2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
28 * Targeted preemption latency for CPU-bound tasks:
29 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
31 * NOTE: this latency value is not the same as the concept of
32 * 'timeslice length' - timeslices in CFS are of variable length
33 * and have no persistent notion like in traditional, time-slice
34 * based scheduling concepts.
36 * (to see the precise effective timeslice length of your workload,
37 * run vmstat and monitor the context-switches (cs) field)
39 unsigned int sysctl_sched_latency = 6000000ULL;
40 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
43 * The initial- and re-scaling of tunables is configurable
44 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
47 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
48 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
49 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 enum sched_tunable_scaling sysctl_sched_tunable_scaling
52 = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
56 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
58 unsigned int sysctl_sched_min_granularity = 750000ULL;
59 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
62 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
64 static unsigned int sched_nr_latency = 8;
67 * After fork, child runs first. If set to 0 (default) then
68 * parent will (try to) run first.
70 unsigned int sysctl_sched_child_runs_first __read_mostly;
73 * SCHED_OTHER wake-up granularity.
74 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
81 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
83 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
86 * The exponential sliding window over which load is averaged for shares
90 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
92 #ifdef CONFIG_CFS_BANDWIDTH
94 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
95 * each time a cfs_rq requests quota.
97 * Note: in the case that the slice exceeds the runtime remaining (either due
98 * to consumption or the quota being specified to be smaller than the slice)
99 * we will always only issue the remaining available time.
101 * default: 5 msec, units: microseconds
103 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
106 static const struct sched_class fair_sched_class;
108 /**************************************************************
109 * CFS operations on generic schedulable entities:
112 #ifdef CONFIG_FAIR_GROUP_SCHED
114 /* cpu runqueue to which this cfs_rq is attached */
115 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
120 /* An entity is a task if it doesn't "own" a runqueue */
121 #define entity_is_task(se) (!se->my_q)
123 static inline struct task_struct *task_of(struct sched_entity *se)
125 #ifdef CONFIG_SCHED_DEBUG
126 WARN_ON_ONCE(!entity_is_task(se));
128 return container_of(se, struct task_struct, se);
131 /* Walk up scheduling entities hierarchy */
132 #define for_each_sched_entity(se) \
133 for (; se; se = se->parent)
135 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
140 /* runqueue on which this entity is (to be) queued */
141 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
146 /* runqueue "owned" by this group */
147 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
152 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
154 if (!cfs_rq->on_list) {
156 * Ensure we either appear before our parent (if already
157 * enqueued) or force our parent to appear after us when it is
158 * enqueued. The fact that we always enqueue bottom-up
159 * reduces this to two cases.
161 if (cfs_rq->tg->parent &&
162 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
163 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
164 &rq_of(cfs_rq)->leaf_cfs_rq_list);
166 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
167 &rq_of(cfs_rq)->leaf_cfs_rq_list);
174 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
176 if (cfs_rq->on_list) {
177 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
182 /* Iterate thr' all leaf cfs_rq's on a runqueue */
183 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
184 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
186 /* Do the two (enqueued) entities belong to the same group ? */
188 is_same_group(struct sched_entity *se, struct sched_entity *pse)
190 if (se->cfs_rq == pse->cfs_rq)
196 static inline struct sched_entity *parent_entity(struct sched_entity *se)
201 /* return depth at which a sched entity is present in the hierarchy */
202 static inline int depth_se(struct sched_entity *se)
206 for_each_sched_entity(se)
213 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
215 int se_depth, pse_depth;
218 * preemption test can be made between sibling entities who are in the
219 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
220 * both tasks until we find their ancestors who are siblings of common
224 /* First walk up until both entities are at same depth */
225 se_depth = depth_se(*se);
226 pse_depth = depth_se(*pse);
228 while (se_depth > pse_depth) {
230 *se = parent_entity(*se);
233 while (pse_depth > se_depth) {
235 *pse = parent_entity(*pse);
238 while (!is_same_group(*se, *pse)) {
239 *se = parent_entity(*se);
240 *pse = parent_entity(*pse);
244 #else /* !CONFIG_FAIR_GROUP_SCHED */
246 static inline struct task_struct *task_of(struct sched_entity *se)
248 return container_of(se, struct task_struct, se);
251 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 return container_of(cfs_rq, struct rq, cfs);
256 #define entity_is_task(se) 1
258 #define for_each_sched_entity(se) \
259 for (; se; se = NULL)
261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
263 return &task_rq(p)->cfs;
266 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268 struct task_struct *p = task_of(se);
269 struct rq *rq = task_rq(p);
274 /* runqueue "owned" by this group */
275 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
280 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
284 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
289 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
292 is_same_group(struct sched_entity *se, struct sched_entity *pse)
297 static inline struct sched_entity *parent_entity(struct sched_entity *se)
303 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
310 unsigned long delta_exec);
312 /**************************************************************
313 * Scheduling class tree data structure manipulation methods:
316 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
318 s64 delta = (s64)(vruntime - min_vruntime);
320 min_vruntime = vruntime;
325 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
327 s64 delta = (s64)(vruntime - min_vruntime);
329 min_vruntime = vruntime;
334 static inline int entity_before(struct sched_entity *a,
335 struct sched_entity *b)
337 return (s64)(a->vruntime - b->vruntime) < 0;
340 static void update_min_vruntime(struct cfs_rq *cfs_rq)
342 u64 vruntime = cfs_rq->min_vruntime;
345 vruntime = cfs_rq->curr->vruntime;
347 if (cfs_rq->rb_leftmost) {
348 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
353 vruntime = se->vruntime;
355 vruntime = min_vruntime(vruntime, se->vruntime);
358 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
361 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
366 * Enqueue an entity into the rb-tree:
368 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
370 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
371 struct rb_node *parent = NULL;
372 struct sched_entity *entry;
376 * Find the right place in the rbtree:
380 entry = rb_entry(parent, struct sched_entity, run_node);
382 * We dont care about collisions. Nodes with
383 * the same key stay together.
385 if (entity_before(se, entry)) {
386 link = &parent->rb_left;
388 link = &parent->rb_right;
394 * Maintain a cache of leftmost tree entries (it is frequently
398 cfs_rq->rb_leftmost = &se->run_node;
400 rb_link_node(&se->run_node, parent, link);
401 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
404 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
406 if (cfs_rq->rb_leftmost == &se->run_node) {
407 struct rb_node *next_node;
409 next_node = rb_next(&se->run_node);
410 cfs_rq->rb_leftmost = next_node;
413 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
416 static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
418 struct rb_node *left = cfs_rq->rb_leftmost;
423 return rb_entry(left, struct sched_entity, run_node);
426 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
428 struct rb_node *next = rb_next(&se->run_node);
433 return rb_entry(next, struct sched_entity, run_node);
436 #ifdef CONFIG_SCHED_DEBUG
437 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
439 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
444 return rb_entry(last, struct sched_entity, run_node);
447 /**************************************************************
448 * Scheduling class statistics methods:
451 int sched_proc_update_handler(struct ctl_table *table, int write,
452 void __user *buffer, size_t *lenp,
455 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
456 int factor = get_update_sysctl_factor();
461 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
462 sysctl_sched_min_granularity);
464 #define WRT_SYSCTL(name) \
465 (normalized_sysctl_##name = sysctl_##name / (factor))
466 WRT_SYSCTL(sched_min_granularity);
467 WRT_SYSCTL(sched_latency);
468 WRT_SYSCTL(sched_wakeup_granularity);
478 static inline unsigned long
479 calc_delta_fair(unsigned long delta, struct sched_entity *se)
481 if (unlikely(se->load.weight != NICE_0_LOAD))
482 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
488 * The idea is to set a period in which each task runs once.
490 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
491 * this period because otherwise the slices get too small.
493 * p = (nr <= nl) ? l : l*nr/nl
495 static u64 __sched_period(unsigned long nr_running)
497 u64 period = sysctl_sched_latency;
498 unsigned long nr_latency = sched_nr_latency;
500 if (unlikely(nr_running > nr_latency)) {
501 period = sysctl_sched_min_granularity;
502 period *= nr_running;
509 * We calculate the wall-time slice from the period by taking a part
510 * proportional to the weight.
514 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
516 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
518 for_each_sched_entity(se) {
519 struct load_weight *load;
520 struct load_weight lw;
522 cfs_rq = cfs_rq_of(se);
523 load = &cfs_rq->load;
525 if (unlikely(!se->on_rq)) {
528 update_load_add(&lw, se->load.weight);
531 slice = calc_delta_mine(slice, se->load.weight, load);
537 * We calculate the vruntime slice of a to be inserted task
541 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 return calc_delta_fair(sched_slice(cfs_rq, se), se);
546 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
547 static void update_cfs_shares(struct cfs_rq *cfs_rq);
550 * Update the current task's runtime statistics. Skip current tasks that
551 * are not in our scheduling class.
554 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
555 unsigned long delta_exec)
557 unsigned long delta_exec_weighted;
559 schedstat_set(curr->statistics.exec_max,
560 max((u64)delta_exec, curr->statistics.exec_max));
562 curr->sum_exec_runtime += delta_exec;
563 schedstat_add(cfs_rq, exec_clock, delta_exec);
564 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
566 curr->vruntime += delta_exec_weighted;
567 update_min_vruntime(cfs_rq);
569 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
570 cfs_rq->load_unacc_exec_time += delta_exec;
574 static void update_curr(struct cfs_rq *cfs_rq)
576 struct sched_entity *curr = cfs_rq->curr;
577 u64 now = rq_of(cfs_rq)->clock_task;
578 unsigned long delta_exec;
584 * Get the amount of time the current task was running
585 * since the last time we changed load (this cannot
586 * overflow on 32 bits):
588 delta_exec = (unsigned long)(now - curr->exec_start);
592 __update_curr(cfs_rq, curr, delta_exec);
593 curr->exec_start = now;
595 if (entity_is_task(curr)) {
596 struct task_struct *curtask = task_of(curr);
598 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
599 cpuacct_charge(curtask, delta_exec);
600 account_group_exec_runtime(curtask, delta_exec);
603 account_cfs_rq_runtime(cfs_rq, delta_exec);
607 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
609 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
613 * Task is being enqueued - update stats:
615 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
618 * Are we enqueueing a waiting task? (for current tasks
619 * a dequeue/enqueue event is a NOP)
621 if (se != cfs_rq->curr)
622 update_stats_wait_start(cfs_rq, se);
626 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
629 rq_of(cfs_rq)->clock - se->statistics.wait_start));
630 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
631 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
632 rq_of(cfs_rq)->clock - se->statistics.wait_start);
633 #ifdef CONFIG_SCHEDSTATS
634 if (entity_is_task(se)) {
635 trace_sched_stat_wait(task_of(se),
636 rq_of(cfs_rq)->clock - se->statistics.wait_start);
639 schedstat_set(se->statistics.wait_start, 0);
643 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
646 * Mark the end of the wait period if dequeueing a
649 if (se != cfs_rq->curr)
650 update_stats_wait_end(cfs_rq, se);
654 * We are picking a new current task - update its stats:
657 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
660 * We are starting a new run period:
662 se->exec_start = rq_of(cfs_rq)->clock_task;
665 /**************************************************
666 * Scheduling class queueing methods:
669 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
671 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
673 cfs_rq->task_weight += weight;
677 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
683 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
685 update_load_add(&cfs_rq->load, se->load.weight);
686 if (!parent_entity(se))
687 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
688 if (entity_is_task(se)) {
689 add_cfs_task_weight(cfs_rq, se->load.weight);
690 list_add(&se->group_node, &cfs_rq->tasks);
692 cfs_rq->nr_running++;
696 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
698 update_load_sub(&cfs_rq->load, se->load.weight);
699 if (!parent_entity(se))
700 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
701 if (entity_is_task(se)) {
702 add_cfs_task_weight(cfs_rq, -se->load.weight);
703 list_del_init(&se->group_node);
705 cfs_rq->nr_running--;
708 #ifdef CONFIG_FAIR_GROUP_SCHED
709 /* we need this in update_cfs_load and load-balance functions below */
710 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
712 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
715 struct task_group *tg = cfs_rq->tg;
718 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
719 load_avg -= cfs_rq->load_contribution;
721 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
722 atomic_add(load_avg, &tg->load_weight);
723 cfs_rq->load_contribution += load_avg;
727 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
729 u64 period = sysctl_sched_shares_window;
731 unsigned long load = cfs_rq->load.weight;
733 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
736 now = rq_of(cfs_rq)->clock_task;
737 delta = now - cfs_rq->load_stamp;
739 /* truncate load history at 4 idle periods */
740 if (cfs_rq->load_stamp > cfs_rq->load_last &&
741 now - cfs_rq->load_last > 4 * period) {
742 cfs_rq->load_period = 0;
743 cfs_rq->load_avg = 0;
747 cfs_rq->load_stamp = now;
748 cfs_rq->load_unacc_exec_time = 0;
749 cfs_rq->load_period += delta;
751 cfs_rq->load_last = now;
752 cfs_rq->load_avg += delta * load;
755 /* consider updating load contribution on each fold or truncate */
756 if (global_update || cfs_rq->load_period > period
757 || !cfs_rq->load_period)
758 update_cfs_rq_load_contribution(cfs_rq, global_update);
760 while (cfs_rq->load_period > period) {
762 * Inline assembly required to prevent the compiler
763 * optimising this loop into a divmod call.
764 * See __iter_div_u64_rem() for another example of this.
766 asm("" : "+rm" (cfs_rq->load_period));
767 cfs_rq->load_period /= 2;
768 cfs_rq->load_avg /= 2;
771 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
772 list_del_leaf_cfs_rq(cfs_rq);
775 static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
780 * Use this CPU's actual weight instead of the last load_contribution
781 * to gain a more accurate current total weight. See
782 * update_cfs_rq_load_contribution().
