7, 4/7, 1/7, 0 }
2772 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2773 * task used to run on and the CPU the waker is running on), we need to
2774 * compute the effect of waking a task on either CPU and, in case of a sync
2775 * wakeup, compute the effect of the current task going to sleep.
2777 * So for a change of @wl to the local @cpu with an overall group weight change
2778 * of @wl we can compute the new shares distribution (s'_i) using:
2780 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2782 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2783 * differences in waking a task to CPU 0. The additional task changes the
2784 * weight and shares distributions like:
2786 * rw'_i = { 3, 4, 1, 0 }
2787 * s'_i = { 3/8, 4/8, 1/8, 0 }
2789 * We can then compute the difference in effective weight by using:
2791 * dw_i = S * (s'_i - s_i) (3)
2793 * Where 'S' is the group weight as seen by its parent.
2795 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2796 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2797 * 4/7) times the weight of the group.
2799 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2801 struct sched_entity *se = tg->se[cpu];
2803 if (!tg->parent) /* the trivial, non-cgroup case */
2806 for_each_sched_entity(se) {
2812 * W = @wg + \Sum rw_j
2814 W = wg + calc_tg_weight(tg, se->my_q);
2819 w = se->my_q->load.weight + wl;
2822 * wl = S * s'_i; see (2)
2825 wl = (w * tg->shares) / W;
2830 * Per the above, wl is the new se->load.weight value; since
2831 * those are clipped to [MIN_SHARES, ...) do so now. See
2832 * calc_cfs_shares().
2834 if (wl < MIN_SHARES)
2838 * wl = dw_i = S * (s'_i - s_i); see (3)
2840 wl -= se->load.weight;
2843 * Recursively apply this logic to all parent groups to compute
2844 * the final effective load change on the root group. Since
2845 * only the @tg group gets extra weight, all parent groups can
2846 * only redistribute existing shares. @wl is the shift in shares
2847 * resulting from this level per the above.
2856 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2857 unsigned long wl, unsigned long wg)
2864 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2866 s64 this_load, load;
2867 int idx, this_cpu, prev_cpu;
2868 unsigned long tl_per_task;
2869 struct task_group *tg;
2870 unsigned long weight;
2874 this_cpu = smp_processor_id();
2875 prev_cpu = task_cpu(p);
2876 load = source_load(prev_cpu, idx);
2877 this_load = target_load(this_cpu, idx);
2880 * If sync wakeup then subtract the (maximum possible)
2881 * effect of the currently running task from the load
2882 * of the current CPU:
2885 tg = task_group(current);
2886 weight = current->se.load.weight;
2888 this_load += effective_load(tg, this_cpu, -weight, -weight);
2889 load += effective_load(tg, prev_cpu, 0, -weight);
2893 weight = p->se.load.weight;
2896 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2897 * due to the sync cause above having dropped this_load to 0, we'll
2898 * always have an imbalance, but there's really nothing you can do
2899 * about that, so that's good too.
2901 * Otherwise check if either cpus are near enough in load to allow this
2902 * task to be woken on this_cpu.
2904 if (this_load > 0) {
2905 s64 this_eff_load, prev_eff_load;
2907 this_eff_load = 100;
2908 this_eff_load *= power_of(prev_cpu);
2909 this_eff_load *= this_load +
2910 effective_load(tg, this_cpu, weight, weight);
2912 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2913 prev_eff_load *= power_of(this_cpu);
2914 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2916 balanced = this_eff_load <= prev_eff_load;
2921 * If the currently running task will sleep within
2922 * a reasonable amount of time then attract this newly
2925 if (sync && balanced)
2928 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2929 tl_per_task = cpu_avg_load_per_task(this_cpu);
2932 (this_load <= load &&
2933 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2935 * This domain has SD_WAKE_AFFINE and
2936 * p is cache cold in this domain, and
2937 * there is no bad imbalance.
2939 schedstat_inc(sd, ttwu_move_affine);
2940 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2948 * find_idlest_group finds and returns the least busy CPU group within the
2951 static struct sched_group *
2952 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2953 int this_cpu, int load_idx)
2955 struct sched_group *idlest = NULL, *group = sd->groups;
2956 unsigned long min_load = ULONG_MAX, this_load = 0;
2957 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2960 unsigned long load, avg_load;
2964 /* Skip over this group if it has no CPUs allowed */
2965 if (!cpumask_intersects(sched_group_cpus(group),
2966 tsk_cpus_allowed(p)))
2969 local_group = cpumask_test_cpu(this_cpu,
2970 sched_group_cpus(group));
2972 /* Tally up the load of all CPUs in the group */
2975 for_each_cpu(i, sched_group_cpus(group)) {
2976 /* Bias balancing toward cpus of our domain */
2978 load = source_load(i, load_idx);
2980 load = target_load(i, load_idx);
2985 /* Adjust by relative CPU power of the group */
2986 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2989 this_load = avg_load;
2990 } else if (avg_load < min_load) {
2991 min_load = avg_load;
2994 } while (group = group->next, group != sd->groups);
2996 if (!idlest || 100*this_load < imbalance*min_load)
3002 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3005 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3007 unsigned long load, min_load = ULONG_MAX;
3011 /* Traverse only the allowed CPUs */
3012 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3013 load = weighted_cpuload(i);
3015 if (load < min_load || (load == min_load && i == this_cpu)) {
3025 * Try and locate an idle CPU in the sched_domain.
3027 static int select_idle_sibling(struct task_struct *p, int target)
3029 int cpu = smp_processor_id();
3030 int prev_cpu = task_cpu(p);
3031 struct sched_domain *sd;
3032 struct sched_group *sg;
3036 * If the task is going to be woken-up on this cpu and if it is
3037 * already idle, then it is the right target.
3039 if (target == cpu && idle_cpu(cpu))
3043 * If the task is going to be woken-up on the cpu where it previously
3044 * ran and if it is currently idle, then it the right target.
3046 if (target == prev_cpu && idle_cpu(prev_cpu))
3050 * Otherwise, iterate the domains and find an elegible idle cpu.
3052 sd = rcu_dereference(per_cpu(sd_llc, target));
3053 for_each_lower_domain(sd) {
3056 if (!cpumask_intersects(sched_group_cpus(sg),
3057 tsk_cpus_allowed(p)))
3060 for_each_cpu(i, sched_group_cpus(sg)) {
3065 target = cpumask_first_and(sched_group_cpus(sg),
3066 tsk_cpus_allowed(p));
3070 } while (sg != sd->groups);
3077 * sched_balance_self: balance the current task (running on cpu) in domains
3078 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3081 * Balance, ie. select the least loaded group.
3083 * Returns the target CPU number, or the same CPU if no balancing is needed.
3085 * preempt must be disabled.
3088 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3090 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3091 int cpu = smp_processor_id();
3092 int prev_cpu = task_cpu(p);
3094 int want_affine = 0;
3095 int sync = wake_flags & WF_SYNC;
3097 if (p->nr_cpus_allowed == 1)
3100 if (sd_flag & SD_BALANCE_WAKE) {
3101 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3107 for_each_domain(cpu, tmp) {
3108 if (!(tmp->flags & SD_LOAD_BALANCE))
3112 * If both cpu and prev_cpu are part of this domain,
3113 * cpu is a valid SD_WAKE_AFFINE target.
3115 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3116 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3121 if (tmp->flags & sd_flag)
3126 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3129 new_cpu = select_idle_sibling(p, prev_cpu);
3134 int load_idx = sd->forkexec_idx;
3135 struct sched_group *group;
3138 if (!(sd->flags & sd_flag)) {
3143 if (sd_flag & SD_BALANCE_WAKE)
3144 load_idx = sd->wake_idx;
3146 group = find_idlest_group(sd, p, cpu, load_idx);
3152 new_cpu = find_idlest_cpu(group, p, cpu);
3153 if (new_cpu == -1 || new_cpu == cpu) {
3154 /* Now try balancing at a lower domain level of cpu */
3159 /* Now try balancing at a lower domain level of new_cpu */
3161 weight = sd->span_weight;
3163 for_each_domain(cpu, tmp) {
3164 if (weight <= tmp->span_weight)
3166 if (tmp->flags & sd_flag)
3169 /* while loop will break here if sd == NULL */
3178 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3179 * removed when useful for applications beyond shares distribution (e.g.
3182 #ifdef CONFIG_FAIR_GROUP_SCHED
3184 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3185 * cfs_rq_of(p) references at time of call are still valid and identify the
3186 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3187 * other assumptions, including the state of rq->lock, should be made.
3190 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3192 struct sched_entity *se = &p->se;
3193 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3196 * Load tracking: accumulate removed load so that it can be processed
3197 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3198 * to blocked load iff they have a positive decay-count. It can never
3199 * be negative here since on-rq tasks have decay-count == 0.
