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);
1270 static inline void update_rq_runnable_avg(struct rq *rq, int runnable)
1272 __update_entity_runnable_avg(rq->clock_task, &rq->avg, runnable);
1273 __update_tg_runnable_avg(&rq->avg, &rq->cfs);
1276 /* Add the load generated by se into cfs_rq's child load-average */
1277 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1278 struct sched_entity *se,
1282 * We track migrations using entity decay_count <= 0, on a wake-up
1283 * migration we use a negative decay count to track the remote decays
1284 * accumulated while sleeping.
1286 if (unlikely(se->avg.decay_count <= 0)) {
1287 se->avg.last_runnable_update = rq_of(cfs_rq)->clock_task;
1288 if (se->avg.decay_count) {
1290 * In a wake-up migration we have to approximate the
1291 * time sleeping. This is because we can't synchronize
1292 * clock_task between the two cpus, and it is not
1293 * guaranteed to be read-safe. Instead, we can
1294 * approximate this using our carried decays, which are
1295 * explicitly atomically readable.
1297 se->avg.last_runnable_update -= (-se->avg.decay_count)
1299 update_entity_load_avg(se, 0);
1300 /* Indicate that we're now synchronized and on-rq */
1301 se->avg.decay_count = 0;
1305 __synchronize_entity_decay(se);
1308 /* migrated tasks did not contribute to our blocked load */
1310 subtract_blocked_load_contrib(cfs_rq, se->avg.load_avg_contrib);
1311 update_entity_load_avg(se, 0);
1314 cfs_rq->runnable_load_avg += se->avg.load_avg_contrib;
1315 /* we force update consideration on load-balancer moves */
1316 update_cfs_rq_blocked_load(cfs_rq, !wakeup);
1320 * Remove se's load from this cfs_rq child load-average, if the entity is
1321 * transitioning to a blocked state we track its projected decay using
1324 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1325 struct sched_entity *se,
1328 update_entity_load_avg(se, 1);
1329 /* we force update consideration on load-balancer moves */
1330 update_cfs_rq_blocked_load(cfs_rq, !sleep);
1332 cfs_rq->runnable_load_avg -= se->avg.load_avg_contrib;
1334 cfs_rq->blocked_load_avg += se->avg.load_avg_contrib;
1335 se->avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
1336 } /* migrations, e.g. sleep=0 leave decay_count == 0 */
1339 static inline void update_entity_load_avg(struct sched_entity *se,
1340 int update_cfs_rq) {}
1341 static inline void update_rq_runnable_avg(struct rq *rq, int runnable) {}
1342 static inline void enqueue_entity_load_avg(struct cfs_rq *cfs_rq,
1343 struct sched_entity *se,
1345 static inline void dequeue_entity_load_avg(struct cfs_rq *cfs_rq,
1346 struct sched_entity *se,
1348 static inline void update_cfs_rq_blocked_load(struct cfs_rq *cfs_rq,
1349 int force_update) {}
1352 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
1354 #ifdef CONFIG_SCHEDSTATS
1355 struct task_struct *tsk = NULL;
1357 if (entity_is_task(se))
1360 if (se->statistics.sleep_start) {
1361 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
1366 if (unlikely(delta > se->statistics.sleep_max))
1367 se->statistics.sleep_max = delta;
1369 se->statistics.sleep_start = 0;
1370 se->statistics.sum_sleep_runtime += delta;
1373 account_scheduler_latency(tsk, delta >> 10, 1);
1374 trace_sched_stat_sleep(tsk, delta);
1377 if (se->statistics.block_start) {
1378 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1383 if (unlikely(delta > se->statistics.block_max))
1384 se->statistics.block_max = delta;
1386 se->statistics.block_start = 0;
1387 se->statistics.sum_sleep_runtime += delta;
1390 if (tsk->in_iowait) {
1391 se->statistics.iowait_sum += delta;
1392 se->statistics.iowait_count++;
1393 trace_sched_stat_iowait(tsk, delta);
1396 trace_sched_stat_blocked(tsk, delta);
1399 * Blocking time is in units of nanosecs, so shift by
1400 * 20 to get a milliseconds-range estimation of the
1401 * amount of time that the task spent sleeping:
1403 if (unlikely(prof_on == SLEEP_PROFILING)) {
1404 profile_hits(SLEEP_PROFILING,
1405 (void *)get_wchan(tsk),
1408 account_scheduler_latency(tsk, delta >> 10, 0);
1414 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1416 #ifdef CONFIG_SCHED_DEBUG
1417 s64 d = se->vruntime - cfs_rq->min_vruntime;
1422 if (d > 3*sysctl_sched_latency)
1423 schedstat_inc(cfs_rq, nr_spread_over);
1428 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1430 u64 vruntime = cfs_rq->min_vruntime;
1433 * The 'current' period is already promised to the current tasks,
1434 * however the extra weight of the new task will slow them down a
1435 * little, place the new task so that it fits in the slot that
1436 * stays open at the end.
1438 if (initial && sched_feat(START_DEBIT))
1439 vruntime += sched_vslice(cfs_rq, se);
1441 /* sleeps up to a single latency don't count. */
1443 unsigned long thresh = sysctl_sched_latency;
1446 * Halve their sleep time's effect, to allow
1447 * for a gentler effect of sleepers:
1449 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1455 /* ensure we never gain time by being placed backwards. */
1456 vruntime = max_vruntime(se->vruntime, vruntime);
1458 se->vruntime = vruntime;
1461 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1464 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1467 * Update the normalized vruntime before updating min_vruntime
1468 * through callig update_curr().
1470 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1471 se->vruntime += cfs_rq->min_vruntime;
1474 * Update run-time statistics of the 'current'.
1476 update_curr(cfs_rq);
1477 enqueue_entity_load_avg(cfs_rq, se, flags & ENQUEUE_WAKEUP);
1478 account_entity_enqueue(cfs_rq, se);
1479 update_cfs_shares(cfs_rq);
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);
1552 dequeue_entity_load_avg(cfs_rq, se, flags & DEQUEUE_SLEEP);
1554 update_stats_dequeue(cfs_rq, se);
1555 if (flags & DEQUEUE_SLEEP) {
1556 #ifdef CONFIG_SCHEDSTATS
1557 if (entity_is_task(se)) {
1558 struct task_struct *tsk = task_of(se);
1560 if (tsk->state & TASK_INTERRUPTIBLE)
1561 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1562 if (tsk->state & TASK_UNINTERRUPTIBLE)
1563 se->statistics.block_start = rq_of(cfs_rq)->clock;
1568 clear_buddies(cfs_rq, se);
1570 if (se != cfs_rq->curr)
1571 __dequeue_entity(cfs_rq, se);
1573 account_entity_dequeue(cfs_rq, se);
1576 * Normalize the entity after updating the min_vruntime because the
1577 * update can refer to the ->curr item and we need to reflect this
1578 * movement in our normalized position.
1580 if (!(flags & DEQUEUE_SLEEP))
1581 se->vruntime -= cfs_rq->min_vruntime;
1583 /* return excess runtime on last dequeue */
1584 return_cfs_rq_runtime(cfs_rq);
1586 update_min_vruntime(cfs_rq);
1587 update_cfs_shares(cfs_rq);
1591 * Preempt the current task with a newly woken task if needed:
1594 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1596 unsigned long ideal_runtime, delta_exec;
1597 struct sched_entity *se;
1600 ideal_runtime = sched_slice(cfs_rq, curr);
1601 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1602 if (delta_exec > ideal_runtime) {
1603 resched_task(rq_of(cfs_rq)->curr);
1605 * The current task ran long enough, ensure it doesn't get
1606 * re-elected due to buddy favours.
1608 clear_buddies(cfs_rq, curr);
1613 * Ensure that a task that missed wakeup preemption by a
1614 * narrow margin doesn't have to wait for a full slice.
1615 * This also mitigates buddy induced latencies under load.
1617 if (delta_exec < sysctl_sched_min_granularity)
1620 se = __pick_first_entity(cfs_rq);
1621 delta = curr->vruntime - se->vruntime;
1626 if (delta > ideal_runtime)
1627 resched_task(rq_of(cfs_rq)->curr);
1631 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1633 /* 'current' is not kept within the tree. */
1636 * Any task has to be enqueued before it get to execute on
1637 * a CPU. So account for the time it spent waiting on the
1640 update_stats_wait_end(cfs_rq, se);
1641 __dequeue_entity(cfs_rq, se);
1644 update_stats_curr_start(cfs_rq, se);
1646 #ifdef CONFIG_SCHEDSTATS
1648 * Track our maximum slice length, if the CPU's load is at
1649 * least twice that of our own weight (i.e. dont track it
1650 * when there are only lesser-weight tasks around):
1652 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1653 se->statistics.slice_max = max(se->statistics.slice_max,
1654 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1657 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1661 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1664 * Pick the next process, keeping these things in mind, in this order:
1665 * 1) keep things fair between processes/task groups
1666 * 2) pick the "next" process, since someone really wants that to run
1667 * 3) pick the "last" process, for cache locality
1668 * 4) do not run the "skip" process, if something else is available
1670 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1672 struct sched_entity *se = __pick_first_entity(cfs_rq);
1673 struct sched_entity *left = se;
1676 * Avoid running the skip buddy, if running something else can
1677 * be done without getting too unfair.
1679 if (cfs_rq->skip == se) {
1680 struct sched_entity *second = __pick_next_entity(se);
1681 if (second && wakeup_preempt_entity(second, left) < 1)
1686 * Prefer last buddy, try to return the CPU to a preempted task.
1688 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1692 * Someone really wants this to run. If it's not unfair, run it.
1694 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1697 clear_buddies(cfs_rq, se);
1702 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1704 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1707 * If still on the runqueue then deactivate_task()
1708 * was not called and update_curr() has to be done:
1711 update_curr(cfs_rq);
1713 /* throttle cfs_rqs exceeding runtime */
1714 check_cfs_rq_runtime(cfs_rq);
1716 check_spread(cfs_rq, prev);
1718 update_stats_wait_start(cfs_rq, prev);
1719 /* Put 'current' back into the tree. */
1720 __enqueue_entity(cfs_rq, prev);
1721 /* in !on_rq case, update occurred at dequeue */
1722 update_entity_load_avg(prev, 1);
1724 cfs_rq->curr = NULL;
1728 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1731 * Update run-time statistics of the 'current'.
1733 update_curr(cfs_rq);
1736 * Ensure that runnable average is periodically updated.
1738 update_entity_load_avg(curr, 1);
1739 update_cfs_rq_blocked_load(cfs_rq, 1);
1741 #ifdef CONFIG_SCHED_HRTICK
1743 * queued ticks are scheduled to match the slice, so don't bother
1744 * validating it and just reschedule.
1747 resched_task(rq_of(cfs_rq)->curr);
1751 * don't let the period tick interfere with the hrtick preemption
1753 if (!sched_feat(DOUBLE_TICK) &&
1754 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1758 if (cfs_rq->nr_running > 1)
1759 check_preempt_tick(cfs_rq, curr);
1763 /**************************************************
1764 * CFS bandwidth control machinery
1767 #ifdef CONFIG_CFS_BANDWIDTH
1769 #ifdef HAVE_JUMP_LABEL
1770 static struct static_key __cfs_bandwidth_used;
1772 static inline bool cfs_bandwidth_used(void)
1774 return static_key_false(&__cfs_bandwidth_used);
1777 void account_cfs_bandwidth_used(int enabled, int was_enabled)
1779 /* only need to count groups transitioning between enabled/!enabled */
1780 if (enabled && !was_enabled)
1781 static_key_slow_inc(&__cfs_bandwidth_used);
1782 else if (!enabled && was_enabled)
1783 static_key_slow_dec(&__cfs_bandwidth_used);
1785 #else /* HAVE_JUMP_LABEL */
1786 static bool cfs_bandwidth_used(void)
1791 void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1792 #endif /* HAVE_JUMP_LABEL */
1795 * default period for cfs group bandwidth.
1796 * default: 0.1s, units: nanoseconds
1798 static inline u64 default_cfs_period(void)
1800 return 100000000ULL;
1803 static inline u64 sched_cfs_bandwidth_slice(void)
1805 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1809 * Replenish runtime according to assigned quota and update expiration time.
1810 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1811 * additional synchronization around rq->lock.
1813 * requires cfs_b->lock
1815 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1819 if (cfs_b->quota == RUNTIME_INF)
1822 now = sched_clock_cpu(smp_processor_id());
1823 cfs_b->runtime = cfs_b->quota;
1824 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1827 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1829 return &tg->cfs_bandwidth;
1832 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
1833 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
1835 if (unlikely(cfs_rq->throttle_count))
1836 return cfs_rq->throttled_clock_task;
1838 return rq_of(cfs_rq)->clock_task - cfs_rq->throttled_clock_task_time;
1841 /* returns 0 on failure to allocate runtime */
1842 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1844 struct task_group *tg = cfs_rq->tg;
1845 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1846 u64 amount = 0, min_amount, expires;
1848 /* note: this is a positive sum as runtime_remaining <= 0 */
1849 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1851 raw_spin_lock(&cfs_b->lock);
1852 if (cfs_b->quota == RUNTIME_INF)
1853 amount = min_amount;
1856 * If the bandwidth pool has become inactive, then at least one
1857 * period must have elapsed since the last consumption.
