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);
1442 static void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1444 struct rq *rq = rq_of(cfs_rq);
1445 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1446 struct sched_entity *se;
1450 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1452 cfs_rq->throttled = 0;
1453 raw_spin_lock(&cfs_b->lock);
1454 list_del_rcu(&cfs_rq->throttled_list);
1455 raw_spin_unlock(&cfs_b->lock);
1457 if (!cfs_rq->load.weight)
1460 task_delta = cfs_rq->h_nr_running;
1461 for_each_sched_entity(se) {
1465 cfs_rq = cfs_rq_of(se);
1467 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1468 cfs_rq->h_nr_running += task_delta;
1470 if (cfs_rq_throttled(cfs_rq))
1475 rq->nr_running += task_delta;
1477 /* determine whether we need to wake up potentially idle cpu */
1478 if (rq->curr == rq->idle && rq->cfs.nr_running)
1479 resched_task(rq->curr);
1482 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1483 u64 remaining, u64 expires)
1485 struct cfs_rq *cfs_rq;
1486 u64 runtime = remaining;
1489 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1491 struct rq *rq = rq_of(cfs_rq);
1493 raw_spin_lock(&rq->lock);
1494 if (!cfs_rq_throttled(cfs_rq))
1497 runtime = -cfs_rq->runtime_remaining + 1;
1498 if (runtime > remaining)
1499 runtime = remaining;
1500 remaining -= runtime;
1502 cfs_rq->runtime_remaining += runtime;
1503 cfs_rq->runtime_expires = expires;
1505 /* we check whether we're throttled above */
1506 if (cfs_rq->runtime_remaining > 0)
1507 unthrottle_cfs_rq(cfs_rq);
1510 raw_spin_unlock(&rq->lock);
1521 * Responsible for refilling a task_group's bandwidth and unthrottling its
1522 * cfs_rqs as appropriate. If there has been no activity within the last
1523 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1524 * used to track this state.
1526 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1528 u64 runtime, runtime_expires;
1529 int idle = 1, throttled;
1531 raw_spin_lock(&cfs_b->lock);
1532 /* no need to continue the timer with no bandwidth constraint */
1533 if (cfs_b->quota == RUNTIME_INF)
1536 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1537 /* idle depends on !throttled (for the case of a large deficit) */
1538 idle = cfs_b->idle && !throttled;
1540 /* if we're going inactive then everything else can be deferred */
1544 __refill_cfs_bandwidth_runtime(cfs_b);
1547 /* mark as potentially idle for the upcoming period */
1553 * There are throttled entities so we must first use the new bandwidth
1554 * to unthrottle them before making it generally available. This
1555 * ensures that all existing debts will be paid before a new cfs_rq is
1558 runtime = cfs_b->runtime;
1559 runtime_expires = cfs_b->runtime_expires;
1563 * This check is repeated as we are holding onto the new bandwidth
1564 * while we unthrottle. This can potentially race with an unthrottled
1565 * group trying to acquire new bandwidth from the global pool.
1567 while (throttled && runtime > 0) {
1568 raw_spin_unlock(&cfs_b->lock);
1569 /* we can't nest cfs_b->lock while distributing bandwidth */
1570 runtime = distribute_cfs_runtime(cfs_b, runtime,
1572 raw_spin_lock(&cfs_b->lock);
1574 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1577 /* return (any) remaining runtime */
1578 cfs_b->runtime = runtime;
1580 * While we are ensured activity in the period following an
1581 * unthrottle, this also covers the case in which the new bandwidth is
1582 * insufficient to cover the existing bandwidth deficit. (Forcing the
1583 * timer to remain active while there are any throttled entities.)