3201 if (se->avg.decay_count) {
3202 se->avg.decay_count = -__synchronize_entity_decay(se);
3203 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3207 #endif /* CONFIG_SMP */
3209 static unsigned long
3210 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3212 unsigned long gran = sysctl_sched_wakeup_granularity;
3215 * Since its curr running now, convert the gran from real-time
3216 * to virtual-time in his units.
3218 * By using 'se' instead of 'curr' we penalize light tasks, so
3219 * they get preempted easier. That is, if 'se' < 'curr' then
3220 * the resulting gran will be larger, therefore penalizing the
3221 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3222 * be smaller, again penalizing the lighter task.
3224 * This is especially important for buddies when the leftmost
3225 * task is higher priority than the buddy.
3227 return calc_delta_fair(gran, se);
3231 * Should 'se' preempt 'curr'.
3245 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3247 s64 gran, vdiff = curr->vruntime - se->vruntime;
3252 gran = wakeup_gran(curr, se);
3259 static void set_last_buddy(struct sched_entity *se)
3261 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3264 for_each_sched_entity(se)
3265 cfs_rq_of(se)->last = se;
3268 static void set_next_buddy(struct sched_entity *se)
3270 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3273 for_each_sched_entity(se)
3274 cfs_rq_of(se)->next = se;
3277 static void set_skip_buddy(struct sched_entity *se)
3279 for_each_sched_entity(se)
3280 cfs_rq_of(se)->skip = se;
3284 * Preempt the current task with a newly woken task if needed:
3286 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3288 struct task_struct *curr = rq->curr;
3289 struct sched_entity *se = &curr->se, *pse = &p->se;
3290 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3291 int scale = cfs_rq->nr_running >= sched_nr_latency;
3292 int next_buddy_marked = 0;
3294 if (unlikely(se == pse))
3298 * This is possible from callers such as move_task(), in which we
3299 * unconditionally check_prempt_curr() after an enqueue (which may have
3300 * lead to a throttle). This both saves work and prevents false
3301 * next-buddy nomination below.
3303 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3306 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3307 set_next_buddy(pse);
3308 next_buddy_marked = 1;
3312 * We can come here with TIF_NEED_RESCHED already set from new task
3315 * Note: this also catches the edge-case of curr being in a throttled
3316 * group (e.g. via set_curr_task), since update_curr() (in the
3317 * enqueue of curr) will have resulted in resched being set. This
3318 * prevents us from potentially nominating it as a false LAST_BUDDY
3321 if (test_tsk_need_resched(curr))
3324 /* Idle tasks are by definition preempted by non-idle tasks. */
3325 if (unlikely(curr->policy == SCHED_IDLE) &&
3326 likely(p->policy != SCHED_IDLE))
3330 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3331 * is driven by the tick):
3333 if (unlikely(p->policy != SCHED_NORMAL))
3336 find_matching_se(&se, &pse);
3337 update_curr(cfs_rq_of(se));
3339 if (wakeup_preempt_entity(se, pse) == 1) {
3341 * Bias pick_next to pick the sched entity that is
3342 * triggering this preemption.
3344 if (!next_buddy_marked)
3345 set_next_buddy(pse);
3354 * Only set the backward buddy when the current task is still
3355 * on the rq. This can happen when a wakeup gets interleaved
3356 * with schedule on the ->pre_schedule() or idle_balance()
3357 * point, either of which can * drop the rq lock.
3359 * Also, during early boot the idle thread is in the fair class,
3360 * for obvious reasons its a bad idea to schedule back to it.
3362 if (unlikely(!se->on_rq || curr == rq->idle))
3365 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3369 static struct task_struct *pick_next_task_fair(struct rq *rq)
3371 struct task_struct *p;
3372 struct cfs_rq *cfs_rq = &rq->cfs;
3373 struct sched_entity *se;
3375 if (!cfs_rq->nr_running)
3379 se = pick_next_entity(cfs_rq);
3380 set_next_entity(cfs_rq, se);
3381 cfs_rq = group_cfs_rq(se);
3385 if (hrtick_enabled(rq))
3386 hrtick_start_fair(rq, p);
3392 * Account for a descheduled task:
3394 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3396 struct sched_entity *se = &prev->se;
3397 struct cfs_rq *cfs_rq;
3399 for_each_sched_entity(se) {
3400 cfs_rq = cfs_rq_of(se);
3401 put_prev_entity(cfs_rq, se);
3406 * sched_yield() is very simple
3408 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3410 static void yield_task_fair(struct rq *rq)
3412 struct task_struct *curr = rq->curr;
3413 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3414 struct sched_entity *se = &curr->se;
3417 * Are we the only task in the tree?
3419 if (unlikely(rq->nr_running == 1))
3422 clear_buddies(cfs_rq, se);
3424 if (curr->policy != SCHED_BATCH) {
3425 update_rq_clock(rq);
3427 * Update run-time statistics of the 'current'.
3429 update_curr(cfs_rq);
3431 * Tell update_rq_clock() that we've just updated,
3432 * so we don't do microscopic update in schedule()
3433 * and double the fastpath cost.
3435 rq->skip_clock_update = 1;
3441 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3443 struct sched_entity *se = &p->se;
3445 /* throttled hierarchies are not runnable */
3446 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3449 /* Tell the scheduler that we'd really like pse to run next. */
3452 yield_task_fair(rq);
3458 /**************************************************
3459 * Fair scheduling class load-balancing methods:
3462 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3464 #define LBF_ALL_PINNED 0x01
3465 #define LBF_NEED_BREAK 0x02
3466 #define LBF_SOME_PINNED 0x04
3469 struct sched_domain *sd;
3477 struct cpumask *dst_grpmask;
3479 enum cpu_idle_type idle;
3481 /* The set of CPUs under consideration for load-balancing */
3482 struct cpumask *cpus;
3487 unsigned int loop_break;
3488 unsigned int loop_max;
3492 * move_task - move a task from one runqueue to another runqueue.
3493 * Both runqueues must be locked.
3495 static void move_task(struct task_struct *p, struct lb_env *env)
3497 deactivate_task(env->src_rq, p, 0);
3498 set_task_cpu(p, env->dst_cpu);
3499 activate_task(env->dst_rq, p, 0);
3500 check_preempt_curr(env->dst_rq, p, 0);
3504 * Is this task likely cache-hot:
3507 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3511 if (p->sched_class != &fair_sched_class)
3514 if (unlikely(p->policy == SCHED_IDLE))
3518 * Buddy candidates are cache hot:
3520 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3521 (&p->se == cfs_rq_of(&p->se)->next ||
3522 &p->se == cfs_rq_of(&p->se)->last))
3525 if (sysctl_sched_migration_cost == -1)
3527 if (sysctl_sched_migration_cost == 0)
3530 delta = now - p->se.exec_start;
3532 return delta < (s64)sysctl_sched_migration_cost;
3536 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3539 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3541 int tsk_cache_hot = 0;
3543 * We do not migrate tasks that are:
3544 * 1) running (obviously), or
3545 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3546 * 3) are cache-hot on their current CPU.
3548 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3551 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3554 * Remember if this task can be migrated to any other cpu in
3555 * our sched_group. We may want to revisit it if we couldn't
3556 * meet load balance goals by pulling other tasks on src_cpu.
3558 * Also avoid computing new_dst_cpu if we have already computed
3559 * one in current iteration.
3561 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3564 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3565 tsk_cpus_allowed(p));
3566 if (new_dst_cpu < nr_cpu_ids) {
3567 env->flags |= LBF_SOME_PINNED;
3568 env->new_dst_cpu = new_dst_cpu;
3573 /* Record that we found atleast one task that could run on dst_cpu */
3574 env->flags &= ~LBF_ALL_PINNED;
3576 if (task_running(env->src_rq, p)) {
3577 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3582 * Aggressive migration if:
3583 * 1) task is cache cold, or
3584 * 2) too many balance attempts have failed.
3587 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3588 if (!tsk_cache_hot ||
3589 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3590 #ifdef CONFIG_SCHEDSTATS
3591 if (tsk_cache_hot) {
3592 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3593 schedstat_inc(p, se.statistics.nr_forced_migrations);
3599 if (tsk_cache_hot) {
3600 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3607 * move_one_task tries to move exactly one task from busiest to this_rq, as
3608 * part of active balancing operations within "domain".
3609 * Returns 1 if successful and 0 otherwise.
3611 * Called with both runqueues locked.
3613 static int move_one_task(struct lb_env *env)
3615 struct task_struct *p, *n;
3617 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3618 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3621 if (!can_migrate_task(p, env))
3626 * Right now, this is only the second place move_task()
3627 * is called, so we can safely collect move_task()
3628 * stats here rather than inside move_task().