1858 * Refresh the global state and ensure bandwidth timer becomes
1861 if (!cfs_b->timer_active) {
1862 __refill_cfs_bandwidth_runtime(cfs_b);
1863 __start_cfs_bandwidth(cfs_b);
1866 if (cfs_b->runtime > 0) {
1867 amount = min(cfs_b->runtime, min_amount);
1868 cfs_b->runtime -= amount;
1872 expires = cfs_b->runtime_expires;
1873 raw_spin_unlock(&cfs_b->lock);
1875 cfs_rq->runtime_remaining += amount;
1877 * we may have advanced our local expiration to account for allowed
1878 * spread between our sched_clock and the one on which runtime was
1881 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1882 cfs_rq->runtime_expires = expires;
1884 return cfs_rq->runtime_remaining > 0;
1888 * Note: This depends on the synchronization provided by sched_clock and the
1889 * fact that rq->clock snapshots this value.
1891 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1893 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1894 struct rq *rq = rq_of(cfs_rq);
1896 /* if the deadline is ahead of our clock, nothing to do */
1897 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1900 if (cfs_rq->runtime_remaining < 0)
1904 * If the local deadline has passed we have to consider the
1905 * possibility that our sched_clock is 'fast' and the global deadline
1906 * has not truly expired.
1908 * Fortunately we can check determine whether this the case by checking
1909 * whether the global deadline has advanced.
1912 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1913 /* extend local deadline, drift is bounded above by 2 ticks */
1914 cfs_rq->runtime_expires += TICK_NSEC;
1916 /* global deadline is ahead, expiration has passed */
1917 cfs_rq->runtime_remaining = 0;
1921 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1922 unsigned long delta_exec)
1924 /* dock delta_exec before expiring quota (as it could span periods) */
1925 cfs_rq->runtime_remaining -= delta_exec;
1926 expire_cfs_rq_runtime(cfs_rq);
1928 if (likely(cfs_rq->runtime_remaining > 0))
1932 * if we're unable to extend our runtime we resched so that the active
1933 * hierarchy can be throttled
1935 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1936 resched_task(rq_of(cfs_rq)->curr);
1939 static __always_inline
1940 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1942 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1945 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1948 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1950 return cfs_bandwidth_used() && cfs_rq->throttled;
1953 /* check whether cfs_rq, or any parent, is throttled */
1954 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1956 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1960 * Ensure that neither of the group entities corresponding to src_cpu or
1961 * dest_cpu are members of a throttled hierarchy when performing group
1962 * load-balance operations.
1964 static inline int throttled_lb_pair(struct task_group *tg,
1965 int src_cpu, int dest_cpu)
1967 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1969 src_cfs_rq = tg->cfs_rq[src_cpu];
1970 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1972 return throttled_hierarchy(src_cfs_rq) ||
1973 throttled_hierarchy(dest_cfs_rq);
1976 /* updated child weight may affect parent so we have to do this bottom up */
1977 static int tg_unthrottle_up(struct task_group *tg, void *data)
1979 struct rq *rq = data;
1980 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1982 cfs_rq->throttle_count--;
1984 if (!cfs_rq->throttle_count) {
1985 /* adjust cfs_rq_clock_task() */
1986 cfs_rq->throttled_clock_task_time += rq->clock_task -
1987 cfs_rq->throttled_clock_task;
1994 static int tg_throttle_down(struct task_group *tg, void *data)
1996 struct rq *rq = data;
1997 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1999 /* group is entering throttled state, stop time */
2000 if (!cfs_rq->throttle_count)
2001 cfs_rq->throttled_clock_task = rq->clock_task;
2002 cfs_rq->throttle_count++;
2007 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
2009 struct rq *rq = rq_of(cfs_rq);
2010 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2011 struct sched_entity *se;
2012 long task_delta, dequeue = 1;
2014 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2016 /* freeze hierarchy runnable averages while throttled */
2018 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
2021 task_delta = cfs_rq->h_nr_running;
2022 for_each_sched_entity(se) {
2023 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
2024 /* throttled entity or throttle-on-deactivate */
2029 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
2030 qcfs_rq->h_nr_running -= task_delta;
2032 if (qcfs_rq->load.weight)
2037 rq->nr_running -= task_delta;
2039 cfs_rq->throttled = 1;
2040 cfs_rq->throttled_clock = rq->clock;
2041 raw_spin_lock(&cfs_b->lock);
2042 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
2043 raw_spin_unlock(&cfs_b->lock);
2046 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
2048 struct rq *rq = rq_of(cfs_rq);
2049 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2050 struct sched_entity *se;
2054 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
2056 cfs_rq->throttled = 0;
2057 raw_spin_lock(&cfs_b->lock);
2058 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_clock;
2059 list_del_rcu(&cfs_rq->throttled_list);
2060 raw_spin_unlock(&cfs_b->lock);
2062 update_rq_clock(rq);
2063 /* update hierarchical throttle state */
2064 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
2066 if (!cfs_rq->load.weight)
2069 task_delta = cfs_rq->h_nr_running;
2070 for_each_sched_entity(se) {
2074 cfs_rq = cfs_rq_of(se);
2076 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
2077 cfs_rq->h_nr_running += task_delta;
2079 if (cfs_rq_throttled(cfs_rq))
2084 rq->nr_running += task_delta;
2086 /* determine whether we need to wake up potentially idle cpu */
2087 if (rq->curr == rq->idle && rq->cfs.nr_running)
2088 resched_task(rq->curr);
2091 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
2092 u64 remaining, u64 expires)
2094 struct cfs_rq *cfs_rq;
2095 u64 runtime = remaining;
2098 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
2100 struct rq *rq = rq_of(cfs_rq);
2102 raw_spin_lock(&rq->lock);
2103 if (!cfs_rq_throttled(cfs_rq))
2106 runtime = -cfs_rq->runtime_remaining + 1;
2107 if (runtime > remaining)
2108 runtime = remaining;
2109 remaining -= runtime;
2111 cfs_rq->runtime_remaining += runtime;
2112 cfs_rq->runtime_expires = expires;
2114 /* we check whether we're throttled above */
2115 if (cfs_rq->runtime_remaining > 0)
2116 unthrottle_cfs_rq(cfs_rq);
2119 raw_spin_unlock(&rq->lock);
2130 * Responsible for refilling a task_group's bandwidth and unthrottling its
2131 * cfs_rqs as appropriate. If there has been no activity within the last
2132 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
2133 * used to track this state.
2135 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
2137 u64 runtime, runtime_expires;
2138 int idle = 1, throttled;
2140 raw_spin_lock(&cfs_b->lock);
2141 /* no need to continue the timer with no bandwidth constraint */
2142 if (cfs_b->quota == RUNTIME_INF)
2145 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2146 /* idle depends on !throttled (for the case of a large deficit) */
2147 idle = cfs_b->idle && !throttled;
2148 cfs_b->nr_periods += overrun;
2150 /* if we're going inactive then everything else can be deferred */
2154 __refill_cfs_bandwidth_runtime(cfs_b);
2157 /* mark as potentially idle for the upcoming period */
2162 /* account preceding periods in which throttling occurred */
2163 cfs_b->nr_throttled += overrun;
2166 * There are throttled entities so we must first use the new bandwidth
2167 * to unthrottle them before making it generally available. This
2168 * ensures that all existing debts will be paid before a new cfs_rq is
2171 runtime = cfs_b->runtime;
2172 runtime_expires = cfs_b->runtime_expires;
2176 * This check is repeated as we are holding onto the new bandwidth
2177 * while we unthrottle. This can potentially race with an unthrottled
2178 * group trying to acquire new bandwidth from the global pool.
2180 while (throttled && runtime > 0) {
2181 raw_spin_unlock(&cfs_b->lock);
2182 /* we can't nest cfs_b->lock while distributing bandwidth */
2183 runtime = distribute_cfs_runtime(cfs_b, runtime,
2185 raw_spin_lock(&cfs_b->lock);
2187 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
2190 /* return (any) remaining runtime */
2191 cfs_b->runtime = runtime;
2193 * While we are ensured activity in the period following an
2194 * unthrottle, this also covers the case in which the new bandwidth is
2195 * insufficient to cover the existing bandwidth deficit. (Forcing the
2196 * timer to remain active while there are any throttled entities.)
2201 cfs_b->timer_active = 0;
2202 raw_spin_unlock(&cfs_b->lock);
2207 /* a cfs_rq won't donate quota below this amount */
2208 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
2209 /* minimum remaining period time to redistribute slack quota */
2210 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
2211 /* how long we wait to gather additional slack before distributing */
2212 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
2214 /* are we near the end of the current quota period? */
2215 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
2217 struct hrtimer *refresh_timer = &cfs_b->period_timer;
2220 /* if the call-back is running a quota refresh is already occurring */
2221 if (hrtimer_callback_running(refresh_timer))
2224 /* is a quota refresh about to occur? */
2225 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
2226 if (remaining < min_expire)
2232 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
2234 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
2236 /* if there's a quota refresh soon don't bother with slack */
2237 if (runtime_refresh_within(cfs_b, min_left))
2240 start_bandwidth_timer(&cfs_b->slack_timer,
2241 ns_to_ktime(cfs_bandwidth_slack_period));
2244 /* we know any runtime found here is valid as update_curr() precedes return */
2245 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2247 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2248 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
2250 if (slack_runtime <= 0)
2253 raw_spin_lock(&cfs_b->lock);
2254 if (cfs_b->quota != RUNTIME_INF &&
2255 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
2256 cfs_b->runtime += slack_runtime;
2258 /* we are under rq->lock, defer unthrottling using a timer */
2259 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
2260 !list_empty(&cfs_b->throttled_cfs_rq))
2261 start_cfs_slack_bandwidth(cfs_b);
2263 raw_spin_unlock(&cfs_b->lock);
2265 /* even if it's not valid for return we don't want to try again */
2266 cfs_rq->runtime_remaining -= slack_runtime;
2269 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2271 if (!cfs_bandwidth_used())
2274 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
2277 __return_cfs_rq_runtime(cfs_rq);
2281 * This is done with a timer (instead of inline with bandwidth return) since
2282 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
2284 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
2286 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
2289 /* confirm we're still not at a refresh boundary */
2290 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
2293 raw_spin_lock(&cfs_b->lock);
2294 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
2295 runtime = cfs_b->runtime;
2298 expires = cfs_b->runtime_expires;
2299 raw_spin_unlock(&cfs_b->lock);
2304 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
2306 raw_spin_lock(&cfs_b->lock);
2307 if (expires == cfs_b->runtime_expires)
2308 cfs_b->runtime = runtime;
2309 raw_spin_unlock(&cfs_b->lock);
2313 * When a group wakes up we want to make sure that its quota is not already
2314 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
2315 * runtime as update_curr() throttling can not not trigger until it's on-rq.
2317 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
2319 if (!cfs_bandwidth_used())
2322 /* an active group must be handled by the update_curr()->put() path */
2323 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
2326 /* ensure the group is not already throttled */
2327 if (cfs_rq_throttled(cfs_rq))
2330 /* update runtime allocation */
2331 account_cfs_rq_runtime(cfs_rq, 0);
2332 if (cfs_rq->runtime_remaining <= 0)
2333 throttle_cfs_rq(cfs_rq);
2336 /* conditionally throttle active cfs_rq's from put_prev_entity() */
2337 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2339 if (!cfs_bandwidth_used())
2342 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
2346 * it's possible for a throttled entity to be forced into a running
2347 * state (e.g. set_curr_task), in this case we're finished.
2349 if (cfs_rq_throttled(cfs_rq))
2352 throttle_cfs_rq(cfs_rq);
2355 static inline u64 default_cfs_period(void);
2356 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
2357 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
2359 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
2361 struct cfs_bandwidth *cfs_b =
2362 container_of(timer, struct cfs_bandwidth, slack_timer);
2363 do_sched_cfs_slack_timer(cfs_b);
2365 return HRTIMER_NORESTART;
2368 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
2370 struct cfs_bandwidth *cfs_b =
2371 container_of(timer, struct cfs_bandwidth, period_timer);
2377 now = hrtimer_cb_get_time(timer);
2378 overrun = hrtimer_forward(timer, now, cfs_b->period);
2383 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2386 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2389 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2391 raw_spin_lock_init(&cfs_b->lock);
2393 cfs_b->quota = RUNTIME_INF;
2394 cfs_b->period = ns_to_ktime(default_cfs_period());
2396 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2397 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2398 cfs_b->period_timer.function = sched_cfs_period_timer;
2399 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2400 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2403 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2405 cfs_rq->runtime_enabled = 0;
2406 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2409 /* requires cfs_b->lock, may release to reprogram timer */
2410 void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2413 * The timer may be active because we're trying to set a new bandwidth
2414 * period or because we're racing with the tear-down path
2415 * (timer_active==0 becomes visible before the hrtimer call-back
2416 * terminates). In either case we ensure that it's re-programmed
2418 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2419 raw_spin_unlock(&cfs_b->lock);
2420 /* ensure cfs_b->lock is available while we wait */
2421 hrtimer_cancel(&cfs_b->period_timer);
2423 raw_spin_lock(&cfs_b->lock);
2424 /* if someone else restarted the timer then we're done */
2425 if (cfs_b->timer_active)
2429 cfs_b->timer_active = 1;
2430 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2433 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2435 hrtimer_cancel(&cfs_b->period_timer);
2436 hrtimer_cancel(&cfs_b->slack_timer);
2439 static void unthrottle_offline_cfs_rqs(struct rq *rq)
2441 struct cfs_rq *cfs_rq;
2443 for_each_leaf_cfs_rq(rq, cfs_rq) {
2444 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2446 if (!cfs_rq->runtime_enabled)
2450 * clock_task is not advancing so we just need to make sure
2451 * there's some valid quota amount
2453 cfs_rq->runtime_remaining = cfs_b->quota;
2454 if (cfs_rq_throttled(cfs_rq))
2455 unthrottle_cfs_rq(cfs_rq);
2459 #else /* CONFIG_CFS_BANDWIDTH */
2460 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
2462 return rq_of(cfs_rq)->clock_task;
2465 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2466 unsigned long delta_exec) {}
2467 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2468 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2469 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2471 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2476 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2481 static inline int throttled_lb_pair(struct task_group *tg,
2482 int src_cpu, int dest_cpu)
2487 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2489 #ifdef CONFIG_FAIR_GROUP_SCHED
2490 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2493 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2497 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2498 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2500 #endif /* CONFIG_CFS_BANDWIDTH */
2502 /**************************************************
2503 * CFS operations on tasks:
2506 #ifdef CONFIG_SCHED_HRTICK
2507 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2509 struct sched_entity *se = &p->se;
2510 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2512 WARN_ON(task_rq(p) != rq);
2514 if (cfs_rq->nr_running > 1) {
2515 u64 slice = sched_slice(cfs_rq, se);
2516 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2517 s64 delta = slice - ran;
2526 * Don't schedule slices shorter than 10000ns, that just
2527 * doesn't make sense. Rely on vruntime for fairness.