1588 cfs_b->timer_active = 0;
1589 raw_spin_unlock(&cfs_b->lock);
1594 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1595 unsigned long delta_exec) {}
1597 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1603 /**************************************************
1604 * CFS operations on tasks:
1607 #ifdef CONFIG_SCHED_HRTICK
1608 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1610 struct sched_entity *se = &p->se;
1611 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1613 WARN_ON(task_rq(p) != rq);
1615 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1616 u64 slice = sched_slice(cfs_rq, se);
1617 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1618 s64 delta = slice - ran;
1627 * Don't schedule slices shorter than 10000ns, that just
1628 * doesn't make sense. Rely on vruntime for fairness.
1631 delta = max_t(s64, 10000LL, delta);
1633 hrtick_start(rq, delta);
1638 * called from enqueue/dequeue and updates the hrtick when the
1639 * current task is from our class and nr_running is low enough
1642 static void hrtick_update(struct rq *rq)
1644 struct task_struct *curr = rq->curr;
1646 if (curr->sched_class != &fair_sched_class)
1649 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1650 hrtick_start_fair(rq, curr);
1652 #else /* !CONFIG_SCHED_HRTICK */
1654 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1658 static inline void hrtick_update(struct rq *rq)
1664 * The enqueue_task method is called before nr_running is
1665 * increased. Here we update the fair scheduling stats and
1666 * then put the task into the rbtree:
1669 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1671 struct cfs_rq *cfs_rq;
1672 struct sched_entity *se = &p->se;
1674 for_each_sched_entity(se) {
1677 cfs_rq = cfs_rq_of(se);
1678 enqueue_entity(cfs_rq, se, flags);
1681 * end evaluation on encountering a throttled cfs_rq
1683 * note: in the case of encountering a throttled cfs_rq we will
1684 * post the final h_nr_running increment below.
1686 if (cfs_rq_throttled(cfs_rq))
1688 cfs_rq->h_nr_running++;
1690 flags = ENQUEUE_WAKEUP;
1693 for_each_sched_entity(se) {
1694 cfs_rq = cfs_rq_of(se);
1695 cfs_rq->h_nr_running++;
1697 if (cfs_rq_throttled(cfs_rq))
1700 update_cfs_load(cfs_rq, 0);
1701 update_cfs_shares(cfs_rq);
1709 static void set_next_buddy(struct sched_entity *se);
1712 * The dequeue_task method is called before nr_running is
1713 * decreased. We remove the task from the rbtree and
1714 * update the fair scheduling stats:
1716 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1718 struct cfs_rq *cfs_rq;
1719 struct sched_entity *se = &p->se;
1720 int task_sleep = flags & DEQUEUE_SLEEP;
1722 for_each_sched_entity(se) {
1723 cfs_rq = cfs_rq_of(se);
1724 dequeue_entity(cfs_rq, se, flags);
1727 * end evaluation on encountering a throttled cfs_rq
1729 * note: in the case of encountering a throttled cfs_rq we will
1730 * post the final h_nr_running decrement below.
1732 if (cfs_rq_throttled(cfs_rq))
1734 cfs_rq->h_nr_running--;
1736 /* Don't dequeue parent if it has other entities besides us */
1737 if (cfs_rq->load.weight) {
1739 * Bias pick_next to pick a task from this cfs_rq, as
1740 * p is sleeping when it is within its sched_slice.
1742 if (task_sleep && parent_entity(se))
1743 set_next_buddy(parent_entity(se));
1745 /* avoid re-evaluating load for this entity */
1746 se = parent_entity(se);
1749 flags |= DEQUEUE_SLEEP;
1752 for_each_sched_entity(se) {
1753 cfs_rq = cfs_rq_of(se);
1754 cfs_rq->h_nr_running--;
1756 if (cfs_rq_throttled(cfs_rq))
1759 update_cfs_load(cfs_rq, 0);
1760 update_cfs_shares(cfs_rq);
1770 static void task_waking_fair(struct task_struct *p)
1772 struct sched_entity *se = &p->se;
1773 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1776 #ifndef CONFIG_64BIT
1777 u64 min_vruntime_copy;
1780 min_vruntime_copy = cfs_rq->min_vruntime_copy;
1782 min_vruntime = cfs_rq->min_vruntime;
1783 } while (min_vruntime != min_vruntime_copy);
1785 min_vruntime = cfs_rq->min_vruntime;
1788 se->vruntime -= min_vruntime;
1791 #ifdef CONFIG_FAIR_GROUP_SCHED
1793 * effective_load() calculates the load change as seen from the root_task_group
1795 * Adding load to a group doesn't make a group heavier, but can cause movement
1796 * of group shares between cpus. Assuming the shares were perfectly aligned one
1797 * can calculate the shift in shares.