3630 schedstat_inc(env->sd, lb_gained[env->idle]);
3636 static unsigned long task_h_load(struct task_struct *p);
3638 static const unsigned int sched_nr_migrate_break = 32;
3641 * move_tasks tries to move up to imbalance weighted load from busiest to
3642 * this_rq, as part of a balancing operation within domain "sd".
3643 * Returns 1 if successful and 0 otherwise.
3645 * Called with both runqueues locked.
3647 static int move_tasks(struct lb_env *env)
3649 struct list_head *tasks = &env->src_rq->cfs_tasks;
3650 struct task_struct *p;
3654 if (env->imbalance <= 0)
3657 while (!list_empty(tasks)) {
3658 p = list_first_entry(tasks, struct task_struct, se.group_node);
3661 /* We've more or less seen every task there is, call it quits */
3662 if (env->loop > env->loop_max)
3665 /* take a breather every nr_migrate tasks */
3666 if (env->loop > env->loop_break) {
3667 env->loop_break += sched_nr_migrate_break;
3668 env->flags |= LBF_NEED_BREAK;
3672 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3675 load = task_h_load(p);
3677 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3680 if ((load / 2) > env->imbalance)
3683 if (!can_migrate_task(p, env))
3688 env->imbalance -= load;
3690 #ifdef CONFIG_PREEMPT
3692 * NEWIDLE balancing is a source of latency, so preemptible
3693 * kernels will stop after the first task is pulled to minimize
3694 * the critical section.
3696 if (env->idle == CPU_NEWLY_IDLE)
3701 * We only want to steal up to the prescribed amount of
3704 if (env->imbalance <= 0)
3709 list_move_tail(&p->se.group_node, tasks);
3713 * Right now, this is one of only two places move_task() is called,
3714 * so we can safely collect move_task() stats here rather than
3715 * inside move_task().
3717 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3722 #ifdef CONFIG_FAIR_GROUP_SCHED
3724 * update tg->load_weight by folding this cpu's load_avg
3726 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
3728 struct sched_entity *se = tg->se[cpu];
3729 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
3731 /* throttled entities do not contribute to load */
3732 if (throttled_hierarchy(cfs_rq))
3735 update_cfs_rq_blocked_load(cfs_rq, 1);
3738 update_entity_load_avg(se, 1);
3740 * We pivot on our runnable average having decayed to zero for
3741 * list removal. This generally implies that all our children
3742 * have also been removed (modulo rounding error or bandwidth
3743 * control); however, such cases are rare and we can fix these
3746 * TODO: fix up out-of-order children on enqueue.
3748 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
3749 list_del_leaf_cfs_rq(cfs_rq);
3751 struct rq *rq = rq_of(cfs_rq);
3752 update_rq_runnable_avg(rq, rq->nr_running);
3756 static void update_blocked_averages(int cpu)
3758 struct rq *rq = cpu_rq(cpu);
3759 struct cfs_rq *cfs_rq;
3760 unsigned long flags;
3762 raw_spin_lock_irqsave(&rq->lock, flags);
3763 update_rq_clock(rq);
3765 * Iterates the task_group tree in a bottom up fashion, see
3766 * list_add_leaf_cfs_rq() for details.
3768 for_each_leaf_cfs_rq(rq, cfs_rq) {
3770 * Note: We may want to consider periodically releasing
3771 * rq->lock about these updates so that creating many task
3772 * groups does not result in continually extending hold time.
3774 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
3777 raw_spin_unlock_irqrestore(&rq->lock, flags);
3781 * Compute the cpu's hierarchical load factor for each task group.
3782 * This needs to be done in a top-down fashion because the load of a child
3783 * group is a fraction of its parents load.
3785 static int tg_load_down(struct task_group *tg, void *data)
3788 long cpu = (long)data;
3791 load = cpu_rq(cpu)->load.weight;
3793 load = tg->parent->cfs_rq[cpu]->h_load;
3794 load *= tg->se[cpu]->load.weight;
3795 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3798 tg->cfs_rq[cpu]->h_load = load;
3803 static void update_h_load(long cpu)
3805 struct rq *rq = cpu_rq(cpu);
3806 unsigned long now = jiffies;
3808 if (rq->h_load_throttle == now)
3811 rq->h_load_throttle = now;
3814 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3818 static unsigned long task_h_load(struct task_struct *p)
3820 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3823 load = p->se.load.weight;
3824 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3829 static inline void update_blocked_averages(int cpu)
3833 static inline void update_h_load(long cpu)
3837 static unsigned long task_h_load(struct task_struct *p)
3839 return p->se.load.weight;
3843 /********** Helpers for find_busiest_group ************************/
3845 * sd_lb_stats - Structure to store the statistics of a sched_domain
3846 * during load balancing.
3848 struct sd_lb_stats {
3849 struct sched_group *busiest; /* Busiest group in this sd */
3850 struct sched_group *this; /* Local group in this sd */
3851 unsigned long total_load; /* Total load of all groups in sd */
3852 unsigned long total_pwr; /* Total power of all groups in sd */
3853 unsigned long avg_load; /* Average load across all groups in sd */
3855 /** Statistics of this group */
3856 unsigned long this_load;
3857 unsigned long this_load_per_task;
3858 unsigned long this_nr_running;
3859 unsigned long this_has_capacity;
3860 unsigned int this_idle_cpus;
3862 /* Statistics of the busiest group */
3863 unsigned int busiest_idle_cpus;
3864 unsigned long max_load;
3865 unsigned long busiest_load_per_task;
3866 unsigned long busiest_nr_running;
3867 unsigned long busiest_group_capacity;
3868 unsigned long busiest_has_capacity;
3869 unsigned int busiest_group_weight;
3871 int group_imb; /* Is there imbalance in this sd */
3875 * sg_lb_stats - stats of a sched_group required for load_balancing
3877 struct sg_lb_stats {
3878 unsigned long avg_load; /*Avg load across the CPUs of the group */
3879 unsigned long group_load; /* Total load over the CPUs of the group */
3880 unsigned long sum_nr_running; /* Nr tasks running in the group */
3881 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3882 unsigned long group_capacity;
3883 unsigned long idle_cpus;
3884 unsigned long group_weight;
3885 int group_imb; /* Is there an imbalance in the group ? */
3886 int group_has_capacity; /* Is there extra capacity in the group? */
3890 * get_sd_load_idx - Obtain the load index for a given sched domain.
3891 * @sd: The sched_domain whose load_idx is to be obtained.
3892 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3894 static inline int get_sd_load_idx(struct sched_domain *sd,
3895 enum cpu_idle_type idle)
3901 load_idx = sd->busy_idx;
3904 case CPU_NEWLY_IDLE:
3905 load_idx = sd->newidle_idx;
3908 load_idx = sd->idle_idx;
3915 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3917 return SCHED_POWER_SCALE;
3920 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3922 return default_scale_freq_power(sd, cpu);
3925 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3927 unsigned long weight = sd->span_weight;
3928 unsigned long smt_gain = sd->smt_gain;
3935 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3937 return default_scale_smt_power(sd, cpu);
3940 unsigned long scale_rt_power(int cpu)
3942 struct rq *rq = cpu_rq(cpu);
3943 u64 total, available, age_stamp, avg;
3946 * Since we're reading these variables without serialization make sure
3947 * we read them once before doing sanity checks on them.
3949 age_stamp = ACCESS_ONCE(rq->age_stamp);
3950 avg = ACCESS_ONCE(rq->rt_avg);
3952 total = sched_avg_period() + (rq->clock - age_stamp);
3954 if (unlikely(total < avg)) {
3955 /* Ensures that power won't end up being negative */
3958 available = total - avg;
3961 if (unlikely((s64)total < SCHED_POWER_SCALE))
3962 total = SCHED_POWER_SCALE;
3964 total >>= SCHED_POWER_SHIFT;
3966 return div_u64(available, total);
3969 static void update_cpu_power(struct sched_domain *sd, int cpu)
3971 unsigned long weight = sd->span_weight;
3972 unsigned long power = SCHED_POWER_SCALE;
3973 struct sched_group *sdg = sd->groups;
3975 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3976 if (sched_feat(ARCH_POWER))
3977 power *= arch_scale_smt_power(sd, cpu);
3979 power *= default_scale_smt_power(sd, cpu);
3981 power >>= SCHED_POWER_SHIFT;
3984 sdg->sgp->power_orig = power;
3986 if (sched_feat(ARCH_POWER))
3987 power *= arch_scale_freq_power(sd, cpu);
3989 power *= default_scale_freq_power(sd, cpu);
3991 power >>= SCHED_POWER_SHIFT;
3993 power *= scale_rt_power(cpu);
3994 power >>= SCHED_POWER_SHIFT;
3999 cpu_rq(cpu)->cpu_power = power;
4000 sdg->sgp->power = power;
4003 void update_group_power(struct sched_domain *sd, int cpu)
4005 struct sched_domain *child = sd->child;
4006 struct sched_group *group, *sdg = sd->groups;
4007 unsigned long power;
4008 unsigned long interval;
4010 interval = msecs_to_jiffies(sd->balance_interval);
4011 interval = clamp(interval, 1UL, max_load_balance_interval);
4012 sdg->sgp->next_update = jiffies + interval;
4015 update_cpu_power(sd, cpu);
4021 if (child->flags & SD_OVERLAP) {
4023 * SD_OVERLAP domains cannot assume that child groups
4024 * span the current group.