2530 delta = max_t(s64, 10000LL, delta);
2532 hrtick_start(rq, delta);
2537 * called from enqueue/dequeue and updates the hrtick when the
2538 * current task is from our class and nr_running is low enough
2541 static void hrtick_update(struct rq *rq)
2543 struct task_struct *curr = rq->curr;
2545 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2548 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2549 hrtick_start_fair(rq, curr);
2551 #else /* !CONFIG_SCHED_HRTICK */
2553 hrtick_start_fair(struct rq *rq, struct task_struct *p)
2557 static inline void hrtick_update(struct rq *rq)
2563 * The enqueue_task method is called before nr_running is
2564 * increased. Here we update the fair scheduling stats and
2565 * then put the task into the rbtree:
2568 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2570 struct cfs_rq *cfs_rq;
2571 struct sched_entity *se = &p->se;
2573 for_each_sched_entity(se) {
2576 cfs_rq = cfs_rq_of(se);
2577 enqueue_entity(cfs_rq, se, flags);
2580 * end evaluation on encountering a throttled cfs_rq
2582 * note: in the case of encountering a throttled cfs_rq we will
2583 * post the final h_nr_running increment below.
2585 if (cfs_rq_throttled(cfs_rq))
2587 cfs_rq->h_nr_running++;
2589 flags = ENQUEUE_WAKEUP;
2592 for_each_sched_entity(se) {
2593 cfs_rq = cfs_rq_of(se);
2594 cfs_rq->h_nr_running++;
2596 if (cfs_rq_throttled(cfs_rq))
2599 update_cfs_shares(cfs_rq);
2600 update_entity_load_avg(se, 1);
2604 update_rq_runnable_avg(rq, rq->nr_running);
2610 static void set_next_buddy(struct sched_entity *se);
2613 * The dequeue_task method is called before nr_running is
2614 * decreased. We remove the task from the rbtree and
2615 * update the fair scheduling stats:
2617 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2619 struct cfs_rq *cfs_rq;
2620 struct sched_entity *se = &p->se;
2621 int task_sleep = flags & DEQUEUE_SLEEP;
2623 for_each_sched_entity(se) {
2624 cfs_rq = cfs_rq_of(se);
2625 dequeue_entity(cfs_rq, se, flags);
2628 * end evaluation on encountering a throttled cfs_rq
2630 * note: in the case of encountering a throttled cfs_rq we will
2631 * post the final h_nr_running decrement below.
2633 if (cfs_rq_throttled(cfs_rq))
2635 cfs_rq->h_nr_running--;
2637 /* Don't dequeue parent if it has other entities besides us */
2638 if (cfs_rq->load.weight) {
2640 * Bias pick_next to pick a task from this cfs_rq, as
2641 * p is sleeping when it is within its sched_slice.
2643 if (task_sleep && parent_entity(se))
2644 set_next_buddy(parent_entity(se));
2646 /* avoid re-evaluating load for this entity */
2647 se = parent_entity(se);
2650 flags |= DEQUEUE_SLEEP;
2653 for_each_sched_entity(se) {
2654 cfs_rq = cfs_rq_of(se);
2655 cfs_rq->h_nr_running--;
2657 if (cfs_rq_throttled(cfs_rq))
2660 update_cfs_shares(cfs_rq);
2661 update_entity_load_avg(se, 1);
2666 update_rq_runnable_avg(rq, 1);
2672 /* Used instead of source_load when we know the type == 0 */
2673 static unsigned long weighted_cpuload(const int cpu)
2675 return cpu_rq(cpu)->load.weight;
2679 * Return a low guess at the load of a migration-source cpu weighted
2680 * according to the scheduling class and "nice" value.
2682 * We want to under-estimate the load of migration sources, to
2683 * balance conservatively.
2685 static unsigned long source_load(int cpu, int type)
2687 struct rq *rq = cpu_rq(cpu);
2688 unsigned long total = weighted_cpuload(cpu);
2690 if (type == 0 || !sched_feat(LB_BIAS))
2693 return min(rq->cpu_load[type-1], total);
2697 * Return a high guess at the load of a migration-target cpu weighted
2698 * according to the scheduling class and "nice" value.
2700 static unsigned long target_load(int cpu, int type)
2702 struct rq *rq = cpu_rq(cpu);
2703 unsigned long total = weighted_cpuload(cpu);
2705 if (type == 0 || !sched_feat(LB_BIAS))
2708 return max(rq->cpu_load[type-1], total);
2711 static unsigned long power_of(int cpu)
2713 return cpu_rq(cpu)->cpu_power;
2716 static unsigned long cpu_avg_load_per_task(int cpu)
2718 struct rq *rq = cpu_rq(cpu);
2719 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2722 return rq->load.weight / nr_running;
2728 static void task_waking_fair(struct task_struct *p)
2730 struct sched_entity *se = &p->se;
2731 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2734 #ifndef CONFIG_64BIT
2735 u64 min_vruntime_copy;
2738 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2740 min_vruntime = cfs_rq->min_vruntime;
2741 } while (min_vruntime != min_vruntime_copy);
2743 min_vruntime = cfs_rq->min_vruntime;
2746 se->vruntime -= min_vruntime;
2749 #ifdef CONFIG_FAIR_GROUP_SCHED
2751 * effective_load() calculates the load change as seen from the root_task_group
2753 * Adding load to a group doesn't make a group heavier, but can cause movement
2754 * of group shares between cpus. Assuming the shares were perfectly aligned one
2755 * can calculate the shift in shares.
2757 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2758 * on this @cpu and results in a total addition (subtraction) of @wg to the
2759 * total group weight.
2761 * Given a runqueue weight distribution (rw_i) we can compute a shares
2762 * distribution (s_i) using:
2764 * s_i = rw_i / \Sum rw_j (1)
2766 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2767 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2768 * shares distribution (s_i):
2770 * rw_i = { 2, 4, 1, 0 }
2771 * s_i = { 2/7, 4/7, 1/7, 0 }
2773 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2774 * task used to run on and the CPU the waker is running on), we need to
2775 * compute the effect of waking a task on either CPU and, in case of a sync
2776 * wakeup, compute the effect of the current task going to sleep.
2778 * So for a change of @wl to the local @cpu with an overall group weight change
2779 * of @wl we can compute the new shares distribution (s'_i) using:
2781 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2783 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2784 * differences in waking a task to CPU 0. The additional task changes the
2785 * weight and shares distributions like:
2787 * rw'_i = { 3, 4, 1, 0 }
2788 * s'_i = { 3/8, 4/8, 1/8, 0 }
2790 * We can then compute the difference in effective weight by using:
2792 * dw_i = S * (s'_i - s_i) (3)
2794 * Where 'S' is the group weight as seen by its parent.
2796 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2797 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2798 * 4/7) times the weight of the group.
2800 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2802 struct sched_entity *se = tg->se[cpu];
2804 if (!tg->parent) /* the trivial, non-cgroup case */
2807 for_each_sched_entity(se) {
2813 * W = @wg + \Sum rw_j
2815 W = wg + calc_tg_weight(tg, se->my_q);
2820 w = se->my_q->load.weight + wl;
2823 * wl = S * s'_i; see (2)
2826 wl = (w * tg->shares) / W;
2831 * Per the above, wl is the new se->load.weight value; since
2832 * those are clipped to [MIN_SHARES, ...) do so now. See
2833 * calc_cfs_shares().
2835 if (wl < MIN_SHARES)
2839 * wl = dw_i = S * (s'_i - s_i); see (3)
2841 wl -= se->load.weight;
2844 * Recursively apply this logic to all parent groups to compute
2845 * the final effective load change on the root group. Since
2846 * only the @tg group gets extra weight, all parent groups can
2847 * only redistribute existing shares. @wl is the shift in shares
2848 * resulting from this level per the above.
2857 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2858 unsigned long wl, unsigned long wg)
2865 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2867 s64 this_load, load;
2868 int idx, this_cpu, prev_cpu;
2869 unsigned long tl_per_task;
2870 struct task_group *tg;
2871 unsigned long weight;
2875 this_cpu = smp_processor_id();
2876 prev_cpu = task_cpu(p);
2877 load = source_load(prev_cpu, idx);
2878 this_load = target_load(this_cpu, idx);
2881 * If sync wakeup then subtract the (maximum possible)
2882 * effect of the currently running task from the load
2883 * of the current CPU:
2886 tg = task_group(current);
2887 weight = current->se.load.weight;
2889 this_load += effective_load(tg, this_cpu, -weight, -weight);
2890 load += effective_load(tg, prev_cpu, 0, -weight);
2894 weight = p->se.load.weight;
2897 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2898 * due to the sync cause above having dropped this_load to 0, we'll
2899 * always have an imbalance, but there's really nothing you can do
2900 * about that, so that's good too.
2902 * Otherwise check if either cpus are near enough in load to allow this
2903 * task to be woken on this_cpu.
2905 if (this_load > 0) {
2906 s64 this_eff_load, prev_eff_load;
2908 this_eff_load = 100;
2909 this_eff_load *= power_of(prev_cpu);
2910 this_eff_load *= this_load +
2911 effective_load(tg, this_cpu, weight, weight);
2913 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2914 prev_eff_load *= power_of(this_cpu);
2915 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2917 balanced = this_eff_load <= prev_eff_load;
2922 * If the currently running task will sleep within
2923 * a reasonable amount of time then attract this newly
2926 if (sync && balanced)
2929 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2930 tl_per_task = cpu_avg_load_per_task(this_cpu);
2933 (this_load <= load &&
2934 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2936 * This domain has SD_WAKE_AFFINE and
2937 * p is cache cold in this domain, and
2938 * there is no bad imbalance.
2940 schedstat_inc(sd, ttwu_move_affine);
2941 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2949 * find_idlest_group finds and returns the least busy CPU group within the
2952 static struct sched_group *
2953 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2954 int this_cpu, int load_idx)
2956 struct sched_group *idlest = NULL, *group = sd->groups;
2957 unsigned long min_load = ULONG_MAX, this_load = 0;
2958 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2961 unsigned long load, avg_load;
2965 /* Skip over this group if it has no CPUs allowed */
2966 if (!cpumask_intersects(sched_group_cpus(group),
2967 tsk_cpus_allowed(p)))
2970 local_group = cpumask_test_cpu(this_cpu,
2971 sched_group_cpus(group));
2973 /* Tally up the load of all CPUs in the group */
2976 for_each_cpu(i, sched_group_cpus(group)) {
2977 /* Bias balancing toward cpus of our domain */
2979 load = source_load(i, load_idx);
2981 load = target_load(i, load_idx);
2986 /* Adjust by relative CPU power of the group */
2987 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2990 this_load = avg_load;
2991 } else if (avg_load < min_load) {
2992 min_load = avg_load;
2995 } while (group = group->next, group != sd->groups);
2997 if (!idlest || 100*this_load < imbalance*min_load)
3003 * find_idlest_cpu - find the idlest cpu among the cpus in group.
3006 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
3008 unsigned long load, min_load = ULONG_MAX;
3012 /* Traverse only the allowed CPUs */
3013 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
3014 load = weighted_cpuload(i);
3016 if (load < min_load || (load == min_load && i == this_cpu)) {
3026 * Try and locate an idle CPU in the sched_domain.
3028 static int select_idle_sibling(struct task_struct *p, int target)
3030 int cpu = smp_processor_id();
3031 int prev_cpu = task_cpu(p);
3032 struct sched_domain *sd;
3033 struct sched_group *sg;
3037 * If the task is going to be woken-up on this cpu and if it is
3038 * already idle, then it is the right target.
3040 if (target == cpu && idle_cpu(cpu))
3044 * If the task is going to be woken-up on the cpu where it previously
3045 * ran and if it is currently idle, then it the right target.
3047 if (target == prev_cpu && idle_cpu(prev_cpu))
3051 * Otherwise, iterate the domains and find an elegible idle cpu.
3053 sd = rcu_dereference(per_cpu(sd_llc, target));
3054 for_each_lower_domain(sd) {
3057 if (!cpumask_intersects(sched_group_cpus(sg),
3058 tsk_cpus_allowed(p)))
3061 for_each_cpu(i, sched_group_cpus(sg)) {
3066 target = cpumask_first_and(sched_group_cpus(sg),
3067 tsk_cpus_allowed(p));
3071 } while (sg != sd->groups);
3078 * sched_balance_self: balance the current task (running on cpu) in domains
3079 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
3082 * Balance, ie. select the least loaded group.