1799 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1801 struct sched_entity *se = tg->se[cpu];
1806 for_each_sched_entity(se) {
1810 w = se->my_q->load.weight;
1812 /* use this cpu's instantaneous contribution */
1813 lw = atomic_read(&tg->load_weight);
1814 lw -= se->my_q->load_contribution;
1819 if (lw > 0 && wl < lw)
1820 wl = (wl * tg->shares) / lw;
1824 /* zero point is MIN_SHARES */
1825 if (wl < MIN_SHARES)
1827 wl -= se->load.weight;
1835 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1836 unsigned long wl, unsigned long wg)
1843 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1845 s64 this_load, load;
1846 int idx, this_cpu, prev_cpu;
1847 unsigned long tl_per_task;
1848 struct task_group *tg;
1849 unsigned long weight;
1853 this_cpu = smp_processor_id();
1854 prev_cpu = task_cpu(p);
1855 load = source_load(prev_cpu, idx);
1856 this_load = target_load(this_cpu, idx);
1859 * If sync wakeup then subtract the (maximum possible)
1860 * effect of the currently running task from the load
1861 * of the current CPU:
1864 tg = task_group(current);
1865 weight = current->se.load.weight;
1867 this_load += effective_load(tg, this_cpu, -weight, -weight);
1868 load += effective_load(tg, prev_cpu, 0, -weight);
1872 weight = p->se.load.weight;
1875 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1876 * due to the sync cause above having dropped this_load to 0, we'll
1877 * always have an imbalance, but there's really nothing you can do
1878 * about that, so that's good too.
1880 * Otherwise check if either cpus are near enough in load to allow this
1881 * task to be woken on this_cpu.
1883 if (this_load > 0) {
1884 s64 this_eff_load, prev_eff_load;
1886 this_eff_load = 100;
1887 this_eff_load *= power_of(prev_cpu);
1888 this_eff_load *= this_load +
1889 effective_load(tg, this_cpu, weight, weight);
1891 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1892 prev_eff_load *= power_of(this_cpu);
1893 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1895 balanced = this_eff_load <= prev_eff_load;
1900 * If the currently running task will sleep within
1901 * a reasonable amount of time then attract this newly
1904 if (sync && balanced)
1907 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1908 tl_per_task = cpu_avg_load_per_task(this_cpu);
1911 (this_load <= load &&
1912 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1914 * This domain has SD_WAKE_AFFINE and
1915 * p is cache cold in this domain, and
1916 * there is no bad imbalance.
1918 schedstat_inc(sd, ttwu_move_affine);
1919 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1927 * find_idlest_group finds and returns the least busy CPU group within the
1930 static struct sched_group *
1931 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1932 int this_cpu, int load_idx)
1934 struct sched_group *idlest = NULL, *group = sd->groups;
1935 unsigned long min_load = ULONG_MAX, this_load = 0;
1936 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1939 unsigned long load, avg_load;
1943 /* Skip over this group if it has no CPUs allowed */
1944 if (!cpumask_intersects(sched_group_cpus(group),
1948 local_group = cpumask_test_cpu(this_cpu,
1949 sched_group_cpus(group));
1951 /* Tally up the load of all CPUs in the group */
1954 for_each_cpu(i, sched_group_cpus(group)) {
1955 /* Bias balancing toward cpus of our domain */
1957 load = source_load(i, load_idx);
1959 load = target_load(i, load_idx);
1964 /* Adjust by relative CPU power of the group */
1965 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
1968 this_load = avg_load;
1969 } else if (avg_load < min_load) {
1970 min_load = avg_load;
1973 } while (group = group->next, group != sd->groups);
1975 if (!idlest || 100*this_load < imbalance*min_load)
1981 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1984 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1986 unsigned long load, min_load = ULONG_MAX;
1990 /* Traverse only the allowed CPUs */
1991 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1992 load = weighted_cpuload(i);
1994 if (load < min_load || (load == min_load && i == this_cpu)) {
2004 * Try and locate an idle CPU in the sched_domain.
