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
710 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
713 struct task_group *tg = cfs_rq->tg;
716 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
717 load_avg -= cfs_rq->load_contribution;
719 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
720 atomic_add(load_avg, &tg->load_weight);
721 cfs_rq->load_contribution += load_avg;
725 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
727 u64 period = sysctl_sched_shares_window;
729 unsigned long load = cfs_rq->load.weight;
731 if (cfs_rq->tg == &root_task_group)
734 now = rq_of(cfs_rq)->clock_task;
735 delta = now - cfs_rq->load_stamp;
737 /* truncate load history at 4 idle periods */
738 if (cfs_rq->load_stamp > cfs_rq->load_last &&
739 now - cfs_rq->load_last > 4 * period) {
740 cfs_rq->load_period = 0;
741 cfs_rq->load_avg = 0;
745 cfs_rq->load_stamp = now;
746 cfs_rq->load_unacc_exec_time = 0;
747 cfs_rq->load_period += delta;
749 cfs_rq->load_last = now;
750 cfs_rq->load_avg += delta * load;
753 /* consider updating load contribution on each fold or truncate */
754 if (global_update || cfs_rq->load_period > period
755 || !cfs_rq->load_period)
756 update_cfs_rq_load_contribution(cfs_rq, global_update);
758 while (cfs_rq->load_period > period) {
760 * Inline assembly required to prevent the compiler
761 * optimising this loop into a divmod call.
762 * See __iter_div_u64_rem() for another example of this.
764 asm("" : "+rm" (cfs_rq->load_period));
765 cfs_rq->load_period /= 2;
766 cfs_rq->load_avg /= 2;
769 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
770 list_del_leaf_cfs_rq(cfs_rq);
773 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
775 long load_weight, load, shares;
777 load = cfs_rq->load.weight;
779 load_weight = atomic_read(&tg->load_weight);
781 load_weight -= cfs_rq->load_contribution;
783 shares = (tg->shares * load);
785 shares /= load_weight;
787 if (shares < MIN_SHARES)
789 if (shares > tg->shares)
795 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
797 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
798 update_cfs_load(cfs_rq, 0);
799 update_cfs_shares(cfs_rq);
802 # else /* CONFIG_SMP */
803 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
807 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
812 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
815 # endif /* CONFIG_SMP */
816 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
817 unsigned long weight)
820 /* commit outstanding execution time */
821 if (cfs_rq->curr == se)
823 account_entity_dequeue(cfs_rq, se);
826 update_load_set(&se->load, weight);
829 account_entity_enqueue(cfs_rq, se);
832 static void update_cfs_shares(struct cfs_rq *cfs_rq)
834 struct task_group *tg;
835 struct sched_entity *se;
839 se = tg->se[cpu_of(rq_of(cfs_rq))];
843 if (likely(se->load.weight == tg->shares))
846 shares = calc_cfs_shares(cfs_rq, tg);
848 reweight_entity(cfs_rq_of(se), se, shares);
850 #else /* CONFIG_FAIR_GROUP_SCHED */
851 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
855 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
859 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
862 #endif /* CONFIG_FAIR_GROUP_SCHED */
864 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
866 #ifdef CONFIG_SCHEDSTATS
867 struct task_struct *tsk = NULL;
869 if (entity_is_task(se))
872 if (se->statistics.sleep_start) {
873 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
878 if (unlikely(delta > se->statistics.sleep_max))
879 se->statistics.sleep_max = delta;
881 se->statistics.sleep_start = 0;
882 se->statistics.sum_sleep_runtime += delta;
885 account_scheduler_latency(tsk, delta >> 10, 1);
886 trace_sched_stat_sleep(tsk, delta);
889 if (se->statistics.block_start) {
890 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
895 if (unlikely(delta > se->statistics.block_max))
896 se->statistics.block_max = delta;
898 se->statistics.block_start = 0;
899 se->statistics.sum_sleep_runtime += delta;
902 if (tsk->in_iowait) {
903 se->statistics.iowait_sum += delta;
904 se->statistics.iowait_count++;
905 trace_sched_stat_iowait(tsk, delta);
909 * Blocking time is in units of nanosecs, so shift by
910 * 20 to get a milliseconds-range estimation of the
911 * amount of time that the task spent sleeping:
913 if (unlikely(prof_on == SLEEP_PROFILING)) {
914 profile_hits(SLEEP_PROFILING,
915 (void *)get_wchan(tsk),
918 account_scheduler_latency(tsk, delta >> 10, 0);
924 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
926 #ifdef CONFIG_SCHED_DEBUG
927 s64 d = se->vruntime - cfs_rq->min_vruntime;
932 if (d > 3*sysctl_sched_latency)
933 schedstat_inc(cfs_rq, nr_spread_over);
938 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
940 u64 vruntime = cfs_rq->min_vruntime;
943 * The 'current' period is already promised to the current tasks,
944 * however the extra weight of the new task will slow them down a
945 * little, place the new task so that it fits in the slot that
946 * stays open at the end.
948 if (initial && sched_feat(START_DEBIT))
949 vruntime += sched_vslice(cfs_rq, se);
951 /* sleeps up to a single latency don't count. */
953 unsigned long thresh = sysctl_sched_latency;
956 * Halve their sleep time's effect, to allow
957 * for a gentler effect of sleepers:
959 if (sched_feat(GENTLE_FAIR_SLEEPERS))
965 /* ensure we never gain time by being placed backwards. */
966 vruntime = max_vruntime(se->vruntime, vruntime);
968 se->vruntime = vruntime;
972 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
975 * Update the normalized vruntime before updating min_vruntime
976 * through callig update_curr().
978 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
979 se->vruntime += cfs_rq->min_vruntime;
982 * Update run-time statistics of the 'current'.
985 update_cfs_load(cfs_rq, 0);
986 account_entity_enqueue(cfs_rq, se);
987 update_cfs_shares(cfs_rq);
989 if (flags & ENQUEUE_WAKEUP) {
990 place_entity(cfs_rq, se, 0);
991 enqueue_sleeper(cfs_rq, se);
994 update_stats_enqueue(cfs_rq, se);
995 check_spread(cfs_rq, se);
996 if (se != cfs_rq->curr)
997 __enqueue_entity(cfs_rq, se);
1000 if (cfs_rq->nr_running == 1)
1001 list_add_leaf_cfs_rq(cfs_rq);
1004 static void __clear_buddies_last(struct sched_entity *se)
1006 for_each_sched_entity(se) {
1007 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1008 if (cfs_rq->last == se)
1009 cfs_rq->last = NULL;
1015 static void __clear_buddies_next(struct sched_entity *se)
1017 for_each_sched_entity(se) {
1018 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1019 if (cfs_rq->next == se)
1020 cfs_rq->next = NULL;
1026 static void __clear_buddies_skip(struct sched_entity *se)
1028 for_each_sched_entity(se) {
1029 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1030 if (cfs_rq->skip == se)
1031 cfs_rq->skip = NULL;
1037 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1039 if (cfs_rq->last == se)
1040 __clear_buddies_last(se);
1042 if (cfs_rq->next == se)
1043 __clear_buddies_next(se);
1045 if (cfs_rq->skip == se)
1046 __clear_buddies_skip(se);
1050 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1053 * Update run-time statistics of the 'current'.
1055 update_curr(cfs_rq);
1057 update_stats_dequeue(cfs_rq, se);
1058 if (flags & DEQUEUE_SLEEP) {
1059 #ifdef CONFIG_SCHEDSTATS
1060 if (entity_is_task(se)) {
1061 struct task_struct *tsk = task_of(se);
1063 if (tsk->state & TASK_INTERRUPTIBLE)
1064 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1065 if (tsk->state & TASK_UNINTERRUPTIBLE)
1066 se->statistics.block_start = rq_of(cfs_rq)->clock;
1071 clear_buddies(cfs_rq, se);
1073 if (se != cfs_rq->curr)
1074 __dequeue_entity(cfs_rq, se);
1076 update_cfs_load(cfs_rq, 0);
1077 account_entity_dequeue(cfs_rq, se);
1080 * Normalize the entity after updating the min_vruntime because the
1081 * update can refer to the ->curr item and we need to reflect this
1082 * movement in our normalized position.
1084 if (!(flags & DEQUEUE_SLEEP))
1085 se->vruntime -= cfs_rq->min_vruntime;
1087 update_min_vruntime(cfs_rq);
1088 update_cfs_shares(cfs_rq);
1092 * Preempt the current task with a newly woken task if needed:
1095 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1097 unsigned long ideal_runtime, delta_exec;
1099 ideal_runtime = sched_slice(cfs_rq, curr);
1100 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1101 if (delta_exec > ideal_runtime) {
1102 resched_task(rq_of(cfs_rq)->curr);
1104 * The current task ran long enough, ensure it doesn't get
1105 * re-elected due to buddy favours.
1107 clear_buddies(cfs_rq, curr);
1112 * Ensure that a task that missed wakeup preemption by a
1113 * narrow margin doesn't have to wait for a full slice.
1114 * This also mitigates buddy induced latencies under load.
1116 if (delta_exec < sysctl_sched_min_granularity)
1119 if (cfs_rq->nr_running > 1) {
1120 struct sched_entity *se = __pick_first_entity(cfs_rq);
1121 s64 delta = curr->vruntime - se->vruntime;
1126 if (delta > ideal_runtime)
1127 resched_task(rq_of(cfs_rq)->curr);
1132 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1134 /* 'current' is not kept within the tree. */
1137 * Any task has to be enqueued before it get to execute on
1138 * a CPU. So account for the time it spent waiting on the
1141 update_stats_wait_end(cfs_rq, se);
1142 __dequeue_entity(cfs_rq, se);
1145 update_stats_curr_start(cfs_rq, se);
1147 #ifdef CONFIG_SCHEDSTATS
1149 * Track our maximum slice length, if the CPU's load is at
1150 * least twice that of our own weight (i.e. dont track it
1151 * when there are only lesser-weight tasks around):
1153 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1154 se->statistics.slice_max = max(se->statistics.slice_max,
1155 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1158 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1162 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1165 * Pick the next process, keeping these things in mind, in this order:
1166 * 1) keep things fair between processes/task groups
1167 * 2) pick the "next" process, since someone really wants that to run
1168 * 3) pick the "last" process, for cache locality
1169 * 4) do not run the "skip" process, if something else is available
1171 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1173 struct sched_entity *se = __pick_first_entity(cfs_rq);
1174 struct sched_entity *left = se;
1177 * Avoid running the skip buddy, if running something else can
1178 * be done without getting too unfair.
1180 if (cfs_rq->skip == se) {
1181 struct sched_entity *second = __pick_next_entity(se);
1182 if (second && wakeup_preempt_entity(second, left) < 1)
1187 * Prefer last buddy, try to return the CPU to a preempted task.
1189 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1193 * Someone really wants this to run. If it's not unfair, run it.