4027 for_each_cpu(cpu, sched_group_cpus(sdg))
4028 power += power_of(cpu);
4031 * !SD_OVERLAP domains can assume that child groups
4032 * span the current group.
4035 group = child->groups;
4037 power += group->sgp->power;
4038 group = group->next;
4039 } while (group != child->groups);
4042 sdg->sgp->power_orig = sdg->sgp->power = power;
4046 * Try and fix up capacity for tiny siblings, this is needed when
4047 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4048 * which on its own isn't powerful enough.
4050 * See update_sd_pick_busiest() and check_asym_packing().
4053 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4056 * Only siblings can have significantly less than SCHED_POWER_SCALE
4058 if (!(sd->flags & SD_SHARE_CPUPOWER))
4062 * If ~90% of the cpu_power is still there, we're good.
4064 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4071 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4072 * @env: The load balancing environment.
4073 * @group: sched_group whose statistics are to be updated.
4074 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4075 * @local_group: Does group contain this_cpu.
4076 * @balance: Should we balance.
4077 * @sgs: variable to hold the statistics for this group.
4079 static inline void update_sg_lb_stats(struct lb_env *env,
4080 struct sched_group *group, int load_idx,
4081 int local_group, int *balance, struct sg_lb_stats *sgs)
4083 unsigned long nr_running, max_nr_running, min_nr_running;
4084 unsigned long load, max_cpu_load, min_cpu_load;
4085 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4086 unsigned long avg_load_per_task = 0;
4090 balance_cpu = group_balance_cpu(group);
4092 /* Tally up the load of all CPUs in the group */
4094 min_cpu_load = ~0UL;
4096 min_nr_running = ~0UL;
4098 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4099 struct rq *rq = cpu_rq(i);
4101 nr_running = rq->nr_running;
4103 /* Bias balancing toward cpus of our domain */
4105 if (idle_cpu(i) && !first_idle_cpu &&
4106 cpumask_test_cpu(i, sched_group_mask(group))) {
4111 load = target_load(i, load_idx);
4113 load = source_load(i, load_idx);
4114 if (load > max_cpu_load)
4115 max_cpu_load = load;
4116 if (min_cpu_load > load)
4117 min_cpu_load = load;
4119 if (nr_running > max_nr_running)
4120 max_nr_running = nr_running;
4121 if (min_nr_running > nr_running)
4122 min_nr_running = nr_running;
4125 sgs->group_load += load;
4126 sgs->sum_nr_running += nr_running;
4127 sgs->sum_weighted_load += weighted_cpuload(i);
4133 * First idle cpu or the first cpu(busiest) in this sched group
4134 * is eligible for doing load balancing at this and above
4135 * domains. In the newly idle case, we will allow all the cpu's
4136 * to do the newly idle load balance.
4139 if (env->idle != CPU_NEWLY_IDLE) {
4140 if (balance_cpu != env->dst_cpu) {
4144 update_group_power(env->sd, env->dst_cpu);
4145 } else if (time_after_eq(jiffies, group->sgp->next_update))
4146 update_group_power(env->sd, env->dst_cpu);
4149 /* Adjust by relative CPU power of the group */
4150 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4153 * Consider the group unbalanced when the imbalance is larger
4154 * than the average weight of a task.
4156 * APZ: with cgroup the avg task weight can vary wildly and
4157 * might not be a suitable number - should we keep a
4158 * normalized nr_running number somewhere that negates
4161 if (sgs->sum_nr_running)
4162 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4164 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4165 (max_nr_running - min_nr_running) > 1)
4168 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4170 if (!sgs->group_capacity)
4171 sgs->group_capacity = fix_small_capacity(env->sd, group);
4172 sgs->group_weight = group->group_weight;
4174 if (sgs->group_capacity > sgs->sum_nr_running)
4175 sgs->group_has_capacity = 1;
4179 * update_sd_pick_busiest - return 1 on busiest group
4180 * @env: The load balancing environment.
4181 * @sds: sched_domain statistics
4182 * @sg: sched_group candidate to be checked for being the busiest
4183 * @sgs: sched_group statistics
4185 * Determine if @sg is a busier group than the previously selected
4188 static bool update_sd_pick_busiest(struct lb_env *env,
4189 struct sd_lb_stats *sds,
4190 struct sched_group *sg,
4191 struct sg_lb_stats *sgs)
4193 if (sgs->avg_load <= sds->max_load)
4196 if (sgs->sum_nr_running > sgs->group_capacity)
4203 * ASYM_PACKING needs to move all the work to the lowest
4204 * numbered CPUs in the group, therefore mark all groups
4205 * higher than ourself as busy.
4207 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4208 env->dst_cpu < group_first_cpu(sg)) {
4212 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4220 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4221 * @env: The load balancing environment.
4222 * @balance: Should we balance.
4223 * @sds: variable to hold the statistics for this sched_domain.
4225 static inline void update_sd_lb_stats(struct lb_env *env,
4226 int *balance, struct sd_lb_stats *sds)
4228 struct sched_domain *child = env->sd->child;
4229 struct sched_group *sg = env->sd->groups;
4230 struct sg_lb_stats sgs;
4231 int load_idx, prefer_sibling = 0;
4233 if (child && child->flags & SD_PREFER_SIBLING)
4236 load_idx = get_sd_load_idx(env->sd, env->idle);
4241 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4242 memset(&sgs, 0, sizeof(sgs));
4243 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4245 if (local_group && !(*balance))
4248 sds->total_load += sgs.group_load;
4249 sds->total_pwr += sg->sgp->power;
4252 * In case the child domain prefers tasks go to siblings
4253 * first, lower the sg capacity to one so that we'll try
4254 * and move all the excess tasks away. We lower the capacity
4255 * of a group only if the local group has the capacity to fit
4256 * these excess tasks, i.e. nr_running < group_capacity. The
4257 * extra check prevents the case where you always pull from the
4258 * heaviest group when it is already under-utilized (possible
4259 * with a large weight task outweighs the tasks on the system).
4261 if (prefer_sibling && !local_group && sds->this_has_capacity)
4262 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4265 sds->this_load = sgs.avg_load;
4267 sds->this_nr_running = sgs.sum_nr_running;
4268 sds->this_load_per_task = sgs.sum_weighted_load;
4269 sds->this_has_capacity = sgs.group_has_capacity;
4270 sds->this_idle_cpus = sgs.idle_cpus;
4271 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4272 sds->max_load = sgs.avg_load;
4274 sds->busiest_nr_running = sgs.sum_nr_running;
4275 sds->busiest_idle_cpus = sgs.idle_cpus;
4276 sds->busiest_group_capacity = sgs.group_capacity;
4277 sds->busiest_load_per_task = sgs.sum_weighted_load;
4278 sds->busiest_has_capacity = sgs.group_has_capacity;
4279 sds->busiest_group_weight = sgs.group_weight;
4280 sds->group_imb = sgs.group_imb;
4284 } while (sg != env->sd->groups);
4288 * check_asym_packing - Check to see if the group is packed into the
4291 * This is primarily intended to used at the sibling level. Some
4292 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4293 * case of POWER7, it can move to lower SMT modes only when higher
4294 * threads are idle. When in lower SMT modes, the threads will
4295 * perform better since they share less core resources. Hence when we
4296 * have idle threads, we want them to be the higher ones.
4298 * This packing function is run on idle threads. It checks to see if
4299 * the busiest CPU in this domain (core in the P7 case) has a higher
4300 * CPU number than the packing function is being run on. Here we are
4301 * assuming lower CPU number will be equivalent to lower a SMT thread
4304 * Returns 1 when packing is required and a task should be moved to
4305 * this CPU. The amount of the imbalance is returned in *imbalance.
4307 * @env: The load balancing environment.
4308 * @sds: Statistics of the sched_domain which is to be packed
4310 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4314 if (!(env->sd->flags & SD_ASYM_PACKING))
4320 busiest_cpu = group_first_cpu(sds->busiest);
4321 if (env->dst_cpu > busiest_cpu)
4324 env->imbalance = DIV_ROUND_CLOSEST(
4325 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4331 * fix_small_imbalance - Calculate the minor imbalance that exists
4332 * amongst the groups of a sched_domain, during
4334 * @env: The load balancing environment.