2006 static int select_idle_sibling(struct task_struct *p, int target)
2008 int cpu = smp_processor_id();
2009 int prev_cpu = task_cpu(p);
2010 struct sched_domain *sd;
2014 * If the task is going to be woken-up on this cpu and if it is
2015 * already idle, then it is the right target.
2017 if (target == cpu && idle_cpu(cpu))
2021 * If the task is going to be woken-up on the cpu where it previously
2022 * ran and if it is currently idle, then it the right target.
2024 if (target == prev_cpu && idle_cpu(prev_cpu))
2028 * Otherwise, iterate the domains and find an elegible idle cpu.
2031 for_each_domain(target, sd) {
2032 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
2035 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
2043 * Lets stop looking for an idle sibling when we reached
2044 * the domain that spans the current cpu and prev_cpu.
2046 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
2047 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
2056 * sched_balance_self: balance the current task (running on cpu) in domains
2057 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2060 * Balance, ie. select the least loaded group.
2062 * Returns the target CPU number, or the same CPU if no balancing is needed.
2064 * preempt must be disabled.
2067 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2069 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2070 int cpu = smp_processor_id();
2071 int prev_cpu = task_cpu(p);
2073 int want_affine = 0;
2075 int sync = wake_flags & WF_SYNC;
2077 if (sd_flag & SD_BALANCE_WAKE) {
2078 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
2084 for_each_domain(cpu, tmp) {
2085 if (!(tmp->flags & SD_LOAD_BALANCE))
2089 * If power savings logic is enabled for a domain, see if we
2090 * are not overloaded, if so, don't balance wider.
2092 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2093 unsigned long power = 0;
2094 unsigned long nr_running = 0;
2095 unsigned long capacity;
2098 for_each_cpu(i, sched_domain_span(tmp)) {
2099 power += power_of(i);
2100 nr_running += cpu_rq(i)->cfs.nr_running;
2103 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2105 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2108 if (nr_running < capacity)
2113 * If both cpu and prev_cpu are part of this domain,
2114 * cpu is a valid SD_WAKE_AFFINE target.
2116 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2117 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2122 if (!want_sd && !want_affine)
2125 if (!(tmp->flags & sd_flag))
2133 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2136 new_cpu = select_idle_sibling(p, prev_cpu);
2141 int load_idx = sd->forkexec_idx;
2142 struct sched_group *group;
2145 if (!(sd->flags & sd_flag)) {
2150 if (sd_flag & SD_BALANCE_WAKE)
2151 load_idx = sd->wake_idx;
2153 group = find_idlest_group(sd, p, cpu, load_idx);
2159 new_cpu = find_idlest_cpu(group, p, cpu);
2160 if (new_cpu == -1 || new_cpu == cpu) {
2161 /* Now try balancing at a lower domain level of cpu */
2166 /* Now try balancing at a lower domain level of new_cpu */
2168 weight = sd->span_weight;
2170 for_each_domain(cpu, tmp) {
2171 if (weight <= tmp->span_weight)
2173 if (tmp->flags & sd_flag)
2176 /* while loop will break here if sd == NULL */
2183 #endif /* CONFIG_SMP */
2185 static unsigned long
2186 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2188 unsigned long gran = sysctl_sched_wakeup_granularity;
2191 * Since its curr running now, convert the gran from real-time
2192 * to virtual-time in his units.