784 tg_weight = atomic_read(&tg->load_weight);
785 tg_weight -= cfs_rq->load_contribution;
786 tg_weight += cfs_rq->load.weight;
791 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
793 long tg_weight, load, shares;
795 tg_weight = calc_tg_weight(tg, cfs_rq);
796 load = cfs_rq->load.weight;
798 shares = (tg->shares * load);
802 if (shares < MIN_SHARES)
804 if (shares > tg->shares)
810 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
812 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
813 update_cfs_load(cfs_rq, 0);
814 update_cfs_shares(cfs_rq);
817 # else /* CONFIG_SMP */
818 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
822 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
827 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
830 # endif /* CONFIG_SMP */
831 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
832 unsigned long weight)
835 /* commit outstanding execution time */
836 if (cfs_rq->curr == se)
838 account_entity_dequeue(cfs_rq, se);
841 update_load_set(&se->load, weight);
844 account_entity_enqueue(cfs_rq, se);
847 static void update_cfs_shares(struct cfs_rq *cfs_rq)
849 struct task_group *tg;
850 struct sched_entity *se;
854 se = tg->se[cpu_of(rq_of(cfs_rq))];
855 if (!se || throttled_hierarchy(cfs_rq))
858 if (likely(se->load.weight == tg->shares))
861 shares = calc_cfs_shares(cfs_rq, tg);
863 reweight_entity(cfs_rq_of(se), se, shares);
865 #else /* CONFIG_FAIR_GROUP_SCHED */
866 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
870 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
874 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
877 #endif /* CONFIG_FAIR_GROUP_SCHED */
879 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
881 #ifdef CONFIG_SCHEDSTATS
882 struct task_struct *tsk = NULL;
884 if (entity_is_task(se))
887 if (se->statistics.sleep_start) {
888 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
893 if (unlikely(delta > se->statistics.sleep_max))
894 se->statistics.sleep_max = delta;
896 se->statistics.sleep_start = 0;
897 se->statistics.sum_sleep_runtime += delta;
900 account_scheduler_latency(tsk, delta >> 10, 1);
901 trace_sched_stat_sleep(tsk, delta);
904 if (se->statistics.block_start) {
905 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
910 if (unlikely(delta > se->statistics.block_max))
911 se->statistics.block_max = delta;
913 se->statistics.block_start = 0;
914 se->statistics.sum_sleep_runtime += delta;
917 if (tsk->in_iowait) {
918 se->statistics.iowait_sum += delta;
919 se->statistics.iowait_count++;
920 trace_sched_stat_iowait(tsk, delta);
924 * Blocking time is in units of nanosecs, so shift by
925 * 20 to get a milliseconds-range estimation of the
926 * amount of time that the task spent sleeping:
928 if (unlikely(prof_on == SLEEP_PROFILING)) {
929 profile_hits(SLEEP_PROFILING,
930 (void *)get_wchan(tsk),
933 account_scheduler_latency(tsk, delta >> 10, 0);
939 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
941 #ifdef CONFIG_SCHED_DEBUG
942 s64 d = se->vruntime - cfs_rq->min_vruntime;
947 if (d > 3*sysctl_sched_latency)
948 schedstat_inc(cfs_rq, nr_spread_over);
953 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
955 u64 vruntime = cfs_rq->min_vruntime;
958 * The 'current' period is already promised to the current tasks,
959 * however the extra weight of the new task will slow them down a
960 * little, place the new task so that it fits in the slot that
961 * stays open at the end.
963 if (initial && sched_feat(START_DEBIT))
964 vruntime += sched_vslice(cfs_rq, se);
966 /* sleeps up to a single latency don't count. */
968 unsigned long thresh = sysctl_sched_latency;
971 * Halve their sleep time's effect, to allow
972 * for a gentler effect of sleepers:
974 if (sched_feat(GENTLE_FAIR_SLEEPERS))
980 /* ensure we never gain time by being placed backwards. */
981 vruntime = max_vruntime(se->vruntime, vruntime);
983 se->vruntime = vruntime;
986 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
989 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
992 * Update the normalized vruntime before updating min_vruntime
993 * through callig update_curr().
995 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
996 se->vruntime += cfs_rq->min_vruntime;
999 * Update run-time statistics of the 'current'.
1001 update_curr(cfs_rq);
1002 update_cfs_load(cfs_rq, 0);
1003 account_entity_enqueue(cfs_rq, se);
1004 update_cfs_shares(cfs_rq);
1006 if (flags & ENQUEUE_WAKEUP) {
1007 place_entity(cfs_rq, se, 0);
1008 enqueue_sleeper(cfs_rq, se);
1011 update_stats_enqueue(cfs_rq, se);
1012 check_spread(cfs_rq, se);
1013 if (se != cfs_rq->curr)
1014 __enqueue_entity(cfs_rq, se);
1017 if (cfs_rq->nr_running == 1) {
1018 list_add_leaf_cfs_rq(cfs_rq);
1019 check_enqueue_throttle(cfs_rq);
1023 static void __clear_buddies_last(struct sched_entity *se)
1025 for_each_sched_entity(se) {
1026 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1027 if (cfs_rq->last == se)
1028 cfs_rq->last = NULL;
1034 static void __clear_buddies_next(struct sched_entity *se)
1036 for_each_sched_entity(se) {
1037 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1038 if (cfs_rq->next == se)
1039 cfs_rq->next = NULL;
1045 static void __clear_buddies_skip(struct sched_entity *se)
1047 for_each_sched_entity(se) {
1048 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1049 if (cfs_rq->skip == se)
1050 cfs_rq->skip = NULL;
1056 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1058 if (cfs_rq->last == se)
1059 __clear_buddies_last(se);
1061 if (cfs_rq->next == se)
1062 __clear_buddies_next(se);
1064 if (cfs_rq->skip == se)
1065 __clear_buddies_skip(se);
1068 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1071 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1074 * Update run-time statistics of the 'current'.
1076 update_curr(cfs_rq);
1078 update_stats_dequeue(cfs_rq, se);
1079 if (flags & DEQUEUE_SLEEP) {
1080 #ifdef CONFIG_SCHEDSTATS
1081 if (entity_is_task(se)) {
1082 struct task_struct *tsk = task_of(se);
1084 if (tsk->state & TASK_INTERRUPTIBLE)
1085 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1086 if (tsk->state & TASK_UNINTERRUPTIBLE)
1087 se->statistics.block_start = rq_of(cfs_rq)->clock;
1092 clear_buddies(cfs_rq, se);
1094 if (se != cfs_rq->curr)
1095 __dequeue_entity(cfs_rq, se);
1097 update_cfs_load(cfs_rq, 0);
1098 account_entity_dequeue(cfs_rq, se);
1101 * Normalize the entity after updating the min_vruntime because the
1102 * update can refer to the ->curr item and we need to reflect this
1103 * movement in our normalized position.
1105 if (!(flags & DEQUEUE_SLEEP))
1106 se->vruntime -= cfs_rq->min_vruntime;
1108 /* return excess runtime on last dequeue */
1109 return_cfs_rq_runtime(cfs_rq);
1111 update_min_vruntime(cfs_rq);
1112 update_cfs_shares(cfs_rq);
1116 * Preempt the current task with a newly woken task if needed:
1119 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1121 unsigned long ideal_runtime, delta_exec;
1122 struct sched_entity *se;
1125 ideal_runtime = sched_slice(cfs_rq, curr);
1126 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1127 if (delta_exec > ideal_runtime) {
1128 resched_task(rq_of(cfs_rq)->curr);
1130 * The current task ran long enough, ensure it doesn't get
1131 * re-elected due to buddy favours.
1133 clear_buddies(cfs_rq, curr);
1138 * Ensure that a task that missed wakeup preemption by a
1139 * narrow margin doesn't have to wait for a full slice.
1140 * This also mitigates buddy induced latencies under load.
1142 if (delta_exec < sysctl_sched_min_granularity)
1145 se = __pick_first_entity(cfs_rq);
1146 delta = curr->vruntime - se->vruntime;
1151 if (delta > ideal_runtime)
1152 resched_task(rq_of(cfs_rq)->curr);
1156 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1158 /* 'current' is not kept within the tree. */
1161 * Any task has to be enqueued before it get to execute on
1162 * a CPU. So account for the time it spent waiting on the
1165 update_stats_wait_end(cfs_rq, se);
1166 __dequeue_entity(cfs_rq, se);
1169 update_stats_curr_start(cfs_rq, se);
1171 #ifdef CONFIG_SCHEDSTATS
1173 * Track our maximum slice length, if the CPU's load is at
1174 * least twice that of our own weight (i.e. dont track it
1175 * when there are only lesser-weight tasks around):
1177 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1178 se->statistics.slice_max = max(se->statistics.slice_max,
1179 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1182 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1186 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1189 * Pick the next process, keeping these things in mind, in this order:
1190 * 1) keep things fair between processes/task groups
1191 * 2) pick the "next" process, since someone really wants that to run
1192 * 3) pick the "last" process, for cache locality
1193 * 4) do not run the "skip" process, if something else is available
1195 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1197 struct sched_entity *se = __pick_first_entity(cfs_rq);
1198 struct sched_entity *left = se;
1201 * Avoid running the skip buddy, if running something else can
1202 * be done without getting too unfair.
1204 if (cfs_rq->skip == se) {
1205 struct sched_entity *second = __pick_next_entity(se);
1206 if (second && wakeup_preempt_entity(second, left) < 1)
1211 * Prefer last buddy, try to return the CPU to a preempted task.
1213 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1217 * Someone really wants this to run. If it's not unfair, run it.
1219 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1222 clear_buddies(cfs_rq, se);
1227 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1229 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1232 * If still on the runqueue then deactivate_task()
1233 * was not called and update_curr() has to be done:
1236 update_curr(cfs_rq);
1238 /* throttle cfs_rqs exceeding runtime */
1239 check_cfs_rq_runtime(cfs_rq);
1241 check_spread(cfs_rq, prev);
1243 update_stats_wait_start(cfs_rq, prev);
1244 /* Put 'current' back into the tree. */
1245 __enqueue_entity(cfs_rq, prev);
1247 cfs_rq->curr = NULL;
1251 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1254 * Update run-time statistics of the 'current'.
1256 update_curr(cfs_rq);
1259 * Update share accounting for long-running entities.
1261 update_entity_shares_tick(cfs_rq);
1263 #ifdef CONFIG_SCHED_HRTICK
1265 * queued ticks are scheduled to match the slice, so don't bother
1266 * validating it and just reschedule.
1269 resched_task(rq_of(cfs_rq)->curr);
1273 * don't let the period tick interfere with the hrtick preemption
1275 if (!sched_feat(DOUBLE_TICK) &&
1276 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1280 if (cfs_rq->nr_running > 1)
1281 check_preempt_tick(cfs_rq, curr);
1285 /**************************************************
1286 * CFS bandwidth control machinery
1289 #ifdef CONFIG_CFS_BANDWIDTH
1291 * default period for cfs group bandwidth.
1292 * default: 0.1s, units: nanoseconds
1294 static inline u64 default_cfs_period(void)
1296 return 100000000ULL;
1299 static inline u64 sched_cfs_bandwidth_slice(void)
1301 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1305 * Replenish runtime according to assigned quota and update expiration time.
1306 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1307 * additional synchronization around rq->lock.
1309 * requires cfs_b->lock
1311 static void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1315 if (cfs_b->quota == RUNTIME_INF)
1318 now = sched_clock_cpu(smp_processor_id());
1319 cfs_b->runtime = cfs_b->quota;
1320 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1323 /* returns 0 on failure to allocate runtime */
1324 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1326 struct task_group *tg = cfs_rq->tg;
1327 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1328 u64 amount = 0, min_amount, expires;
1330 /* note: this is a positive sum as runtime_remaining <= 0 */
1331 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1333 raw_spin_lock(&cfs_b->lock);
1334 if (cfs_b->quota == RUNTIME_INF)
1335 amount = min_amount;
1338 * If the bandwidth pool has become inactive, then at least one
1339 * period must have elapsed since the last consumption.
1340 * Refresh the global state and ensure bandwidth timer becomes
1343 if (!cfs_b->timer_active) {
1344 __refill_cfs_bandwidth_runtime(cfs_b);
1345 __start_cfs_bandwidth(cfs_b);
1348 if (cfs_b->runtime > 0) {
1349 amount = min(cfs_b->runtime, min_amount);
1350 cfs_b->runtime -= amount;
1354 expires = cfs_b->runtime_expires;
1355 raw_spin_unlock(&cfs_b->lock);
1357 cfs_rq->runtime_remaining += amount;
1359 * we may have advanced our local expiration to account for allowed
1360 * spread between our sched_clock and the one on which runtime was
1363 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1364 cfs_rq->runtime_expires = expires;
1366 return cfs_rq->runtime_remaining > 0;
1370 * Note: This depends on the synchronization provided by sched_clock and the
1371 * fact that rq->clock snapshots this value.
1373 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1375 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1376 struct rq *rq = rq_of(cfs_rq);
1378 /* if the deadline is ahead of our clock, nothing to do */
1379 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1382 if (cfs_rq->runtime_remaining < 0)
1386 * If the local deadline has passed we have to consider the
1387 * possibility that our sched_clock is 'fast' and the global deadline
1388 * has not truly expired.
1390 * Fortunately we can check determine whether this the case by checking
1391 * whether the global deadline has advanced.
1394 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1395 /* extend local deadline, drift is bounded above by 2 ticks */
1396 cfs_rq->runtime_expires += TICK_NSEC;
1398 /* global deadline is ahead, expiration has passed */
1399 cfs_rq->runtime_remaining = 0;
1403 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1404 unsigned long delta_exec)
1406 /* dock delta_exec before expiring quota (as it could span periods) */
1407 cfs_rq->runtime_remaining -= delta_exec;
1408 expire_cfs_rq_runtime(cfs_rq);
1410 if (likely(cfs_rq->runtime_remaining > 0))
1414 * if we're unable to extend our runtime we resched so that the active
1415 * hierarchy can be throttled
1417 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1418 resched_task(rq_of(cfs_rq)->curr);
1421 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1422 unsigned long delta_exec)
1424 if (!cfs_rq->runtime_enabled)
1427 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1430 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1432 return cfs_rq->throttled;
1435 /* check whether cfs_rq, or any parent, is throttled */
1436 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1438 return cfs_rq->throttle_count;
1442 * Ensure that neither of the group entities corresponding to src_cpu or
1443 * dest_cpu are members of a throttled hierarchy when performing group
1444 * load-balance operations.