1195 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1198 clear_buddies(cfs_rq, se);
1203 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1206 * If still on the runqueue then deactivate_task()
1207 * was not called and update_curr() has to be done:
1210 update_curr(cfs_rq);
1212 check_spread(cfs_rq, prev);
1214 update_stats_wait_start(cfs_rq, prev);
1215 /* Put 'current' back into the tree. */
1216 __enqueue_entity(cfs_rq, prev);
1218 cfs_rq->curr = NULL;
1222 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1225 * Update run-time statistics of the 'current'.
1227 update_curr(cfs_rq);
1230 * Update share accounting for long-running entities.
1232 update_entity_shares_tick(cfs_rq);
1234 #ifdef CONFIG_SCHED_HRTICK
1236 * queued ticks are scheduled to match the slice, so don't bother
1237 * validating it and just reschedule.
1240 resched_task(rq_of(cfs_rq)->curr);
1244 * don't let the period tick interfere with the hrtick preemption
1246 if (!sched_feat(DOUBLE_TICK) &&
1247 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1251 if (cfs_rq->nr_running > 1)
1252 check_preempt_tick(cfs_rq, curr);
1256 /**************************************************
1257 * CFS bandwidth control machinery
1260 #ifdef CONFIG_CFS_BANDWIDTH
1262 * default period for cfs group bandwidth.
1263 * default: 0.1s, units: nanoseconds
1265 static inline u64 default_cfs_period(void)
1267 return 100000000ULL;
1270 static inline u64 sched_cfs_bandwidth_slice(void)
1272 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1276 * Replenish runtime according to assigned quota and update expiration time.
1277 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1278 * additional synchronization around rq->lock.
1280 * requires cfs_b->lock
1282 static void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1286 if (cfs_b->quota == RUNTIME_INF)
1289 now = sched_clock_cpu(smp_processor_id());
1290 cfs_b->runtime = cfs_b->quota;
1291 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1294 /* returns 0 on failure to allocate runtime */
1295 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1297 struct task_group *tg = cfs_rq->tg;
1298 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1299 u64 amount = 0, min_amount, expires;
1301 /* note: this is a positive sum as runtime_remaining <= 0 */
1302 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1304 raw_spin_lock(&cfs_b->lock);
1305 if (cfs_b->quota == RUNTIME_INF)
1306 amount = min_amount;
1309 * If the bandwidth pool has become inactive, then at least one
1310 * period must have elapsed since the last consumption.
1311 * Refresh the global state and ensure bandwidth timer becomes
1314 if (!cfs_b->timer_active) {
1315 __refill_cfs_bandwidth_runtime(cfs_b);
1316 __start_cfs_bandwidth(cfs_b);
1319 if (cfs_b->runtime > 0) {
1320 amount = min(cfs_b->runtime, min_amount);
1321 cfs_b->runtime -= amount;
1325 expires = cfs_b->runtime_expires;
1326 raw_spin_unlock(&cfs_b->lock);
1328 cfs_rq->runtime_remaining += amount;
1330 * we may have advanced our local expiration to account for allowed
1331 * spread between our sched_clock and the one on which runtime was
1334 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1335 cfs_rq->runtime_expires = expires;
1337 return cfs_rq->runtime_remaining > 0;
1341 * Note: This depends on the synchronization provided by sched_clock and the
1342 * fact that rq->clock snapshots this value.
1344 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1346 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1347 struct rq *rq = rq_of(cfs_rq);
1349 /* if the deadline is ahead of our clock, nothing to do */
1350 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1353 if (cfs_rq->runtime_remaining < 0)
1357 * If the local deadline has passed we have to consider the
1358 * possibility that our sched_clock is 'fast' and the global deadline
1359 * has not truly expired.
1361 * Fortunately we can check determine whether this the case by checking
1362 * whether the global deadline has advanced.
1365 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1366 /* extend local deadline, drift is bounded above by 2 ticks */
1367 cfs_rq->runtime_expires += TICK_NSEC;
1369 /* global deadline is ahead, expiration has passed */
1370 cfs_rq->runtime_remaining = 0;
1374 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1375 unsigned long delta_exec)
1377 /* dock delta_exec before expiring quota (as it could span periods) */
1378 cfs_rq->runtime_remaining -= delta_exec;
1379 expire_cfs_rq_runtime(cfs_rq);
1381 if (likely(cfs_rq->runtime_remaining > 0))
1385 * if we're unable to extend our runtime we resched so that the active
1386 * hierarchy can be throttled
1388 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1389 resched_task(rq_of(cfs_rq)->curr);
1392 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1393 unsigned long delta_exec)
1395 if (!cfs_rq->runtime_enabled)
1398 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1401 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1403 return cfs_rq->throttled;
1406 static __used void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1408 struct rq *rq = rq_of(cfs_rq);
1409 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1410 struct sched_entity *se;
1411 long task_delta, dequeue = 1;
1413 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1415 /* account load preceding throttle */
1416 update_cfs_load(cfs_rq, 0);
1418 task_delta = cfs_rq->h_nr_running;
1419 for_each_sched_entity(se) {
1420 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1421 /* throttled entity or throttle-on-deactivate */
1426 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1427 qcfs_rq->h_nr_running -= task_delta;
1429 if (qcfs_rq->load.weight)
1434 rq->nr_running -= task_delta;
1436 cfs_rq->throttled = 1;
1437 raw_spin_lock(&cfs_b->lock);
1438 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1439 raw_spin_unlock(&cfs_b->lock);
1443 * Responsible for refilling a task_group's bandwidth and unthrottling its
1444 * cfs_rqs as appropriate. If there has been no activity within the last
1445 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1446 * used to track this state.
1448 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1452 raw_spin_lock(&cfs_b->lock);
1453 /* no need to continue the timer with no bandwidth constraint */
1454 if (cfs_b->quota == RUNTIME_INF)
1458 /* if we're going inactive then everything else can be deferred */
1462 __refill_cfs_bandwidth_runtime(cfs_b);
1465 /* mark as potentially idle for the upcoming period */
1469 cfs_b->timer_active = 0;
1470 raw_spin_unlock(&cfs_b->lock);
1475 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1476 unsigned long delta_exec) {}
1478 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1484 /**************************************************
1485 * CFS operations on tasks:
1488 #ifdef CONFIG_SCHED_HRTICK
1489 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1491 struct sched_entity *se = &p->se;
1492 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1494 WARN_ON(task_rq(p) != rq);
1496 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1497 u64 slice = sched_slice(cfs_rq, se);
1498 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1499 s64 delta = slice - ran;
1508 * Don't schedule slices shorter than 10000ns, that just
1509 * doesn't make sense. Rely on vruntime for fairness.
1512 delta = max_t(s64, 10000LL, delta);
1514 hrtick_start(rq, delta);
1519 * called from enqueue/dequeue and updates the hrtick when the
1520 * current task is from our class and nr_running is low enough
1523 static void hrtick_update(struct rq *rq)
1525 struct task_struct *curr = rq->curr;
1527 if (curr->sched_class != &fair_sched_class)
1530 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1531 hrtick_start_fair(rq, curr);
1533 #else /* !CONFIG_SCHED_HRTICK */
1535 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1539 static inline void hrtick_update(struct rq *rq)
1545 * The enqueue_task method is called before nr_running is
1546 * increased. Here we update the fair scheduling stats and
1547 * then put the task into the rbtree:
1550 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1552 struct cfs_rq *cfs_rq;
1553 struct sched_entity *se = &p->se;
1555 for_each_sched_entity(se) {
1558 cfs_rq = cfs_rq_of(se);
1559 enqueue_entity(cfs_rq, se, flags);
1562 * end evaluation on encountering a throttled cfs_rq
1564 * note: in the case of encountering a throttled cfs_rq we will
1565 * post the final h_nr_running increment below.
1567 if (cfs_rq_throttled(cfs_rq))
1569 cfs_rq->h_nr_running++;
1571 flags = ENQUEUE_WAKEUP;
1574 for_each_sched_entity(se) {
1575 cfs_rq = cfs_rq_of(se);
1576 cfs_rq->h_nr_running++;
1578 if (cfs_rq_throttled(cfs_rq))
1581 update_cfs_load(cfs_rq, 0);
1582 update_cfs_shares(cfs_rq);
1590 static void set_next_buddy(struct sched_entity *se);
1593 * The dequeue_task method is called before nr_running is
1594 * decreased. We remove the task from the rbtree and
1595 * update the fair scheduling stats:
1597 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1599 struct cfs_rq *cfs_rq;
1600 struct sched_entity *se = &p->se;
1601 int task_sleep = flags & DEQUEUE_SLEEP;
1603 for_each_sched_entity(se) {
1604 cfs_rq = cfs_rq_of(se);
1605 dequeue_entity(cfs_rq, se, flags);
1608 * end evaluation on encountering a throttled cfs_rq
1610 * note: in the case of encountering a throttled cfs_rq we will
1611 * post the final h_nr_running decrement below.
1613 if (cfs_rq_throttled(cfs_rq))
1615 cfs_rq->h_nr_running--;
1617 /* Don't dequeue parent if it has other entities besides us */
1618 if (cfs_rq->load.weight) {
1620 * Bias pick_next to pick a task from this cfs_rq, as
1621 * p is sleeping when it is within its sched_slice.
1623 if (task_sleep && parent_entity(se))
1624 set_next_buddy(parent_entity(se));
1626 /* avoid re-evaluating load for this entity */
1627 se = parent_entity(se);
1630 flags |= DEQUEUE_SLEEP;
1633 for_each_sched_entity(se) {
1634 cfs_rq = cfs_rq_of(se);
1635 cfs_rq->h_nr_running--;
1637 if (cfs_rq_throttled(cfs_rq))
1640 update_cfs_load(cfs_rq, 0);
1641 update_cfs_shares(cfs_rq);
1651 static void task_waking_fair(struct task_struct *p)
1653 struct sched_entity *se = &p->se;
1654 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1657 #ifndef CONFIG_64BIT
1658 u64 min_vruntime_copy;
1661 min_vruntime_copy = cfs_rq->min_vruntime_copy;
1663 min_vruntime = cfs_rq->min_vruntime;
1664 } while (min_vruntime != min_vruntime_copy);
1666 min_vruntime = cfs_rq->min_vruntime;
1669 se->vruntime -= min_vruntime;
1672 #ifdef CONFIG_FAIR_GROUP_SCHED
1674 * effective_load() calculates the load change as seen from the root_task_group
1676 * Adding load to a group doesn't make a group heavier, but can cause movement
1677 * of group shares between cpus. Assuming the shares were perfectly aligned one
1678 * can calculate the shift in shares.