3084 * Returns the target CPU number, or the same CPU if no balancing is needed.
3086 * preempt must be disabled.
3089 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
3091 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
3092 int cpu = smp_processor_id();
3093 int prev_cpu = task_cpu(p);
3095 int want_affine = 0;
3096 int sync = wake_flags & WF_SYNC;
3098 if (p->nr_cpus_allowed == 1)
3101 if (sd_flag & SD_BALANCE_WAKE) {
3102 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
3108 for_each_domain(cpu, tmp) {
3109 if (!(tmp->flags & SD_LOAD_BALANCE))
3113 * If both cpu and prev_cpu are part of this domain,
3114 * cpu is a valid SD_WAKE_AFFINE target.
3116 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
3117 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
3122 if (tmp->flags & sd_flag)
3127 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
3130 new_cpu = select_idle_sibling(p, prev_cpu);
3135 int load_idx = sd->forkexec_idx;
3136 struct sched_group *group;
3139 if (!(sd->flags & sd_flag)) {
3144 if (sd_flag & SD_BALANCE_WAKE)
3145 load_idx = sd->wake_idx;
3147 group = find_idlest_group(sd, p, cpu, load_idx);
3153 new_cpu = find_idlest_cpu(group, p, cpu);
3154 if (new_cpu == -1 || new_cpu == cpu) {
3155 /* Now try balancing at a lower domain level of cpu */
3160 /* Now try balancing at a lower domain level of new_cpu */
3162 weight = sd->span_weight;
3164 for_each_domain(cpu, tmp) {
3165 if (weight <= tmp->span_weight)
3167 if (tmp->flags & sd_flag)
3170 /* while loop will break here if sd == NULL */
3179 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
3180 * removed when useful for applications beyond shares distribution (e.g.
3183 #ifdef CONFIG_FAIR_GROUP_SCHED
3185 * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
3186 * cfs_rq_of(p) references at time of call are still valid and identify the
3187 * previous cpu. However, the caller only guarantees p->pi_lock is held; no
3188 * other assumptions, including the state of rq->lock, should be made.
3191 migrate_task_rq_fair(struct task_struct *p, int next_cpu)
3193 struct sched_entity *se = &p->se;
3194 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3197 * Load tracking: accumulate removed load so that it can be processed
3198 * when we next update owning cfs_rq under rq->lock. Tasks contribute
3199 * to blocked load iff they have a positive decay-count. It can never
3200 * be negative here since on-rq tasks have decay-count == 0.
3202 if (se->avg.decay_count) {
3203 se->avg.decay_count = -__synchronize_entity_decay(se);
3204 atomic64_add(se->avg.load_avg_contrib, &cfs_rq->removed_load);
3208 #endif /* CONFIG_SMP */
3210 static unsigned long
3211 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
3213 unsigned long gran = sysctl_sched_wakeup_granularity;
3216 * Since its curr running now, convert the gran from real-time
3217 * to virtual-time in his units.
3219 * By using 'se' instead of 'curr' we penalize light tasks, so
3220 * they get preempted easier. That is, if 'se' < 'curr' then
3221 * the resulting gran will be larger, therefore penalizing the
3222 * lighter, if otoh 'se' > 'curr' then the resulting gran will
3223 * be smaller, again penalizing the lighter task.
3225 * This is especially important for buddies when the leftmost
3226 * task is higher priority than the buddy.
3228 return calc_delta_fair(gran, se);
3232 * Should 'se' preempt 'curr'.
3246 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
3248 s64 gran, vdiff = curr->vruntime - se->vruntime;
3253 gran = wakeup_gran(curr, se);
3260 static void set_last_buddy(struct sched_entity *se)
3262 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3265 for_each_sched_entity(se)
3266 cfs_rq_of(se)->last = se;
3269 static void set_next_buddy(struct sched_entity *se)
3271 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
3274 for_each_sched_entity(se)
3275 cfs_rq_of(se)->next = se;
3278 static void set_skip_buddy(struct sched_entity *se)
3280 for_each_sched_entity(se)
3281 cfs_rq_of(se)->skip = se;
3285 * Preempt the current task with a newly woken task if needed:
3287 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
3289 struct task_struct *curr = rq->curr;
3290 struct sched_entity *se = &curr->se, *pse = &p->se;
3291 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3292 int scale = cfs_rq->nr_running >= sched_nr_latency;
3293 int next_buddy_marked = 0;
3295 if (unlikely(se == pse))
3299 * This is possible from callers such as move_task(), in which we
3300 * unconditionally check_prempt_curr() after an enqueue (which may have
3301 * lead to a throttle). This both saves work and prevents false
3302 * next-buddy nomination below.
3304 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
3307 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
3308 set_next_buddy(pse);
3309 next_buddy_marked = 1;
3313 * We can come here with TIF_NEED_RESCHED already set from new task
3316 * Note: this also catches the edge-case of curr being in a throttled
3317 * group (e.g. via set_curr_task), since update_curr() (in the
3318 * enqueue of curr) will have resulted in resched being set. This
3319 * prevents us from potentially nominating it as a false LAST_BUDDY
3322 if (test_tsk_need_resched(curr))
3325 /* Idle tasks are by definition preempted by non-idle tasks. */
3326 if (unlikely(curr->policy == SCHED_IDLE) &&
3327 likely(p->policy != SCHED_IDLE))
3331 * Batch and idle tasks do not preempt non-idle tasks (their preemption
3332 * is driven by the tick):
3334 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
3337 find_matching_se(&se, &pse);
3338 update_curr(cfs_rq_of(se));
3340 if (wakeup_preempt_entity(se, pse) == 1) {
3342 * Bias pick_next to pick the sched entity that is
3343 * triggering this preemption.
3345 if (!next_buddy_marked)
3346 set_next_buddy(pse);
3355 * Only set the backward buddy when the current task is still
3356 * on the rq. This can happen when a wakeup gets interleaved
3357 * with schedule on the ->pre_schedule() or idle_balance()
3358 * point, either of which can * drop the rq lock.
3360 * Also, during early boot the idle thread is in the fair class,
3361 * for obvious reasons its a bad idea to schedule back to it.
3363 if (unlikely(!se->on_rq || curr == rq->idle))
3366 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
3370 static struct task_struct *pick_next_task_fair(struct rq *rq)
3372 struct task_struct *p;
3373 struct cfs_rq *cfs_rq = &rq->cfs;
3374 struct sched_entity *se;
3376 if (!cfs_rq->nr_running)
3380 se = pick_next_entity(cfs_rq);
3381 set_next_entity(cfs_rq, se);
3382 cfs_rq = group_cfs_rq(se);
3386 if (hrtick_enabled(rq))
3387 hrtick_start_fair(rq, p);
3393 * Account for a descheduled task:
3395 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3397 struct sched_entity *se = &prev->se;
3398 struct cfs_rq *cfs_rq;
3400 for_each_sched_entity(se) {
3401 cfs_rq = cfs_rq_of(se);
3402 put_prev_entity(cfs_rq, se);
3407 * sched_yield() is very simple
3409 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3411 static void yield_task_fair(struct rq *rq)
3413 struct task_struct *curr = rq->curr;
3414 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3415 struct sched_entity *se = &curr->se;
3418 * Are we the only task in the tree?
3420 if (unlikely(rq->nr_running == 1))
3423 clear_buddies(cfs_rq, se);
3425 if (curr->policy != SCHED_BATCH) {
3426 update_rq_clock(rq);
3428 * Update run-time statistics of the 'current'.
3430 update_curr(cfs_rq);
3432 * Tell update_rq_clock() that we've just updated,
3433 * so we don't do microscopic update in schedule()
3434 * and double the fastpath cost.
3436 rq->skip_clock_update = 1;
3442 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3444 struct sched_entity *se = &p->se;
3446 /* throttled hierarchies are not runnable */
3447 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3450 /* Tell the scheduler that we'd really like pse to run next. */
3453 yield_task_fair(rq);
3459 /**************************************************
3460 * Fair scheduling class load-balancing methods.
3464 * The purpose of load-balancing is to achieve the same basic fairness the
3465 * per-cpu scheduler provides, namely provide a proportional amount of compute
3466 * time to each task. This is expressed in the following equation:
3468 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
3470 * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
3471 * W_i,0 is defined as:
3473 * W_i,0 = \Sum_j w_i,j (2)
3475 * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
3476 * is derived from the nice value as per prio_to_weight[].
3478 * The weight average is an exponential decay average of the instantaneous
3481 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
3483 * P_i is the cpu power (or compute capacity) of cpu i, typically it is the
3484 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
3485 * can also include other factors [XXX].
3487 * To achieve this balance we define a measure of imbalance which follows
3488 * directly from (1):
3490 * imb_i,j = max{ avg(W/P), W_i/P_i } - min{ avg(W/P), W_j/P_j } (4)
3492 * We them move tasks around to minimize the imbalance. In the continuous
3493 * function space it is obvious this converges, in the discrete case we get
3494 * a few fun cases generally called infeasible weight scenarios.
3497 * - infeasible weights;
3498 * - local vs global optima in the discrete case. ]
3503 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
3504 * for all i,j solution, we create a tree of cpus that follows the hardware
3505 * topology where each level pairs two lower groups (or better). This results
3506 * in O(log n) layers. Furthermore we reduce the number of cpus going up the
3507 * tree to only the first of the previous level and we decrease the frequency
3508 * of load-balance at each level inv. proportional to the number of cpus in
3514 * \Sum { --- * --- * 2^i } = O(n) (5)
3516 * `- size of each group
3517 * | | `- number of cpus doing load-balance
3519 * `- sum over all levels
3521 * Coupled with a limit on how many tasks we can migrate every balance pass,
3522 * this makes (5) the runtime complexity of the balancer.
3524 * An important property here is that each CPU is still (indirectly) connected
3525 * to every other cpu in at most O(log n) steps:
3527 * The adjacency matrix of the resulting graph is given by:
3530 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
3533 * And you'll find that:
3535 * A^(log_2 n)_i,j != 0 for all i,j (7)
3537 * Showing there's indeed a path between every cpu in at most O(log n) steps.
3538 * The task movement gives a factor of O(m), giving a convergence complexity
3541 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
3546 * In order to avoid CPUs going idle while there's still work to do, new idle
3547 * balancing is more aggressive and has the newly idle cpu iterate up the domain
3548 * tree itself instead of relying on other CPUs to bring it work.
3550 * This adds some complexity to both (5) and (8) but it reduces the total idle
3558 * Cgroups make a horror show out of (2), instead of a simple sum we get:
3561 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
3566 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
3568 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
3570 * The big problem is S_k, its a global sum needed to compute a local (W_i)
3573 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
3574 * rewrite all of this once again.]
3577 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3579 #define LBF_ALL_PINNED 0x01
3580 #define LBF_NEED_BREAK 0x02
3581 #define LBF_SOME_PINNED 0x04
3584 struct sched_domain *sd;
3592 struct cpumask *dst_grpmask;
3594 enum cpu_idle_type idle;
3596 /* The set of CPUs under consideration for load-balancing */
3597 struct cpumask *cpus;
3602 unsigned int loop_break;
3603 unsigned int loop_max;
3607 * move_task - move a task from one runqueue to another runqueue.
3608 * Both runqueues must be locked.
3610 static void move_task(struct task_struct *p, struct lb_env *env)
3612 deactivate_task(env->src_rq, p, 0);
3613 set_task_cpu(p, env->dst_cpu);
3614 activate_task(env->dst_rq, p, 0);
3615 check_preempt_curr(env->dst_rq, p, 0);
3619 * Is this task likely cache-hot:
3622 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3626 if (p->sched_class != &fair_sched_class)
3629 if (unlikely(p->policy == SCHED_IDLE))
3633 * Buddy candidates are cache hot:
3635 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3636 (&p->se == cfs_rq_of(&p->se)->next ||
3637 &p->se == cfs_rq_of(&p->se)->last))
3640 if (sysctl_sched_migration_cost == -1)
3642 if (sysctl_sched_migration_cost == 0)
3645 delta = now - p->se.exec_start;
3647 return delta < (s64)sysctl_sched_migration_cost;
3651 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3654 int can_migrate_task(struct task_struct *p, struct lb_env *env)
3656 int tsk_cache_hot = 0;
3658 * We do not migrate tasks that are:
3659 * 1) running (obviously), or
3660 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3661 * 3) are cache-hot on their current CPU.
3663 if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
3666 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3669 * Remember if this task can be migrated to any other cpu in
3670 * our sched_group. We may want to revisit it if we couldn't
3671 * meet load balance goals by pulling other tasks on src_cpu.
3673 * Also avoid computing new_dst_cpu if we have already computed
3674 * one in current iteration.
3676 if (!env->dst_grpmask || (env->flags & LBF_SOME_PINNED))
3679 new_dst_cpu = cpumask_first_and(env->dst_grpmask,
3680 tsk_cpus_allowed(p));
3681 if (new_dst_cpu < nr_cpu_ids) {
3682 env->flags |= LBF_SOME_PINNED;
3683 env->new_dst_cpu = new_dst_cpu;
3688 /* Record that we found atleast one task that could run on dst_cpu */
3689 env->flags &= ~LBF_ALL_PINNED;
3691 if (task_running(env->src_rq, p)) {
3692 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3697 * Aggressive migration if:
3698 * 1) task is cache cold, or
3699 * 2) too many balance attempts have failed.