1446 static inline int throttled_lb_pair(struct task_group *tg,
1447 int src_cpu, int dest_cpu)
1449 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1451 src_cfs_rq = tg->cfs_rq[src_cpu];
1452 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1454 return throttled_hierarchy(src_cfs_rq) ||
1455 throttled_hierarchy(dest_cfs_rq);
1458 /* updated child weight may affect parent so we have to do this bottom up */
1459 static int tg_unthrottle_up(struct task_group *tg, void *data)
1461 struct rq *rq = data;
1462 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1464 cfs_rq->throttle_count--;
1466 if (!cfs_rq->throttle_count) {
1467 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1469 /* leaving throttled state, advance shares averaging windows */
1470 cfs_rq->load_stamp += delta;
1471 cfs_rq->load_last += delta;
1473 /* update entity weight now that we are on_rq again */
1474 update_cfs_shares(cfs_rq);
1481 static int tg_throttle_down(struct task_group *tg, void *data)
1483 struct rq *rq = data;
1484 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1486 /* group is entering throttled state, record last load */
1487 if (!cfs_rq->throttle_count)
1488 update_cfs_load(cfs_rq, 0);
1489 cfs_rq->throttle_count++;
1494 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1496 struct rq *rq = rq_of(cfs_rq);
1497 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1498 struct sched_entity *se;
1499 long task_delta, dequeue = 1;
1501 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1503 /* account load preceding throttle */
1505 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1508 task_delta = cfs_rq->h_nr_running;
1509 for_each_sched_entity(se) {
1510 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1511 /* throttled entity or throttle-on-deactivate */
1516 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1517 qcfs_rq->h_nr_running -= task_delta;
1519 if (qcfs_rq->load.weight)
1524 rq->nr_running -= task_delta;
1526 cfs_rq->throttled = 1;
1527 cfs_rq->throttled_timestamp = rq->clock;
1528 raw_spin_lock(&cfs_b->lock);
1529 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1530 if (!cfs_b->timer_active)
1531 __start_cfs_bandwidth(cfs_b);
1532 raw_spin_unlock(&cfs_b->lock);
1535 static void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1537 struct rq *rq = rq_of(cfs_rq);
1538 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1539 struct sched_entity *se;
1543 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1545 cfs_rq->throttled = 0;
1546 raw_spin_lock(&cfs_b->lock);
1547 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1548 list_del_rcu(&cfs_rq->throttled_list);
1549 raw_spin_unlock(&cfs_b->lock);
1550 cfs_rq->throttled_timestamp = 0;
1552 update_rq_clock(rq);
1553 /* update hierarchical throttle state */
1554 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1556 if (!cfs_rq->load.weight)
1559 task_delta = cfs_rq->h_nr_running;
1560 for_each_sched_entity(se) {
1564 cfs_rq = cfs_rq_of(se);
1566 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1567 cfs_rq->h_nr_running += task_delta;
1569 if (cfs_rq_throttled(cfs_rq))
1574 rq->nr_running += task_delta;
1576 /* determine whether we need to wake up potentially idle cpu */
1577 if (rq->curr == rq->idle && rq->cfs.nr_running)
1578 resched_task(rq->curr);
1581 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1582 u64 remaining, u64 expires)
1584 struct cfs_rq *cfs_rq;
1585 u64 runtime = remaining;
1588 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1590 struct rq *rq = rq_of(cfs_rq);
1592 raw_spin_lock(&rq->lock);
1593 if (!cfs_rq_throttled(cfs_rq))
1596 runtime = -cfs_rq->runtime_remaining + 1;
1597 if (runtime > remaining)
1598 runtime = remaining;
1599 remaining -= runtime;
1601 cfs_rq->runtime_remaining += runtime;
1602 cfs_rq->runtime_expires = expires;
1604 /* we check whether we're throttled above */
1605 if (cfs_rq->runtime_remaining > 0)
1606 unthrottle_cfs_rq(cfs_rq);
1609 raw_spin_unlock(&rq->lock);
1620 * Responsible for refilling a task_group's bandwidth and unthrottling its
1621 * cfs_rqs as appropriate. If there has been no activity within the last
1622 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1623 * used to track this state.
1625 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1627 u64 runtime, runtime_expires;
1628 int idle = 1, throttled;
1630 raw_spin_lock(&cfs_b->lock);
1631 /* no need to continue the timer with no bandwidth constraint */
1632 if (cfs_b->quota == RUNTIME_INF)
1635 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1636 /* idle depends on !throttled (for the case of a large deficit) */
1637 idle = cfs_b->idle && !throttled;
1638 cfs_b->nr_periods += overrun;
1640 /* if we're going inactive then everything else can be deferred */
1644 __refill_cfs_bandwidth_runtime(cfs_b);
1647 /* mark as potentially idle for the upcoming period */
1652 /* account preceding periods in which throttling occurred */
1653 cfs_b->nr_throttled += overrun;
1656 * There are throttled entities so we must first use the new bandwidth
1657 * to unthrottle them before making it generally available. This
1658 * ensures that all existing debts will be paid before a new cfs_rq is
1661 runtime = cfs_b->runtime;
1662 runtime_expires = cfs_b->runtime_expires;
1666 * This check is repeated as we are holding onto the new bandwidth
1667 * while we unthrottle. This can potentially race with an unthrottled
1668 * group trying to acquire new bandwidth from the global pool.
1670 while (throttled && runtime > 0) {
1671 raw_spin_unlock(&cfs_b->lock);
1672 /* we can't nest cfs_b->lock while distributing bandwidth */
1673 runtime = distribute_cfs_runtime(cfs_b, runtime,
1675 raw_spin_lock(&cfs_b->lock);
1677 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1680 /* return (any) remaining runtime */
1681 cfs_b->runtime = runtime;
1683 * While we are ensured activity in the period following an
1684 * unthrottle, this also covers the case in which the new bandwidth is
1685 * insufficient to cover the existing bandwidth deficit. (Forcing the
1686 * timer to remain active while there are any throttled entities.)
1691 cfs_b->timer_active = 0;
1692 raw_spin_unlock(&cfs_b->lock);
1697 /* a cfs_rq won't donate quota below this amount */
1698 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1699 /* minimum remaining period time to redistribute slack quota */
1700 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1701 /* how long we wait to gather additional slack before distributing */
1702 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1704 /* are we near the end of the current quota period? */
1705 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
1707 struct hrtimer *refresh_timer = &cfs_b->period_timer;
1710 /* if the call-back is running a quota refresh is already occurring */
1711 if (hrtimer_callback_running(refresh_timer))
1714 /* is a quota refresh about to occur? */
1715 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
1716 if (remaining < min_expire)
1722 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
1724 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
1726 /* if there's a quota refresh soon don't bother with slack */
1727 if (runtime_refresh_within(cfs_b, min_left))
1730 start_bandwidth_timer(&cfs_b->slack_timer,
1731 ns_to_ktime(cfs_bandwidth_slack_period));
1734 /* we know any runtime found here is valid as update_curr() precedes return */
1735 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1737 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1738 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
1740 if (slack_runtime <= 0)
1743 raw_spin_lock(&cfs_b->lock);
1744 if (cfs_b->quota != RUNTIME_INF &&
1745 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
1746 cfs_b->runtime += slack_runtime;
1748 /* we are under rq->lock, defer unthrottling using a timer */
1749 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
1750 !list_empty(&cfs_b->throttled_cfs_rq))
1751 start_cfs_slack_bandwidth(cfs_b);
1753 raw_spin_unlock(&cfs_b->lock);
1755 /* even if it's not valid for return we don't want to try again */
1756 cfs_rq->runtime_remaining -= slack_runtime;
1759 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1761 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
1764 __return_cfs_rq_runtime(cfs_rq);
1768 * This is done with a timer (instead of inline with bandwidth return) since
1769 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
1771 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
1773 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
1776 /* confirm we're still not at a refresh boundary */
1777 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
1780 raw_spin_lock(&cfs_b->lock);
1781 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
1782 runtime = cfs_b->runtime;
1785 expires = cfs_b->runtime_expires;
1786 raw_spin_unlock(&cfs_b->lock);
1791 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
1793 raw_spin_lock(&cfs_b->lock);
1794 if (expires == cfs_b->runtime_expires)
1795 cfs_b->runtime = runtime;
1796 raw_spin_unlock(&cfs_b->lock);
1800 * When a group wakes up we want to make sure that its quota is not already
1801 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1802 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1804 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
1806 /* an active group must be handled by the update_curr()->put() path */
1807 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
1810 /* ensure the group is not already throttled */
1811 if (cfs_rq_throttled(cfs_rq))
1814 /* update runtime allocation */
1815 account_cfs_rq_runtime(cfs_rq, 0);
1816 if (cfs_rq->runtime_remaining <= 0)
1817 throttle_cfs_rq(cfs_rq);
1820 /* conditionally throttle active cfs_rq's from put_prev_entity() */
1821 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1823 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
1827 * it's possible for a throttled entity to be forced into a running
1828 * state (e.g. set_curr_task), in this case we're finished.
1830 if (cfs_rq_throttled(cfs_rq))
1833 throttle_cfs_rq(cfs_rq);
1836 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1837 unsigned long delta_exec) {}
1838 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
1839 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
1840 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
1842 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1847 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1852 static inline int throttled_lb_pair(struct task_group *tg,
1853 int src_cpu, int dest_cpu)
1859 /**************************************************
1860 * CFS operations on tasks:
1863 #ifdef CONFIG_SCHED_HRTICK
1864 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1866 struct sched_entity *se = &p->se;
1867 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1869 WARN_ON(task_rq(p) != rq);
1871 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1872 u64 slice = sched_slice(cfs_rq, se);
1873 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1874 s64 delta = slice - ran;
1883 * Don't schedule slices shorter than 10000ns, that just
1884 * doesn't make sense. Rely on vruntime for fairness.
1887 delta = max_t(s64, 10000LL, delta);
1889 hrtick_start(rq, delta);
1894 * called from enqueue/dequeue and updates the hrtick when the
1895 * current task is from our class and nr_running is low enough
1898 static void hrtick_update(struct rq *rq)
1900 struct task_struct *curr = rq->curr;
1902 if (curr->sched_class != &fair_sched_class)
1905 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1906 hrtick_start_fair(rq, curr);
1908 #else /* !CONFIG_SCHED_HRTICK */
1910 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1914 static inline void hrtick_update(struct rq *rq)
1920 * The enqueue_task method is called before nr_running is
1921 * increased. Here we update the fair scheduling stats and
1922 * then put the task into the rbtree:
1925 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1927 struct cfs_rq *cfs_rq;
1928 struct sched_entity *se = &p->se;
1930 for_each_sched_entity(se) {
1933 cfs_rq = cfs_rq_of(se);
1934 enqueue_entity(cfs_rq, se, flags);
1937 * end evaluation on encountering a throttled cfs_rq
1939 * note: in the case of encountering a throttled cfs_rq we will
1940 * post the final h_nr_running increment below.
1942 if (cfs_rq_throttled(cfs_rq))
1944 cfs_rq->h_nr_running++;
1946 flags = ENQUEUE_WAKEUP;
1949 for_each_sched_entity(se) {
1950 cfs_rq = cfs_rq_of(se);
1951 cfs_rq->h_nr_running++;
1953 if (cfs_rq_throttled(cfs_rq))
1956 update_cfs_load(cfs_rq, 0);
1957 update_cfs_shares(cfs_rq);
1965 static void set_next_buddy(struct sched_entity *se);
1968 * The dequeue_task method is called before nr_running is
1969 * decreased. We remove the task from the rbtree and
1970 * update the fair scheduling stats:
1972 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1974 struct cfs_rq *cfs_rq;
1975 struct sched_entity *se = &p->se;
1976 int task_sleep = flags & DEQUEUE_SLEEP;
1978 for_each_sched_entity(se) {
1979 cfs_rq = cfs_rq_of(se);
1980 dequeue_entity(cfs_rq, se, flags);
1983 * end evaluation on encountering a throttled cfs_rq
1985 * note: in the case of encountering a throttled cfs_rq we will
1986 * post the final h_nr_running decrement below.
1988 if (cfs_rq_throttled(cfs_rq))
1990 cfs_rq->h_nr_running--;
1992 /* Don't dequeue parent if it has other entities besides us */
1993 if (cfs_rq->load.weight) {
1995 * Bias pick_next to pick a task from this cfs_rq, as
1996 * p is sleeping when it is within its sched_slice.
1998 if (task_sleep && parent_entity(se))
1999 set_next_buddy(parent_entity(se));
2001 /* avoid re-evaluating load for this entity */
2002 se = parent_entity(se);
2005 flags |= DEQUEUE_SLEEP;
2008 for_each_sched_entity(se) {
2009 cfs_rq = cfs_rq_of(se);
2010 cfs_rq->h_nr_running--;
2012 if (cfs_rq_throttled(cfs_rq))
2015 update_cfs_load(cfs_rq, 0);
2016 update_cfs_shares(cfs_rq);
2026 static void task_waking_fair(struct task_struct *p)
2028 struct sched_entity *se = &p->se;
2029 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2032 #ifndef CONFIG_64BIT
2033 u64 min_vruntime_copy;
2036 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2038 min_vruntime = cfs_rq->min_vruntime;
2039 } while (min_vruntime != min_vruntime_copy);
2041 min_vruntime = cfs_rq->min_vruntime;
2044 se->vruntime -= min_vruntime;
2047 #ifdef CONFIG_FAIR_GROUP_SCHED
2049 * effective_load() calculates the load change as seen from the root_task_group
2051 * Adding load to a group doesn't make a group heavier, but can cause movement
2052 * of group shares between cpus. Assuming the shares were perfectly aligned one
2053 * can calculate the shift in shares.
2055 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2056 * on this @cpu and results in a total addition (subtraction) of @wg to the
2057 * total group weight.