1680 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1682 struct sched_entity *se = tg->se[cpu];
1687 for_each_sched_entity(se) {
1691 w = se->my_q->load.weight;
1693 /* use this cpu's instantaneous contribution */
1694 lw = atomic_read(&tg->load_weight);
1695 lw -= se->my_q->load_contribution;
1700 if (lw > 0 && wl < lw)
1701 wl = (wl * tg->shares) / lw;
1705 /* zero point is MIN_SHARES */
1706 if (wl < MIN_SHARES)
1708 wl -= se->load.weight;
1716 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1717 unsigned long wl, unsigned long wg)
1724 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1726 s64 this_load, load;
1727 int idx, this_cpu, prev_cpu;
1728 unsigned long tl_per_task;
1729 struct task_group *tg;
1730 unsigned long weight;
1734 this_cpu = smp_processor_id();
1735 prev_cpu = task_cpu(p);
1736 load = source_load(prev_cpu, idx);
1737 this_load = target_load(this_cpu, idx);
1740 * If sync wakeup then subtract the (maximum possible)
1741 * effect of the currently running task from the load
1742 * of the current CPU:
1745 tg = task_group(current);
1746 weight = current->se.load.weight;
1748 this_load += effective_load(tg, this_cpu, -weight, -weight);
1749 load += effective_load(tg, prev_cpu, 0, -weight);
1753 weight = p->se.load.weight;
1756 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1757 * due to the sync cause above having dropped this_load to 0, we'll
1758 * always have an imbalance, but there's really nothing you can do
1759 * about that, so that's good too.
1761 * Otherwise check if either cpus are near enough in load to allow this
1762 * task to be woken on this_cpu.
1764 if (this_load > 0) {
1765 s64 this_eff_load, prev_eff_load;
1767 this_eff_load = 100;
1768 this_eff_load *= power_of(prev_cpu);
1769 this_eff_load *= this_load +
1770 effective_load(tg, this_cpu, weight, weight);
1772 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1773 prev_eff_load *= power_of(this_cpu);
1774 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1776 balanced = this_eff_load <= prev_eff_load;
1781 * If the currently running task will sleep within
1782 * a reasonable amount of time then attract this newly
1785 if (sync && balanced)
1788 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1789 tl_per_task = cpu_avg_load_per_task(this_cpu);
1792 (this_load <= load &&
1793 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1795 * This domain has SD_WAKE_AFFINE and
1796 * p is cache cold in this domain, and
1797 * there is no bad imbalance.
1799 schedstat_inc(sd, ttwu_move_affine);
1800 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1808 * find_idlest_group finds and returns the least busy CPU group within the
1811 static struct sched_group *
1812 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1813 int this_cpu, int load_idx)
1815 struct sched_group *idlest = NULL, *group = sd->groups;
1816 unsigned long min_load = ULONG_MAX, this_load = 0;
1817 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1820 unsigned long load, avg_load;
1824 /* Skip over this group if it has no CPUs allowed */
1825 if (!cpumask_intersects(sched_group_cpus(group),
1829 local_group = cpumask_test_cpu(this_cpu,
1830 sched_group_cpus(group));
1832 /* Tally up the load of all CPUs in the group */
1835 for_each_cpu(i, sched_group_cpus(group)) {
1836 /* Bias balancing toward cpus of our domain */
1838 load = source_load(i, load_idx);
1840 load = target_load(i, load_idx);
1845 /* Adjust by relative CPU power of the group */
1846 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
1849 this_load = avg_load;
1850 } else if (avg_load < min_load) {
1851 min_load = avg_load;
1854 } while (group = group->next, group != sd->groups);
1856 if (!idlest || 100*this_load < imbalance*min_load)
1862 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1865 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1867 unsigned long load, min_load = ULONG_MAX;
1871 /* Traverse only the allowed CPUs */
1872 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1873 load = weighted_cpuload(i);
1875 if (load < min_load || (load == min_load && i == this_cpu)) {
1885 * Try and locate an idle CPU in the sched_domain.
1887 static int select_idle_sibling(struct task_struct *p, int target)
1889 int cpu = smp_processor_id();
1890 int prev_cpu = task_cpu(p);
1891 struct sched_domain *sd;
1895 * If the task is going to be woken-up on this cpu and if it is
1896 * already idle, then it is the right target.
1898 if (target == cpu && idle_cpu(cpu))
1902 * If the task is going to be woken-up on the cpu where it previously
1903 * ran and if it is currently idle, then it the right target.
1905 if (target == prev_cpu && idle_cpu(prev_cpu))
1909 * Otherwise, iterate the domains and find an elegible idle cpu.
1912 for_each_domain(target, sd) {
1913 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1916 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1924 * Lets stop looking for an idle sibling when we reached
1925 * the domain that spans the current cpu and prev_cpu.
1927 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1928 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1937 * sched_balance_self: balance the current task (running on cpu) in domains
1938 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1941 * Balance, ie. select the least loaded group.
1943 * Returns the target CPU number, or the same CPU if no balancing is needed.
1945 * preempt must be disabled.
1948 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1950 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1951 int cpu = smp_processor_id();
1952 int prev_cpu = task_cpu(p);
1954 int want_affine = 0;
1956 int sync = wake_flags & WF_SYNC;
1958 if (sd_flag & SD_BALANCE_WAKE) {
1959 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1965 for_each_domain(cpu, tmp) {
1966 if (!(tmp->flags & SD_LOAD_BALANCE))
1970 * If power savings logic is enabled for a domain, see if we
1971 * are not overloaded, if so, don't balance wider.
1973 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1974 unsigned long power = 0;
1975 unsigned long nr_running = 0;
1976 unsigned long capacity;
1979 for_each_cpu(i, sched_domain_span(tmp)) {
1980 power += power_of(i);
1981 nr_running += cpu_rq(i)->cfs.nr_running;
1984 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
1986 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1989 if (nr_running < capacity)
1994 * If both cpu and prev_cpu are part of this domain,
1995 * cpu is a valid SD_WAKE_AFFINE target.
1997 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1998 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2003 if (!want_sd && !want_affine)
2006 if (!(tmp->flags & sd_flag))
2014 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2017 new_cpu = select_idle_sibling(p, prev_cpu);
2022 int load_idx = sd->forkexec_idx;
2023 struct sched_group *group;
2026 if (!(sd->flags & sd_flag)) {
2031 if (sd_flag & SD_BALANCE_WAKE)
2032 load_idx = sd->wake_idx;
2034 group = find_idlest_group(sd, p, cpu, load_idx);
2040 new_cpu = find_idlest_cpu(group, p, cpu);
2041 if (new_cpu == -1 || new_cpu == cpu) {
2042 /* Now try balancing at a lower domain level of cpu */
2047 /* Now try balancing at a lower domain level of new_cpu */
2049 weight = sd->span_weight;
2051 for_each_domain(cpu, tmp) {
2052 if (weight <= tmp->span_weight)
2054 if (tmp->flags & sd_flag)
2057 /* while loop will break here if sd == NULL */
2064 #endif /* CONFIG_SMP */
2066 static unsigned long
2067 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2069 unsigned long gran = sysctl_sched_wakeup_granularity;
2072 * Since its curr running now, convert the gran from real-time
2073 * to virtual-time in his units.
2075 * By using 'se' instead of 'curr' we penalize light tasks, so
2076 * they get preempted easier. That is, if 'se' < 'curr' then
2077 * the resulting gran will be larger, therefore penalizing the
2078 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2079 * be smaller, again penalizing the lighter task.
2081 * This is especially important for buddies when the leftmost
2082 * task is higher priority than the buddy.
2084 return calc_delta_fair(gran, se);
2088 * Should 'se' preempt 'curr'.
2102 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2104 s64 gran, vdiff = curr->vruntime - se->vruntime;
2109 gran = wakeup_gran(curr, se);
2116 static void set_last_buddy(struct sched_entity *se)
2118 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2121 for_each_sched_entity(se)
2122 cfs_rq_of(se)->last = se;
2125 static void set_next_buddy(struct sched_entity *se)
2127 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2130 for_each_sched_entity(se)
2131 cfs_rq_of(se)->next = se;
2134 static void set_skip_buddy(struct sched_entity *se)
2136 for_each_sched_entity(se)
2137 cfs_rq_of(se)->skip = se;
2141 * Preempt the current task with a newly woken task if needed:
2143 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2145 struct task_struct *curr = rq->curr;
2146 struct sched_entity *se = &curr->se, *pse = &p->se;
2147 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2148 int scale = cfs_rq->nr_running >= sched_nr_latency;
2149 int next_buddy_marked = 0;
2151 if (unlikely(se == pse))
2154 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2155 set_next_buddy(pse);
2156 next_buddy_marked = 1;
2160 * We can come here with TIF_NEED_RESCHED already set from new task
2163 if (test_tsk_need_resched(curr))
2166 /* Idle tasks are by definition preempted by non-idle tasks. */
2167 if (unlikely(curr->policy == SCHED_IDLE) &&
2168 likely(p->policy != SCHED_IDLE))
2172 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2173 * is driven by the tick):
2175 if (unlikely(p->policy != SCHED_NORMAL))
2178 find_matching_se(&se, &pse);
2179 update_curr(cfs_rq_of(se));
2181 if (wakeup_preempt_entity(se, pse) == 1) {
2183 * Bias pick_next to pick the sched entity that is
2184 * triggering this preemption.
2186 if (!next_buddy_marked)
2187 set_next_buddy(pse);
2196 * Only set the backward buddy when the current task is still
2197 * on the rq. This can happen when a wakeup gets interleaved
2198 * with schedule on the ->pre_schedule() or idle_balance()
2199 * point, either of which can * drop the rq lock.
2201 * Also, during early boot the idle thread is in the fair class,
2202 * for obvious reasons its a bad idea to schedule back to it.
2204 if (unlikely(!se->on_rq || curr == rq->idle))
2207 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2211 static struct task_struct *pick_next_task_fair(struct rq *rq)
2213 struct task_struct *p;
2214 struct cfs_rq *cfs_rq = &rq->cfs;
2215 struct sched_entity *se;
2217 if (!cfs_rq->nr_running)
2221 se = pick_next_entity(cfs_rq);
2222 set_next_entity(cfs_rq, se);
2223 cfs_rq = group_cfs_rq(se);
2227 hrtick_start_fair(rq, p);
2233 * Account for a descheduled task:
2235 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
2237 struct sched_entity *se = &prev->se;
2238 struct cfs_rq *cfs_rq;
2240 for_each_sched_entity(se) {
2241 cfs_rq = cfs_rq_of(se);
2242 put_prev_entity(cfs_rq, se);
2247 * sched_yield() is very simple
2249 * The magic of dealing with the ->skip buddy is in pick_next_entity.
2251 static void yield_task_fair(struct rq *rq)
2253 struct task_struct *curr = rq->curr;
2254 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2255 struct sched_entity *se = &curr->se;
2258 * Are we the only task in the tree?
2260 if (unlikely(rq->nr_running == 1))
2263 clear_buddies(cfs_rq, se);
2265 if (curr->policy != SCHED_BATCH) {
2266 update_rq_clock(rq);
2268 * Update run-time statistics of the 'current'.
2270 update_curr(cfs_rq);
2276 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2278 struct sched_entity *se = &p->se;
2283 /* Tell the scheduler that we'd really like pse to run next. */
2286 yield_task_fair(rq);
2292 /**************************************************
2293 * Fair scheduling class load-balancing methods:
2297 * pull_task - move a task from a remote runqueue to the local runqueue.
2298 * Both runqueues must be locked.
2300 static void pull_task(struct rq *src_rq, struct task_struct *p,
2301 struct rq *this_rq, int this_cpu)
2303 deactivate_task(src_rq, p, 0);
2304 set_task_cpu(p, this_cpu);
2305 activate_task(this_rq, p, 0);
2306 check_preempt_curr(this_rq, p, 0);
2310 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2313 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2314 struct sched_domain *sd, enum cpu_idle_type idle,
2317 int tsk_cache_hot = 0;
2319 * We do not migrate tasks that are:
2320 * 1) running (obviously), or
2321 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2322 * 3) are cache-hot on their current CPU.