2194 * By using 'se' instead of 'curr' we penalize light tasks, so
2195 * they get preempted easier. That is, if 'se' < 'curr' then
2196 * the resulting gran will be larger, therefore penalizing the
2197 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2198 * be smaller, again penalizing the lighter task.
2200 * This is especially important for buddies when the leftmost
2201 * task is higher priority than the buddy.
2203 return calc_delta_fair(gran, se);
2207 * Should 'se' preempt 'curr'.
2221 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2223 s64 gran, vdiff = curr->vruntime - se->vruntime;
2228 gran = wakeup_gran(curr, se);
2235 static void set_last_buddy(struct sched_entity *se)
2237 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2240 for_each_sched_entity(se)
2241 cfs_rq_of(se)->last = se;
2244 static void set_next_buddy(struct sched_entity *se)
2246 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2249 for_each_sched_entity(se)
2250 cfs_rq_of(se)->next = se;
2253 static void set_skip_buddy(struct sched_entity *se)
2255 for_each_sched_entity(se)
2256 cfs_rq_of(se)->skip = se;
2260 * Preempt the current task with a newly woken task if needed:
2262 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2264 struct task_struct *curr = rq->curr;
2265 struct sched_entity *se = &curr->se, *pse = &p->se;
2266 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2267 int scale = cfs_rq->nr_running >= sched_nr_latency;
2268 int next_buddy_marked = 0;
2270 if (unlikely(se == pse))
2273 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2274 set_next_buddy(pse);
2275 next_buddy_marked = 1;
2279 * We can come here with TIF_NEED_RESCHED already set from new task
2282 if (test_tsk_need_resched(curr))
2285 /* Idle tasks are by definition preempted by non-idle tasks. */
2286 if (unlikely(curr->policy == SCHED_IDLE) &&
2287 likely(p->policy != SCHED_IDLE))
2291 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2292 * is driven by the tick):
2294 if (unlikely(p->policy != SCHED_NORMAL))
2297 find_matching_se(&se, &pse);
2298 update_curr(cfs_rq_of(se));
2300 if (wakeup_preempt_entity(se, pse) == 1) {
2302 * Bias pick_next to pick the sched entity that is
2303 * triggering this preemption.
2305 if (!next_buddy_marked)
2306 set_next_buddy(pse);
2315 * Only set the backward buddy when the current task is still
2316 * on the rq. This can happen when a wakeup gets interleaved
2317 * with schedule on the ->pre_schedule() or idle_balance()
2318 * point, either of which can * drop the rq lock.
2320 * Also, during early boot the idle thread is in the fair class,
2321 * for obvious reasons its a bad idea to schedule back to it.
2323 if (unlikely(!se->on_rq || curr == rq->idle))
2326 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2330 static struct task_struct *pick_next_task_fair(struct rq *rq)
2332 struct task_struct *p;
2333 struct cfs_rq *cfs_rq = &rq->cfs;
2334 struct sched_entity *se;
2336 if (!cfs_rq->nr_running)
2340 se = pick_next_entity(cfs_rq);
2341 set_next_entity(cfs_rq, se);
2342 cfs_rq = group_cfs_rq(se);
2346 hrtick_start_fair(rq, p);
2352 * Account for a descheduled task:
2354 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
2356 struct sched_entity *se = &prev->se;
2357 struct cfs_rq *cfs_rq;
2359 for_each_sched_entity(se) {
2360 cfs_rq = cfs_rq_of(se);
2361 put_prev_entity(cfs_rq, se);
2366 * sched_yield() is very simple
2368 * The magic of dealing with the ->skip buddy is in pick_next_entity.
2370 static void yield_task_fair(struct rq *rq)
2372 struct task_struct *curr = rq->curr;
2373 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2374 struct sched_entity *se = &curr->se;
2377 * Are we the only task in the tree?