2059 * Given a runqueue weight distribution (rw_i) we can compute a shares
2060 * distribution (s_i) using:
2062 * s_i = rw_i / \Sum rw_j (1)
2064 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2065 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2066 * shares distribution (s_i):
2068 * rw_i = { 2, 4, 1, 0 }
2069 * s_i = { 2/7, 4/7, 1/7, 0 }
2071 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2072 * task used to run on and the CPU the waker is running on), we need to
2073 * compute the effect of waking a task on either CPU and, in case of a sync
2074 * wakeup, compute the effect of the current task going to sleep.
2076 * So for a change of @wl to the local @cpu with an overall group weight change
2077 * of @wl we can compute the new shares distribution (s'_i) using:
2079 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2081 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2082 * differences in waking a task to CPU 0. The additional task changes the
2083 * weight and shares distributions like:
2085 * rw'_i = { 3, 4, 1, 0 }
2086 * s'_i = { 3/8, 4/8, 1/8, 0 }
2088 * We can then compute the difference in effective weight by using:
2090 * dw_i = S * (s'_i - s_i) (3)
2092 * Where 'S' is the group weight as seen by its parent.
2094 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2095 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2096 * 4/7) times the weight of the group.
2098 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2100 struct sched_entity *se = tg->se[cpu];
2102 if (!tg->parent) /* the trivial, non-cgroup case */
2105 for_each_sched_entity(se) {
2111 * W = @wg + \Sum rw_j
2113 W = wg + calc_tg_weight(tg, se->my_q);
2118 w = se->my_q->load.weight + wl;
2121 * wl = S * s'_i; see (2)
2124 wl = (w * tg->shares) / W;
2129 * Per the above, wl is the new se->load.weight value; since
2130 * those are clipped to [MIN_SHARES, ...) do so now. See
2131 * calc_cfs_shares().
2133 if (wl < MIN_SHARES)
2137 * wl = dw_i = S * (s'_i - s_i); see (3)
2139 wl -= se->load.weight;
2142 * Recursively apply this logic to all parent groups to compute
2143 * the final effective load change on the root group. Since
2144 * only the @tg group gets extra weight, all parent groups can
2145 * only redistribute existing shares. @wl is the shift in shares
2146 * resulting from this level per the above.
2155 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2156 unsigned long wl, unsigned long wg)
2163 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2165 s64 this_load, load;
2166 int idx, this_cpu, prev_cpu;
2167 unsigned long tl_per_task;
2168 struct task_group *tg;
2169 unsigned long weight;
2173 this_cpu = smp_processor_id();
2174 prev_cpu = task_cpu(p);
2175 load = source_load(prev_cpu, idx);
2176 this_load = target_load(this_cpu, idx);
2179 * If sync wakeup then subtract the (maximum possible)
2180 * effect of the currently running task from the load
2181 * of the current CPU:
2184 tg = task_group(current);
2185 weight = current->se.load.weight;
2187 this_load += effective_load(tg, this_cpu, -weight, -weight);
2188 load += effective_load(tg, prev_cpu, 0, -weight);
2192 weight = p->se.load.weight;
2195 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2196 * due to the sync cause above having dropped this_load to 0, we'll
2197 * always have an imbalance, but there's really nothing you can do
2198 * about that, so that's good too.
2200 * Otherwise check if either cpus are near enough in load to allow this
2201 * task to be woken on this_cpu.
2203 if (this_load > 0) {
2204 s64 this_eff_load, prev_eff_load;
2206 this_eff_load = 100;
2207 this_eff_load *= power_of(prev_cpu);
2208 this_eff_load *= this_load +
2209 effective_load(tg, this_cpu, weight, weight);
2211 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2212 prev_eff_load *= power_of(this_cpu);
2213 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2215 balanced = this_eff_load <= prev_eff_load;
2220 * If the currently running task will sleep within
2221 * a reasonable amount of time then attract this newly
2224 if (sync && balanced)
2227 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2228 tl_per_task = cpu_avg_load_per_task(this_cpu);
2231 (this_load <= load &&
2232 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2234 * This domain has SD_WAKE_AFFINE and
2235 * p is cache cold in this domain, and
2236 * there is no bad imbalance.
2238 schedstat_inc(sd, ttwu_move_affine);
2239 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2247 * find_idlest_group finds and returns the least busy CPU group within the
2250 static struct sched_group *
2251 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2252 int this_cpu, int load_idx)
2254 struct sched_group *idlest = NULL, *group = sd->groups;
2255 unsigned long min_load = ULONG_MAX, this_load = 0;
2256 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2259 unsigned long load, avg_load;
2263 /* Skip over this group if it has no CPUs allowed */
2264 if (!cpumask_intersects(sched_group_cpus(group),
2265 tsk_cpus_allowed(p)))
2268 local_group = cpumask_test_cpu(this_cpu,
2269 sched_group_cpus(group));
2271 /* Tally up the load of all CPUs in the group */
2274 for_each_cpu(i, sched_group_cpus(group)) {
2275 /* Bias balancing toward cpus of our domain */
2277 load = source_load(i, load_idx);
2279 load = target_load(i, load_idx);
2284 /* Adjust by relative CPU power of the group */
2285 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2288 this_load = avg_load;
2289 } else if (avg_load < min_load) {
2290 min_load = avg_load;
2293 } while (group = group->next, group != sd->groups);
2295 if (!idlest || 100*this_load < imbalance*min_load)
2301 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2304 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2306 unsigned long load, min_load = ULONG_MAX;
2310 /* Traverse only the allowed CPUs */
2311 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2312 load = weighted_cpuload(i);
2314 if (load < min_load || (load == min_load && i == this_cpu)) {
2324 * Try and locate an idle CPU in the sched_domain.
2326 static int select_idle_sibling(struct task_struct *p, int target)
2328 int cpu = smp_processor_id();
2329 int prev_cpu = task_cpu(p);
2330 struct sched_domain *sd;
2331 struct sched_group *sg;
2335 * If the task is going to be woken-up on this cpu and if it is
2336 * already idle, then it is the right target.
2338 if (target == cpu && idle_cpu(cpu))
2342 * If the task is going to be woken-up on the cpu where it previously
2343 * ran and if it is currently idle, then it the right target.
2345 if (target == prev_cpu && idle_cpu(prev_cpu))
2349 * Otherwise, iterate the domains and find an elegible idle cpu.
2353 for_each_domain(target, sd) {
2354 if (!smt && (sd->flags & SD_SHARE_CPUPOWER))
2357 if (smt && !(sd->flags & SD_SHARE_CPUPOWER))
2360 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
2365 if (!cpumask_intersects(sched_group_cpus(sg),
2366 tsk_cpus_allowed(p)))
2369 for_each_cpu(i, sched_group_cpus(sg)) {
2374 target = cpumask_first_and(sched_group_cpus(sg),
2375 tsk_cpus_allowed(p));
2379 } while (sg != sd->groups);
2392 * sched_balance_self: balance the current task (running on cpu) in domains
2393 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2396 * Balance, ie. select the least loaded group.
2398 * Returns the target CPU number, or the same CPU if no balancing is needed.
2400 * preempt must be disabled.
2403 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2405 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2406 int cpu = smp_processor_id();
2407 int prev_cpu = task_cpu(p);
2409 int want_affine = 0;
2411 int sync = wake_flags & WF_SYNC;
2413 if (sd_flag & SD_BALANCE_WAKE) {
2414 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2420 for_each_domain(cpu, tmp) {
2421 if (!(tmp->flags & SD_LOAD_BALANCE))
2425 * If power savings logic is enabled for a domain, see if we
2426 * are not overloaded, if so, don't balance wider.
2428 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2429 unsigned long power = 0;
2430 unsigned long nr_running = 0;
2431 unsigned long capacity;
2434 for_each_cpu(i, sched_domain_span(tmp)) {
2435 power += power_of(i);
2436 nr_running += cpu_rq(i)->cfs.nr_running;
2439 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2441 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2444 if (nr_running < capacity)
2449 * If both cpu and prev_cpu are part of this domain,
2450 * cpu is a valid SD_WAKE_AFFINE target.
2452 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2453 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2458 if (!want_sd && !want_affine)
2461 if (!(tmp->flags & sd_flag))
2469 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2472 new_cpu = select_idle_sibling(p, prev_cpu);
2477 int load_idx = sd->forkexec_idx;
2478 struct sched_group *group;
2481 if (!(sd->flags & sd_flag)) {
2486 if (sd_flag & SD_BALANCE_WAKE)
2487 load_idx = sd->wake_idx;
2489 group = find_idlest_group(sd, p, cpu, load_idx);
2495 new_cpu = find_idlest_cpu(group, p, cpu);
2496 if (new_cpu == -1 || new_cpu == cpu) {
2497 /* Now try balancing at a lower domain level of cpu */
2502 /* Now try balancing at a lower domain level of new_cpu */
2504 weight = sd->span_weight;
2506 for_each_domain(cpu, tmp) {
2507 if (weight <= tmp->span_weight)
2509 if (tmp->flags & sd_flag)
2512 /* while loop will break here if sd == NULL */
2519 #endif /* CONFIG_SMP */
2521 static unsigned long
2522 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2524 unsigned long gran = sysctl_sched_wakeup_granularity;
2527 * Since its curr running now, convert the gran from real-time
2528 * to virtual-time in his units.
2530 * By using 'se' instead of 'curr' we penalize light tasks, so
2531 * they get preempted easier. That is, if 'se' < 'curr' then
2532 * the resulting gran will be larger, therefore penalizing the
2533 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2534 * be smaller, again penalizing the lighter task.
2536 * This is especially important for buddies when the leftmost
2537 * task is higher priority than the buddy.
2539 return calc_delta_fair(gran, se);
2543 * Should 'se' preempt 'curr'.
2557 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2559 s64 gran, vdiff = curr->vruntime - se->vruntime;
2564 gran = wakeup_gran(curr, se);
2571 static void set_last_buddy(struct sched_entity *se)
2573 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2576 for_each_sched_entity(se)
2577 cfs_rq_of(se)->last = se;
2580 static void set_next_buddy(struct sched_entity *se)
2582 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2585 for_each_sched_entity(se)
2586 cfs_rq_of(se)->next = se;
2589 static void set_skip_buddy(struct sched_entity *se)
2591 for_each_sched_entity(se)
2592 cfs_rq_of(se)->skip = se;
2596 * Preempt the current task with a newly woken task if needed:
2598 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2600 struct task_struct *curr = rq->curr;
2601 struct sched_entity *se = &curr->se, *pse = &p->se;
2602 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2603 int scale = cfs_rq->nr_running >= sched_nr_latency;
2604 int next_buddy_marked = 0;
2606 if (unlikely(se == pse))
2610 * This is possible from callers such as pull_task(), in which we
2611 * unconditionally check_prempt_curr() after an enqueue (which may have
2612 * lead to a throttle). This both saves work and prevents false
2613 * next-buddy nomination below.
2615 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
2618 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2619 set_next_buddy(pse);
2620 next_buddy_marked = 1;
2624 * We can come here with TIF_NEED_RESCHED already set from new task
2627 * Note: this also catches the edge-case of curr being in a throttled
2628 * group (e.g. via set_curr_task), since update_curr() (in the
2629 * enqueue of curr) will have resulted in resched being set. This
2630 * prevents us from potentially nominating it as a false LAST_BUDDY
2633 if (test_tsk_need_resched(curr))
2636 /* Idle tasks are by definition preempted by non-idle tasks. */
2637 if (unlikely(curr->policy == SCHED_IDLE) &&
2638 likely(p->policy != SCHED_IDLE))
2642 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2643 * is driven by the tick):
2645 if (unlikely(p->policy != SCHED_NORMAL))
2648 find_matching_se(&se, &pse);
2649 update_curr(cfs_rq_of(se));
2651 if (wakeup_preempt_entity(se, pse) == 1) {
2653 * Bias pick_next to pick the sched entity that is
2654 * triggering this preemption.
2656 if (!next_buddy_marked)
2657 set_next_buddy(pse);
2666 * Only set the backward buddy when the current task is still
2667 * on the rq. This can happen when a wakeup gets interleaved
2668 * with schedule on the ->pre_schedule() or idle_balance()
2669 * point, either of which can * drop the rq lock.
2671 * Also, during early boot the idle thread is in the fair class,
2672 * for obvious reasons its a bad idea to schedule back to it.
2674 if (unlikely(!se->on_rq || curr == rq->idle))
2677 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2681 static struct task_struct *pick_next_task_fair(struct rq *rq)
2683 struct task_struct *p;
2684 struct cfs_rq *cfs_rq = &rq->cfs;
2685 struct sched_entity *se;
2687 if (!cfs_rq->nr_running)
2691 se = pick_next_entity(cfs_rq);
2692 set_next_entity(cfs_rq, se);
2693 cfs_rq = group_cfs_rq(se);
2697 hrtick_start_fair(rq, p);
2703 * Account for a descheduled task:
2705 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
2707 struct sched_entity *se = &prev->se;
2708 struct cfs_rq *cfs_rq;
2710 for_each_sched_entity(se) {
2711 cfs_rq = cfs_rq_of(se);
2712 put_prev_entity(cfs_rq, se);
2717 * sched_yield() is very simple
2719 * The magic of dealing with the ->skip buddy is in pick_next_entity.
2721 static void yield_task_fair(struct rq *rq)
2723 struct task_struct *curr = rq->curr;
2724 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2725 struct sched_entity *se = &curr->se;
2728 * Are we the only task in the tree?
2730 if (unlikely(rq->nr_running == 1))
2733 clear_buddies(cfs_rq, se);
2735 if (curr->policy != SCHED_BATCH) {
2736 update_rq_clock(rq);
2738 * Update run-time statistics of the 'current'.
2740 update_curr(cfs_rq);
2746 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2748 struct sched_entity *se = &p->se;
2750 /* throttled hierarchies are not runnable */
2751 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
2754 /* Tell the scheduler that we'd really like pse to run next. */
2757 yield_task_fair(rq);
2763 /**************************************************
2764 * Fair scheduling class load-balancing methods:
2768 * pull_task - move a task from a remote runqueue to the local runqueue.
2769 * Both runqueues must be locked.