2324 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2325 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2330 if (task_running(rq, p)) {
2331 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2336 * Aggressive migration if:
2337 * 1) task is cache cold, or
2338 * 2) too many balance attempts have failed.
2341 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2342 if (!tsk_cache_hot ||
2343 sd->nr_balance_failed > sd->cache_nice_tries) {
2344 #ifdef CONFIG_SCHEDSTATS
2345 if (tsk_cache_hot) {
2346 schedstat_inc(sd, lb_hot_gained[idle]);
2347 schedstat_inc(p, se.statistics.nr_forced_migrations);
2353 if (tsk_cache_hot) {
2354 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2361 * move_one_task tries to move exactly one task from busiest to this_rq, as
2362 * part of active balancing operations within "domain".
2363 * Returns 1 if successful and 0 otherwise.
2365 * Called with both runqueues locked.
2368 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2369 struct sched_domain *sd, enum cpu_idle_type idle)
2371 struct task_struct *p, *n;
2372 struct cfs_rq *cfs_rq;
2375 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2376 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2378 if (!can_migrate_task(p, busiest, this_cpu,
2382 pull_task(busiest, p, this_rq, this_cpu);
2384 * Right now, this is only the second place pull_task()
2385 * is called, so we can safely collect pull_task()
2386 * stats here rather than inside pull_task().
2388 schedstat_inc(sd, lb_gained[idle]);
2396 static unsigned long
2397 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2398 unsigned long max_load_move, struct sched_domain *sd,
2399 enum cpu_idle_type idle, int *all_pinned,
2400 struct cfs_rq *busiest_cfs_rq)
2402 int loops = 0, pulled = 0;
2403 long rem_load_move = max_load_move;
2404 struct task_struct *p, *n;
2406 if (max_load_move == 0)
2409 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2410 if (loops++ > sysctl_sched_nr_migrate)
2413 if ((p->se.load.weight >> 1) > rem_load_move ||
2414 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2418 pull_task(busiest, p, this_rq, this_cpu);
2420 rem_load_move -= p->se.load.weight;
2422 #ifdef CONFIG_PREEMPT
2424 * NEWIDLE balancing is a source of latency, so preemptible
2425 * kernels will stop after the first task is pulled to minimize
2426 * the critical section.
2428 if (idle == CPU_NEWLY_IDLE)
2433 * We only want to steal up to the prescribed amount of
2436 if (rem_load_move <= 0)
2441 * Right now, this is one of only two places pull_task() is called,
2442 * so we can safely collect pull_task() stats here rather than
2443 * inside pull_task().
2445 schedstat_add(sd, lb_gained[idle], pulled);
2447 return max_load_move - rem_load_move;
2450 #ifdef CONFIG_FAIR_GROUP_SCHED
2452 * update tg->load_weight by folding this cpu's load_avg
2454 static int update_shares_cpu(struct task_group *tg, int cpu)
2456 struct cfs_rq *cfs_rq;
2457 unsigned long flags;
2464 cfs_rq = tg->cfs_rq[cpu];
2466 raw_spin_lock_irqsave(&rq->lock, flags);
2468 update_rq_clock(rq);
2469 update_cfs_load(cfs_rq, 1);
2472 * We need to update shares after updating tg->load_weight in
2473 * order to adjust the weight of groups with long running tasks.
2475 update_cfs_shares(cfs_rq);
2477 raw_spin_unlock_irqrestore(&rq->lock, flags);
2482 static void update_shares(int cpu)
2484 struct cfs_rq *cfs_rq;
2485 struct rq *rq = cpu_rq(cpu);
2489 * Iterates the task_group tree in a bottom up fashion, see
2490 * list_add_leaf_cfs_rq() for details.
2492 for_each_leaf_cfs_rq(rq, cfs_rq)
2493 update_shares_cpu(cfs_rq->tg, cpu);
2498 * Compute the cpu's hierarchical load factor for each task group.
2499 * This needs to be done in a top-down fashion because the load of a child
2500 * group is a fraction of its parents load.
2502 static int tg_load_down(struct task_group *tg, void *data)
2505 long cpu = (long)data;
2508 load = cpu_rq(cpu)->load.weight;
2510 load = tg->parent->cfs_rq[cpu]->h_load;
2511 load *= tg->se[cpu]->load.weight;
2512 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2515 tg->cfs_rq[cpu]->h_load = load;
2520 static void update_h_load(long cpu)
2522 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2525 static unsigned long
2526 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2527 unsigned long max_load_move,
2528 struct sched_domain *sd, enum cpu_idle_type idle,
2531 long rem_load_move = max_load_move;
2532 struct cfs_rq *busiest_cfs_rq;
2535 update_h_load(cpu_of(busiest));
2537 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
2538 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2539 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2540 u64 rem_load, moved_load;
2545 if (!busiest_cfs_rq->task_weight)
2548 rem_load = (u64)rem_load_move * busiest_weight;
2549 rem_load = div_u64(rem_load, busiest_h_load + 1);
2551 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2552 rem_load, sd, idle, all_pinned,
2558 moved_load *= busiest_h_load;
2559 moved_load = div_u64(moved_load, busiest_weight + 1);
2561 rem_load_move -= moved_load;
2562 if (rem_load_move < 0)
2567 return max_load_move - rem_load_move;
2570 static inline void update_shares(int cpu)
2574 static unsigned long
2575 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2576 unsigned long max_load_move,
2577 struct sched_domain *sd, enum cpu_idle_type idle,
2580 return balance_tasks(this_rq, this_cpu, busiest,
2581 max_load_move, sd, idle, all_pinned,
2587 * move_tasks tries to move up to max_load_move weighted load from busiest to
2588 * this_rq, as part of a balancing operation within domain "sd".
2589 * Returns 1 if successful and 0 otherwise.
2591 * Called with both runqueues locked.
2593 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2594 unsigned long max_load_move,
2595 struct sched_domain *sd, enum cpu_idle_type idle,
2598 unsigned long total_load_moved = 0, load_moved;
2601 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2602 max_load_move - total_load_moved,
2603 sd, idle, all_pinned);
2605 total_load_moved += load_moved;
2607 #ifdef CONFIG_PREEMPT
2609 * NEWIDLE balancing is a source of latency, so preemptible
2610 * kernels will stop after the first task is pulled to minimize
2611 * the critical section.
2613 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2616 if (raw_spin_is_contended(&this_rq->lock) ||
2617 raw_spin_is_contended(&busiest->lock))
2620 } while (load_moved && max_load_move > total_load_moved);
2622 return total_load_moved > 0;
2625 /********** Helpers for find_busiest_group ************************/
2627 * sd_lb_stats - Structure to store the statistics of a sched_domain
2628 * during load balancing.
2630 struct sd_lb_stats {
2631 struct sched_group *busiest; /* Busiest group in this sd */
2632 struct sched_group *this; /* Local group in this sd */
2633 unsigned long total_load; /* Total load of all groups in sd */
2634 unsigned long total_pwr; /* Total power of all groups in sd */
2635 unsigned long avg_load; /* Average load across all groups in sd */
2637 /** Statistics of this group */
2638 unsigned long this_load;
2639 unsigned long this_load_per_task;
2640 unsigned long this_nr_running;
2641 unsigned long this_has_capacity;
2642 unsigned int this_idle_cpus;
2644 /* Statistics of the busiest group */
2645 unsigned int busiest_idle_cpus;
2646 unsigned long max_load;
2647 unsigned long busiest_load_per_task;
2648 unsigned long busiest_nr_running;
2649 unsigned long busiest_group_capacity;
2650 unsigned long busiest_has_capacity;
2651 unsigned int busiest_group_weight;
2653 int group_imb; /* Is there imbalance in this sd */
2654 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2655 int power_savings_balance; /* Is powersave balance needed for this sd */
2656 struct sched_group *group_min; /* Least loaded group in sd */
2657 struct sched_group *group_leader; /* Group which relieves group_min */
2658 unsigned long min_load_per_task; /* load_per_task in group_min */
2659 unsigned long leader_nr_running; /* Nr running of group_leader */
2660 unsigned long min_nr_running; /* Nr running of group_min */
2665 * sg_lb_stats - stats of a sched_group required for load_balancing
2667 struct sg_lb_stats {
2668 unsigned long avg_load; /*Avg load across the CPUs of the group */
2669 unsigned long group_load; /* Total load over the CPUs of the group */
2670 unsigned long sum_nr_running; /* Nr tasks running in the group */
2671 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2672 unsigned long group_capacity;
2673 unsigned long idle_cpus;
2674 unsigned long group_weight;
2675 int group_imb; /* Is there an imbalance in the group ? */
2676 int group_has_capacity; /* Is there extra capacity in the group? */
2680 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2681 * @group: The group whose first cpu is to be returned.
2683 static inline unsigned int group_first_cpu(struct sched_group *group)
2685 return cpumask_first(sched_group_cpus(group));
2689 * get_sd_load_idx - Obtain the load index for a given sched domain.
2690 * @sd: The sched_domain whose load_idx is to be obtained.
2691 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2693 static inline int get_sd_load_idx(struct sched_domain *sd,
2694 enum cpu_idle_type idle)
2700 load_idx = sd->busy_idx;
2703 case CPU_NEWLY_IDLE:
2704 load_idx = sd->newidle_idx;
2707 load_idx = sd->idle_idx;
2715 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2717 * init_sd_power_savings_stats - Initialize power savings statistics for
2718 * the given sched_domain, during load balancing.
2720 * @sd: Sched domain whose power-savings statistics are to be initialized.
2721 * @sds: Variable containing the statistics for sd.
2722 * @idle: Idle status of the CPU at which we're performing load-balancing.
2724 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2725 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2728 * Busy processors will not participate in power savings
2731 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2732 sds->power_savings_balance = 0;
2734 sds->power_savings_balance = 1;
2735 sds->min_nr_running = ULONG_MAX;
2736 sds->leader_nr_running = 0;
2741 * update_sd_power_savings_stats - Update the power saving stats for a
2742 * sched_domain while performing load balancing.
2744 * @group: sched_group belonging to the sched_domain under consideration.
2745 * @sds: Variable containing the statistics of the sched_domain
2746 * @local_group: Does group contain the CPU for which we're performing
2748 * @sgs: Variable containing the statistics of the group.
2750 static inline void update_sd_power_savings_stats(struct sched_group *group,
2751 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2754 if (!sds->power_savings_balance)
2758 * If the local group is idle or completely loaded
2759 * no need to do power savings balance at this domain
2761 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2762 !sds->this_nr_running))
2763 sds->power_savings_balance = 0;
2766 * If a group is already running at full capacity or idle,
2767 * don't include that group in power savings calculations
2769 if (!sds->power_savings_balance ||
2770 sgs->sum_nr_running >= sgs->group_capacity ||
2771 !sgs->sum_nr_running)
2775 * Calculate the group which has the least non-idle load.