2379 if (unlikely(rq->nr_running == 1))
2382 clear_buddies(cfs_rq, se);
2384 if (curr->policy != SCHED_BATCH) {
2385 update_rq_clock(rq);
2387 * Update run-time statistics of the 'current'.
2389 update_curr(cfs_rq);
2395 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2397 struct sched_entity *se = &p->se;
2402 /* Tell the scheduler that we'd really like pse to run next. */
2405 yield_task_fair(rq);
2411 /**************************************************
2412 * Fair scheduling class load-balancing methods:
2416 * pull_task - move a task from a remote runqueue to the local runqueue.
2417 * Both runqueues must be locked.
2419 static void pull_task(struct rq *src_rq, struct task_struct *p,
2420 struct rq *this_rq, int this_cpu)
2422 deactivate_task(src_rq, p, 0);
2423 set_task_cpu(p, this_cpu);
2424 activate_task(this_rq, p, 0);
2425 check_preempt_curr(this_rq, p, 0);
2429 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2432 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2433 struct sched_domain *sd, enum cpu_idle_type idle,
2436 int tsk_cache_hot = 0;
2438 * We do not migrate tasks that are:
2439 * 1) running (obviously), or
2440 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2441 * 3) are cache-hot on their current CPU.
2443 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2444 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2449 if (task_running(rq, p)) {
2450 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2455 * Aggressive migration if:
2456 * 1) task is cache cold, or
2457 * 2) too many balance attempts have failed.
2460 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2461 if (!tsk_cache_hot ||
2462 sd->nr_balance_failed > sd->cache_nice_tries) {
2463 #ifdef CONFIG_SCHEDSTATS
2464 if (tsk_cache_hot) {
2465 schedstat_inc(sd, lb_hot_gained[idle]);
2466 schedstat_inc(p, se.statistics.nr_forced_migrations);
2472 if (tsk_cache_hot) {
2473 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2480 * move_one_task tries to move exactly one task from busiest to this_rq, as
2481 * part of active balancing operations within "domain".
2482 * Returns 1 if successful and 0 otherwise.
2484 * Called with both runqueues locked.
2487 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2488 struct sched_domain *sd, enum cpu_idle_type idle)
2490 struct task_struct *p, *n;
2491 struct cfs_rq *cfs_rq;
2494 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2495 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2497 if (!can_migrate_task(p, busiest, this_cpu,
2501 pull_task(busiest, p, this_rq, this_cpu);
2503 * Right now, this is only the second place pull_task()
2504 * is called, so we can safely collect pull_task()
2505 * stats here rather than inside pull_task().
2507 schedstat_inc(sd, lb_gained[idle]);
2515 static unsigned long
2516 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2517 unsigned long max_load_move, struct sched_domain *sd,
2518 enum cpu_idle_type idle, int *all_pinned,
2519 struct cfs_rq *busiest_cfs_rq)
2521 int loops = 0, pulled = 0;
2522 long rem_load_move = max_load_move;
2523 struct task_struct *p, *n;
2525 if (max_load_move == 0)
2528 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2529 if (loops++ > sysctl_sched_nr_migrate)
2532 if ((p->se.load.weight >> 1) > rem_load_move ||
2533 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2537 pull_task(busiest, p, this_rq, this_cpu);
2539 rem_load_move -= p->se.load.weight;
2541 #ifdef CONFIG_PREEMPT
2543 * NEWIDLE balancing is a source of latency, so preemptible
2544 * kernels will stop after the first task is pulled to minimize
2545 * the critical section.
2547 if (idle == CPU_NEWLY_IDLE)
2552 * We only want to steal up to the prescribed amount of
2555 if (rem_load_move <= 0)
2560 * Right now, this is one of only two places pull_task() is called,
2561 * so we can safely collect pull_task() stats here rather than
2562 * inside pull_task().