2771 static void pull_task(struct rq *src_rq, struct task_struct *p,
2772 struct rq *this_rq, int this_cpu)
2774 deactivate_task(src_rq, p, 0);
2775 set_task_cpu(p, this_cpu);
2776 activate_task(this_rq, p, 0);
2777 check_preempt_curr(this_rq, p, 0);
2781 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2784 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2785 struct sched_domain *sd, enum cpu_idle_type idle,
2788 int tsk_cache_hot = 0;
2790 * We do not migrate tasks that are:
2791 * 1) running (obviously), or
2792 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2793 * 3) are cache-hot on their current CPU.
2794 * 4) p->pi_lock is held.
2796 if (!cpumask_test_cpu(this_cpu, tsk_cpus_allowed(p))) {
2797 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2802 if (task_running(rq, p)) {
2803 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2808 * rt -> fair class change may be in progress. If we sneak in should
2809 * double_lock_balance() release rq->lock, and move the task, we will
2810 * cause switched_to_fair() to meet a passed but no longer valid rq.
2812 if (raw_spin_is_locked(&p->pi_lock))
2816 * Aggressive migration if:
2817 * 1) task is cache cold, or
2818 * 2) too many balance attempts have failed.
2821 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2822 if (!tsk_cache_hot ||
2823 sd->nr_balance_failed > sd->cache_nice_tries) {
2824 #ifdef CONFIG_SCHEDSTATS
2825 if (tsk_cache_hot) {
2826 schedstat_inc(sd, lb_hot_gained[idle]);
2827 schedstat_inc(p, se.statistics.nr_forced_migrations);
2833 if (tsk_cache_hot) {
2834 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2841 * move_one_task tries to move exactly one task from busiest to this_rq, as
2842 * part of active balancing operations within "domain".
2843 * Returns 1 if successful and 0 otherwise.
2845 * Called with both runqueues locked.
2848 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2849 struct sched_domain *sd, enum cpu_idle_type idle)
2851 struct task_struct *p, *n;
2852 struct cfs_rq *cfs_rq;
2855 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2856 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2857 if (throttled_lb_pair(task_group(p),
2858 busiest->cpu, this_cpu))
2861 if (!can_migrate_task(p, busiest, this_cpu,
2865 pull_task(busiest, p, this_rq, this_cpu);
2867 * Right now, this is only the second place pull_task()
2868 * is called, so we can safely collect pull_task()
2869 * stats here rather than inside pull_task().
2871 schedstat_inc(sd, lb_gained[idle]);
2879 static unsigned long
2880 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2881 unsigned long max_load_move, struct sched_domain *sd,
2882 enum cpu_idle_type idle, int *all_pinned,
2883 struct cfs_rq *busiest_cfs_rq)
2885 int loops = 0, pulled = 0;
2886 long rem_load_move = max_load_move;
2887 struct task_struct *p, *n;
2889 if (max_load_move == 0)
2892 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2893 if (loops++ > sysctl_sched_nr_migrate)
2896 if ((p->se.load.weight >> 1) > rem_load_move ||
2897 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2901 pull_task(busiest, p, this_rq, this_cpu);
2903 rem_load_move -= p->se.load.weight;
2905 #ifdef CONFIG_PREEMPT
2907 * NEWIDLE balancing is a source of latency, so preemptible
2908 * kernels will stop after the first task is pulled to minimize
2909 * the critical section.
2911 if (idle == CPU_NEWLY_IDLE)
2916 * We only want to steal up to the prescribed amount of
2919 if (rem_load_move <= 0)
2924 * Right now, this is one of only two places pull_task() is called,
2925 * so we can safely collect pull_task() stats here rather than
2926 * inside pull_task().
2928 schedstat_add(sd, lb_gained[idle], pulled);
2930 return max_load_move - rem_load_move;
2933 #ifdef CONFIG_FAIR_GROUP_SCHED
2935 * update tg->load_weight by folding this cpu's load_avg
2937 static int update_shares_cpu(struct task_group *tg, int cpu)
2939 struct cfs_rq *cfs_rq;
2940 unsigned long flags;
2947 cfs_rq = tg->cfs_rq[cpu];
2949 raw_spin_lock_irqsave(&rq->lock, flags);
2951 update_rq_clock(rq);
2952 update_cfs_load(cfs_rq, 1);
2955 * We need to update shares after updating tg->load_weight in
2956 * order to adjust the weight of groups with long running tasks.
2958 update_cfs_shares(cfs_rq);
2960 raw_spin_unlock_irqrestore(&rq->lock, flags);
2965 static void update_shares(int cpu)
2967 struct cfs_rq *cfs_rq;
2968 struct rq *rq = cpu_rq(cpu);
2972 * Iterates the task_group tree in a bottom up fashion, see
2973 * list_add_leaf_cfs_rq() for details.
2975 for_each_leaf_cfs_rq(rq, cfs_rq) {
2976 /* throttled entities do not contribute to load */
2977 if (throttled_hierarchy(cfs_rq))
2980 update_shares_cpu(cfs_rq->tg, cpu);
2986 * Compute the cpu's hierarchical load factor for each task group.
2987 * This needs to be done in a top-down fashion because the load of a child
2988 * group is a fraction of its parents load.
2990 static int tg_load_down(struct task_group *tg, void *data)
2993 long cpu = (long)data;
2996 load = cpu_rq(cpu)->load.weight;
2998 load = tg->parent->cfs_rq[cpu]->h_load;
2999 load *= tg->se[cpu]->load.weight;
3000 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3003 tg->cfs_rq[cpu]->h_load = load;
3008 static void update_h_load(long cpu)
3010 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3013 static unsigned long
3014 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
3015 unsigned long max_load_move,
3016 struct sched_domain *sd, enum cpu_idle_type idle,
3019 long rem_load_move = max_load_move;
3020 struct cfs_rq *busiest_cfs_rq;
3023 update_h_load(cpu_of(busiest));
3025 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
3026 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
3027 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
3028 u64 rem_load, moved_load;
3031 * empty group or part of a throttled hierarchy
3033 if (!busiest_cfs_rq->task_weight ||
3034 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu))
3037 rem_load = (u64)rem_load_move * busiest_weight;
3038 rem_load = div_u64(rem_load, busiest_h_load + 1);
3040 moved_load = balance_tasks(this_rq, this_cpu, busiest,
3041 rem_load, sd, idle, all_pinned,
3047 moved_load *= busiest_h_load;
3048 moved_load = div_u64(moved_load, busiest_weight + 1);
3050 rem_load_move -= moved_load;
3051 if (rem_load_move < 0)
3056 return max_load_move - rem_load_move;
3059 static inline void update_shares(int cpu)
3063 static unsigned long
3064 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
3065 unsigned long max_load_move,
3066 struct sched_domain *sd, enum cpu_idle_type idle,
3069 return balance_tasks(this_rq, this_cpu, busiest,
3070 max_load_move, sd, idle, all_pinned,
3076 * move_tasks tries to move up to max_load_move weighted load from busiest to
3077 * this_rq, as part of a balancing operation within domain "sd".
3078 * Returns 1 if successful and 0 otherwise.
3080 * Called with both runqueues locked.
3082 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3083 unsigned long max_load_move,
3084 struct sched_domain *sd, enum cpu_idle_type idle,
3087 unsigned long total_load_moved = 0, load_moved;
3090 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
3091 max_load_move - total_load_moved,
3092 sd, idle, all_pinned);
3094 total_load_moved += load_moved;
3096 #ifdef CONFIG_PREEMPT
3098 * NEWIDLE balancing is a source of latency, so preemptible
3099 * kernels will stop after the first task is pulled to minimize
3100 * the critical section.
3102 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3105 if (raw_spin_is_contended(&this_rq->lock) ||
3106 raw_spin_is_contended(&busiest->lock))
3109 } while (load_moved && max_load_move > total_load_moved);
3111 return total_load_moved > 0;
3114 /********** Helpers for find_busiest_group ************************/
3116 * sd_lb_stats - Structure to store the statistics of a sched_domain
3117 * during load balancing.
3119 struct sd_lb_stats {
3120 struct sched_group *busiest; /* Busiest group in this sd */
3121 struct sched_group *this; /* Local group in this sd */
3122 unsigned long total_load; /* Total load of all groups in sd */
3123 unsigned long total_pwr; /* Total power of all groups in sd */
3124 unsigned long avg_load; /* Average load across all groups in sd */
3126 /** Statistics of this group */
3127 unsigned long this_load;
3128 unsigned long this_load_per_task;
3129 unsigned long this_nr_running;
3130 unsigned long this_has_capacity;
3131 unsigned int this_idle_cpus;
3133 /* Statistics of the busiest group */
3134 unsigned int busiest_idle_cpus;
3135 unsigned long max_load;
3136 unsigned long busiest_load_per_task;
3137 unsigned long busiest_nr_running;
3138 unsigned long busiest_group_capacity;
3139 unsigned long busiest_has_capacity;
3140 unsigned int busiest_group_weight;
3142 int group_imb; /* Is there imbalance in this sd */
3143 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3144 int power_savings_balance; /* Is powersave balance needed for this sd */
3145 struct sched_group *group_min; /* Least loaded group in sd */
3146 struct sched_group *group_leader; /* Group which relieves group_min */
3147 unsigned long min_load_per_task; /* load_per_task in group_min */
3148 unsigned long leader_nr_running; /* Nr running of group_leader */
3149 unsigned long min_nr_running; /* Nr running of group_min */
3154 * sg_lb_stats - stats of a sched_group required for load_balancing
3156 struct sg_lb_stats {
3157 unsigned long avg_load; /*Avg load across the CPUs of the group */
3158 unsigned long group_load; /* Total load over the CPUs of the group */
3159 unsigned long sum_nr_running; /* Nr tasks running in the group */
3160 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3161 unsigned long group_capacity;
3162 unsigned long idle_cpus;
3163 unsigned long group_weight;
3164 int group_imb; /* Is there an imbalance in the group ? */
3165 int group_has_capacity; /* Is there extra capacity in the group? */
3169 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3170 * @group: The group whose first cpu is to be returned.
3172 static inline unsigned int group_first_cpu(struct sched_group *group)
3174 return cpumask_first(sched_group_cpus(group));
3178 * get_sd_load_idx - Obtain the load index for a given sched domain.
3179 * @sd: The sched_domain whose load_idx is to be obtained.
3180 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3182 static inline int get_sd_load_idx(struct sched_domain *sd,
3183 enum cpu_idle_type idle)
3189 load_idx = sd->busy_idx;
3192 case CPU_NEWLY_IDLE:
3193 load_idx = sd->newidle_idx;
3196 load_idx = sd->idle_idx;
3204 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3206 * init_sd_power_savings_stats - Initialize power savings statistics for
3207 * the given sched_domain, during load balancing.
3209 * @sd: Sched domain whose power-savings statistics are to be initialized.
3210 * @sds: Variable containing the statistics for sd.
3211 * @idle: Idle status of the CPU at which we're performing load-balancing.
3213 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3214 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3217 * Busy processors will not participate in power savings
3220 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3221 sds->power_savings_balance = 0;
3223 sds->power_savings_balance = 1;
3224 sds->min_nr_running = ULONG_MAX;
3225 sds->leader_nr_running = 0;
3230 * update_sd_power_savings_stats - Update the power saving stats for a
3231 * sched_domain while performing load balancing.
3233 * @group: sched_group belonging to the sched_domain under consideration.
3234 * @sds: Variable containing the statistics of the sched_domain
3235 * @local_group: Does group contain the CPU for which we're performing
3237 * @sgs: Variable containing the statistics of the group.
3239 static inline void update_sd_power_savings_stats(struct sched_group *group,
3240 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3243 if (!sds->power_savings_balance)
3247 * If the local group is idle or completely loaded
3248 * no need to do power savings balance at this domain
3250 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3251 !sds->this_nr_running))
3252 sds->power_savings_balance = 0;
3255 * If a group is already running at full capacity or idle,
3256 * don't include that group in power savings calculations
3258 if (!sds->power_savings_balance ||
3259 sgs->sum_nr_running >= sgs->group_capacity ||
3260 !sgs->sum_nr_running)
3264 * Calculate the group which has the least non-idle load.
3265 * This is the group from where we need to pick up the load
3268 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3269 (sgs->sum_nr_running == sds->min_nr_running &&
3270 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3271 sds->group_min = group;
3272 sds->min_nr_running = sgs->sum_nr_running;
3273 sds->min_load_per_task = sgs->sum_weighted_load /
3274 sgs->sum_nr_running;
3278 * Calculate the group which is almost near its
3279 * capacity but still has some space to pick up some load
3280 * from other group and save more power
3282 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3285 if (sgs->sum_nr_running > sds->leader_nr_running ||
3286 (sgs->sum_nr_running == sds->leader_nr_running &&
3287 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3288 sds->group_leader = group;
3289 sds->leader_nr_running = sgs->sum_nr_running;
3294 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3295 * @sds: Variable containing the statistics of the sched_domain
3296 * under consideration.
3297 * @this_cpu: Cpu at which we're currently performing load-balancing.
3298 * @imbalance: Variable to store the imbalance.
3301 * Check if we have potential to perform some power-savings balance.
3302 * If yes, set the busiest group to be the least loaded group in the
3303 * sched_domain, so that it's CPUs can be put to idle.
3305 * Returns 1 if there is potential to perform power-savings balance.