2776 * This is the group from where we need to pick up the load
2779 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2780 (sgs->sum_nr_running == sds->min_nr_running &&
2781 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2782 sds->group_min = group;
2783 sds->min_nr_running = sgs->sum_nr_running;
2784 sds->min_load_per_task = sgs->sum_weighted_load /
2785 sgs->sum_nr_running;
2789 * Calculate the group which is almost near its
2790 * capacity but still has some space to pick up some load
2791 * from other group and save more power
2793 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2796 if (sgs->sum_nr_running > sds->leader_nr_running ||
2797 (sgs->sum_nr_running == sds->leader_nr_running &&
2798 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2799 sds->group_leader = group;
2800 sds->leader_nr_running = sgs->sum_nr_running;
2805 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2806 * @sds: Variable containing the statistics of the sched_domain
2807 * under consideration.
2808 * @this_cpu: Cpu at which we're currently performing load-balancing.
2809 * @imbalance: Variable to store the imbalance.
2812 * Check if we have potential to perform some power-savings balance.
2813 * If yes, set the busiest group to be the least loaded group in the
2814 * sched_domain, so that it's CPUs can be put to idle.
2816 * Returns 1 if there is potential to perform power-savings balance.
2819 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2820 int this_cpu, unsigned long *imbalance)
2822 if (!sds->power_savings_balance)
2825 if (sds->this != sds->group_leader ||
2826 sds->group_leader == sds->group_min)
2829 *imbalance = sds->min_load_per_task;
2830 sds->busiest = sds->group_min;
2835 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2836 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2837 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2842 static inline void update_sd_power_savings_stats(struct sched_group *group,
2843 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2848 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2849 int this_cpu, unsigned long *imbalance)
2853 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2856 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2858 return SCHED_POWER_SCALE;
2861 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2863 return default_scale_freq_power(sd, cpu);
2866 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2868 unsigned long weight = sd->span_weight;
2869 unsigned long smt_gain = sd->smt_gain;
2876 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2878 return default_scale_smt_power(sd, cpu);
2881 unsigned long scale_rt_power(int cpu)
2883 struct rq *rq = cpu_rq(cpu);
2884 u64 total, available;
2886 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2888 if (unlikely(total < rq->rt_avg)) {
2889 /* Ensures that power won't end up being negative */
2892 available = total - rq->rt_avg;
2895 if (unlikely((s64)total < SCHED_POWER_SCALE))
2896 total = SCHED_POWER_SCALE;
2898 total >>= SCHED_POWER_SHIFT;
2900 return div_u64(available, total);
2903 static void update_cpu_power(struct sched_domain *sd, int cpu)
2905 unsigned long weight = sd->span_weight;
2906 unsigned long power = SCHED_POWER_SCALE;
2907 struct sched_group *sdg = sd->groups;
2909 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2910 if (sched_feat(ARCH_POWER))
2911 power *= arch_scale_smt_power(sd, cpu);
2913 power *= default_scale_smt_power(sd, cpu);
2915 power >>= SCHED_POWER_SHIFT;
2918 sdg->sgp->power_orig = power;
2920 if (sched_feat(ARCH_POWER))
2921 power *= arch_scale_freq_power(sd, cpu);
2923 power *= default_scale_freq_power(sd, cpu);
2925 power >>= SCHED_POWER_SHIFT;
2927 power *= scale_rt_power(cpu);
2928 power >>= SCHED_POWER_SHIFT;
2933 cpu_rq(cpu)->cpu_power = power;
2934 sdg->sgp->power = power;
2937 static void update_group_power(struct sched_domain *sd, int cpu)
2939 struct sched_domain *child = sd->child;
2940 struct sched_group *group, *sdg = sd->groups;
2941 unsigned long power;
2944 update_cpu_power(sd, cpu);
2950 group = child->groups;
2952 power += group->sgp->power;
2953 group = group->next;
2954 } while (group != child->groups);
2956 sdg->sgp->power = power;
2960 * Try and fix up capacity for tiny siblings, this is needed when
2961 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2962 * which on its own isn't powerful enough.
2964 * See update_sd_pick_busiest() and check_asym_packing().
2967 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2970 * Only siblings can have significantly less than SCHED_POWER_SCALE
2972 if (!(sd->flags & SD_SHARE_CPUPOWER))
2976 * If ~90% of the cpu_power is still there, we're good.
2978 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
2985 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2986 * @sd: The sched_domain whose statistics are to be updated.
2987 * @group: sched_group whose statistics are to be updated.
2988 * @this_cpu: Cpu for which load balance is currently performed.
2989 * @idle: Idle status of this_cpu
2990 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2991 * @local_group: Does group contain this_cpu.
2992 * @cpus: Set of cpus considered for load balancing.
2993 * @balance: Should we balance.
2994 * @sgs: variable to hold the statistics for this group.
2996 static inline void update_sg_lb_stats(struct sched_domain *sd,
2997 struct sched_group *group, int this_cpu,
2998 enum cpu_idle_type idle, int load_idx,
2999 int local_group, const struct cpumask *cpus,
3000 int *balance, struct sg_lb_stats *sgs)
3002 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3004 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3005 unsigned long avg_load_per_task = 0;
3008 balance_cpu = group_first_cpu(group);
3010 /* Tally up the load of all CPUs in the group */
3012 min_cpu_load = ~0UL;
3015 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3016 struct rq *rq = cpu_rq(i);
3018 /* Bias balancing toward cpus of our domain */
3020 if (idle_cpu(i) && !first_idle_cpu) {
3025 load = target_load(i, load_idx);
3027 load = source_load(i, load_idx);
3028 if (load > max_cpu_load) {
3029 max_cpu_load = load;
3030 max_nr_running = rq->nr_running;
3032 if (min_cpu_load > load)
3033 min_cpu_load = load;
3036 sgs->group_load += load;
3037 sgs->sum_nr_running += rq->nr_running;
3038 sgs->sum_weighted_load += weighted_cpuload(i);
3044 * First idle cpu or the first cpu(busiest) in this sched group
3045 * is eligible for doing load balancing at this and above
3046 * domains. In the newly idle case, we will allow all the cpu's
3047 * to do the newly idle load balance.
3049 if (idle != CPU_NEWLY_IDLE && local_group) {
3050 if (balance_cpu != this_cpu) {
3054 update_group_power(sd, this_cpu);
3057 /* Adjust by relative CPU power of the group */
3058 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3061 * Consider the group unbalanced when the imbalance is larger
3062 * than the average weight of a task.
3064 * APZ: with cgroup the avg task weight can vary wildly and
3065 * might not be a suitable number - should we keep a
3066 * normalized nr_running number somewhere that negates
3069 if (sgs->sum_nr_running)
3070 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3072 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3075 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3077 if (!sgs->group_capacity)
3078 sgs->group_capacity = fix_small_capacity(sd, group);
3079 sgs->group_weight = group->group_weight;
3081 if (sgs->group_capacity > sgs->sum_nr_running)
3082 sgs->group_has_capacity = 1;
3086 * update_sd_pick_busiest - return 1 on busiest group
3087 * @sd: sched_domain whose statistics are to be checked
3088 * @sds: sched_domain statistics
3089 * @sg: sched_group candidate to be checked for being the busiest
3090 * @sgs: sched_group statistics
3091 * @this_cpu: the current cpu
3093 * Determine if @sg is a busier group than the previously selected
3096 static bool update_sd_pick_busiest(struct sched_domain *sd,
3097 struct sd_lb_stats *sds,
3098 struct sched_group *sg,
3099 struct sg_lb_stats *sgs,
3102 if (sgs->avg_load <= sds->max_load)
3105 if (sgs->sum_nr_running > sgs->group_capacity)
3112 * ASYM_PACKING needs to move all the work to the lowest
3113 * numbered CPUs in the group, therefore mark all groups
3114 * higher than ourself as busy.
3116 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3117 this_cpu < group_first_cpu(sg)) {
3121 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3129 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3130 * @sd: sched_domain whose statistics are to be updated.
3131 * @this_cpu: Cpu for which load balance is currently performed.
3132 * @idle: Idle status of this_cpu
3133 * @cpus: Set of cpus considered for load balancing.
3134 * @balance: Should we balance.
3135 * @sds: variable to hold the statistics for this sched_domain.
3137 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3138 enum cpu_idle_type idle, const struct cpumask *cpus,
3139 int *balance, struct sd_lb_stats *sds)
3141 struct sched_domain *child = sd->child;
3142 struct sched_group *sg = sd->groups;
3143 struct sg_lb_stats sgs;
3144 int load_idx, prefer_sibling = 0;
3146 if (child && child->flags & SD_PREFER_SIBLING)
3149 init_sd_power_savings_stats(sd, sds, idle);
3150 load_idx = get_sd_load_idx(sd, idle);
3155 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3156 memset(&sgs, 0, sizeof(sgs));
3157 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3158 local_group, cpus, balance, &sgs);
3160 if (local_group && !(*balance))
3163 sds->total_load += sgs.group_load;
3164 sds->total_pwr += sg->sgp->power;
3167 * In case the child domain prefers tasks go to siblings
3168 * first, lower the sg capacity to one so that we'll try
3169 * and move all the excess tasks away. We lower the capacity
3170 * of a group only if the local group has the capacity to fit
3171 * these excess tasks, i.e. nr_running < group_capacity. The
3172 * extra check prevents the case where you always pull from the
3173 * heaviest group when it is already under-utilized (possible
3174 * with a large weight task outweighs the tasks on the system).
3176 if (prefer_sibling && !local_group && sds->this_has_capacity)
3177 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3180 sds->this_load = sgs.avg_load;
3182 sds->this_nr_running = sgs.sum_nr_running;
3183 sds->this_load_per_task = sgs.sum_weighted_load;
3184 sds->this_has_capacity = sgs.group_has_capacity;
3185 sds->this_idle_cpus = sgs.idle_cpus;
3186 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
3187 sds->max_load = sgs.avg_load;
3189 sds->busiest_nr_running = sgs.sum_nr_running;
3190 sds->busiest_idle_cpus = sgs.idle_cpus;
3191 sds->busiest_group_capacity = sgs.group_capacity;
3192 sds->busiest_load_per_task = sgs.sum_weighted_load;
3193 sds->busiest_has_capacity = sgs.group_has_capacity;
3194 sds->busiest_group_weight = sgs.group_weight;
3195 sds->group_imb = sgs.group_imb;
3198 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
3200 } while (sg != sd->groups);
3203 int __weak arch_sd_sibling_asym_packing(void)
3205 return 0*SD_ASYM_PACKING;
3209 * check_asym_packing - Check to see if the group is packed into the
3212 * This is primarily intended to used at the sibling level. Some
3213 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3214 * case of POWER7, it can move to lower SMT modes only when higher
3215 * threads are idle. When in lower SMT modes, the threads will
3216 * perform better since they share less core resources. Hence when we
3217 * have idle threads, we want them to be the higher ones.