3702 tsk_cache_hot = task_hot(p, env->src_rq->clock_task, env->sd);
3703 if (!tsk_cache_hot ||
3704 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3705 #ifdef CONFIG_SCHEDSTATS
3706 if (tsk_cache_hot) {
3707 schedstat_inc(env->sd, lb_hot_gained[env->idle]);
3708 schedstat_inc(p, se.statistics.nr_forced_migrations);
3714 if (tsk_cache_hot) {
3715 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3722 * move_one_task tries to move exactly one task from busiest to this_rq, as
3723 * part of active balancing operations within "domain".
3724 * Returns 1 if successful and 0 otherwise.
3726 * Called with both runqueues locked.
3728 static int move_one_task(struct lb_env *env)
3730 struct task_struct *p, *n;
3732 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3733 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3736 if (!can_migrate_task(p, env))
3741 * Right now, this is only the second place move_task()
3742 * is called, so we can safely collect move_task()
3743 * stats here rather than inside move_task().
3745 schedstat_inc(env->sd, lb_gained[env->idle]);
3751 static unsigned long task_h_load(struct task_struct *p);
3753 static const unsigned int sched_nr_migrate_break = 32;
3756 * move_tasks tries to move up to imbalance weighted load from busiest to
3757 * this_rq, as part of a balancing operation within domain "sd".
3758 * Returns 1 if successful and 0 otherwise.
3760 * Called with both runqueues locked.
3762 static int move_tasks(struct lb_env *env)
3764 struct list_head *tasks = &env->src_rq->cfs_tasks;
3765 struct task_struct *p;
3769 if (env->imbalance <= 0)
3772 while (!list_empty(tasks)) {
3773 p = list_first_entry(tasks, struct task_struct, se.group_node);
3776 /* We've more or less seen every task there is, call it quits */
3777 if (env->loop > env->loop_max)
3780 /* take a breather every nr_migrate tasks */
3781 if (env->loop > env->loop_break) {
3782 env->loop_break += sched_nr_migrate_break;
3783 env->flags |= LBF_NEED_BREAK;
3787 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3790 load = task_h_load(p);
3792 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3795 if ((load / 2) > env->imbalance)
3798 if (!can_migrate_task(p, env))
3803 env->imbalance -= load;
3805 #ifdef CONFIG_PREEMPT
3807 * NEWIDLE balancing is a source of latency, so preemptible
3808 * kernels will stop after the first task is pulled to minimize
3809 * the critical section.
3811 if (env->idle == CPU_NEWLY_IDLE)
3816 * We only want to steal up to the prescribed amount of
3819 if (env->imbalance <= 0)
3824 list_move_tail(&p->se.group_node, tasks);
3828 * Right now, this is one of only two places move_task() is called,
3829 * so we can safely collect move_task() stats here rather than
3830 * inside move_task().
3832 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3837 #ifdef CONFIG_FAIR_GROUP_SCHED
3839 * update tg->load_weight by folding this cpu's load_avg
3841 static void __update_blocked_averages_cpu(struct task_group *tg, int cpu)
3843 struct sched_entity *se = tg->se[cpu];
3844 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];
3846 /* throttled entities do not contribute to load */
3847 if (throttled_hierarchy(cfs_rq))
3850 update_cfs_rq_blocked_load(cfs_rq, 1);
3853 update_entity_load_avg(se, 1);
3855 * We pivot on our runnable average having decayed to zero for
3856 * list removal. This generally implies that all our children
3857 * have also been removed (modulo rounding error or bandwidth
3858 * control); however, such cases are rare and we can fix these
3861 * TODO: fix up out-of-order children on enqueue.
3863 if (!se->avg.runnable_avg_sum && !cfs_rq->nr_running)
3864 list_del_leaf_cfs_rq(cfs_rq);
3866 struct rq *rq = rq_of(cfs_rq);
3867 update_rq_runnable_avg(rq, rq->nr_running);
3871 static void update_blocked_averages(int cpu)
3873 struct rq *rq = cpu_rq(cpu);
3874 struct cfs_rq *cfs_rq;
3875 unsigned long flags;
3877 raw_spin_lock_irqsave(&rq->lock, flags);
3878 update_rq_clock(rq);
3880 * Iterates the task_group tree in a bottom up fashion, see
3881 * list_add_leaf_cfs_rq() for details.
3883 for_each_leaf_cfs_rq(rq, cfs_rq) {
3885 * Note: We may want to consider periodically releasing
3886 * rq->lock about these updates so that creating many task
3887 * groups does not result in continually extending hold time.
3889 __update_blocked_averages_cpu(cfs_rq->tg, rq->cpu);
3892 raw_spin_unlock_irqrestore(&rq->lock, flags);
3896 * Compute the cpu's hierarchical load factor for each task group.
3897 * This needs to be done in a top-down fashion because the load of a child
3898 * group is a fraction of its parents load.
3900 static int tg_load_down(struct task_group *tg, void *data)
3903 long cpu = (long)data;
3906 load = cpu_rq(cpu)->load.weight;
3908 load = tg->parent->cfs_rq[cpu]->h_load;
3909 load *= tg->se[cpu]->load.weight;
3910 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3913 tg->cfs_rq[cpu]->h_load = load;
3918 static void update_h_load(long cpu)
3920 struct rq *rq = cpu_rq(cpu);
3921 unsigned long now = jiffies;
3923 if (rq->h_load_throttle == now)
3926 rq->h_load_throttle = now;
3929 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3933 static unsigned long task_h_load(struct task_struct *p)
3935 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3938 load = p->se.load.weight;
3939 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3944 static inline void update_blocked_averages(int cpu)
3948 static inline void update_h_load(long cpu)
3952 static unsigned long task_h_load(struct task_struct *p)
3954 return p->se.load.weight;
3958 /********** Helpers for find_busiest_group ************************/
3960 * sd_lb_stats - Structure to store the statistics of a sched_domain
3961 * during load balancing.
3963 struct sd_lb_stats {
3964 struct sched_group *busiest; /* Busiest group in this sd */
3965 struct sched_group *this; /* Local group in this sd */
3966 unsigned long total_load; /* Total load of all groups in sd */
3967 unsigned long total_pwr; /* Total power of all groups in sd */
3968 unsigned long avg_load; /* Average load across all groups in sd */
3970 /** Statistics of this group */
3971 unsigned long this_load;
3972 unsigned long this_load_per_task;
3973 unsigned long this_nr_running;
3974 unsigned long this_has_capacity;
3975 unsigned int this_idle_cpus;
3977 /* Statistics of the busiest group */
3978 unsigned int busiest_idle_cpus;
3979 unsigned long max_load;
3980 unsigned long busiest_load_per_task;
3981 unsigned long busiest_nr_running;
3982 unsigned long busiest_group_capacity;
3983 unsigned long busiest_has_capacity;
3984 unsigned int busiest_group_weight;
3986 int group_imb; /* Is there imbalance in this sd */
3990 * sg_lb_stats - stats of a sched_group required for load_balancing
3992 struct sg_lb_stats {
3993 unsigned long avg_load; /*Avg load across the CPUs of the group */
3994 unsigned long group_load; /* Total load over the CPUs of the group */
3995 unsigned long sum_nr_running; /* Nr tasks running in the group */
3996 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3997 unsigned long group_capacity;
3998 unsigned long idle_cpus;
3999 unsigned long group_weight;
4000 int group_imb; /* Is there an imbalance in the group ? */
4001 int group_has_capacity; /* Is there extra capacity in the group? */
4005 * get_sd_load_idx - Obtain the load index for a given sched domain.
4006 * @sd: The sched_domain whose load_idx is to be obtained.
4007 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
4009 static inline int get_sd_load_idx(struct sched_domain *sd,
4010 enum cpu_idle_type idle)
4016 load_idx = sd->busy_idx;
4019 case CPU_NEWLY_IDLE:
4020 load_idx = sd->newidle_idx;
4023 load_idx = sd->idle_idx;
4030 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
4032 return SCHED_POWER_SCALE;
4035 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
4037 return default_scale_freq_power(sd, cpu);
4040 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
4042 unsigned long weight = sd->span_weight;
4043 unsigned long smt_gain = sd->smt_gain;
4050 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
4052 return default_scale_smt_power(sd, cpu);
4055 unsigned long scale_rt_power(int cpu)
4057 struct rq *rq = cpu_rq(cpu);
4058 u64 total, available, age_stamp, avg;
4061 * Since we're reading these variables without serialization make sure
4062 * we read them once before doing sanity checks on them.
4064 age_stamp = ACCESS_ONCE(rq->age_stamp);
4065 avg = ACCESS_ONCE(rq->rt_avg);
4067 total = sched_avg_period() + (rq->clock - age_stamp);
4069 if (unlikely(total < avg)) {
4070 /* Ensures that power won't end up being negative */
4073 available = total - avg;
4076 if (unlikely((s64)total < SCHED_POWER_SCALE))
4077 total = SCHED_POWER_SCALE;
4079 total >>= SCHED_POWER_SHIFT;
4081 return div_u64(available, total);
4084 static void update_cpu_power(struct sched_domain *sd, int cpu)
4086 unsigned long weight = sd->span_weight;
4087 unsigned long power = SCHED_POWER_SCALE;
4088 struct sched_group *sdg = sd->groups;
4090 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
4091 if (sched_feat(ARCH_POWER))
4092 power *= arch_scale_smt_power(sd, cpu);
4094 power *= default_scale_smt_power(sd, cpu);
4096 power >>= SCHED_POWER_SHIFT;
4099 sdg->sgp->power_orig = power;
4101 if (sched_feat(ARCH_POWER))
4102 power *= arch_scale_freq_power(sd, cpu);
4104 power *= default_scale_freq_power(sd, cpu);
4106 power >>= SCHED_POWER_SHIFT;
4108 power *= scale_rt_power(cpu);
4109 power >>= SCHED_POWER_SHIFT;
4114 cpu_rq(cpu)->cpu_power = power;
4115 sdg->sgp->power = power;
4118 void update_group_power(struct sched_domain *sd, int cpu)
4120 struct sched_domain *child = sd->child;
4121 struct sched_group *group, *sdg = sd->groups;
4122 unsigned long power;
4123 unsigned long interval;
4125 interval = msecs_to_jiffies(sd->balance_interval);
4126 interval = clamp(interval, 1UL, max_load_balance_interval);
4127 sdg->sgp->next_update = jiffies + interval;
4130 update_cpu_power(sd, cpu);
4136 if (child->flags & SD_OVERLAP) {
4138 * SD_OVERLAP domains cannot assume that child groups
4139 * span the current group.
4142 for_each_cpu(cpu, sched_group_cpus(sdg))
4143 power += power_of(cpu);
4146 * !SD_OVERLAP domains can assume that child groups
4147 * span the current group.
4150 group = child->groups;
4152 power += group->sgp->power;
4153 group = group->next;
4154 } while (group != child->groups);
4157 sdg->sgp->power_orig = sdg->sgp->power = power;
4161 * Try and fix up capacity for tiny siblings, this is needed when
4162 * things like SD_ASYM_PACKING need f_b_g to select another sibling
4163 * which on its own isn't powerful enough.
4165 * See update_sd_pick_busiest() and check_asym_packing().
4168 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
4171 * Only siblings can have significantly less than SCHED_POWER_SCALE
4173 if (!(sd->flags & SD_SHARE_CPUPOWER))
4177 * If ~90% of the cpu_power is still there, we're good.
4179 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
4186 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
4187 * @env: The load balancing environment.
4188 * @group: sched_group whose statistics are to be updated.
4189 * @load_idx: Load index of sched_domain of this_cpu for load calc.
4190 * @local_group: Does group contain this_cpu.
4191 * @balance: Should we balance.
4192 * @sgs: variable to hold the statistics for this group.
4194 static inline void update_sg_lb_stats(struct lb_env *env,
4195 struct sched_group *group, int load_idx,
4196 int local_group, int *balance, struct sg_lb_stats *sgs)
4198 unsigned long nr_running, max_nr_running, min_nr_running;
4199 unsigned long load, max_cpu_load, min_cpu_load;
4200 unsigned int balance_cpu = -1, first_idle_cpu = 0;
4201 unsigned long avg_load_per_task = 0;
4205 balance_cpu = group_balance_cpu(group);
4207 /* Tally up the load of all CPUs in the group */
4209 min_cpu_load = ~0UL;
4211 min_nr_running = ~0UL;
4213 for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
4214 struct rq *rq = cpu_rq(i);
4216 nr_running = rq->nr_running;
4218 /* Bias balancing toward cpus of our domain */
4220 if (idle_cpu(i) && !first_idle_cpu &&
4221 cpumask_test_cpu(i, sched_group_mask(group))) {
4226 load = target_load(i, load_idx);
4228 load = source_load(i, load_idx);
4229 if (load > max_cpu_load)
4230 max_cpu_load = load;
4231 if (min_cpu_load > load)
4232 min_cpu_load = load;
4234 if (nr_running > max_nr_running)
4235 max_nr_running = nr_running;
4236 if (min_nr_running > nr_running)
4237 min_nr_running = nr_running;
4240 sgs->group_load += load;
4241 sgs->sum_nr_running += nr_running;
4242 sgs->sum_weighted_load += weighted_cpuload(i);
4248 * First idle cpu or the first cpu(busiest) in this sched group
4249 * is eligible for doing load balancing at this and above
4250 * domains. In the newly idle case, we will allow all the cpu's
4251 * to do the newly idle load balance.
4254 if (env->idle != CPU_NEWLY_IDLE) {
4255 if (balance_cpu != env->dst_cpu) {
4259 update_group_power(env->sd, env->dst_cpu);
4260 } else if (time_after_eq(jiffies, group->sgp->next_update))
4261 update_group_power(env->sd, env->dst_cpu);
4264 /* Adjust by relative CPU power of the group */
4265 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
4268 * Consider the group unbalanced when the imbalance is larger
4269 * than the average weight of a task.