3308 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3309 int this_cpu, unsigned long *imbalance)
3311 if (!sds->power_savings_balance)
3314 if (sds->this != sds->group_leader ||
3315 sds->group_leader == sds->group_min)
3318 *imbalance = sds->min_load_per_task;
3319 sds->busiest = sds->group_min;
3324 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3325 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3326 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3331 static inline void update_sd_power_savings_stats(struct sched_group *group,
3332 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3337 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3338 int this_cpu, unsigned long *imbalance)
3342 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3345 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3347 return SCHED_POWER_SCALE;
3350 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3352 return default_scale_freq_power(sd, cpu);
3355 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3357 unsigned long weight = sd->span_weight;
3358 unsigned long smt_gain = sd->smt_gain;
3365 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3367 return default_scale_smt_power(sd, cpu);
3370 unsigned long scale_rt_power(int cpu)
3372 struct rq *rq = cpu_rq(cpu);
3373 u64 total, available;
3375 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3377 if (unlikely(total < rq->rt_avg)) {
3378 /* Ensures that power won't end up being negative */
3381 available = total - rq->rt_avg;
3384 if (unlikely((s64)total < SCHED_POWER_SCALE))
3385 total = SCHED_POWER_SCALE;
3387 total >>= SCHED_POWER_SHIFT;
3389 return div_u64(available, total);
3392 static void update_cpu_power(struct sched_domain *sd, int cpu)
3394 unsigned long weight = sd->span_weight;
3395 unsigned long power = SCHED_POWER_SCALE;
3396 struct sched_group *sdg = sd->groups;
3398 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3399 if (sched_feat(ARCH_POWER))
3400 power *= arch_scale_smt_power(sd, cpu);
3402 power *= default_scale_smt_power(sd, cpu);
3404 power >>= SCHED_POWER_SHIFT;
3407 sdg->sgp->power_orig = power;
3409 if (sched_feat(ARCH_POWER))
3410 power *= arch_scale_freq_power(sd, cpu);
3412 power *= default_scale_freq_power(sd, cpu);
3414 power >>= SCHED_POWER_SHIFT;
3416 power *= scale_rt_power(cpu);
3417 power >>= SCHED_POWER_SHIFT;
3422 cpu_rq(cpu)->cpu_power = power;
3423 sdg->sgp->power = power;
3426 static void update_group_power(struct sched_domain *sd, int cpu)
3428 struct sched_domain *child = sd->child;
3429 struct sched_group *group, *sdg = sd->groups;
3430 unsigned long power;
3433 update_cpu_power(sd, cpu);
3439 group = child->groups;
3441 power += group->sgp->power;
3442 group = group->next;
3443 } while (group != child->groups);
3445 sdg->sgp->power = power;
3449 * Try and fix up capacity for tiny siblings, this is needed when
3450 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3451 * which on its own isn't powerful enough.
3453 * See update_sd_pick_busiest() and check_asym_packing().
3456 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3459 * Only siblings can have significantly less than SCHED_POWER_SCALE
3461 if (!(sd->flags & SD_SHARE_CPUPOWER))
3465 * If ~90% of the cpu_power is still there, we're good.
3467 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3474 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3475 * @sd: The sched_domain whose statistics are to be updated.
3476 * @group: sched_group whose statistics are to be updated.
3477 * @this_cpu: Cpu for which load balance is currently performed.
3478 * @idle: Idle status of this_cpu
3479 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3480 * @local_group: Does group contain this_cpu.
3481 * @cpus: Set of cpus considered for load balancing.
3482 * @balance: Should we balance.
3483 * @sgs: variable to hold the statistics for this group.
3485 static inline void update_sg_lb_stats(struct sched_domain *sd,
3486 struct sched_group *group, int this_cpu,
3487 enum cpu_idle_type idle, int load_idx,
3488 int local_group, const struct cpumask *cpus,
3489 int *balance, struct sg_lb_stats *sgs)
3491 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3493 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3494 unsigned long avg_load_per_task = 0;
3497 balance_cpu = group_first_cpu(group);
3499 /* Tally up the load of all CPUs in the group */
3501 min_cpu_load = ~0UL;
3504 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3505 struct rq *rq = cpu_rq(i);
3507 /* Bias balancing toward cpus of our domain */
3509 if (idle_cpu(i) && !first_idle_cpu) {
3514 load = target_load(i, load_idx);
3516 load = source_load(i, load_idx);
3517 if (load > max_cpu_load) {
3518 max_cpu_load = load;
3519 max_nr_running = rq->nr_running;
3521 if (min_cpu_load > load)
3522 min_cpu_load = load;
3525 sgs->group_load += load;
3526 sgs->sum_nr_running += rq->nr_running;
3527 sgs->sum_weighted_load += weighted_cpuload(i);
3533 * First idle cpu or the first cpu(busiest) in this sched group
3534 * is eligible for doing load balancing at this and above
3535 * domains. In the newly idle case, we will allow all the cpu's
3536 * to do the newly idle load balance.
3538 if (idle != CPU_NEWLY_IDLE && local_group) {
3539 if (balance_cpu != this_cpu) {
3543 update_group_power(sd, this_cpu);
3546 /* Adjust by relative CPU power of the group */
3547 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3550 * Consider the group unbalanced when the imbalance is larger
3551 * than the average weight of a task.
3553 * APZ: with cgroup the avg task weight can vary wildly and
3554 * might not be a suitable number - should we keep a
3555 * normalized nr_running number somewhere that negates
3558 if (sgs->sum_nr_running)
3559 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3561 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3564 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3566 if (!sgs->group_capacity)
3567 sgs->group_capacity = fix_small_capacity(sd, group);
3568 sgs->group_weight = group->group_weight;
3570 if (sgs->group_capacity > sgs->sum_nr_running)
3571 sgs->group_has_capacity = 1;
3575 * update_sd_pick_busiest - return 1 on busiest group
3576 * @sd: sched_domain whose statistics are to be checked
3577 * @sds: sched_domain statistics
3578 * @sg: sched_group candidate to be checked for being the busiest
3579 * @sgs: sched_group statistics
3580 * @this_cpu: the current cpu
3582 * Determine if @sg is a busier group than the previously selected
3585 static bool update_sd_pick_busiest(struct sched_domain *sd,
3586 struct sd_lb_stats *sds,
3587 struct sched_group *sg,
3588 struct sg_lb_stats *sgs,
3591 if (sgs->avg_load <= sds->max_load)
3594 if (sgs->sum_nr_running > sgs->group_capacity)
3601 * ASYM_PACKING needs to move all the work to the lowest
3602 * numbered CPUs in the group, therefore mark all groups
3603 * higher than ourself as busy.
3605 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3606 this_cpu < group_first_cpu(sg)) {
3610 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3618 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3619 * @sd: sched_domain whose statistics are to be updated.
3620 * @this_cpu: Cpu for which load balance is currently performed.
3621 * @idle: Idle status of this_cpu
3622 * @cpus: Set of cpus considered for load balancing.
3623 * @balance: Should we balance.
3624 * @sds: variable to hold the statistics for this sched_domain.
3626 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3627 enum cpu_idle_type idle, const struct cpumask *cpus,
3628 int *balance, struct sd_lb_stats *sds)
3630 struct sched_domain *child = sd->child;
3631 struct sched_group *sg = sd->groups;
3632 struct sg_lb_stats sgs;
3633 int load_idx, prefer_sibling = 0;
3635 if (child && child->flags & SD_PREFER_SIBLING)
3638 init_sd_power_savings_stats(sd, sds, idle);
3639 load_idx = get_sd_load_idx(sd, idle);
3644 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3645 memset(&sgs, 0, sizeof(sgs));
3646 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3647 local_group, cpus, balance, &sgs);
3649 if (local_group && !(*balance))
3652 sds->total_load += sgs.group_load;
3653 sds->total_pwr += sg->sgp->power;
3656 * In case the child domain prefers tasks go to siblings
3657 * first, lower the sg capacity to one so that we'll try
3658 * and move all the excess tasks away. We lower the capacity
3659 * of a group only if the local group has the capacity to fit
3660 * these excess tasks, i.e. nr_running < group_capacity. The
3661 * extra check prevents the case where you always pull from the
3662 * heaviest group when it is already under-utilized (possible
3663 * with a large weight task outweighs the tasks on the system).
3665 if (prefer_sibling && !local_group && sds->this_has_capacity)
3666 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3669 sds->this_load = sgs.avg_load;
3671 sds->this_nr_running = sgs.sum_nr_running;
3672 sds->this_load_per_task = sgs.sum_weighted_load;
3673 sds->this_has_capacity = sgs.group_has_capacity;
3674 sds->this_idle_cpus = sgs.idle_cpus;
3675 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
3676 sds->max_load = sgs.avg_load;
3678 sds->busiest_nr_running = sgs.sum_nr_running;
3679 sds->busiest_idle_cpus = sgs.idle_cpus;
3680 sds->busiest_group_capacity = sgs.group_capacity;
3681 sds->busiest_load_per_task = sgs.sum_weighted_load;
3682 sds->busiest_has_capacity = sgs.group_has_capacity;
3683 sds->busiest_group_weight = sgs.group_weight;
3684 sds->group_imb = sgs.group_imb;
3687 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
3689 } while (sg != sd->groups);
3692 int __weak arch_sd_sibling_asym_packing(void)
3694 return 0*SD_ASYM_PACKING;
3698 * check_asym_packing - Check to see if the group is packed into the
3701 * This is primarily intended to used at the sibling level. Some
3702 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3703 * case of POWER7, it can move to lower SMT modes only when higher
3704 * threads are idle. When in lower SMT modes, the threads will
3705 * perform better since they share less core resources. Hence when we
3706 * have idle threads, we want them to be the higher ones.
3708 * This packing function is run on idle threads. It checks to see if
3709 * the busiest CPU in this domain (core in the P7 case) has a higher
3710 * CPU number than the packing function is being run on. Here we are
3711 * assuming lower CPU number will be equivalent to lower a SMT thread
3714 * Returns 1 when packing is required and a task should be moved to
3715 * this CPU. The amount of the imbalance is returned in *imbalance.
3717 * @sd: The sched_domain whose packing is to be checked.
3718 * @sds: Statistics of the sched_domain which is to be packed
3719 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3720 * @imbalance: returns amount of imbalanced due to packing.
3722 static int check_asym_packing(struct sched_domain *sd,
3723 struct sd_lb_stats *sds,
3724 int this_cpu, unsigned long *imbalance)
3728 if (!(sd->flags & SD_ASYM_PACKING))
3734 busiest_cpu = group_first_cpu(sds->busiest);
3735 if (this_cpu > busiest_cpu)
3738 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
3744 * fix_small_imbalance - Calculate the minor imbalance that exists
3745 * amongst the groups of a sched_domain, during
3747 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3748 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3749 * @imbalance: Variable to store the imbalance.
3751 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3752 int this_cpu, unsigned long *imbalance)
3754 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3755 unsigned int imbn = 2;
3756 unsigned long scaled_busy_load_per_task;
3758 if (sds->this_nr_running) {
3759 sds->this_load_per_task /= sds->this_nr_running;
3760 if (sds->busiest_load_per_task >
3761 sds->this_load_per_task)
3764 sds->this_load_per_task =
3765 cpu_avg_load_per_task(this_cpu);
3767 scaled_busy_load_per_task = sds->busiest_load_per_task
3768 * SCHED_POWER_SCALE;
3769 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3771 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3772 (scaled_busy_load_per_task * imbn)) {
3773 *imbalance = sds->busiest_load_per_task;
3778 * OK, we don't have enough imbalance to justify moving tasks,
3779 * however we may be able to increase total CPU power used by
3783 pwr_now += sds->busiest->sgp->power *
3784 min(sds->busiest_load_per_task, sds->max_load);
3785 pwr_now += sds->this->sgp->power *
3786 min(sds->this_load_per_task, sds->this_load);
3787 pwr_now /= SCHED_POWER_SCALE;
3789 /* Amount of load we'd subtract */
3790 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3791 sds->busiest->sgp->power;
3792 if (sds->max_load > tmp)
3793 pwr_move += sds->busiest->sgp->power *
3794 min(sds->busiest_load_per_task, sds->max_load - tmp);
3796 /* Amount of load we'd add */
3797 if (sds->max_load * sds->busiest->sgp->power <
3798 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3799 tmp = (sds->max_load * sds->busiest->sgp->power) /
3800 sds->this->sgp->power;
3802 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3803 sds->this->sgp->power;
3804 pwr_move += sds->this->sgp->power *
3805 min(sds->this_load_per_task, sds->this_load + tmp);
3806 pwr_move /= SCHED_POWER_SCALE;
3808 /* Move if we gain throughput */
3809 if (pwr_move > pwr_now)
3810 *imbalance = sds->busiest_load_per_task;
3814 * calculate_imbalance - Calculate the amount of imbalance present within the
3815 * groups of a given sched_domain during load balance.
3816 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3817 * @this_cpu: Cpu for which currently load balance is being performed.
3818 * @imbalance: The variable to store the imbalance.
3820 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3821 unsigned long *imbalance)
3823 unsigned long max_pull, load_above_capacity = ~0UL;
3825 sds->busiest_load_per_task /= sds->busiest_nr_running;
3826 if (sds->group_imb) {
3827 sds->busiest_load_per_task =
3828 min(sds->busiest_load_per_task, sds->avg_load);
3832 * In the presence of smp nice balancing, certain scenarios can have
3833 * max load less than avg load(as we skip the groups at or below
3834 * its cpu_power, while calculating max_load..)
3836 if (sds->max_load < sds->avg_load) {
3838 return fix_small_imbalance(sds, this_cpu, imbalance);
3841 if (!sds->group_imb) {
3843 * Don't want to pull so many tasks that a group would go idle.
3845 load_above_capacity = (sds->busiest_nr_running -
3846 sds->busiest_group_capacity);
3848 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3850 load_above_capacity /= sds->busiest->sgp->power;
3854 * We're trying to get all the cpus to the average_load, so we don't
3855 * want to push ourselves above the average load, nor do we wish to
3856 * reduce the max loaded cpu below the average load. At the same time,
3857 * we also don't want to reduce the group load below the group capacity
3858 * (so that we can implement power-savings policies etc). Thus we look
3859 * for the minimum possible imbalance.
3860 * Be careful of negative numbers as they'll appear as very large values
3861 * with unsigned longs.
3863 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3865 /* How much load to actually move to equalise the imbalance */
3866 *imbalance = min(max_pull * sds->busiest->sgp->power,
3867 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3868 / SCHED_POWER_SCALE;
3871 * if *imbalance is less than the average load per runnable task
3872 * there is no guarantee that any tasks will be moved so we'll have
3873 * a think about bumping its value to force at least one task to be
3876 if (*imbalance < sds->busiest_load_per_task)
3877 return fix_small_imbalance(sds, this_cpu, imbalance);
3881 /******* find_busiest_group() helpers end here *********************/
3884 * find_busiest_group - Returns the busiest group within the sched_domain
3885 * if there is an imbalance. If there isn't an imbalance, and
3886 * the user has opted for power-savings, it returns a group whose
3887 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3888 * such a group exists.