3219 * This packing function is run on idle threads. It checks to see if
3220 * the busiest CPU in this domain (core in the P7 case) has a higher
3221 * CPU number than the packing function is being run on. Here we are
3222 * assuming lower CPU number will be equivalent to lower a SMT thread
3225 * Returns 1 when packing is required and a task should be moved to
3226 * this CPU. The amount of the imbalance is returned in *imbalance.
3228 * @sd: The sched_domain whose packing is to be checked.
3229 * @sds: Statistics of the sched_domain which is to be packed
3230 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3231 * @imbalance: returns amount of imbalanced due to packing.
3233 static int check_asym_packing(struct sched_domain *sd,
3234 struct sd_lb_stats *sds,
3235 int this_cpu, unsigned long *imbalance)
3239 if (!(sd->flags & SD_ASYM_PACKING))
3245 busiest_cpu = group_first_cpu(sds->busiest);
3246 if (this_cpu > busiest_cpu)
3249 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
3255 * fix_small_imbalance - Calculate the minor imbalance that exists
3256 * amongst the groups of a sched_domain, during
3258 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3259 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3260 * @imbalance: Variable to store the imbalance.
3262 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3263 int this_cpu, unsigned long *imbalance)
3265 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3266 unsigned int imbn = 2;
3267 unsigned long scaled_busy_load_per_task;
3269 if (sds->this_nr_running) {
3270 sds->this_load_per_task /= sds->this_nr_running;
3271 if (sds->busiest_load_per_task >
3272 sds->this_load_per_task)
3275 sds->this_load_per_task =
3276 cpu_avg_load_per_task(this_cpu);
3278 scaled_busy_load_per_task = sds->busiest_load_per_task
3279 * SCHED_POWER_SCALE;
3280 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3282 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3283 (scaled_busy_load_per_task * imbn)) {
3284 *imbalance = sds->busiest_load_per_task;
3289 * OK, we don't have enough imbalance to justify moving tasks,
3290 * however we may be able to increase total CPU power used by
3294 pwr_now += sds->busiest->sgp->power *
3295 min(sds->busiest_load_per_task, sds->max_load);
3296 pwr_now += sds->this->sgp->power *
3297 min(sds->this_load_per_task, sds->this_load);
3298 pwr_now /= SCHED_POWER_SCALE;
3300 /* Amount of load we'd subtract */
3301 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3302 sds->busiest->sgp->power;
3303 if (sds->max_load > tmp)
3304 pwr_move += sds->busiest->sgp->power *
3305 min(sds->busiest_load_per_task, sds->max_load - tmp);
3307 /* Amount of load we'd add */
3308 if (sds->max_load * sds->busiest->sgp->power <
3309 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3310 tmp = (sds->max_load * sds->busiest->sgp->power) /
3311 sds->this->sgp->power;
3313 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3314 sds->this->sgp->power;
3315 pwr_move += sds->this->sgp->power *
3316 min(sds->this_load_per_task, sds->this_load + tmp);
3317 pwr_move /= SCHED_POWER_SCALE;
3319 /* Move if we gain throughput */
3320 if (pwr_move > pwr_now)
3321 *imbalance = sds->busiest_load_per_task;
3325 * calculate_imbalance - Calculate the amount of imbalance present within the
3326 * groups of a given sched_domain during load balance.
3327 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3328 * @this_cpu: Cpu for which currently load balance is being performed.
3329 * @imbalance: The variable to store the imbalance.
3331 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3332 unsigned long *imbalance)
3334 unsigned long max_pull, load_above_capacity = ~0UL;
3336 sds->busiest_load_per_task /= sds->busiest_nr_running;
3337 if (sds->group_imb) {
3338 sds->busiest_load_per_task =
3339 min(sds->busiest_load_per_task, sds->avg_load);
3343 * In the presence of smp nice balancing, certain scenarios can have
3344 * max load less than avg load(as we skip the groups at or below
3345 * its cpu_power, while calculating max_load..)
3347 if (sds->max_load < sds->avg_load) {
3349 return fix_small_imbalance(sds, this_cpu, imbalance);
3352 if (!sds->group_imb) {
3354 * Don't want to pull so many tasks that a group would go idle.
3356 load_above_capacity = (sds->busiest_nr_running -
3357 sds->busiest_group_capacity);
3359 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3361 load_above_capacity /= sds->busiest->sgp->power;
3365 * We're trying to get all the cpus to the average_load, so we don't
3366 * want to push ourselves above the average load, nor do we wish to
3367 * reduce the max loaded cpu below the average load. At the same time,
3368 * we also don't want to reduce the group load below the group capacity
3369 * (so that we can implement power-savings policies etc). Thus we look
3370 * for the minimum possible imbalance.
3371 * Be careful of negative numbers as they'll appear as very large values
3372 * with unsigned longs.
3374 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3376 /* How much load to actually move to equalise the imbalance */
3377 *imbalance = min(max_pull * sds->busiest->sgp->power,
3378 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3379 / SCHED_POWER_SCALE;
3382 * if *imbalance is less than the average load per runnable task
3383 * there is no guarantee that any tasks will be moved so we'll have
3384 * a think about bumping its value to force at least one task to be
3387 if (*imbalance < sds->busiest_load_per_task)
3388 return fix_small_imbalance(sds, this_cpu, imbalance);
3392 /******* find_busiest_group() helpers end here *********************/
3395 * find_busiest_group - Returns the busiest group within the sched_domain
3396 * if there is an imbalance. If there isn't an imbalance, and
3397 * the user has opted for power-savings, it returns a group whose
3398 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3399 * such a group exists.
3401 * Also calculates the amount of weighted load which should be moved
3402 * to restore balance.
3404 * @sd: The sched_domain whose busiest group is to be returned.
3405 * @this_cpu: The cpu for which load balancing is currently being performed.
3406 * @imbalance: Variable which stores amount of weighted load which should
3407 * be moved to restore balance/put a group to idle.
3408 * @idle: The idle status of this_cpu.
3409 * @cpus: The set of CPUs under consideration for load-balancing.
3410 * @balance: Pointer to a variable indicating if this_cpu
3411 * is the appropriate cpu to perform load balancing at this_level.
3413 * Returns: - the busiest group if imbalance exists.
3414 * - If no imbalance and user has opted for power-savings balance,
3415 * return the least loaded group whose CPUs can be
3416 * put to idle by rebalancing its tasks onto our group.
3418 static struct sched_group *
3419 find_busiest_group(struct sched_domain *sd, int this_cpu,
3420 unsigned long *imbalance, enum cpu_idle_type idle,
3421 const struct cpumask *cpus, int *balance)
3423 struct sd_lb_stats sds;
3425 memset(&sds, 0, sizeof(sds));
3428 * Compute the various statistics relavent for load balancing at
3431 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3434 * this_cpu is not the appropriate cpu to perform load balancing at
3440 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3441 check_asym_packing(sd, &sds, this_cpu, imbalance))
3444 /* There is no busy sibling group to pull tasks from */
3445 if (!sds.busiest || sds.busiest_nr_running == 0)
3448 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3451 * If the busiest group is imbalanced the below checks don't
3452 * work because they assumes all things are equal, which typically
3453 * isn't true due to cpus_allowed constraints and the like.
3458 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3459 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3460 !sds.busiest_has_capacity)
3464 * If the local group is more busy than the selected busiest group
3465 * don't try and pull any tasks.
3467 if (sds.this_load >= sds.max_load)
3471 * Don't pull any tasks if this group is already above the domain
3474 if (sds.this_load >= sds.avg_load)
3477 if (idle == CPU_IDLE) {
3479 * This cpu is idle. If the busiest group load doesn't
3480 * have more tasks than the number of available cpu's and
3481 * there is no imbalance between this and busiest group
3482 * wrt to idle cpu's, it is balanced.
3484 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3485 sds.busiest_nr_running <= sds.busiest_group_weight)
3489 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3490 * imbalance_pct to be conservative.
3492 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3497 /* Looks like there is an imbalance. Compute it */
3498 calculate_imbalance(&sds, this_cpu, imbalance);
3503 * There is no obvious imbalance. But check if we can do some balancing
3506 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3514 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3517 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3518 enum cpu_idle_type idle, unsigned long imbalance,
3519 const struct cpumask *cpus)
3521 struct rq *busiest = NULL, *rq;
3522 unsigned long max_load = 0;
3525 for_each_cpu(i, sched_group_cpus(group)) {
3526 unsigned long power = power_of(i);
3527 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3532 capacity = fix_small_capacity(sd, group);
3534 if (!cpumask_test_cpu(i, cpus))
3538 wl = weighted_cpuload(i);
3541 * When comparing with imbalance, use weighted_cpuload()
3542 * which is not scaled with the cpu power.
3544 if (capacity && rq->nr_running == 1 && wl > imbalance)
3548 * For the load comparisons with the other cpu's, consider
3549 * the weighted_cpuload() scaled with the cpu power, so that
3550 * the load can be moved away from the cpu that is potentially
3551 * running at a lower capacity.
3553 wl = (wl * SCHED_POWER_SCALE) / power;
3555 if (wl > max_load) {
3565 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3566 * so long as it is large enough.
3568 #define MAX_PINNED_INTERVAL 512
3570 /* Working cpumask for load_balance and load_balance_newidle. */
3571 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3573 static int need_active_balance(struct sched_domain *sd, int idle,
3574 int busiest_cpu, int this_cpu)
3576 if (idle == CPU_NEWLY_IDLE) {
3579 * ASYM_PACKING needs to force migrate tasks from busy but
3580 * higher numbered CPUs in order to pack all tasks in the
3581 * lowest numbered CPUs.
3583 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3587 * The only task running in a non-idle cpu can be moved to this
3588 * cpu in an attempt to completely freeup the other CPU
3591 * The package power saving logic comes from
3592 * find_busiest_group(). If there are no imbalance, then
3593 * f_b_g() will return NULL. However when sched_mc={1,2} then
3594 * f_b_g() will select a group from which a running task may be
3595 * pulled to this cpu in order to make the other package idle.
3596 * If there is no opportunity to make a package idle and if
3597 * there are no imbalance, then f_b_g() will return NULL and no
3598 * action will be taken in load_balance_newidle().
3600 * Under normal task pull operation due to imbalance, there
3601 * will be more than one task in the source run queue and
3602 * move_tasks() will succeed. ld_moved will be true and this
3603 * active balance code will not be triggered.
3605 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3609 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3612 static int active_load_balance_cpu_stop(void *data);
3615 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3616 * tasks if there is an imbalance.
3618 static int load_balance(int this_cpu, struct rq *this_rq,
3619 struct sched_domain *sd, enum cpu_idle_type idle,
3622 int ld_moved, all_pinned = 0, active_balance = 0;
3623 struct sched_group *group;
3624 unsigned long imbalance;
3626 unsigned long flags;
3627 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3629 cpumask_copy(cpus, cpu_active_mask);
3631 schedstat_inc(sd, lb_count[idle]);
3634 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
3641 schedstat_inc(sd, lb_nobusyg[idle]);
3645 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3647 schedstat_inc(sd, lb_nobusyq[idle]);
3651 BUG_ON(busiest == this_rq);
3653 schedstat_add(sd, lb_imbalance[idle], imbalance);
3656 if (busiest->nr_running > 1) {
3658 * Attempt to move tasks. If find_busiest_group has found
3659 * an imbalance but busiest->nr_running <= 1, the group is
3660 * still unbalanced. ld_moved simply stays zero, so it is
3661 * correctly treated as an imbalance.