4335 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4338 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4340 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4341 unsigned int imbn = 2;
4342 unsigned long scaled_busy_load_per_task;
4344 if (sds->this_nr_running) {
4345 sds->this_load_per_task /= sds->this_nr_running;
4346 if (sds->busiest_load_per_task >
4347 sds->this_load_per_task)
4350 sds->this_load_per_task =
4351 cpu_avg_load_per_task(env->dst_cpu);
4354 scaled_busy_load_per_task = sds->busiest_load_per_task
4355 * SCHED_POWER_SCALE;
4356 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4358 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4359 (scaled_busy_load_per_task * imbn)) {
4360 env->imbalance = sds->busiest_load_per_task;
4365 * OK, we don't have enough imbalance to justify moving tasks,
4366 * however we may be able to increase total CPU power used by
4370 pwr_now += sds->busiest->sgp->power *
4371 min(sds->busiest_load_per_task, sds->max_load);
4372 pwr_now += sds->this->sgp->power *
4373 min(sds->this_load_per_task, sds->this_load);
4374 pwr_now /= SCHED_POWER_SCALE;
4376 /* Amount of load we'd subtract */
4377 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4378 sds->busiest->sgp->power;
4379 if (sds->max_load > tmp)
4380 pwr_move += sds->busiest->sgp->power *
4381 min(sds->busiest_load_per_task, sds->max_load - tmp);
4383 /* Amount of load we'd add */
4384 if (sds->max_load * sds->busiest->sgp->power <
4385 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4386 tmp = (sds->max_load * sds->busiest->sgp->power) /
4387 sds->this->sgp->power;
4389 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4390 sds->this->sgp->power;
4391 pwr_move += sds->this->sgp->power *
4392 min(sds->this_load_per_task, sds->this_load + tmp);
4393 pwr_move /= SCHED_POWER_SCALE;
4395 /* Move if we gain throughput */
4396 if (pwr_move > pwr_now)
4397 env->imbalance = sds->busiest_load_per_task;
4401 * calculate_imbalance - Calculate the amount of imbalance present within the
4402 * groups of a given sched_domain during load balance.
4403 * @env: load balance environment
4404 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4406 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4408 unsigned long max_pull, load_above_capacity = ~0UL;
4410 sds->busiest_load_per_task /= sds->busiest_nr_running;
4411 if (sds->group_imb) {
4412 sds->busiest_load_per_task =
4413 min(sds->busiest_load_per_task, sds->avg_load);
4417 * In the presence of smp nice balancing, certain scenarios can have
4418 * max load less than avg load(as we skip the groups at or below
4419 * its cpu_power, while calculating max_load..)
4421 if (sds->max_load < sds->avg_load) {
4423 return fix_small_imbalance(env, sds);
4426 if (!sds->group_imb) {
4428 * Don't want to pull so many tasks that a group would go idle.
4430 load_above_capacity = (sds->busiest_nr_running -
4431 sds->busiest_group_capacity);
4433 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4435 load_above_capacity /= sds->busiest->sgp->power;
4439 * We're trying to get all the cpus to the average_load, so we don't
4440 * want to push ourselves above the average load, nor do we wish to
4441 * reduce the max loaded cpu below the average load. At the same time,
4442 * we also don't want to reduce the group load below the group capacity
4443 * (so that we can implement power-savings policies etc). Thus we look
4444 * for the minimum possible imbalance.
4445 * Be careful of negative numbers as they'll appear as very large values
4446 * with unsigned longs.
4448 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4450 /* How much load to actually move to equalise the imbalance */
4451 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4452 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4453 / SCHED_POWER_SCALE;
4456 * if *imbalance is less than the average load per runnable task
4457 * there is no guarantee that any tasks will be moved so we'll have
4458 * a think about bumping its value to force at least one task to be
4461 if (env->imbalance < sds->busiest_load_per_task)
4462 return fix_small_imbalance(env, sds);
4466 /******* find_busiest_group() helpers end here *********************/
4469 * find_busiest_group - Returns the busiest group within the sched_domain
4470 * if there is an imbalance. If there isn't an imbalance, and
4471 * the user has opted for power-savings, it returns a group whose
4472 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4473 * such a group exists.
4475 * Also calculates the amount of weighted load which should be moved
4476 * to restore balance.
4478 * @env: The load balancing environment.
4479 * @balance: Pointer to a variable indicating if this_cpu
4480 * is the appropriate cpu to perform load balancing at this_level.
4482 * Returns: - the busiest group if imbalance exists.
4483 * - If no imbalance and user has opted for power-savings balance,
4484 * return the least loaded group whose CPUs can be
4485 * put to idle by rebalancing its tasks onto our group.
4487 static struct sched_group *
4488 find_busiest_group(struct lb_env *env, int *balance)
4490 struct sd_lb_stats sds;
4492 memset(&sds, 0, sizeof(sds));
4495 * Compute the various statistics relavent for load balancing at
4498 update_sd_lb_stats(env, balance, &sds);
4501 * this_cpu is not the appropriate cpu to perform load balancing at
4507 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4508 check_asym_packing(env, &sds))
4511 /* There is no busy sibling group to pull tasks from */
4512 if (!sds.busiest || sds.busiest_nr_running == 0)
4515 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4518 * If the busiest group is imbalanced the below checks don't
4519 * work because they assumes all things are equal, which typically
4520 * isn't true due to cpus_allowed constraints and the like.
4525 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4526 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4527 !sds.busiest_has_capacity)
4531 * If the local group is more busy than the selected busiest group
4532 * don't try and pull any tasks.
4534 if (sds.this_load >= sds.max_load)
4538 * Don't pull any tasks if this group is already above the domain
4541 if (sds.this_load >= sds.avg_load)
4544 if (env->idle == CPU_IDLE) {
4546 * This cpu is idle. If the busiest group load doesn't
4547 * have more tasks than the number of available cpu's and
4548 * there is no imbalance between this and busiest group
4549 * wrt to idle cpu's, it is balanced.
4551 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4552 sds.busiest_nr_running <= sds.busiest_group_weight)
4556 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4557 * imbalance_pct to be conservative.
4559 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4564 /* Looks like there is an imbalance. Compute it */
4565 calculate_imbalance(env, &sds);
4575 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4577 static struct rq *find_busiest_queue(struct lb_env *env,
4578 struct sched_group *group)
4580 struct rq *busiest = NULL, *rq;
4581 unsigned long max_load = 0;
4584 for_each_cpu(i, sched_group_cpus(group)) {
4585 unsigned long power = power_of(i);
4586 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4591 capacity = fix_small_capacity(env->sd, group);
4593 if (!cpumask_test_cpu(i, env->cpus))
4597 wl = weighted_cpuload(i);
4600 * When comparing with imbalance, use weighted_cpuload()
4601 * which is not scaled with the cpu power.
4603 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4607 * For the load comparisons with the other cpu's, consider
4608 * the weighted_cpuload() scaled with the cpu power, so that
4609 * the load can be moved away from the cpu that is potentially
4610 * running at a lower capacity.
4612 wl = (wl * SCHED_POWER_SCALE) / power;
4614 if (wl > max_load) {
4624 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4625 * so long as it is large enough.
4627 #define MAX_PINNED_INTERVAL 512
4629 /* Working cpumask for load_balance and load_balance_newidle. */
4630 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4632 static int need_active_balance(struct lb_env *env)
4634 struct sched_domain *sd = env->sd;
4636 if (env->idle == CPU_NEWLY_IDLE) {
4639 * ASYM_PACKING needs to force migrate tasks from busy but
4640 * higher numbered CPUs in order to pack all tasks in the
4641 * lowest numbered CPUs.
4643 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4647 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4650 static int active_load_balance_cpu_stop(void *data);
4653 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4654 * tasks if there is an imbalance.
4656 static int load_balance(int this_cpu, struct rq *this_rq,
4657 struct sched_domain *sd, enum cpu_idle_type idle,
4660 int ld_moved, cur_ld_moved, active_balance = 0;
4661 int lb_iterations, max_lb_iterations;
4662 struct sched_group *group;
4664 unsigned long flags;
4665 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4667 struct lb_env env = {
4669 .dst_cpu = this_cpu,
4671 .dst_grpmask = sched_group_cpus(sd->groups),
4673 .loop_break = sched_nr_migrate_break,
4677 cpumask_copy(cpus, cpu_active_mask);
4678 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4680 schedstat_inc(sd, lb_count[idle]);
4683 group = find_busiest_group(&env, balance);
4689 schedstat_inc(sd, lb_nobusyg[idle]);
4693 busiest = find_busiest_queue(&env, group);
4695 schedstat_inc(sd, lb_nobusyq[idle]);
4699 BUG_ON(busiest == env.dst_rq);
4701 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4705 if (busiest->nr_running > 1) {
4707 * Attempt to move tasks. If find_busiest_group has found
4708 * an imbalance but busiest->nr_running <= 1, the group is
4709 * still unbalanced. ld_moved simply stays zero, so it is
4710 * correctly treated as an imbalance.