3890 * Also calculates the amount of weighted load which should be moved
3891 * to restore balance.
3893 * @sd: The sched_domain whose busiest group is to be returned.
3894 * @this_cpu: The cpu for which load balancing is currently being performed.
3895 * @imbalance: Variable which stores amount of weighted load which should
3896 * be moved to restore balance/put a group to idle.
3897 * @idle: The idle status of this_cpu.
3898 * @cpus: The set of CPUs under consideration for load-balancing.
3899 * @balance: Pointer to a variable indicating if this_cpu
3900 * is the appropriate cpu to perform load balancing at this_level.
3902 * Returns: - the busiest group if imbalance exists.
3903 * - If no imbalance and user has opted for power-savings balance,
3904 * return the least loaded group whose CPUs can be
3905 * put to idle by rebalancing its tasks onto our group.
3907 static struct sched_group *
3908 find_busiest_group(struct sched_domain *sd, int this_cpu,
3909 unsigned long *imbalance, enum cpu_idle_type idle,
3910 const struct cpumask *cpus, int *balance)
3912 struct sd_lb_stats sds;
3914 memset(&sds, 0, sizeof(sds));
3917 * Compute the various statistics relavent for load balancing at
3920 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3923 * this_cpu is not the appropriate cpu to perform load balancing at
3929 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3930 check_asym_packing(sd, &sds, this_cpu, imbalance))
3933 /* There is no busy sibling group to pull tasks from */
3934 if (!sds.busiest || sds.busiest_nr_running == 0)
3937 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3940 * If the busiest group is imbalanced the below checks don't
3941 * work because they assumes all things are equal, which typically
3942 * isn't true due to cpus_allowed constraints and the like.
3947 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3948 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3949 !sds.busiest_has_capacity)
3953 * If the local group is more busy than the selected busiest group
3954 * don't try and pull any tasks.
3956 if (sds.this_load >= sds.max_load)
3960 * Don't pull any tasks if this group is already above the domain
3963 if (sds.this_load >= sds.avg_load)
3966 if (idle == CPU_IDLE) {
3968 * This cpu is idle. If the busiest group load doesn't
3969 * have more tasks than the number of available cpu's and
3970 * there is no imbalance between this and busiest group
3971 * wrt to idle cpu's, it is balanced.
3973 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3974 sds.busiest_nr_running <= sds.busiest_group_weight)
3978 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3979 * imbalance_pct to be conservative.
3981 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3986 /* Looks like there is an imbalance. Compute it */
3987 calculate_imbalance(&sds, this_cpu, imbalance);
3992 * There is no obvious imbalance. But check if we can do some balancing
3995 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4003 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4006 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
4007 enum cpu_idle_type idle, unsigned long imbalance,
4008 const struct cpumask *cpus)
4010 struct rq *busiest = NULL, *rq;
4011 unsigned long max_load = 0;
4014 for_each_cpu(i, sched_group_cpus(group)) {
4015 unsigned long power = power_of(i);
4016 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4021 capacity = fix_small_capacity(sd, group);
4023 if (!cpumask_test_cpu(i, cpus))
4027 wl = weighted_cpuload(i);
4030 * When comparing with imbalance, use weighted_cpuload()
4031 * which is not scaled with the cpu power.
4033 if (capacity && rq->nr_running == 1 && wl > imbalance)
4037 * For the load comparisons with the other cpu's, consider
4038 * the weighted_cpuload() scaled with the cpu power, so that
4039 * the load can be moved away from the cpu that is potentially
4040 * running at a lower capacity.
4042 wl = (wl * SCHED_POWER_SCALE) / power;
4044 if (wl > max_load) {
4054 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4055 * so long as it is large enough.
4057 #define MAX_PINNED_INTERVAL 512
4059 /* Working cpumask for load_balance and load_balance_newidle. */
4060 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4062 static int need_active_balance(struct sched_domain *sd, int idle,
4063 int busiest_cpu, int this_cpu)
4065 if (idle == CPU_NEWLY_IDLE) {
4068 * ASYM_PACKING needs to force migrate tasks from busy but
4069 * higher numbered CPUs in order to pack all tasks in the
4070 * lowest numbered CPUs.
4072 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
4076 * The only task running in a non-idle cpu can be moved to this
4077 * cpu in an attempt to completely freeup the other CPU
4080 * The package power saving logic comes from
4081 * find_busiest_group(). If there are no imbalance, then
4082 * f_b_g() will return NULL. However when sched_mc={1,2} then
4083 * f_b_g() will select a group from which a running task may be
4084 * pulled to this cpu in order to make the other package idle.
4085 * If there is no opportunity to make a package idle and if
4086 * there are no imbalance, then f_b_g() will return NULL and no
4087 * action will be taken in load_balance_newidle().
4089 * Under normal task pull operation due to imbalance, there
4090 * will be more than one task in the source run queue and
4091 * move_tasks() will succeed. ld_moved will be true and this
4092 * active balance code will not be triggered.
4094 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4098 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4101 static int active_load_balance_cpu_stop(void *data);
4104 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4105 * tasks if there is an imbalance.
4107 static int load_balance(int this_cpu, struct rq *this_rq,
4108 struct sched_domain *sd, enum cpu_idle_type idle,
4111 int ld_moved, all_pinned = 0, active_balance = 0;
4112 struct sched_group *group;
4113 unsigned long imbalance;
4115 unsigned long flags;
4116 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4118 cpumask_copy(cpus, cpu_active_mask);
4120 schedstat_inc(sd, lb_count[idle]);
4123 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
4130 schedstat_inc(sd, lb_nobusyg[idle]);
4134 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
4136 schedstat_inc(sd, lb_nobusyq[idle]);
4140 BUG_ON(busiest == this_rq);
4142 schedstat_add(sd, lb_imbalance[idle], imbalance);
4145 if (busiest->nr_running > 1) {
4147 * Attempt to move tasks. If find_busiest_group has found
4148 * an imbalance but busiest->nr_running <= 1, the group is
4149 * still unbalanced. ld_moved simply stays zero, so it is
4150 * correctly treated as an imbalance.
4153 local_irq_save(flags);
4154 double_rq_lock(this_rq, busiest);
4155 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4156 imbalance, sd, idle, &all_pinned);
4157 double_rq_unlock(this_rq, busiest);
4158 local_irq_restore(flags);
4161 * some other cpu did the load balance for us.
4163 if (ld_moved && this_cpu != smp_processor_id())
4164 resched_cpu(this_cpu);
4166 /* All tasks on this runqueue were pinned by CPU affinity */
4167 if (unlikely(all_pinned)) {
4168 cpumask_clear_cpu(cpu_of(busiest), cpus);
4169 if (!cpumask_empty(cpus))
4176 schedstat_inc(sd, lb_failed[idle]);
4178 * Increment the failure counter only on periodic balance.
4179 * We do not want newidle balance, which can be very
4180 * frequent, pollute the failure counter causing
4181 * excessive cache_hot migrations and active balances.
4183 if (idle != CPU_NEWLY_IDLE)
4184 sd->nr_balance_failed++;
4186 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
4187 raw_spin_lock_irqsave(&busiest->lock, flags);
4189 /* don't kick the active_load_balance_cpu_stop,
4190 * if the curr task on busiest cpu can't be
4193 if (!cpumask_test_cpu(this_cpu,
4194 tsk_cpus_allowed(busiest->curr))) {
4195 raw_spin_unlock_irqrestore(&busiest->lock,
4198 goto out_one_pinned;
4202 * ->active_balance synchronizes accesses to
4203 * ->active_balance_work. Once set, it's cleared
4204 * only after active load balance is finished.
4206 if (!busiest->active_balance) {
4207 busiest->active_balance = 1;
4208 busiest->push_cpu = this_cpu;
4211 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4214 stop_one_cpu_nowait(cpu_of(busiest),
4215 active_load_balance_cpu_stop, busiest,
4216 &busiest->active_balance_work);
4219 * We've kicked active balancing, reset the failure
4222 sd->nr_balance_failed = sd->cache_nice_tries+1;
4225 sd->nr_balance_failed = 0;
4227 if (likely(!active_balance)) {
4228 /* We were unbalanced, so reset the balancing interval */
4229 sd->balance_interval = sd->min_interval;
4232 * If we've begun active balancing, start to back off. This
4233 * case may not be covered by the all_pinned logic if there
4234 * is only 1 task on the busy runqueue (because we don't call
4237 if (sd->balance_interval < sd->max_interval)
4238 sd->balance_interval *= 2;
4244 schedstat_inc(sd, lb_balanced[idle]);
4246 sd->nr_balance_failed = 0;
4249 /* tune up the balancing interval */
4250 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4251 (sd->balance_interval < sd->max_interval))
4252 sd->balance_interval *= 2;
4260 * idle_balance is called by schedule() if this_cpu is about to become
4261 * idle. Attempts to pull tasks from other CPUs.
4263 static void idle_balance(int this_cpu, struct rq *this_rq)
4265 struct sched_domain *sd;
4266 int pulled_task = 0;
4267 unsigned long next_balance = jiffies + HZ;
4269 this_rq->idle_stamp = this_rq->clock;
4271 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4275 * Drop the rq->lock, but keep IRQ/preempt disabled.
4277 raw_spin_unlock(&this_rq->lock);
4279 update_shares(this_cpu);
4281 for_each_domain(this_cpu, sd) {
4282 unsigned long interval;
4285 if (!(sd->flags & SD_LOAD_BALANCE))
4288 if (sd->flags & SD_BALANCE_NEWIDLE) {
4289 /* If we've pulled tasks over stop searching: */
4290 pulled_task = load_balance(this_cpu, this_rq,
4291 sd, CPU_NEWLY_IDLE, &balance);
4294 interval = msecs_to_jiffies(sd->balance_interval);
4295 if (time_after(next_balance, sd->last_balance + interval))
4296 next_balance = sd->last_balance + interval;
4298 this_rq->idle_stamp = 0;
4304 raw_spin_lock(&this_rq->lock);
4306 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4308 * We are going idle. next_balance may be set based on
4309 * a busy processor. So reset next_balance.
4311 this_rq->next_balance = next_balance;
4316 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4317 * running tasks off the busiest CPU onto idle CPUs. It requires at
4318 * least 1 task to be running on each physical CPU where possible, and
4319 * avoids physical / logical imbalances.
4321 static int active_load_balance_cpu_stop(void *data)
4323 struct rq *busiest_rq = data;
4324 int busiest_cpu = cpu_of(busiest_rq);
4325 int target_cpu = busiest_rq->push_cpu;
4326 struct rq *target_rq = cpu_rq(target_cpu);
4327 struct sched_domain *sd;
4329 raw_spin_lock_irq(&busiest_rq->lock);
4331 /* make sure the requested cpu hasn't gone down in the meantime */
4332 if (unlikely(busiest_cpu != smp_processor_id() ||
4333 !busiest_rq->active_balance))
4336 /* Is there any task to move? */
4337 if (busiest_rq->nr_running <= 1)
4341 * This condition is "impossible", if it occurs
4342 * we need to fix it. Originally reported by
4343 * Bjorn Helgaas on a 128-cpu setup.
4345 BUG_ON(busiest_rq == target_rq);
4347 /* move a task from busiest_rq to target_rq */
4348 double_lock_balance(busiest_rq, target_rq);
4350 /* Search for an sd spanning us and the target CPU. */
4352 for_each_domain(target_cpu, sd) {
4353 if ((sd->flags & SD_LOAD_BALANCE) &&
4354 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4359 schedstat_inc(sd, alb_count);
4361 if (move_one_task(target_rq, target_cpu, busiest_rq,
4363 schedstat_inc(sd, alb_pushed);
4365 schedstat_inc(sd, alb_failed);
4368 double_unlock_balance(busiest_rq, target_rq);
4370 busiest_rq->active_balance = 0;
4371 raw_spin_unlock_irq(&busiest_rq->lock);
4377 * idle load balancing details
4378 * - One of the idle CPUs nominates itself as idle load_balancer, while
4380 * - This idle load balancer CPU will also go into tickless mode when
4381 * it is idle, just like all other idle CPUs
4382 * - When one of the busy CPUs notice that there may be an idle rebalancing
4383 * needed, they will kick the idle load balancer, which then does idle
4384 * load balancing for all the idle CPUs.
4387 atomic_t load_balancer;
4388 atomic_t first_pick_cpu;
4389 atomic_t second_pick_cpu;
4390 cpumask_var_t idle_cpus_mask;
4391 cpumask_var_t grp_idle_mask;
4392 unsigned long next_balance; /* in jiffy units */
4393 } nohz ____cacheline_aligned;
4395 int get_nohz_load_balancer(void)
4397 return atomic_read(&nohz.load_balancer);
4400 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4402 * lowest_flag_domain - Return lowest sched_domain containing flag.
4403 * @cpu: The cpu whose lowest level of sched domain is to
4405 * @flag: The flag to check for the lowest sched_domain
4406 * for the given cpu.
4408 * Returns the lowest sched_domain of a cpu which contains the given flag.
4410 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4412 struct sched_domain *sd;
4414 for_each_domain(cpu, sd)
4415 if (sd->flags & flag)
4422 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4423 * @cpu: The cpu whose domains we're iterating over.
4424 * @sd: variable holding the value of the power_savings_sd
4426 * @flag: The flag to filter the sched_domains to be iterated.
4428 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4429 * set, starting from the lowest sched_domain to the highest.
4431 #define for_each_flag_domain(cpu, sd, flag) \
4432 for (sd = lowest_flag_domain(cpu, flag); \
4433 (sd && (sd->flags & flag)); sd = sd->parent)
4436 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4437 * @ilb_group: group to be checked for semi-idleness
4439 * Returns: 1 if the group is semi-idle. 0 otherwise.
4441 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4442 * and atleast one non-idle CPU. This helper function checks if the given
4443 * sched_group is semi-idle or not.
4445 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4447 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
4448 sched_group_cpus(ilb_group));
4451 * A sched_group is semi-idle when it has atleast one busy cpu
4452 * and atleast one idle cpu.