3664 local_irq_save(flags);
3665 double_rq_lock(this_rq, busiest);
3666 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3667 imbalance, sd, idle, &all_pinned);
3668 double_rq_unlock(this_rq, busiest);
3669 local_irq_restore(flags);
3672 * some other cpu did the load balance for us.
3674 if (ld_moved && this_cpu != smp_processor_id())
3675 resched_cpu(this_cpu);
3677 /* All tasks on this runqueue were pinned by CPU affinity */
3678 if (unlikely(all_pinned)) {
3679 cpumask_clear_cpu(cpu_of(busiest), cpus);
3680 if (!cpumask_empty(cpus))
3687 schedstat_inc(sd, lb_failed[idle]);
3689 * Increment the failure counter only on periodic balance.
3690 * We do not want newidle balance, which can be very
3691 * frequent, pollute the failure counter causing
3692 * excessive cache_hot migrations and active balances.
3694 if (idle != CPU_NEWLY_IDLE)
3695 sd->nr_balance_failed++;
3697 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
3698 raw_spin_lock_irqsave(&busiest->lock, flags);
3700 /* don't kick the active_load_balance_cpu_stop,
3701 * if the curr task on busiest cpu can't be
3704 if (!cpumask_test_cpu(this_cpu,
3705 &busiest->curr->cpus_allowed)) {
3706 raw_spin_unlock_irqrestore(&busiest->lock,
3709 goto out_one_pinned;
3713 * ->active_balance synchronizes accesses to
3714 * ->active_balance_work. Once set, it's cleared
3715 * only after active load balance is finished.
3717 if (!busiest->active_balance) {
3718 busiest->active_balance = 1;
3719 busiest->push_cpu = this_cpu;
3722 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3725 stop_one_cpu_nowait(cpu_of(busiest),
3726 active_load_balance_cpu_stop, busiest,
3727 &busiest->active_balance_work);
3730 * We've kicked active balancing, reset the failure
3733 sd->nr_balance_failed = sd->cache_nice_tries+1;
3736 sd->nr_balance_failed = 0;
3738 if (likely(!active_balance)) {
3739 /* We were unbalanced, so reset the balancing interval */
3740 sd->balance_interval = sd->min_interval;
3743 * If we've begun active balancing, start to back off. This
3744 * case may not be covered by the all_pinned logic if there
3745 * is only 1 task on the busy runqueue (because we don't call
3748 if (sd->balance_interval < sd->max_interval)
3749 sd->balance_interval *= 2;
3755 schedstat_inc(sd, lb_balanced[idle]);
3757 sd->nr_balance_failed = 0;
3760 /* tune up the balancing interval */
3761 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3762 (sd->balance_interval < sd->max_interval))
3763 sd->balance_interval *= 2;
3771 * idle_balance is called by schedule() if this_cpu is about to become
3772 * idle. Attempts to pull tasks from other CPUs.
3774 static void idle_balance(int this_cpu, struct rq *this_rq)
3776 struct sched_domain *sd;
3777 int pulled_task = 0;
3778 unsigned long next_balance = jiffies + HZ;
3780 this_rq->idle_stamp = this_rq->clock;
3782 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3786 * Drop the rq->lock, but keep IRQ/preempt disabled.
3788 raw_spin_unlock(&this_rq->lock);
3790 update_shares(this_cpu);
3792 for_each_domain(this_cpu, sd) {
3793 unsigned long interval;
3796 if (!(sd->flags & SD_LOAD_BALANCE))
3799 if (sd->flags & SD_BALANCE_NEWIDLE) {
3800 /* If we've pulled tasks over stop searching: */
3801 pulled_task = load_balance(this_cpu, this_rq,
3802 sd, CPU_NEWLY_IDLE, &balance);
3805 interval = msecs_to_jiffies(sd->balance_interval);
3806 if (time_after(next_balance, sd->last_balance + interval))
3807 next_balance = sd->last_balance + interval;
3809 this_rq->idle_stamp = 0;
3815 raw_spin_lock(&this_rq->lock);
3817 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3819 * We are going idle. next_balance may be set based on
3820 * a busy processor. So reset next_balance.
3822 this_rq->next_balance = next_balance;
3827 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3828 * running tasks off the busiest CPU onto idle CPUs. It requires at
3829 * least 1 task to be running on each physical CPU where possible, and
3830 * avoids physical / logical imbalances.
3832 static int active_load_balance_cpu_stop(void *data)
3834 struct rq *busiest_rq = data;
3835 int busiest_cpu = cpu_of(busiest_rq);
3836 int target_cpu = busiest_rq->push_cpu;
3837 struct rq *target_rq = cpu_rq(target_cpu);
3838 struct sched_domain *sd;
3840 raw_spin_lock_irq(&busiest_rq->lock);
3842 /* make sure the requested cpu hasn't gone down in the meantime */
3843 if (unlikely(busiest_cpu != smp_processor_id() ||
3844 !busiest_rq->active_balance))
3847 /* Is there any task to move? */
3848 if (busiest_rq->nr_running <= 1)
3852 * This condition is "impossible", if it occurs
3853 * we need to fix it. Originally reported by
3854 * Bjorn Helgaas on a 128-cpu setup.
3856 BUG_ON(busiest_rq == target_rq);
3858 /* move a task from busiest_rq to target_rq */
3859 double_lock_balance(busiest_rq, target_rq);
3861 /* Search for an sd spanning us and the target CPU. */
3863 for_each_domain(target_cpu, sd) {
3864 if ((sd->flags & SD_LOAD_BALANCE) &&
3865 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3870 schedstat_inc(sd, alb_count);
3872 if (move_one_task(target_rq, target_cpu, busiest_rq,
3874 schedstat_inc(sd, alb_pushed);
3876 schedstat_inc(sd, alb_failed);
3879 double_unlock_balance(busiest_rq, target_rq);
3881 busiest_rq->active_balance = 0;
3882 raw_spin_unlock_irq(&busiest_rq->lock);
3888 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3890 static void trigger_sched_softirq(void *data)
3892 raise_softirq_irqoff(SCHED_SOFTIRQ);
3895 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3897 csd->func = trigger_sched_softirq;
3904 * idle load balancing details
3905 * - One of the idle CPUs nominates itself as idle load_balancer, while
3907 * - This idle load balancer CPU will also go into tickless mode when
3908 * it is idle, just like all other idle CPUs
3909 * - When one of the busy CPUs notice that there may be an idle rebalancing
3910 * needed, they will kick the idle load balancer, which then does idle
3911 * load balancing for all the idle CPUs.
3914 atomic_t load_balancer;
3915 atomic_t first_pick_cpu;
3916 atomic_t second_pick_cpu;
3917 cpumask_var_t idle_cpus_mask;
3918 cpumask_var_t grp_idle_mask;
3919 unsigned long next_balance; /* in jiffy units */
3920 } nohz ____cacheline_aligned;
3922 int get_nohz_load_balancer(void)
3924 return atomic_read(&nohz.load_balancer);
3927 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3929 * lowest_flag_domain - Return lowest sched_domain containing flag.
3930 * @cpu: The cpu whose lowest level of sched domain is to
3932 * @flag: The flag to check for the lowest sched_domain
3933 * for the given cpu.
3935 * Returns the lowest sched_domain of a cpu which contains the given flag.
3937 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3939 struct sched_domain *sd;
3941 for_each_domain(cpu, sd)
3942 if (sd->flags & flag)
3949 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3950 * @cpu: The cpu whose domains we're iterating over.
3951 * @sd: variable holding the value of the power_savings_sd
3953 * @flag: The flag to filter the sched_domains to be iterated.
3955 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3956 * set, starting from the lowest sched_domain to the highest.
3958 #define for_each_flag_domain(cpu, sd, flag) \
3959 for (sd = lowest_flag_domain(cpu, flag); \
3960 (sd && (sd->flags & flag)); sd = sd->parent)
3963 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3964 * @ilb_group: group to be checked for semi-idleness
3966 * Returns: 1 if the group is semi-idle. 0 otherwise.
3968 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3969 * and atleast one non-idle CPU. This helper function checks if the given
3970 * sched_group is semi-idle or not.
3972 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3974 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3975 sched_group_cpus(ilb_group));
3978 * A sched_group is semi-idle when it has atleast one busy cpu
3979 * and atleast one idle cpu.
3981 if (cpumask_empty(nohz.grp_idle_mask))
3984 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3990 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3991 * @cpu: The cpu which is nominating a new idle_load_balancer.
3993 * Returns: Returns the id of the idle load balancer if it exists,
3994 * Else, returns >= nr_cpu_ids.
3996 * This algorithm picks the idle load balancer such that it belongs to a
3997 * semi-idle powersavings sched_domain. The idea is to try and avoid
3998 * completely idle packages/cores just for the purpose of idle load balancing
3999 * when there are other idle cpu's which are better suited for that job.
4001 static int find_new_ilb(int cpu)
4003 struct sched_domain *sd;
4004 struct sched_group *ilb_group;
4005 int ilb = nr_cpu_ids;
4008 * Have idle load balancer selection from semi-idle packages only
4009 * when power-aware load balancing is enabled
4011 if (!(sched_smt_power_savings || sched_mc_power_savings))
4015 * Optimize for the case when we have no idle CPUs or only one
4016 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4018 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4022 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4023 ilb_group = sd->groups;
4026 if (is_semi_idle_group(ilb_group)) {
4027 ilb = cpumask_first(nohz.grp_idle_mask);
4031 ilb_group = ilb_group->next;
4033 } while (ilb_group != sd->groups);
4041 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4042 static inline int find_new_ilb(int call_cpu)
4049 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4050 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4051 * CPU (if there is one).
4053 static void nohz_balancer_kick(int cpu)
4057 nohz.next_balance++;
4059 ilb_cpu = get_nohz_load_balancer();
4061 if (ilb_cpu >= nr_cpu_ids) {
4062 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4063 if (ilb_cpu >= nr_cpu_ids)
4067 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4068 struct call_single_data *cp;
4070 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4071 cp = &per_cpu(remote_sched_softirq_cb, cpu);
4072 __smp_call_function_single(ilb_cpu, cp, 0);
4078 * This routine will try to nominate the ilb (idle load balancing)
4079 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4080 * load balancing on behalf of all those cpus.
4082 * When the ilb owner becomes busy, we will not have new ilb owner until some
4083 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4084 * idle load balancing by kicking one of the idle CPUs.
4086 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4087 * ilb owner CPU in future (when there is a need for idle load balancing on
4088 * behalf of all idle CPUs).