4271 * APZ: with cgroup the avg task weight can vary wildly and
4272 * might not be a suitable number - should we keep a
4273 * normalized nr_running number somewhere that negates
4276 if (sgs->sum_nr_running)
4277 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
4279 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
4280 (max_nr_running - min_nr_running) > 1)
4283 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
4285 if (!sgs->group_capacity)
4286 sgs->group_capacity = fix_small_capacity(env->sd, group);
4287 sgs->group_weight = group->group_weight;
4289 if (sgs->group_capacity > sgs->sum_nr_running)
4290 sgs->group_has_capacity = 1;
4294 * update_sd_pick_busiest - return 1 on busiest group
4295 * @env: The load balancing environment.
4296 * @sds: sched_domain statistics
4297 * @sg: sched_group candidate to be checked for being the busiest
4298 * @sgs: sched_group statistics
4300 * Determine if @sg is a busier group than the previously selected
4303 static bool update_sd_pick_busiest(struct lb_env *env,
4304 struct sd_lb_stats *sds,
4305 struct sched_group *sg,
4306 struct sg_lb_stats *sgs)
4308 if (sgs->avg_load <= sds->max_load)
4311 if (sgs->sum_nr_running > sgs->group_capacity)
4318 * ASYM_PACKING needs to move all the work to the lowest
4319 * numbered CPUs in the group, therefore mark all groups
4320 * higher than ourself as busy.
4322 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
4323 env->dst_cpu < group_first_cpu(sg)) {
4327 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
4335 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
4336 * @env: The load balancing environment.
4337 * @balance: Should we balance.
4338 * @sds: variable to hold the statistics for this sched_domain.
4340 static inline void update_sd_lb_stats(struct lb_env *env,
4341 int *balance, struct sd_lb_stats *sds)
4343 struct sched_domain *child = env->sd->child;
4344 struct sched_group *sg = env->sd->groups;
4345 struct sg_lb_stats sgs;
4346 int load_idx, prefer_sibling = 0;
4348 if (child && child->flags & SD_PREFER_SIBLING)
4351 load_idx = get_sd_load_idx(env->sd, env->idle);
4356 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
4357 memset(&sgs, 0, sizeof(sgs));
4358 update_sg_lb_stats(env, sg, load_idx, local_group, balance, &sgs);
4360 if (local_group && !(*balance))
4363 sds->total_load += sgs.group_load;
4364 sds->total_pwr += sg->sgp->power;
4367 * In case the child domain prefers tasks go to siblings
4368 * first, lower the sg capacity to one so that we'll try
4369 * and move all the excess tasks away. We lower the capacity
4370 * of a group only if the local group has the capacity to fit
4371 * these excess tasks, i.e. nr_running < group_capacity. The
4372 * extra check prevents the case where you always pull from the
4373 * heaviest group when it is already under-utilized (possible
4374 * with a large weight task outweighs the tasks on the system).
4376 if (prefer_sibling && !local_group && sds->this_has_capacity)
4377 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4380 sds->this_load = sgs.avg_load;
4382 sds->this_nr_running = sgs.sum_nr_running;
4383 sds->this_load_per_task = sgs.sum_weighted_load;
4384 sds->this_has_capacity = sgs.group_has_capacity;
4385 sds->this_idle_cpus = sgs.idle_cpus;
4386 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
4387 sds->max_load = sgs.avg_load;
4389 sds->busiest_nr_running = sgs.sum_nr_running;
4390 sds->busiest_idle_cpus = sgs.idle_cpus;
4391 sds->busiest_group_capacity = sgs.group_capacity;
4392 sds->busiest_load_per_task = sgs.sum_weighted_load;
4393 sds->busiest_has_capacity = sgs.group_has_capacity;
4394 sds->busiest_group_weight = sgs.group_weight;
4395 sds->group_imb = sgs.group_imb;
4399 } while (sg != env->sd->groups);
4403 * check_asym_packing - Check to see if the group is packed into the
4406 * This is primarily intended to used at the sibling level. Some
4407 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4408 * case of POWER7, it can move to lower SMT modes only when higher
4409 * threads are idle. When in lower SMT modes, the threads will
4410 * perform better since they share less core resources. Hence when we
4411 * have idle threads, we want them to be the higher ones.
4413 * This packing function is run on idle threads. It checks to see if
4414 * the busiest CPU in this domain (core in the P7 case) has a higher
4415 * CPU number than the packing function is being run on. Here we are
4416 * assuming lower CPU number will be equivalent to lower a SMT thread
4419 * Returns 1 when packing is required and a task should be moved to
4420 * this CPU. The amount of the imbalance is returned in *imbalance.
4422 * @env: The load balancing environment.
4423 * @sds: Statistics of the sched_domain which is to be packed
4425 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
4429 if (!(env->sd->flags & SD_ASYM_PACKING))
4435 busiest_cpu = group_first_cpu(sds->busiest);
4436 if (env->dst_cpu > busiest_cpu)
4439 env->imbalance = DIV_ROUND_CLOSEST(
4440 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
4446 * fix_small_imbalance - Calculate the minor imbalance that exists
4447 * amongst the groups of a sched_domain, during
4449 * @env: The load balancing environment.
4450 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4453 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4455 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4456 unsigned int imbn = 2;
4457 unsigned long scaled_busy_load_per_task;
4459 if (sds->this_nr_running) {
4460 sds->this_load_per_task /= sds->this_nr_running;
4461 if (sds->busiest_load_per_task >
4462 sds->this_load_per_task)
4465 sds->this_load_per_task =
4466 cpu_avg_load_per_task(env->dst_cpu);
4469 scaled_busy_load_per_task = sds->busiest_load_per_task
4470 * SCHED_POWER_SCALE;
4471 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4473 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4474 (scaled_busy_load_per_task * imbn)) {
4475 env->imbalance = sds->busiest_load_per_task;
4480 * OK, we don't have enough imbalance to justify moving tasks,
4481 * however we may be able to increase total CPU power used by
4485 pwr_now += sds->busiest->sgp->power *
4486 min(sds->busiest_load_per_task, sds->max_load);
4487 pwr_now += sds->this->sgp->power *
4488 min(sds->this_load_per_task, sds->this_load);
4489 pwr_now /= SCHED_POWER_SCALE;
4491 /* Amount of load we'd subtract */
4492 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4493 sds->busiest->sgp->power;
4494 if (sds->max_load > tmp)
4495 pwr_move += sds->busiest->sgp->power *
4496 min(sds->busiest_load_per_task, sds->max_load - tmp);
4498 /* Amount of load we'd add */
4499 if (sds->max_load * sds->busiest->sgp->power <
4500 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4501 tmp = (sds->max_load * sds->busiest->sgp->power) /
4502 sds->this->sgp->power;
4504 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4505 sds->this->sgp->power;
4506 pwr_move += sds->this->sgp->power *
4507 min(sds->this_load_per_task, sds->this_load + tmp);
4508 pwr_move /= SCHED_POWER_SCALE;
4510 /* Move if we gain throughput */
4511 if (pwr_move > pwr_now)
4512 env->imbalance = sds->busiest_load_per_task;
4516 * calculate_imbalance - Calculate the amount of imbalance present within the
4517 * groups of a given sched_domain during load balance.
4518 * @env: load balance environment
4519 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4521 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
4523 unsigned long max_pull, load_above_capacity = ~0UL;
4525 sds->busiest_load_per_task /= sds->busiest_nr_running;
4526 if (sds->group_imb) {
4527 sds->busiest_load_per_task =
4528 min(sds->busiest_load_per_task, sds->avg_load);
4532 * In the presence of smp nice balancing, certain scenarios can have
4533 * max load less than avg load(as we skip the groups at or below
4534 * its cpu_power, while calculating max_load..)
4536 if (sds->max_load < sds->avg_load) {
4538 return fix_small_imbalance(env, sds);
4541 if (!sds->group_imb) {
4543 * Don't want to pull so many tasks that a group would go idle.
4545 load_above_capacity = (sds->busiest_nr_running -
4546 sds->busiest_group_capacity);
4548 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4550 load_above_capacity /= sds->busiest->sgp->power;
4554 * We're trying to get all the cpus to the average_load, so we don't
4555 * want to push ourselves above the average load, nor do we wish to
4556 * reduce the max loaded cpu below the average load. At the same time,
4557 * we also don't want to reduce the group load below the group capacity
4558 * (so that we can implement power-savings policies etc). Thus we look
4559 * for the minimum possible imbalance.
4560 * Be careful of negative numbers as they'll appear as very large values
4561 * with unsigned longs.
4563 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4565 /* How much load to actually move to equalise the imbalance */
4566 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4567 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4568 / SCHED_POWER_SCALE;
4571 * if *imbalance is less than the average load per runnable task
4572 * there is no guarantee that any tasks will be moved so we'll have
4573 * a think about bumping its value to force at least one task to be
4576 if (env->imbalance < sds->busiest_load_per_task)
4577 return fix_small_imbalance(env, sds);
4581 /******* find_busiest_group() helpers end here *********************/
4584 * find_busiest_group - Returns the busiest group within the sched_domain
4585 * if there is an imbalance. If there isn't an imbalance, and
4586 * the user has opted for power-savings, it returns a group whose
4587 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4588 * such a group exists.
4590 * Also calculates the amount of weighted load which should be moved
4591 * to restore balance.
4593 * @env: The load balancing environment.
4594 * @balance: Pointer to a variable indicating if this_cpu
4595 * is the appropriate cpu to perform load balancing at this_level.
4597 * Returns: - the busiest group if imbalance exists.
4598 * - If no imbalance and user has opted for power-savings balance,
4599 * return the least loaded group whose CPUs can be
4600 * put to idle by rebalancing its tasks onto our group.
4602 static struct sched_group *
4603 find_busiest_group(struct lb_env *env, int *balance)
4605 struct sd_lb_stats sds;
4607 memset(&sds, 0, sizeof(sds));
4610 * Compute the various statistics relavent for load balancing at
4613 update_sd_lb_stats(env, balance, &sds);
4616 * this_cpu is not the appropriate cpu to perform load balancing at
4622 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4623 check_asym_packing(env, &sds))
4626 /* There is no busy sibling group to pull tasks from */
4627 if (!sds.busiest || sds.busiest_nr_running == 0)
4630 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4633 * If the busiest group is imbalanced the below checks don't
4634 * work because they assumes all things are equal, which typically
4635 * isn't true due to cpus_allowed constraints and the like.
4640 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4641 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4642 !sds.busiest_has_capacity)
4646 * If the local group is more busy than the selected busiest group
4647 * don't try and pull any tasks.
4649 if (sds.this_load >= sds.max_load)
4653 * Don't pull any tasks if this group is already above the domain
4656 if (sds.this_load >= sds.avg_load)
4659 if (env->idle == CPU_IDLE) {
4661 * This cpu is idle. If the busiest group load doesn't
4662 * have more tasks than the number of available cpu's and
4663 * there is no imbalance between this and busiest group
4664 * wrt to idle cpu's, it is balanced.
4666 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4667 sds.busiest_nr_running <= sds.busiest_group_weight)
4671 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4672 * imbalance_pct to be conservative.
4674 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4679 /* Looks like there is an imbalance. Compute it */
4680 calculate_imbalance(env, &sds);
4690 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4692 static struct rq *find_busiest_queue(struct lb_env *env,
4693 struct sched_group *group)
4695 struct rq *busiest = NULL, *rq;
4696 unsigned long max_load = 0;
4699 for_each_cpu(i, sched_group_cpus(group)) {
4700 unsigned long power = power_of(i);
4701 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4706 capacity = fix_small_capacity(env->sd, group);
4708 if (!cpumask_test_cpu(i, env->cpus))
4712 wl = weighted_cpuload(i);
4715 * When comparing with imbalance, use weighted_cpuload()
4716 * which is not scaled with the cpu power.
4718 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4722 * For the load comparisons with the other cpu's, consider
4723 * the weighted_cpuload() scaled with the cpu power, so that
4724 * the load can be moved away from the cpu that is potentially
4725 * running at a lower capacity.
4727 wl = (wl * SCHED_POWER_SCALE) / power;
4729 if (wl > max_load) {
4739 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4740 * so long as it is large enough.
4742 #define MAX_PINNED_INTERVAL 512
4744 /* Working cpumask for load_balance and load_balance_newidle. */
4745 DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4747 static int need_active_balance(struct lb_env *env)
4749 struct sched_domain *sd = env->sd;
4751 if (env->idle == CPU_NEWLY_IDLE) {
4754 * ASYM_PACKING needs to force migrate tasks from busy but
4755 * higher numbered CPUs in order to pack all tasks in the
4756 * lowest numbered CPUs.
4758 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4762 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4765 static int active_load_balance_cpu_stop(void *data);
4768 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4769 * tasks if there is an imbalance.
4771 static int load_balance(int this_cpu, struct rq *this_rq,
4772 struct sched_domain *sd, enum cpu_idle_type idle,
4775 int ld_moved, cur_ld_moved, active_balance = 0;
4776 int lb_iterations, max_lb_iterations;
4777 struct sched_group *group;
4779 unsigned long flags;
4780 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4782 struct lb_env env = {
4784 .dst_cpu = this_cpu,
4786 .dst_grpmask = sched_group_cpus(sd->groups),
4788 .loop_break = sched_nr_migrate_break,
4792 cpumask_copy(cpus, cpu_active_mask);
4793 max_lb_iterations = cpumask_weight(env.dst_grpmask);
4795 schedstat_inc(sd, lb_count[idle]);
4798 group = find_busiest_group(&env, balance);
4804 schedstat_inc(sd, lb_nobusyg[idle]);
4808 busiest = find_busiest_queue(&env, group);
4810 schedstat_inc(sd, lb_nobusyq[idle]);
4814 BUG_ON(busiest == env.dst_rq);
4816 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4820 if (busiest->nr_running > 1) {
4822 * Attempt to move tasks. If find_busiest_group has found
4823 * an imbalance but busiest->nr_running <= 1, the group is
4824 * still unbalanced. ld_moved simply stays zero, so it is
4825 * correctly treated as an imbalance.