4712 env.flags |= LBF_ALL_PINNED;
4713 env.src_cpu = busiest->cpu;
4714 env.src_rq = busiest;
4715 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4717 update_h_load(env.src_cpu);
4719 local_irq_save(flags);
4720 double_rq_lock(env.dst_rq, busiest);
4723 * cur_ld_moved - load moved in current iteration
4724 * ld_moved - cumulative load moved across iterations
4726 cur_ld_moved = move_tasks(&env);
4727 ld_moved += cur_ld_moved;
4728 double_rq_unlock(env.dst_rq, busiest);
4729 local_irq_restore(flags);
4731 if (env.flags & LBF_NEED_BREAK) {
4732 env.flags &= ~LBF_NEED_BREAK;
4737 * some other cpu did the load balance for us.
4739 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4740 resched_cpu(env.dst_cpu);
4743 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4744 * us and move them to an alternate dst_cpu in our sched_group
4745 * where they can run. The upper limit on how many times we
4746 * iterate on same src_cpu is dependent on number of cpus in our
4749 * This changes load balance semantics a bit on who can move
4750 * load to a given_cpu. In addition to the given_cpu itself
4751 * (or a ilb_cpu acting on its behalf where given_cpu is
4752 * nohz-idle), we now have balance_cpu in a position to move
4753 * load to given_cpu. In rare situations, this may cause
4754 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4755 * _independently_ and at _same_ time to move some load to
4756 * given_cpu) causing exceess load to be moved to given_cpu.
4757 * This however should not happen so much in practice and
4758 * moreover subsequent load balance cycles should correct the
4759 * excess load moved.
4761 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4762 lb_iterations++ < max_lb_iterations) {
4764 env.dst_rq = cpu_rq(env.new_dst_cpu);
4765 env.dst_cpu = env.new_dst_cpu;
4766 env.flags &= ~LBF_SOME_PINNED;
4768 env.loop_break = sched_nr_migrate_break;
4770 * Go back to "more_balance" rather than "redo" since we
4771 * need to continue with same src_cpu.
4776 /* All tasks on this runqueue were pinned by CPU affinity */
4777 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4778 cpumask_clear_cpu(cpu_of(busiest), cpus);
4779 if (!cpumask_empty(cpus)) {
4781 env.loop_break = sched_nr_migrate_break;
4789 schedstat_inc(sd, lb_failed[idle]);
4791 * Increment the failure counter only on periodic balance.
4792 * We do not want newidle balance, which can be very
4793 * frequent, pollute the failure counter causing
4794 * excessive cache_hot migrations and active balances.
4796 if (idle != CPU_NEWLY_IDLE)
4797 sd->nr_balance_failed++;
4799 if (need_active_balance(&env)) {
4800 raw_spin_lock_irqsave(&busiest->lock, flags);
4802 /* don't kick the active_load_balance_cpu_stop,
4803 * if the curr task on busiest cpu can't be
4806 if (!cpumask_test_cpu(this_cpu,
4807 tsk_cpus_allowed(busiest->curr))) {
4808 raw_spin_unlock_irqrestore(&busiest->lock,
4810 env.flags |= LBF_ALL_PINNED;
4811 goto out_one_pinned;
4815 * ->active_balance synchronizes accesses to
4816 * ->active_balance_work. Once set, it's cleared
4817 * only after active load balance is finished.
4819 if (!busiest->active_balance) {
4820 busiest->active_balance = 1;
4821 busiest->push_cpu = this_cpu;
4824 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4826 if (active_balance) {
4827 stop_one_cpu_nowait(cpu_of(busiest),
4828 active_load_balance_cpu_stop, busiest,
4829 &busiest->active_balance_work);
4833 * We've kicked active balancing, reset the failure
4836 sd->nr_balance_failed = sd->cache_nice_tries+1;
4839 sd->nr_balance_failed = 0;
4841 if (likely(!active_balance)) {
4842 /* We were unbalanced, so reset the balancing interval */
4843 sd->balance_interval = sd->min_interval;
4846 * If we've begun active balancing, start to back off. This
4847 * case may not be covered by the all_pinned logic if there
4848 * is only 1 task on the busy runqueue (because we don't call
4851 if (sd->balance_interval < sd->max_interval)
4852 sd->balance_interval *= 2;
4858 schedstat_inc(sd, lb_balanced[idle]);
4860 sd->nr_balance_failed = 0;
4863 /* tune up the balancing interval */
4864 if (((env.flags & LBF_ALL_PINNED) &&
4865 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4866 (sd->balance_interval < sd->max_interval))
4867 sd->balance_interval *= 2;
4875 * idle_balance is called by schedule() if this_cpu is about to become
4876 * idle. Attempts to pull tasks from other CPUs.
4878 void idle_balance(int this_cpu, struct rq *this_rq)
4880 struct sched_domain *sd;
4881 int pulled_task = 0;
4882 unsigned long next_balance = jiffies + HZ;
4884 this_rq->idle_stamp = this_rq->clock;
4886 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4889 update_rq_runnable_avg(this_rq, 1);
4892 * Drop the rq->lock, but keep IRQ/preempt disabled.
4894 raw_spin_unlock(&this_rq->lock);
4896 update_blocked_averages(this_cpu);
4898 for_each_domain(this_cpu, sd) {
4899 unsigned long interval;
4902 if (!(sd->flags & SD_LOAD_BALANCE))
4905 if (sd->flags & SD_BALANCE_NEWIDLE) {
4906 /* If we've pulled tasks over stop searching: */
4907 pulled_task = load_balance(this_cpu, this_rq,
4908 sd, CPU_NEWLY_IDLE, &balance);
4911 interval = msecs_to_jiffies(sd->balance_interval);
4912 if (time_after(next_balance, sd->last_balance + interval))
4913 next_balance = sd->last_balance + interval;
4915 this_rq->idle_stamp = 0;
4921 raw_spin_lock(&this_rq->lock);
4923 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4925 * We are going idle. next_balance may be set based on
4926 * a busy processor. So reset next_balance.
4928 this_rq->next_balance = next_balance;
4933 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4934 * running tasks off the busiest CPU onto idle CPUs. It requires at
4935 * least 1 task to be running on each physical CPU where possible, and
4936 * avoids physical / logical imbalances.
4938 static int active_load_balance_cpu_stop(void *data)
4940 struct rq *busiest_rq = data;
4941 int busiest_cpu = cpu_of(busiest_rq);
4942 int target_cpu = busiest_rq->push_cpu;
4943 struct rq *target_rq = cpu_rq(target_cpu);
4944 struct sched_domain *sd;
4946 raw_spin_lock_irq(&busiest_rq->lock);
4948 /* make sure the requested cpu hasn't gone down in the meantime */
4949 if (unlikely(busiest_cpu != smp_processor_id() ||
4950 !busiest_rq->active_balance))
4953 /* Is there any task to move? */
4954 if (busiest_rq->nr_running <= 1)
4958 * This condition is "impossible", if it occurs
4959 * we need to fix it. Originally reported by
4960 * Bjorn Helgaas on a 128-cpu setup.
4962 BUG_ON(busiest_rq == target_rq);
4964 /* move a task from busiest_rq to target_rq */
4965 double_lock_balance(busiest_rq, target_rq);
4967 /* Search for an sd spanning us and the target CPU. */
4969 for_each_domain(target_cpu, sd) {
4970 if ((sd->flags & SD_LOAD_BALANCE) &&
4971 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4976 struct lb_env env = {
4978 .dst_cpu = target_cpu,
4979 .dst_rq = target_rq,
4980 .src_cpu = busiest_rq->cpu,
4981 .src_rq = busiest_rq,
4985 schedstat_inc(sd, alb_count);
4987 if (move_one_task(&env))
4988 schedstat_inc(sd, alb_pushed);
4990 schedstat_inc(sd, alb_failed);
4993 double_unlock_balance(busiest_rq, target_rq);
4995 busiest_rq->active_balance = 0;
4996 raw_spin_unlock_irq(&busiest_rq->lock);
5002 * idle load balancing details
5003 * - When one of the busy CPUs notice that there may be an idle rebalancing
5004 * needed, they will kick the idle load balancer, which then does idle
5005 * load balancing for all the idle CPUs.
5008 cpumask_var_t idle_cpus_mask;
5010 unsigned long next_balance; /* in jiffy units */
5011 } nohz ____cacheline_aligned;
5013 static inline int find_new_ilb(int call_cpu)
5015 int ilb = cpumask_first(nohz.idle_cpus_mask);
5017 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5024 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5025 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5026 * CPU (if there is one).