4454 if (cpumask_empty(nohz.grp_idle_mask))
4457 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
4463 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4464 * @cpu: The cpu which is nominating a new idle_load_balancer.
4466 * Returns: Returns the id of the idle load balancer if it exists,
4467 * Else, returns >= nr_cpu_ids.
4469 * This algorithm picks the idle load balancer such that it belongs to a
4470 * semi-idle powersavings sched_domain. The idea is to try and avoid
4471 * completely idle packages/cores just for the purpose of idle load balancing
4472 * when there are other idle cpu's which are better suited for that job.
4474 static int find_new_ilb(int cpu)
4476 struct sched_domain *sd;
4477 struct sched_group *ilb_group;
4478 int ilb = nr_cpu_ids;
4481 * Have idle load balancer selection from semi-idle packages only
4482 * when power-aware load balancing is enabled
4484 if (!(sched_smt_power_savings || sched_mc_power_savings))
4488 * Optimize for the case when we have no idle CPUs or only one
4489 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4491 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4495 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4496 ilb_group = sd->groups;
4499 if (is_semi_idle_group(ilb_group)) {
4500 ilb = cpumask_first(nohz.grp_idle_mask);
4504 ilb_group = ilb_group->next;
4506 } while (ilb_group != sd->groups);
4514 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4515 static inline int find_new_ilb(int call_cpu)
4522 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4523 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4524 * CPU (if there is one).
4526 static void nohz_balancer_kick(int cpu)
4530 nohz.next_balance++;
4532 ilb_cpu = get_nohz_load_balancer();
4534 if (ilb_cpu >= nr_cpu_ids) {
4535 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4536 if (ilb_cpu >= nr_cpu_ids)
4540 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4541 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4545 * Use smp_send_reschedule() instead of resched_cpu().
4546 * This way we generate a sched IPI on the target cpu which
4547 * is idle. And the softirq performing nohz idle load balance
4548 * will be run before returning from the IPI.
4550 smp_send_reschedule(ilb_cpu);
4556 * This routine will try to nominate the ilb (idle load balancing)
4557 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4558 * load balancing on behalf of all those cpus.
4560 * When the ilb owner becomes busy, we will not have new ilb owner until some
4561 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4562 * idle load balancing by kicking one of the idle CPUs.
4564 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4565 * ilb owner CPU in future (when there is a need for idle load balancing on
4566 * behalf of all idle CPUs).
4568 void select_nohz_load_balancer(int stop_tick)
4570 int cpu = smp_processor_id();
4573 if (!cpu_active(cpu)) {
4574 if (atomic_read(&nohz.load_balancer) != cpu)
4578 * If we are going offline and still the leader,
4581 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4588 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4590 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4591 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4592 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4593 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4595 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4598 /* make me the ilb owner */
4599 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4604 * Check to see if there is a more power-efficient
4607 new_ilb = find_new_ilb(cpu);
4608 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4609 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4610 resched_cpu(new_ilb);
4616 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4619 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4621 if (atomic_read(&nohz.load_balancer) == cpu)
4622 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4630 static DEFINE_SPINLOCK(balancing);
4632 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4635 * Scale the max load_balance interval with the number of CPUs in the system.
4636 * This trades load-balance latency on larger machines for less cross talk.
4638 static void update_max_interval(void)
4640 max_load_balance_interval = HZ*num_online_cpus()/10;
4644 * It checks each scheduling domain to see if it is due to be balanced,
4645 * and initiates a balancing operation if so.
4647 * Balancing parameters are set up in arch_init_sched_domains.
4649 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4652 struct rq *rq = cpu_rq(cpu);
4653 unsigned long interval;
4654 struct sched_domain *sd;
4655 /* Earliest time when we have to do rebalance again */
4656 unsigned long next_balance = jiffies + 60*HZ;
4657 int update_next_balance = 0;
4663 for_each_domain(cpu, sd) {
4664 if (!(sd->flags & SD_LOAD_BALANCE))
4667 interval = sd->balance_interval;
4668 if (idle != CPU_IDLE)
4669 interval *= sd->busy_factor;
4671 /* scale ms to jiffies */
4672 interval = msecs_to_jiffies(interval);
4673 interval = clamp(interval, 1UL, max_load_balance_interval);
4675 need_serialize = sd->flags & SD_SERIALIZE;
4677 if (need_serialize) {
4678 if (!spin_trylock(&balancing))
4682 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4683 if (load_balance(cpu, rq, sd, idle, &balance)) {
4685 * We've pulled tasks over so either we're no
4688 idle = CPU_NOT_IDLE;
4690 sd->last_balance = jiffies;
4693 spin_unlock(&balancing);
4695 if (time_after(next_balance, sd->last_balance + interval)) {
4696 next_balance = sd->last_balance + interval;
4697 update_next_balance = 1;
4701 * Stop the load balance at this level. There is another
4702 * CPU in our sched group which is doing load balancing more
4711 * next_balance will be updated only when there is a need.
4712 * When the cpu is attached to null domain for ex, it will not be
4715 if (likely(update_next_balance))
4716 rq->next_balance = next_balance;
4721 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4722 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4724 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4726 struct rq *this_rq = cpu_rq(this_cpu);
4730 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
4733 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4734 if (balance_cpu == this_cpu)
4738 * If this cpu gets work to do, stop the load balancing
4739 * work being done for other cpus. Next load
4740 * balancing owner will pick it up.
4742 if (need_resched()) {
4743 this_rq->nohz_balance_kick = 0;
4747 raw_spin_lock_irq(&this_rq->lock);
4748 update_rq_clock(this_rq);
4749 update_idle_cpu_load(this_rq);
4750 raw_spin_unlock_irq(&this_rq->lock);
4752 rebalance_domains(balance_cpu, CPU_IDLE);
4754 rq = cpu_rq(balance_cpu);
4755 if (time_after(this_rq->next_balance, rq->next_balance))
4756 this_rq->next_balance = rq->next_balance;
4758 nohz.next_balance = this_rq->next_balance;
4759 this_rq->nohz_balance_kick = 0;
4763 * Current heuristic for kicking the idle load balancer
4764 * - first_pick_cpu is the one of the busy CPUs. It will kick
4765 * idle load balancer when it has more than one process active. This
4766 * eliminates the need for idle load balancing altogether when we have
4767 * only one running process in the system (common case).
4768 * - If there are more than one busy CPU, idle load balancer may have
4769 * to run for active_load_balance to happen (i.e., two busy CPUs are
4770 * SMT or core siblings and can run better if they move to different
4771 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4772 * which will kick idle load balancer as soon as it has any load.
4774 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4776 unsigned long now = jiffies;
4778 int first_pick_cpu, second_pick_cpu;
4780 if (time_before(now, nohz.next_balance))
4786 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4787 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4789 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4790 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4793 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4794 if (ret == nr_cpu_ids || ret == cpu) {
4795 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4796 if (rq->nr_running > 1)
4799 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4800 if (ret == nr_cpu_ids || ret == cpu) {
4808 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4812 * run_rebalance_domains is triggered when needed from the scheduler tick.
4813 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4815 static void run_rebalance_domains(struct softirq_action *h)
4817 int this_cpu = smp_processor_id();
4818 struct rq *this_rq = cpu_rq(this_cpu);
4819 enum cpu_idle_type idle = this_rq->idle_balance ?
4820 CPU_IDLE : CPU_NOT_IDLE;
4822 rebalance_domains(this_cpu, idle);
4825 * If this cpu has a pending nohz_balance_kick, then do the
4826 * balancing on behalf of the other idle cpus whose ticks are
4829 nohz_idle_balance(this_cpu, idle);
4832 static inline int on_null_domain(int cpu)
4834 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4838 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4840 static inline void trigger_load_balance(struct rq *rq, int cpu)
4842 /* Don't need to rebalance while attached to NULL domain */
4843 if (time_after_eq(jiffies, rq->next_balance) &&
4844 likely(!on_null_domain(cpu)))
4845 raise_softirq(SCHED_SOFTIRQ);
4847 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4848 nohz_balancer_kick(cpu);
4852 static void rq_online_fair(struct rq *rq)
4857 static void rq_offline_fair(struct rq *rq)
4861 /* Ensure any throttled groups are reachable by pick_next_task */
4862 unthrottle_offline_cfs_rqs(rq);
4865 #else /* CONFIG_SMP */
4868 * on UP we do not need to balance between CPUs:
4870 static inline void idle_balance(int cpu, struct rq *rq)
4874 #endif /* CONFIG_SMP */
4877 * scheduler tick hitting a task of our scheduling class:
4879 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4881 struct cfs_rq *cfs_rq;
4882 struct sched_entity *se = &curr->se;
4884 for_each_sched_entity(se) {
4885 cfs_rq = cfs_rq_of(se);
4886 entity_tick(cfs_rq, se, queued);
4891 * called on fork with the child task as argument from the parent's context
4892 * - child not yet on the tasklist
4893 * - preemption disabled
4895 static void task_fork_fair(struct task_struct *p)
4897 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4898 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4899 int this_cpu = smp_processor_id();
4900 struct rq *rq = this_rq();
4901 unsigned long flags;
4903 raw_spin_lock_irqsave(&rq->lock, flags);
4905 update_rq_clock(rq);
4908 * Not only the cpu but also the task_group of the parent might have
4909 * been changed after parent->se.parent,cfs_rq were copied to
4910 * child->se.parent,cfs_rq. So call __set_task_cpu() to make those
4911 * of child point to valid ones.
4914 __set_task_cpu(p, this_cpu);
4917 update_curr(cfs_rq);
4920 se->vruntime = curr->vruntime;
4921 place_entity(cfs_rq, se, 1);
4923 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4925 * Upon rescheduling, sched_class::put_prev_task() will place
4926 * 'current' within the tree based on its new key value.
4928 swap(curr->vruntime, se->vruntime);
4929 resched_task(rq->curr);
4932 se->vruntime -= cfs_rq->min_vruntime;
4934 raw_spin_unlock_irqrestore(&rq->lock, flags);
4938 * Priority of the task has changed. Check to see if we preempt
4942 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4948 * Reschedule if we are currently running on this runqueue and
4949 * our priority decreased, or if we are not currently running on
4950 * this runqueue and our priority is higher than the current's
4952 if (rq->curr == p) {
4953 if (p->prio > oldprio)
4954 resched_task(rq->curr);
4956 check_preempt_curr(rq, p, 0);
4959 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4961 struct sched_entity *se = &p->se;
4962 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4965 * Ensure the task's vruntime is normalized, so that when it's
4966 * switched back to the fair class the enqueue_entity(.flags=0) will
4967 * do the right thing.
4969 * If it's on_rq, then the dequeue_entity(.flags=0) will already
4970 * have normalized the vruntime, if it's !on_rq, then only when
4971 * the task is sleeping will it still have non-normalized vruntime.
4973 if (!p->on_rq && p->state != TASK_RUNNING) {
4975 * Fix up our vruntime so that the current sleep doesn't
4976 * cause 'unlimited' sleep bonus.
4978 place_entity(cfs_rq, se, 0);
4979 se->vruntime -= cfs_rq->min_vruntime;
4984 * We switched to the sched_fair class.
4986 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4992 * We were most likely switched from sched_rt, so
4993 * kick off the schedule if running, otherwise just see
4994 * if we can still preempt the current task.
4997 resched_task(rq->curr);
4999 check_preempt_curr(rq, p, 0);
5002 /* Account for a task changing its policy or group.
5004 * This routine is mostly called to set cfs_rq->curr field when a task
5005 * migrates between groups/classes.
5007 static void set_curr_task_fair(struct rq *rq)
5009 struct sched_entity *se = &rq->curr->se;
5011 for_each_sched_entity(se) {
5012 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5014 set_next_entity(cfs_rq, se);
5015 /* ensure bandwidth has been allocated on our new cfs_rq */
5016 account_cfs_rq_runtime(cfs_rq, 0);
5020 #ifdef CONFIG_FAIR_GROUP_SCHED
5021 static void task_move_group_fair(struct task_struct *p, int on_rq)
5024 * If the task was not on the rq at the time of this cgroup movement
5025 * it must have been asleep, sleeping tasks keep their ->vruntime
5026 * absolute on their old rq until wakeup (needed for the fair sleeper
5027 * bonus in place_entity()).
5029 * If it was on the rq, we've just 'preempted' it, which does convert
5030 * ->vruntime to a relative base.
5032 * Make sure both cases convert their relative position when migrating
5033 * to another cgroup's rq. This does somewhat interfere with the
5034 * fair sleeper stuff for the first placement, but who cares.
5037 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5038 set_task_rq(p, task_cpu(p));
5040 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5044 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5046 struct sched_entity *se = &task->se;
5047 unsigned int rr_interval = 0;
5050 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5053 if (rq->cfs.load.weight)
5054 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
5060 * All the scheduling class methods:
5062 static const struct sched_class fair_sched_class = {
5063 .next = &idle_sched_class,
5064 .enqueue_task = enqueue_task_fair,
5065 .dequeue_task = dequeue_task_fair,
5066 .yield_task = yield_task_fair,
5067 .yield_to_task = yield_to_task_fair,
5069 .check_preempt_curr = check_preempt_wakeup,
5071 .pick_next_task = pick_next_task_fair,
5072 .put_prev_task = put_prev_task_fair,
5075 .select_task_rq = select_task_rq_fair,
5077 .rq_online = rq_online_fair,
5078 .rq_offline = rq_offline_fair,
5080 .task_waking = task_waking_fair,
5083 .set_curr_task = set_curr_task_fair,
5084 .task_tick = task_tick_fair,
5085 .task_fork = task_fork_fair,
5087 .prio_changed = prio_changed_fair,
5088 .switched_from = switched_from_fair,
5089 .switched_to = switched_to_fair,
5091 .get_rr_interval = get_rr_interval_fair,
5093 #ifdef CONFIG_FAIR_GROUP_SCHED
5094 .task_move_group = task_move_group_fair,
5098 #ifdef CONFIG_SCHED_DEBUG
5099 static void print_cfs_stats(struct seq_file *m, int cpu)
5101 struct cfs_rq *cfs_rq;
5104 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5105 print_cfs_rq(m, cpu, cfs_rq);