4827 env.flags |= LBF_ALL_PINNED;
4828 env.src_cpu = busiest->cpu;
4829 env.src_rq = busiest;
4830 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4832 update_h_load(env.src_cpu);
4834 local_irq_save(flags);
4835 double_rq_lock(env.dst_rq, busiest);
4838 * cur_ld_moved - load moved in current iteration
4839 * ld_moved - cumulative load moved across iterations
4841 cur_ld_moved = move_tasks(&env);
4842 ld_moved += cur_ld_moved;
4843 double_rq_unlock(env.dst_rq, busiest);
4844 local_irq_restore(flags);
4846 if (env.flags & LBF_NEED_BREAK) {
4847 env.flags &= ~LBF_NEED_BREAK;
4852 * some other cpu did the load balance for us.
4854 if (cur_ld_moved && env.dst_cpu != smp_processor_id())
4855 resched_cpu(env.dst_cpu);
4858 * Revisit (affine) tasks on src_cpu that couldn't be moved to
4859 * us and move them to an alternate dst_cpu in our sched_group
4860 * where they can run. The upper limit on how many times we
4861 * iterate on same src_cpu is dependent on number of cpus in our
4864 * This changes load balance semantics a bit on who can move
4865 * load to a given_cpu. In addition to the given_cpu itself
4866 * (or a ilb_cpu acting on its behalf where given_cpu is
4867 * nohz-idle), we now have balance_cpu in a position to move
4868 * load to given_cpu. In rare situations, this may cause
4869 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
4870 * _independently_ and at _same_ time to move some load to
4871 * given_cpu) causing exceess load to be moved to given_cpu.
4872 * This however should not happen so much in practice and
4873 * moreover subsequent load balance cycles should correct the
4874 * excess load moved.
4876 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0 &&
4877 lb_iterations++ < max_lb_iterations) {
4879 env.dst_rq = cpu_rq(env.new_dst_cpu);
4880 env.dst_cpu = env.new_dst_cpu;
4881 env.flags &= ~LBF_SOME_PINNED;
4883 env.loop_break = sched_nr_migrate_break;
4885 * Go back to "more_balance" rather than "redo" since we
4886 * need to continue with same src_cpu.
4891 /* All tasks on this runqueue were pinned by CPU affinity */
4892 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4893 cpumask_clear_cpu(cpu_of(busiest), cpus);
4894 if (!cpumask_empty(cpus)) {
4896 env.loop_break = sched_nr_migrate_break;
4904 schedstat_inc(sd, lb_failed[idle]);
4906 * Increment the failure counter only on periodic balance.
4907 * We do not want newidle balance, which can be very
4908 * frequent, pollute the failure counter causing
4909 * excessive cache_hot migrations and active balances.
4911 if (idle != CPU_NEWLY_IDLE)
4912 sd->nr_balance_failed++;
4914 if (need_active_balance(&env)) {
4915 raw_spin_lock_irqsave(&busiest->lock, flags);
4917 /* don't kick the active_load_balance_cpu_stop,
4918 * if the curr task on busiest cpu can't be
4921 if (!cpumask_test_cpu(this_cpu,
4922 tsk_cpus_allowed(busiest->curr))) {
4923 raw_spin_unlock_irqrestore(&busiest->lock,
4925 env.flags |= LBF_ALL_PINNED;
4926 goto out_one_pinned;
4930 * ->active_balance synchronizes accesses to
4931 * ->active_balance_work. Once set, it's cleared
4932 * only after active load balance is finished.
4934 if (!busiest->active_balance) {
4935 busiest->active_balance = 1;
4936 busiest->push_cpu = this_cpu;
4939 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4941 if (active_balance) {
4942 stop_one_cpu_nowait(cpu_of(busiest),
4943 active_load_balance_cpu_stop, busiest,
4944 &busiest->active_balance_work);
4948 * We've kicked active balancing, reset the failure
4951 sd->nr_balance_failed = sd->cache_nice_tries+1;
4954 sd->nr_balance_failed = 0;
4956 if (likely(!active_balance)) {
4957 /* We were unbalanced, so reset the balancing interval */
4958 sd->balance_interval = sd->min_interval;
4961 * If we've begun active balancing, start to back off. This
4962 * case may not be covered by the all_pinned logic if there
4963 * is only 1 task on the busy runqueue (because we don't call
4966 if (sd->balance_interval < sd->max_interval)
4967 sd->balance_interval *= 2;
4973 schedstat_inc(sd, lb_balanced[idle]);
4975 sd->nr_balance_failed = 0;
4978 /* tune up the balancing interval */
4979 if (((env.flags & LBF_ALL_PINNED) &&
4980 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4981 (sd->balance_interval < sd->max_interval))
4982 sd->balance_interval *= 2;
4990 * idle_balance is called by schedule() if this_cpu is about to become
4991 * idle. Attempts to pull tasks from other CPUs.
4993 void idle_balance(int this_cpu, struct rq *this_rq)
4995 struct sched_domain *sd;
4996 int pulled_task = 0;
4997 unsigned long next_balance = jiffies + HZ;
4999 this_rq->idle_stamp = this_rq->clock;
5001 if (this_rq->avg_idle < sysctl_sched_migration_cost)
5004 update_rq_runnable_avg(this_rq, 1);
5007 * Drop the rq->lock, but keep IRQ/preempt disabled.
5009 raw_spin_unlock(&this_rq->lock);
5011 update_blocked_averages(this_cpu);
5013 for_each_domain(this_cpu, sd) {
5014 unsigned long interval;
5017 if (!(sd->flags & SD_LOAD_BALANCE))
5020 if (sd->flags & SD_BALANCE_NEWIDLE) {
5021 /* If we've pulled tasks over stop searching: */
5022 pulled_task = load_balance(this_cpu, this_rq,
5023 sd, CPU_NEWLY_IDLE, &balance);
5026 interval = msecs_to_jiffies(sd->balance_interval);
5027 if (time_after(next_balance, sd->last_balance + interval))
5028 next_balance = sd->last_balance + interval;
5030 this_rq->idle_stamp = 0;
5036 raw_spin_lock(&this_rq->lock);
5038 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
5040 * We are going idle. next_balance may be set based on
5041 * a busy processor. So reset next_balance.
5043 this_rq->next_balance = next_balance;
5048 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
5049 * running tasks off the busiest CPU onto idle CPUs. It requires at
5050 * least 1 task to be running on each physical CPU where possible, and
5051 * avoids physical / logical imbalances.
5053 static int active_load_balance_cpu_stop(void *data)
5055 struct rq *busiest_rq = data;
5056 int busiest_cpu = cpu_of(busiest_rq);
5057 int target_cpu = busiest_rq->push_cpu;
5058 struct rq *target_rq = cpu_rq(target_cpu);
5059 struct sched_domain *sd;
5061 raw_spin_lock_irq(&busiest_rq->lock);
5063 /* make sure the requested cpu hasn't gone down in the meantime */
5064 if (unlikely(busiest_cpu != smp_processor_id() ||
5065 !busiest_rq->active_balance))
5068 /* Is there any task to move? */
5069 if (busiest_rq->nr_running <= 1)
5073 * This condition is "impossible", if it occurs
5074 * we need to fix it. Originally reported by
5075 * Bjorn Helgaas on a 128-cpu setup.
5077 BUG_ON(busiest_rq == target_rq);
5079 /* move a task from busiest_rq to target_rq */
5080 double_lock_balance(busiest_rq, target_rq);
5082 /* Search for an sd spanning us and the target CPU. */
5084 for_each_domain(target_cpu, sd) {
5085 if ((sd->flags & SD_LOAD_BALANCE) &&
5086 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
5091 struct lb_env env = {
5093 .dst_cpu = target_cpu,
5094 .dst_rq = target_rq,
5095 .src_cpu = busiest_rq->cpu,
5096 .src_rq = busiest_rq,
5100 schedstat_inc(sd, alb_count);
5102 if (move_one_task(&env))
5103 schedstat_inc(sd, alb_pushed);
5105 schedstat_inc(sd, alb_failed);
5108 double_unlock_balance(busiest_rq, target_rq);
5110 busiest_rq->active_balance = 0;
5111 raw_spin_unlock_irq(&busiest_rq->lock);
5117 * idle load balancing details
5118 * - When one of the busy CPUs notice that there may be an idle rebalancing
5119 * needed, they will kick the idle load balancer, which then does idle
5120 * load balancing for all the idle CPUs.
5123 cpumask_var_t idle_cpus_mask;
5125 unsigned long next_balance; /* in jiffy units */
5126 } nohz ____cacheline_aligned;
5128 static inline int find_new_ilb(int call_cpu)
5130 int ilb = cpumask_first(nohz.idle_cpus_mask);
5132 if (ilb < nr_cpu_ids && idle_cpu(ilb))
5139 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
5140 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
5141 * CPU (if there is one).
5143 static void nohz_balancer_kick(int cpu)
5147 nohz.next_balance++;
5149 ilb_cpu = find_new_ilb(cpu);
5151 if (ilb_cpu >= nr_cpu_ids)
5154 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
5157 * Use smp_send_reschedule() instead of resched_cpu().
5158 * This way we generate a sched IPI on the target cpu which
5159 * is idle. And the softirq performing nohz idle load balance
5160 * will be run before returning from the IPI.
5162 smp_send_reschedule(ilb_cpu);
5166 static inline void nohz_balance_exit_idle(int cpu)
5168 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
5169 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
5170 atomic_dec(&nohz.nr_cpus);
5171 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5175 static inline void set_cpu_sd_state_busy(void)
5177 struct sched_domain *sd;
5178 int cpu = smp_processor_id();
5180 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5182 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
5185 for_each_domain(cpu, sd)
5186 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
5190 void set_cpu_sd_state_idle(void)
5192 struct sched_domain *sd;
5193 int cpu = smp_processor_id();
5195 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
5197 set_bit(NOHZ_IDLE, nohz_flags(cpu));
5200 for_each_domain(cpu, sd)
5201 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
5206 * This routine will record that the cpu is going idle with tick stopped.
5207 * This info will be used in performing idle load balancing in the future.
5209 void nohz_balance_enter_idle(int cpu)
5212 * If this cpu is going down, then nothing needs to be done.
5214 if (!cpu_active(cpu))
5217 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
5220 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
5221 atomic_inc(&nohz.nr_cpus);
5222 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
5225 static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
5226 unsigned long action, void *hcpu)
5228 switch (action & ~CPU_TASKS_FROZEN) {
5230 nohz_balance_exit_idle(smp_processor_id());
5238 static DEFINE_SPINLOCK(balancing);
5241 * Scale the max load_balance interval with the number of CPUs in the system.
5242 * This trades load-balance latency on larger machines for less cross talk.
5244 void update_max_interval(void)
5246 max_load_balance_interval = HZ*num_online_cpus()/10;
5250 * It checks each scheduling domain to see if it is due to be balanced,
5251 * and initiates a balancing operation if so.
5253 * Balancing parameters are set up in arch_init_sched_domains.
5255 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
5258 struct rq *rq = cpu_rq(cpu);
5259 unsigned long interval;
5260 struct sched_domain *sd;
5261 /* Earliest time when we have to do rebalance again */
5262 unsigned long next_balance = jiffies + 60*HZ;
5263 int update_next_balance = 0;
5266 update_blocked_averages(cpu);
5269 for_each_domain(cpu, sd) {
5270 if (!(sd->flags & SD_LOAD_BALANCE))
5273 interval = sd->balance_interval;
5274 if (idle != CPU_IDLE)
5275 interval *= sd->busy_factor;
5277 /* scale ms to jiffies */
5278 interval = msecs_to_jiffies(interval);
5279 interval = clamp(interval, 1UL, max_load_balance_interval);
5281 need_serialize = sd->flags & SD_SERIALIZE;
5283 if (need_serialize) {
5284 if (!spin_trylock(&balancing))
5288 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5289 if (load_balance(cpu, rq, sd, idle, &balance)) {
5291 * We've pulled tasks over so either we're no
5294 idle = CPU_NOT_IDLE;
5296 sd->last_balance = jiffies;
5299 spin_unlock(&balancing);
5301 if (time_after(next_balance, sd->last_balance + interval)) {
5302 next_balance = sd->last_balance + interval;
5303 update_next_balance = 1;
5307 * Stop the load balance at this level. There is another
5308 * CPU in our sched group which is doing load balancing more
5317 * next_balance will be updated only when there is a need.
5318 * When the cpu is attached to null domain for ex, it will not be
5321 if (likely(update_next_balance))
5322 rq->next_balance = next_balance;
5327 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5328 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5330 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5332 struct rq *this_rq = cpu_rq(this_cpu);
5336 if (idle != CPU_IDLE ||
5337 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
5340 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5341 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
5345 * If this cpu gets work to do, stop the load balancing
5346 * work being done for other cpus. Next load
5347 * balancing owner will pick it up.