5028 static void nohz_balancer_kick(int cpu)
5032 nohz.next_balance++;
5034 ilb_cpu = find_new_ilb(cpu);
5036 if (ilb_cpu >= nr_cpu_ids)
5039 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5042 * Use smp_send_reschedule() instead of resched_cpu().
5043 * This way we generate a sched IPI on the target cpu which
5044 * is idle. And the softirq performing nohz idle load balance
5045 * will be run before returning from the IPI.
5047 smp_send_reschedule(ilb_cpu);
5051 static inline void nohz_balance_exit_idle(int cpu)
5053 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5054 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5055 atomic_dec(&nohz.nr_cpus);
5056 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5060 static inline void set_cpu_sd_state_busy(void)
5062 struct sched_domain *sd;
5063 int cpu = smp_processor_id();
5065 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5067 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
5070 for_each_domain(cpu, sd)
5071 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5075 void set_cpu_sd_state_idle(void)
5077 struct sched_domain *sd;
5078 int cpu = smp_processor_id();
5080 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5082 set_bit(NOHZ_IDLE, nohz_flags(cpu));
5085 for_each_domain(cpu, sd)
5086 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5091 * This routine will record that the cpu is going idle with tick stopped.
5092 * This info will be used in performing idle load balancing in the future.
5094 void nohz_balance_enter_idle(int cpu)
5097 * If this cpu is going down, then nothing needs to be done.
5099 if (!cpu_active(cpu))
5102 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5105 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5106 atomic_inc(&nohz.nr_cpus);
5107 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5110 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5111 unsigned long action, void *hcpu)
5113 switch (action & ~CPU_TASKS_FROZEN) {
5115 nohz_balance_exit_idle(smp_processor_id());
5123 static DEFINE_SPINLOCK(balancing);
5126 * Scale the max load_balance interval with the number of CPUs in the system.
5127 * This trades load-balance latency on larger machines for less cross talk.
5129 void update_max_interval(void)
5131 max_load_balance_interval = HZ*num_online_cpus()/10;
5135 * It checks each scheduling domain to see if it is due to be balanced,
5136 * and initiates a balancing operation if so.
5138 * Balancing parameters are set up in arch_init_sched_domains.
5140 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5143 struct rq *rq = cpu_rq(cpu);
5144 unsigned long interval;
5145 struct sched_domain *sd;
5146 /* Earliest time when we have to do rebalance again */
5147 unsigned long next_balance = jiffies + 60*HZ;
5148 int update_next_balance = 0;
5151 update_blocked_averages(cpu);
5154 for_each_domain(cpu, sd) {
5155 if (!(sd->flags & SD_LOAD_BALANCE))
5158 interval = sd->balance_interval;
5159 if (idle != CPU_IDLE)
5160 interval *= sd->busy_factor;
5162 /* scale ms to jiffies */
5163 interval = msecs_to_jiffies(interval);
5164 interval = clamp(interval, 1UL, max_load_balance_interval);
5166 need_serialize = sd->flags & SD_SERIALIZE;
5168 if (need_serialize) {
5169 if (!spin_trylock(&balancing))
5173 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5174 if (load_balance(cpu, rq, sd, idle, &balance)) {
5176 * We've pulled tasks over so either we're no
5179 idle = CPU_NOT_IDLE;
5181 sd->last_balance = jiffies;
5184 spin_unlock(&balancing);
5186 if (time_after(next_balance, sd->last_balance + interval)) {
5187 next_balance = sd->last_balance + interval;
5188 update_next_balance = 1;
5192 * Stop the load balance at this level. There is another
5193 * CPU in our sched group which is doing load balancing more
5202 * next_balance will be updated only when there is a need.
5203 * When the cpu is attached to null domain for ex, it will not be
5206 if (likely(update_next_balance))
5207 rq->next_balance = next_balance;
5212 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5213 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5215 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5217 struct rq *this_rq = cpu_rq(this_cpu);
5221 if (idle != CPU_IDLE ||
5222 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5225 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5226 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5230 * If this cpu gets work to do, stop the load balancing
5231 * work being done for other cpus. Next load
5232 * balancing owner will pick it up.
5237 rq = cpu_rq(balance_cpu);
5239 raw_spin_lock_irq(&rq->lock);
5240 update_rq_clock(rq);
5241 update_idle_cpu_load(rq);
5242 raw_spin_unlock_irq(&rq->lock);
5244 rebalance_domains(balance_cpu, CPU_IDLE);
5246 if (time_after(this_rq->next_balance, rq->next_balance))
5247 this_rq->next_balance = rq->next_balance;
5249 nohz.next_balance = this_rq->next_balance;
5251 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5255 * Current heuristic for kicking the idle load balancer in the presence
5256 * of an idle cpu is the system.
5257 * - This rq has more than one task.
5258 * - At any scheduler domain level, this cpu's scheduler group has multiple
5259 * busy cpu's exceeding the group's power.
5260 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5261 * domain span are idle.
5263 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5265 unsigned long now = jiffies;
5266 struct sched_domain *sd;
5268 if (unlikely(idle_cpu(cpu)))
5272 * We may be recently in ticked or tickless idle mode. At the first
5273 * busy tick after returning from idle, we will update the busy stats.
5275 set_cpu_sd_state_busy();
5276 nohz_balance_exit_idle(cpu);
5279 * None are in tickless mode and hence no need for NOHZ idle load
5282 if (likely(!atomic_read(&nohz.nr_cpus)))
5285 if (time_before(now, nohz.next_balance))
5288 if (rq->nr_running >= 2)
5292 for_each_domain(cpu, sd) {
5293 struct sched_group *sg = sd->groups;
5294 struct sched_group_power *sgp = sg->sgp;
5295 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5297 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5298 goto need_kick_unlock;
5300 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5301 && (cpumask_first_and(nohz.idle_cpus_mask,
5302 sched_domain_span(sd)) < cpu))
5303 goto need_kick_unlock;
5305 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5317 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5321 * run_rebalance_domains is triggered when needed from the scheduler tick.
5322 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5324 static void run_rebalance_domains(struct softirq_action *h)
5326 int this_cpu = smp_processor_id();
5327 struct rq *this_rq = cpu_rq(this_cpu);
5328 enum cpu_idle_type idle = this_rq->idle_balance ?
5329 CPU_IDLE : CPU_NOT_IDLE;
5331 rebalance_domains(this_cpu, idle);
5334 * If this cpu has a pending nohz_balance_kick, then do the
5335 * balancing on behalf of the other idle cpus whose ticks are
5338 nohz_idle_balance(this_cpu, idle);
5341 static inline int on_null_domain(int cpu)
5343 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5347 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5349 void trigger_load_balance(struct rq *rq, int cpu)
5351 /* Don't need to rebalance while attached to NULL domain */
5352 if (time_after_eq(jiffies, rq->next_balance) &&
5353 likely(!on_null_domain(cpu)))
5354 raise_softirq(SCHED_SOFTIRQ);
5356 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5357 nohz_balancer_kick(cpu);
5361 static void rq_online_fair(struct rq *rq)
5366 static void rq_offline_fair(struct rq *rq)
5370 /* Ensure any throttled groups are reachable by pick_next_task */
5371 unthrottle_offline_cfs_rqs(rq);
5374 #endif /* CONFIG_SMP */
5377 * scheduler tick hitting a task of our scheduling class:
5379 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5381 struct cfs_rq *cfs_rq;
5382 struct sched_entity *se = &curr->se;
5384 for_each_sched_entity(se) {
5385 cfs_rq = cfs_rq_of(se);
5386 entity_tick(cfs_rq, se, queued);
5389 update_rq_runnable_avg(rq, 1);
5393 * called on fork with the child task as argument from the parent's context
5394 * - child not yet on the tasklist
5395 * - preemption disabled
5397 static void task_fork_fair(struct task_struct *p)
5399 struct cfs_rq *cfs_rq;
5400 struct sched_entity *se = &p->se, *curr;
5401 int this_cpu = smp_processor_id();
5402 struct rq *rq = this_rq();
5403 unsigned long flags;
5405 raw_spin_lock_irqsave(&rq->lock, flags);
5407 update_rq_clock(rq);
5409 cfs_rq = task_cfs_rq(current);
5410 curr = cfs_rq->curr;
5412 if (unlikely(task_cpu(p) != this_cpu)) {
5414 __set_task_cpu(p, this_cpu);
5418 update_curr(cfs_rq);
5421 se->vruntime = curr->vruntime;
5422 place_entity(cfs_rq, se, 1);
5424 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5426 * Upon rescheduling, sched_class::put_prev_task() will place
5427 * 'current' within the tree based on its new key value.