4090 void select_nohz_load_balancer(int stop_tick)
4092 int cpu = smp_processor_id();
4095 if (!cpu_active(cpu)) {
4096 if (atomic_read(&nohz.load_balancer) != cpu)
4100 * If we are going offline and still the leader,
4103 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4110 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4112 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4113 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4114 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4115 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4117 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4120 /* make me the ilb owner */
4121 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4126 * Check to see if there is a more power-efficient
4129 new_ilb = find_new_ilb(cpu);
4130 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4131 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4132 resched_cpu(new_ilb);
4138 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4141 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4143 if (atomic_read(&nohz.load_balancer) == cpu)
4144 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4152 static DEFINE_SPINLOCK(balancing);
4154 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4157 * Scale the max load_balance interval with the number of CPUs in the system.
4158 * This trades load-balance latency on larger machines for less cross talk.
4160 static void update_max_interval(void)
4162 max_load_balance_interval = HZ*num_online_cpus()/10;
4166 * It checks each scheduling domain to see if it is due to be balanced,
4167 * and initiates a balancing operation if so.
4169 * Balancing parameters are set up in arch_init_sched_domains.
4171 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4174 struct rq *rq = cpu_rq(cpu);
4175 unsigned long interval;
4176 struct sched_domain *sd;
4177 /* Earliest time when we have to do rebalance again */
4178 unsigned long next_balance = jiffies + 60*HZ;
4179 int update_next_balance = 0;
4185 for_each_domain(cpu, sd) {
4186 if (!(sd->flags & SD_LOAD_BALANCE))
4189 interval = sd->balance_interval;
4190 if (idle != CPU_IDLE)
4191 interval *= sd->busy_factor;
4193 /* scale ms to jiffies */
4194 interval = msecs_to_jiffies(interval);
4195 interval = clamp(interval, 1UL, max_load_balance_interval);
4197 need_serialize = sd->flags & SD_SERIALIZE;
4199 if (need_serialize) {
4200 if (!spin_trylock(&balancing))
4204 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4205 if (load_balance(cpu, rq, sd, idle, &balance)) {
4207 * We've pulled tasks over so either we're no
4210 idle = CPU_NOT_IDLE;
4212 sd->last_balance = jiffies;
4215 spin_unlock(&balancing);
4217 if (time_after(next_balance, sd->last_balance + interval)) {
4218 next_balance = sd->last_balance + interval;
4219 update_next_balance = 1;
4223 * Stop the load balance at this level. There is another
4224 * CPU in our sched group which is doing load balancing more
4233 * next_balance will be updated only when there is a need.
4234 * When the cpu is attached to null domain for ex, it will not be
4237 if (likely(update_next_balance))
4238 rq->next_balance = next_balance;
4243 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4244 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4246 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4248 struct rq *this_rq = cpu_rq(this_cpu);
4252 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
4255 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4256 if (balance_cpu == this_cpu)
4260 * If this cpu gets work to do, stop the load balancing
4261 * work being done for other cpus. Next load
4262 * balancing owner will pick it up.
4264 if (need_resched()) {
4265 this_rq->nohz_balance_kick = 0;
4269 raw_spin_lock_irq(&this_rq->lock);
4270 update_rq_clock(this_rq);
4271 update_cpu_load(this_rq);
4272 raw_spin_unlock_irq(&this_rq->lock);
4274 rebalance_domains(balance_cpu, CPU_IDLE);
4276 rq = cpu_rq(balance_cpu);
4277 if (time_after(this_rq->next_balance, rq->next_balance))
4278 this_rq->next_balance = rq->next_balance;
4280 nohz.next_balance = this_rq->next_balance;
4281 this_rq->nohz_balance_kick = 0;
4285 * Current heuristic for kicking the idle load balancer
4286 * - first_pick_cpu is the one of the busy CPUs. It will kick
4287 * idle load balancer when it has more than one process active. This
4288 * eliminates the need for idle load balancing altogether when we have
4289 * only one running process in the system (common case).
4290 * - If there are more than one busy CPU, idle load balancer may have
4291 * to run for active_load_balance to happen (i.e., two busy CPUs are
4292 * SMT or core siblings and can run better if they move to different
4293 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4294 * which will kick idle load balancer as soon as it has any load.
4296 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4298 unsigned long now = jiffies;
4300 int first_pick_cpu, second_pick_cpu;
4302 if (time_before(now, nohz.next_balance))
4305 if (rq->idle_at_tick)
4308 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4309 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4311 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4312 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4315 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4316 if (ret == nr_cpu_ids || ret == cpu) {
4317 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4318 if (rq->nr_running > 1)
4321 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4322 if (ret == nr_cpu_ids || ret == cpu) {
4330 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4334 * run_rebalance_domains is triggered when needed from the scheduler tick.
4335 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4337 static void run_rebalance_domains(struct softirq_action *h)
4339 int this_cpu = smp_processor_id();
4340 struct rq *this_rq = cpu_rq(this_cpu);
4341 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4342 CPU_IDLE : CPU_NOT_IDLE;
4344 rebalance_domains(this_cpu, idle);
4347 * If this cpu has a pending nohz_balance_kick, then do the
4348 * balancing on behalf of the other idle cpus whose ticks are
4351 nohz_idle_balance(this_cpu, idle);
4354 static inline int on_null_domain(int cpu)
4356 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4360 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4362 static inline void trigger_load_balance(struct rq *rq, int cpu)
4364 /* Don't need to rebalance while attached to NULL domain */
4365 if (time_after_eq(jiffies, rq->next_balance) &&
4366 likely(!on_null_domain(cpu)))
4367 raise_softirq(SCHED_SOFTIRQ);
4369 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4370 nohz_balancer_kick(cpu);
4374 static void rq_online_fair(struct rq *rq)
4379 static void rq_offline_fair(struct rq *rq)
4384 #else /* CONFIG_SMP */
4387 * on UP we do not need to balance between CPUs:
4389 static inline void idle_balance(int cpu, struct rq *rq)
4393 #endif /* CONFIG_SMP */
4396 * scheduler tick hitting a task of our scheduling class:
4398 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4400 struct cfs_rq *cfs_rq;
4401 struct sched_entity *se = &curr->se;
4403 for_each_sched_entity(se) {
4404 cfs_rq = cfs_rq_of(se);
4405 entity_tick(cfs_rq, se, queued);
4410 * called on fork with the child task as argument from the parent's context
4411 * - child not yet on the tasklist
4412 * - preemption disabled
4414 static void task_fork_fair(struct task_struct *p)
4416 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4417 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4418 int this_cpu = smp_processor_id();
4419 struct rq *rq = this_rq();
4420 unsigned long flags;
4422 raw_spin_lock_irqsave(&rq->lock, flags);
4424 update_rq_clock(rq);
4426 if (unlikely(task_cpu(p) != this_cpu)) {
4428 __set_task_cpu(p, this_cpu);
4432 update_curr(cfs_rq);
4435 se->vruntime = curr->vruntime;
4436 place_entity(cfs_rq, se, 1);
4438 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4440 * Upon rescheduling, sched_class::put_prev_task() will place
4441 * 'current' within the tree based on its new key value.
4443 swap(curr->vruntime, se->vruntime);
4444 resched_task(rq->curr);
4447 se->vruntime -= cfs_rq->min_vruntime;
4449 raw_spin_unlock_irqrestore(&rq->lock, flags);
4453 * Priority of the task has changed. Check to see if we preempt
4457 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4463 * Reschedule if we are currently running on this runqueue and
4464 * our priority decreased, or if we are not currently running on
4465 * this runqueue and our priority is higher than the current's
4467 if (rq->curr == p) {
4468 if (p->prio > oldprio)
4469 resched_task(rq->curr);
4471 check_preempt_curr(rq, p, 0);
4474 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4476 struct sched_entity *se = &p->se;
4477 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4480 * Ensure the task's vruntime is normalized, so that when its
4481 * switched back to the fair class the enqueue_entity(.flags=0) will
4482 * do the right thing.
4484 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4485 * have normalized the vruntime, if it was !on_rq, then only when
4486 * the task is sleeping will it still have non-normalized vruntime.
4488 if (!se->on_rq && p->state != TASK_RUNNING) {
4490 * Fix up our vruntime so that the current sleep doesn't
4491 * cause 'unlimited' sleep bonus.
4493 place_entity(cfs_rq, se, 0);
4494 se->vruntime -= cfs_rq->min_vruntime;
4499 * We switched to the sched_fair class.
4501 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4507 * We were most likely switched from sched_rt, so
4508 * kick off the schedule if running, otherwise just see
4509 * if we can still preempt the current task.
4512 resched_task(rq->curr);
4514 check_preempt_curr(rq, p, 0);
4517 /* Account for a task changing its policy or group.
4519 * This routine is mostly called to set cfs_rq->curr field when a task
4520 * migrates between groups/classes.
4522 static void set_curr_task_fair(struct rq *rq)
4524 struct sched_entity *se = &rq->curr->se;
4526 for_each_sched_entity(se) {
4527 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4529 set_next_entity(cfs_rq, se);
4530 /* ensure bandwidth has been allocated on our new cfs_rq */
4531 account_cfs_rq_runtime(cfs_rq, 0);
4535 #ifdef CONFIG_FAIR_GROUP_SCHED
4536 static void task_move_group_fair(struct task_struct *p, int on_rq)
4539 * If the task was not on the rq at the time of this cgroup movement
4540 * it must have been asleep, sleeping tasks keep their ->vruntime
4541 * absolute on their old rq until wakeup (needed for the fair sleeper
4542 * bonus in place_entity()).
4544 * If it was on the rq, we've just 'preempted' it, which does convert
4545 * ->vruntime to a relative base.
4547 * Make sure both cases convert their relative position when migrating
4548 * to another cgroup's rq. This does somewhat interfere with the
4549 * fair sleeper stuff for the first placement, but who cares.
4552 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4553 set_task_rq(p, task_cpu(p));
4555 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4559 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4561 struct sched_entity *se = &task->se;
4562 unsigned int rr_interval = 0;
4565 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4568 if (rq->cfs.load.weight)
4569 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4575 * All the scheduling class methods:
4577 static const struct sched_class fair_sched_class = {
4578 .next = &idle_sched_class,
4579 .enqueue_task = enqueue_task_fair,
4580 .dequeue_task = dequeue_task_fair,
4581 .yield_task = yield_task_fair,
4582 .yield_to_task = yield_to_task_fair,
4584 .check_preempt_curr = check_preempt_wakeup,
4586 .pick_next_task = pick_next_task_fair,
4587 .put_prev_task = put_prev_task_fair,
4590 .select_task_rq = select_task_rq_fair,
4592 .rq_online = rq_online_fair,
4593 .rq_offline = rq_offline_fair,
4595 .task_waking = task_waking_fair,
4598 .set_curr_task = set_curr_task_fair,
4599 .task_tick = task_tick_fair,
4600 .task_fork = task_fork_fair,
4602 .prio_changed = prio_changed_fair,
4603 .switched_from = switched_from_fair,
4604 .switched_to = switched_to_fair,
4606 .get_rr_interval = get_rr_interval_fair,
4608 #ifdef CONFIG_FAIR_GROUP_SCHED
4609 .task_move_group = task_move_group_fair,
4613 #ifdef CONFIG_SCHED_DEBUG
4614 static void print_cfs_stats(struct seq_file *m, int cpu)
4616 struct cfs_rq *cfs_rq;
4619 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4620 print_cfs_rq(m, cpu, cfs_rq);