5352 rq = cpu_rq(balance_cpu);
5354 raw_spin_lock_irq(&rq->lock);
5355 update_rq_clock(rq);
5356 update_idle_cpu_load(rq);
5357 raw_spin_unlock_irq(&rq->lock);
5359 rebalance_domains(balance_cpu, CPU_IDLE);
5361 if (time_after(this_rq->next_balance, rq->next_balance))
5362 this_rq->next_balance = rq->next_balance;
5364 nohz.next_balance = this_rq->next_balance;
5366 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
5370 * Current heuristic for kicking the idle load balancer in the presence
5371 * of an idle cpu is the system.
5372 * - This rq has more than one task.
5373 * - At any scheduler domain level, this cpu's scheduler group has multiple
5374 * busy cpu's exceeding the group's power.
5375 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
5376 * domain span are idle.
5378 static inline int nohz_kick_needed(struct rq *rq, int cpu)
5380 unsigned long now = jiffies;
5381 struct sched_domain *sd;
5383 if (unlikely(idle_cpu(cpu)))
5387 * We may be recently in ticked or tickless idle mode. At the first
5388 * busy tick after returning from idle, we will update the busy stats.
5390 set_cpu_sd_state_busy();
5391 nohz_balance_exit_idle(cpu);
5394 * None are in tickless mode and hence no need for NOHZ idle load
5397 if (likely(!atomic_read(&nohz.nr_cpus)))
5400 if (time_before(now, nohz.next_balance))
5403 if (rq->nr_running >= 2)
5407 for_each_domain(cpu, sd) {
5408 struct sched_group *sg = sd->groups;
5409 struct sched_group_power *sgp = sg->sgp;
5410 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
5412 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
5413 goto need_kick_unlock;
5415 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
5416 && (cpumask_first_and(nohz.idle_cpus_mask,
5417 sched_domain_span(sd)) < cpu))
5418 goto need_kick_unlock;
5420 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
5432 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5436 * run_rebalance_domains is triggered when needed from the scheduler tick.
5437 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5439 static void run_rebalance_domains(struct softirq_action *h)
5441 int this_cpu = smp_processor_id();
5442 struct rq *this_rq = cpu_rq(this_cpu);
5443 enum cpu_idle_type idle = this_rq->idle_balance ?
5444 CPU_IDLE : CPU_NOT_IDLE;
5446 rebalance_domains(this_cpu, idle);
5449 * If this cpu has a pending nohz_balance_kick, then do the
5450 * balancing on behalf of the other idle cpus whose ticks are
5453 nohz_idle_balance(this_cpu, idle);
5456 static inline int on_null_domain(int cpu)
5458 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5462 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5464 void trigger_load_balance(struct rq *rq, int cpu)
5466 /* Don't need to rebalance while attached to NULL domain */
5467 if (time_after_eq(jiffies, rq->next_balance) &&
5468 likely(!on_null_domain(cpu)))
5469 raise_softirq(SCHED_SOFTIRQ);
5471 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5472 nohz_balancer_kick(cpu);
5476 static void rq_online_fair(struct rq *rq)
5481 static void rq_offline_fair(struct rq *rq)
5485 /* Ensure any throttled groups are reachable by pick_next_task */
5486 unthrottle_offline_cfs_rqs(rq);
5489 #endif /* CONFIG_SMP */
5492 * scheduler tick hitting a task of our scheduling class:
5494 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5496 struct cfs_rq *cfs_rq;
5497 struct sched_entity *se = &curr->se;
5499 for_each_sched_entity(se) {
5500 cfs_rq = cfs_rq_of(se);
5501 entity_tick(cfs_rq, se, queued);
5504 update_rq_runnable_avg(rq, 1);
5508 * called on fork with the child task as argument from the parent's context
5509 * - child not yet on the tasklist
5510 * - preemption disabled
5512 static void task_fork_fair(struct task_struct *p)
5514 struct cfs_rq *cfs_rq;
5515 struct sched_entity *se = &p->se, *curr;
5516 int this_cpu = smp_processor_id();
5517 struct rq *rq = this_rq();
5518 unsigned long flags;
5520 raw_spin_lock_irqsave(&rq->lock, flags);
5522 update_rq_clock(rq);
5524 cfs_rq = task_cfs_rq(current);
5525 curr = cfs_rq->curr;
5527 if (unlikely(task_cpu(p) != this_cpu)) {
5529 __set_task_cpu(p, this_cpu);
5533 update_curr(cfs_rq);
5536 se->vruntime = curr->vruntime;
5537 place_entity(cfs_rq, se, 1);
5539 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5541 * Upon rescheduling, sched_class::put_prev_task() will place
5542 * 'current' within the tree based on its new key value.
5544 swap(curr->vruntime, se->vruntime);
5545 resched_task(rq->curr);
5548 se->vruntime -= cfs_rq->min_vruntime;
5550 raw_spin_unlock_irqrestore(&rq->lock, flags);
5554 * Priority of the task has changed. Check to see if we preempt
5558 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5564 * Reschedule if we are currently running on this runqueue and
5565 * our priority decreased, or if we are not currently running on
5566 * this runqueue and our priority is higher than the current's
5568 if (rq->curr == p) {
5569 if (p->prio > oldprio)
5570 resched_task(rq->curr);
5572 check_preempt_curr(rq, p, 0);
5575 static void switched_from_fair(struct rq *rq, struct task_struct *p)
5577 struct sched_entity *se = &p->se;
5578 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5581 * Ensure the task's vruntime is normalized, so that when its
5582 * switched back to the fair class the enqueue_entity(.flags=0) will
5583 * do the right thing.
5585 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5586 * have normalized the vruntime, if it was !on_rq, then only when
5587 * the task is sleeping will it still have non-normalized vruntime.
5589 if (!se->on_rq && p->state != TASK_RUNNING) {
5591 * Fix up our vruntime so that the current sleep doesn't
5592 * cause 'unlimited' sleep bonus.
5594 place_entity(cfs_rq, se, 0);
5595 se->vruntime -= cfs_rq->min_vruntime;
5598 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5600 * Remove our load from contribution when we leave sched_fair
5601 * and ensure we don't carry in an old decay_count if we
5604 if (p->se.avg.decay_count) {
5605 struct cfs_rq *cfs_rq = cfs_rq_of(&p->se);
5606 __synchronize_entity_decay(&p->se);
5607 subtract_blocked_load_contrib(cfs_rq,
5608 p->se.avg.load_avg_contrib);
5614 * We switched to the sched_fair class.
5616 static void switched_to_fair(struct rq *rq, struct task_struct *p)
5622 * We were most likely switched from sched_rt, so
5623 * kick off the schedule if running, otherwise just see
5624 * if we can still preempt the current task.
5627 resched_task(rq->curr);
5629 check_preempt_curr(rq, p, 0);
5632 /* Account for a task changing its policy or group.
5634 * This routine is mostly called to set cfs_rq->curr field when a task
5635 * migrates between groups/classes.
5637 static void set_curr_task_fair(struct rq *rq)
5639 struct sched_entity *se = &rq->curr->se;
5641 for_each_sched_entity(se) {
5642 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5644 set_next_entity(cfs_rq, se);
5645 /* ensure bandwidth has been allocated on our new cfs_rq */
5646 account_cfs_rq_runtime(cfs_rq, 0);
5650 void init_cfs_rq(struct cfs_rq *cfs_rq)
5652 cfs_rq->tasks_timeline = RB_ROOT;
5653 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5654 #ifndef CONFIG_64BIT
5655 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5657 #if defined(CONFIG_FAIR_GROUP_SCHED) && defined(CONFIG_SMP)
5658 atomic64_set(&cfs_rq->decay_counter, 1);
5659 atomic64_set(&cfs_rq->removed_load, 0);
5663 #ifdef CONFIG_FAIR_GROUP_SCHED
5664 static void task_move_group_fair(struct task_struct *p, int on_rq)
5666 struct cfs_rq *cfs_rq;
5668 * If the task was not on the rq at the time of this cgroup movement
5669 * it must have been asleep, sleeping tasks keep their ->vruntime
5670 * absolute on their old rq until wakeup (needed for the fair sleeper
5671 * bonus in place_entity()).
5673 * If it was on the rq, we've just 'preempted' it, which does convert
5674 * ->vruntime to a relative base.
5676 * Make sure both cases convert their relative position when migrating
5677 * to another cgroup's rq. This does somewhat interfere with the
5678 * fair sleeper stuff for the first placement, but who cares.
5681 * When !on_rq, vruntime of the task has usually NOT been normalized.
5682 * But there are some cases where it has already been normalized:
5684 * - Moving a forked child which is waiting for being woken up by
5685 * wake_up_new_task().
5686 * - Moving a task which has been woken up by try_to_wake_up() and
5687 * waiting for actually being woken up by sched_ttwu_pending().
5689 * To prevent boost or penalty in the new cfs_rq caused by delta
5690 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5692 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5696 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5697 set_task_rq(p, task_cpu(p));
5699 cfs_rq = cfs_rq_of(&p->se);
5700 p->se.vruntime += cfs_rq->min_vruntime;
5703 * migrate_task_rq_fair() will have removed our previous
5704 * contribution, but we must synchronize for ongoing future
5707 p->se.avg.decay_count = atomic64_read(&cfs_rq->decay_counter);
5708 cfs_rq->blocked_load_avg += p->se.avg.load_avg_contrib;
5713 void free_fair_sched_group(struct task_group *tg)
5717 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5719 for_each_possible_cpu(i) {
5721 kfree(tg->cfs_rq[i]);
5730 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5732 struct cfs_rq *cfs_rq;
5733 struct sched_entity *se;
5736 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5739 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5743 tg->shares = NICE_0_LOAD;
5745 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5747 for_each_possible_cpu(i) {
5748 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5749 GFP_KERNEL, cpu_to_node(i));
5753 se = kzalloc_node(sizeof(struct sched_entity),
5754 GFP_KERNEL, cpu_to_node(i));
5758 init_cfs_rq(cfs_rq);
5759 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5770 void unregister_fair_sched_group(struct task_group *tg, int cpu)
5772 struct rq *rq = cpu_rq(cpu);
5773 unsigned long flags;
5776 * Only empty task groups can be destroyed; so we can speculatively
5777 * check on_list without danger of it being re-added.
5779 if (!tg->cfs_rq[cpu]->on_list)
5782 raw_spin_lock_irqsave(&rq->lock, flags);
5783 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5784 raw_spin_unlock_irqrestore(&rq->lock, flags);
5787 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5788 struct sched_entity *se, int cpu,
5789 struct sched_entity *parent)
5791 struct rq *rq = cpu_rq(cpu);
5795 init_cfs_rq_runtime(cfs_rq);
5797 tg->cfs_rq[cpu] = cfs_rq;
5800 /* se could be NULL for root_task_group */
5805 se->cfs_rq = &rq->cfs;
5807 se->cfs_rq = parent->my_q;
5810 update_load_set(&se->load, 0);
5811 se->parent = parent;
5814 static DEFINE_MUTEX(shares_mutex);
5816 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5819 unsigned long flags;
5822 * We can't change the weight of the root cgroup.
5827 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5829 mutex_lock(&shares_mutex);
5830 if (tg->shares == shares)
5833 tg->shares = shares;
5834 for_each_possible_cpu(i) {
5835 struct rq *rq = cpu_rq(i);
5836 struct sched_entity *se;
5839 /* Propagate contribution to hierarchy */
5840 raw_spin_lock_irqsave(&rq->lock, flags);
5841 for_each_sched_entity(se)
5842 update_cfs_shares(group_cfs_rq(se));
5843 raw_spin_unlock_irqrestore(&rq->lock, flags);
5847 mutex_unlock(&shares_mutex);
5850 #else /* CONFIG_FAIR_GROUP_SCHED */
5852 void free_fair_sched_group(struct task_group *tg) { }
5854 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5859 void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5861 #endif /* CONFIG_FAIR_GROUP_SCHED */
5864 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5866 struct sched_entity *se = &task->se;
5867 unsigned int rr_interval = 0;
5870 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5873 if (rq->cfs.load.weight)
5874 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5880 * All the scheduling class methods:
5882 const struct sched_class fair_sched_class = {
5883 .next = &idle_sched_class,
5884 .enqueue_task = enqueue_task_fair,
5885 .dequeue_task = dequeue_task_fair,
5886 .yield_task = yield_task_fair,
5887 .yield_to_task = yield_to_task_fair,
5889 .check_preempt_curr = check_preempt_wakeup,
5891 .pick_next_task = pick_next_task_fair,
5892 .put_prev_task = put_prev_task_fair,
5895 .select_task_rq = select_task_rq_fair,
5896 #ifdef CONFIG_FAIR_GROUP_SCHED
5897 .migrate_task_rq = migrate_task_rq_fair,
5899 .rq_online = rq_online_fair,
5900 .rq_offline = rq_offline_fair,
5902 .task_waking = task_waking_fair,
5905 .set_curr_task = set_curr_task_fair,
5906 .task_tick = task_tick_fair,
5907 .task_fork = task_fork_fair,
5909 .prio_changed = prio_changed_fair,
5910 .switched_from = switched_from_fair,
5911 .switched_to = switched_to_fair,
5913 .get_rr_interval = get_rr_interval_fair,
5915 #ifdef CONFIG_FAIR_GROUP_SCHED
5916 .task_move_group = task_move_group_fair,
5920 #ifdef CONFIG_SCHED_DEBUG
5921 void print_cfs_stats(struct seq_file *m, int cpu)
5923 struct cfs_rq *cfs_rq;
5926 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5927 print_cfs_rq(m, cpu, cfs_rq);
5932 __init void init_sched_fair_class(void)
5935 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5938 nohz.next_balance = jiffies;
5939 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5940 cpu_notifier(sched_ilb_notifier, 0);