5429 swap(curr->vruntime, se->vruntime);
5430 resched_task(rq->curr);
5433 se->vruntime -= cfs_rq->min_vruntime;
5435 raw_spin_unlock_irqrestore(&rq->lock, flags);
5439 * Priority of the task has changed. Check to see if we preempt
5443 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5449 * Reschedule if we are currently running on this runqueue and
5450 * our priority decreased, or if we are not currently running on
5451 * this runqueue and our priority is higher than the current's
5453 if (rq->curr == p) {
5454 if (p->prio > oldprio)
5455 resched_task(rq->curr);
5457 check_preempt_curr(rq, p, 0);
5460 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5462 struct sched_entity *se = &p->se;
5463 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5466 * Ensure the task's vruntime is normalized, so that when its
5467 * switched back to the fair class the enqueue_entity(.flags=0) will
5468 * do the right thing.
5470 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5471 * have normalized the vruntime, if it was !on_rq, then only when
5472 * the task is sleeping will it still have non-normalized vruntime.
5474 if (!se->on_rq && p->state != TASK_RUNNING) {
5476 * Fix up our vruntime so that the current sleep doesn't
5477 * cause 'unlimited' sleep bonus.
5479 place_entity(cfs_rq, se, 0);
5480 se->vruntime -= cfs_rq->min_vruntime;
5483 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5485 * Remove our load from contribution when we leave sched_fair
5486 * and ensure we don't carry in an old decay_count if we
5489 if (p->se.avg.decay_count) {
5490 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
5491 __synchronize_entity_decay(&p->se);
5492 subtract_blocked_load_contrib(cfs_rq,
5493 p->se.avg.load_avg_contrib);
5499 * We switched to the sched_fair class.
5501 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5507 * We were most likely switched from sched_rt, so
5508 * kick off the schedule if running, otherwise just see
5509 * if we can still preempt the current task.
5512 resched_task(rq->curr);
5514 check_preempt_curr(rq, p, 0);
5517 /* Account for a task changing its policy or group.
5519 * This routine is mostly called to set cfs_rq->curr field when a task
5520 * migrates between groups/classes.
5522 static void set_curr_task_fair(struct rq *rq)
5524 struct sched_entity *se = &rq->curr->se;
5526 for_each_sched_entity(se) {
5527 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5529 set_next_entity(cfs_rq, se);
5530 /* ensure bandwidth has been allocated on our new cfs_rq */
5531 account_cfs_rq_runtime(cfs_rq, 0);
5535 void init_cfs_rq(struct cfs_rq *cfs_rq)
5537 cfs_rq->tasks_timeline = RB_ROOT;
5538 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5539 #ifndef CONFIG_64BIT
5540 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5542 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5543 atomic64_set(&cfs_rq->decay_counter, 1);
5544 atomic64_set(&cfs_rq->removed_load, 0);
5548 #ifdef CONFIG_FAIR_GROUP_SCHED
5549 static void task_move_group_fair(struct task_struct *p, int on_rq)
5551 struct cfs_rq *cfs_rq;
5553 * If the task was not on the rq at the time of this cgroup movement
5554 * it must have been asleep, sleeping tasks keep their ->vruntime
5555 * absolute on their old rq until wakeup (needed for the fair sleeper
5556 * bonus in place_entity()).
5558 * If it was on the rq, we've just 'preempted' it, which does convert
5559 * ->vruntime to a relative base.
5561 * Make sure both cases convert their relative position when migrating
5562 * to another cgroup's rq. This does somewhat interfere with the
5563 * fair sleeper stuff for the first placement, but who cares.
5566 * When !on_rq, vruntime of the task has usually NOT been normalized.
5567 * But there are some cases where it has already been normalized:
5569 * - Moving a forked child which is waiting for being woken up by
5570 * wake_up_new_task().
5571 * - Moving a task which has been woken up by try_to_wake_up() and
5572 * waiting for actually being woken up by sched_ttwu_pending().
5574 * To prevent boost or penalty in the new cfs_rq caused by delta
5575 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5577 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5581 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5582 set_task_rq(p, task_cpu(p));
5584 cfs_rq = cfs_rq_of(&p->se);
5585 p->se.vruntime += cfs_rq->min_vruntime;
5588 * migrate_task_rq_fair() will have removed our previous
5589 * contribution, but we must synchronize for ongoing future
5592 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
5593 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
5598 void free_fair_sched_group(struct task_group *tg)
5602 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5604 for_each_possible_cpu(i) {
5606 kfree(tg->cfs_rq[i]);
5615 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5617 struct cfs_rq *cfs_rq;
5618 struct sched_entity *se;
5621 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5624 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5628 tg->shares = NICE_0_LOAD;
5630 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5632 for_each_possible_cpu(i) {
5633 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5634 GFP_KERNEL, cpu_to_node(i));
5638 se = kzalloc_node(sizeof(struct sched_entity),
5639 GFP_KERNEL, cpu_to_node(i));
5643 init_cfs_rq(cfs_rq);
5644 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5655 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5657 struct rq *rq = cpu_rq(cpu);
5658 unsigned long flags;
5661 * Only empty task groups can be destroyed; so we can speculatively
5662 * check on_list without danger of it being re-added.
5664 if (!tg->cfs_rq[cpu]->on_list)
5667 raw_spin_lock_irqsave(&rq->lock, flags);
5668 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5669 raw_spin_unlock_irqrestore(&rq->lock, flags);
5672 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5673 struct sched_entity *se, int cpu,
5674 struct sched_entity *parent)
5676 struct rq *rq = cpu_rq(cpu);
5680 init_cfs_rq_runtime(cfs_rq);
5682 tg->cfs_rq[cpu] = cfs_rq;
5685 /* se could be NULL for root_task_group */
5690 se->cfs_rq = &rq->cfs;
5692 se->cfs_rq = parent->my_q;
5695 update_load_set(&se->load, 0);
5696 se->parent = parent;
5699 static DEFINE_MUTEX(shares_mutex);
5701 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5704 unsigned long flags;
5707 * We can't change the weight of the root cgroup.
5712 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5714 mutex_lock(&shares_mutex);
5715 if (tg->shares == shares)
5718 tg->shares = shares;
5719 for_each_possible_cpu(i) {
5720 struct rq *rq = cpu_rq(i);
5721 struct sched_entity *se;
5724 /* Propagate contribution to hierarchy */
5725 raw_spin_lock_irqsave(&rq->lock, flags);
5726 for_each_sched_entity(se) {
5727 update_cfs_shares(group_cfs_rq(se));
5728 /* update contribution to parent */
5729 update_entity_load_avg(se, 1);
5731 raw_spin_unlock_irqrestore(&rq->lock, flags);
5735 mutex_unlock(&shares_mutex);
5738 #else /* CONFIG_FAIR_GROUP_SCHED */
5740 void free_fair_sched_group(struct task_group *tg) { }
5742 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5747 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5749 #endif /* CONFIG_FAIR_GROUP_SCHED */
5752 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5754 struct sched_entity *se = &task->se;
5755 unsigned int rr_interval = 0;
5758 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5761 if (rq->cfs.load.weight)
5762 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5768 * All the scheduling class methods:
5770 const struct sched_class fair_sched_class = {
5771 .next = &idle_sched_class,
5772 .enqueue_task = enqueue_task_fair,
5773 .dequeue_task = dequeue_task_fair,
5774 .yield_task = yield_task_fair,
5775 .yield_to_task = yield_to_task_fair,
5777 .check_preempt_curr = check_preempt_wakeup,
5779 .pick_next_task = pick_next_task_fair,
5780 .put_prev_task = put_prev_task_fair,
5783 .select_task_rq = select_task_rq_fair,
5784 #ifdef CONFIG_FAIR_GROUP_SCHED
5785 .migrate_task_rq = migrate_task_rq_fair,
5787 .rq_online = rq_online_fair,
5788 .rq_offline = rq_offline_fair,
5790 .task_waking = task_waking_fair,
5793 .set_curr_task = set_curr_task_fair,
5794 .task_tick = task_tick_fair,
5795 .task_fork = task_fork_fair,
5797 .prio_changed = prio_changed_fair,
5798 .switched_from = switched_from_fair,
5799 .switched_to = switched_to_fair,
5801 .get_rr_interval = get_rr_interval_fair,
5803 #ifdef CONFIG_FAIR_GROUP_SCHED
5804 .task_move_group = task_move_group_fair,
5808 #ifdef CONFIG_SCHED_DEBUG
5809 void print_cfs_stats(struct seq_file *m, int cpu)
5811 struct cfs_rq *cfs_rq;
5814 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5815 print_cfs_rq(m, cpu, cfs_rq);
5820 __init void init_sched_fair_class(void)
5823 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5826 nohz.next_balance = jiffies;
5827 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5828 cpu_notifier(sched_ilb_notifier, 0);