2564 schedstat_add(sd, lb_gained[idle], pulled);
2566 return max_load_move - rem_load_move;
2569 #ifdef CONFIG_FAIR_GROUP_SCHED
2571 * update tg->load_weight by folding this cpu's load_avg
2573 static int update_shares_cpu(struct task_group *tg, int cpu)
2575 struct cfs_rq *cfs_rq;
2576 unsigned long flags;
2583 cfs_rq = tg->cfs_rq[cpu];
2585 raw_spin_lock_irqsave(&rq->lock, flags);
2587 update_rq_clock(rq);
2588 update_cfs_load(cfs_rq, 1);
2591 * We need to update shares after updating tg->load_weight in
2592 * order to adjust the weight of groups with long running tasks.
2594 update_cfs_shares(cfs_rq);
2596 raw_spin_unlock_irqrestore(&rq->lock, flags);
2601 static void update_shares(int cpu)
2603 struct cfs_rq *cfs_rq;
2604 struct rq *rq = cpu_rq(cpu);
2608 * Iterates the task_group tree in a bottom up fashion, see
2609 * list_add_leaf_cfs_rq() for details.
2611 for_each_leaf_cfs_rq(rq, cfs_rq)
2612 update_shares_cpu(cfs_rq->tg, cpu);
2617 * Compute the cpu's hierarchical load factor for each task group.
2618 * This needs to be done in a top-down fashion because the load of a child
2619 * group is a fraction of its parents load.
2621 static int tg_load_down(struct task_group *tg, void *data)
2624 long cpu = (long)data;
2627 load = cpu_rq(cpu)->load.weight;
2629 load = tg->parent->cfs_rq[cpu]->h_load;
2630 load *= tg->se[cpu]->load.weight;
2631 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2634 tg->cfs_rq[cpu]->h_load = load;
2639 static void update_h_load(long cpu)
2641 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2644 static unsigned long
2645 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2646 unsigned long max_load_move,
2647 struct sched_domain *sd, enum cpu_idle_type idle,
2650 long rem_load_move = max_load_move;
2651 struct cfs_rq *busiest_cfs_rq;
2654 update_h_load(cpu_of(busiest));
2656 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
2657 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2658 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2659 u64 rem_load, moved_load;
2664 if (!busiest_cfs_rq->task_weight)
2667 rem_load = (u64)rem_load_move * busiest_weight;
2668 rem_load = div_u64(rem_load, busiest_h_load + 1);
2670 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2671 rem_load, sd, idle, all_pinned,
2677 moved_load *= busiest_h_load;
2678 moved_load = div_u64(moved_load, busiest_weight + 1);
2680 rem_load_move -= moved_load;
2681 if (rem_load_move < 0)
2686 return max_load_move - rem_load_move;
2689 static inline void update_shares(int cpu)
2693 static unsigned long
2694 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2695 unsigned long max_load_move,
2696 struct sched_domain *sd, enum cpu_idle_type idle,
2699 return balance_tasks(this_rq, this_cpu, busiest,
2700 max_load_move, sd, idle, all_pinned,
2706 * move_tasks tries to move up to max_load_move weighted load from busiest to
2707 * this_rq, as part of a balancing operation within domain "sd".
2708 * Returns 1 if successful and 0 otherwise.
2710 * Called with both runqueues locked.
2712 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2713 unsigned long max_load_move,
2714 struct sched_domain *sd, enum cpu_idle_type idle,
2717 unsigned long total_load_moved = 0, load_moved;
2720 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2721 max_load_move - total_load_moved,
2722 sd, idle, all_pinned);
2724 total_load_moved += load_moved;
2726 #ifdef CONFIG_PREEMPT
2728 * NEWIDLE balancing is a source of latency, so preemptible
2729 * kernels will stop after the first task is pulled to minimize
2730 * the critical section.
2732 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2735 if (raw_spin_is_contended(&this_rq->lock) ||
2736 raw_spin_is_contended(&busiest->lock))
2739 } while (load_moved && max_load_move > total_load_moved);
2741 return total_load_moved > 0;
2744 /********** Helpers for find_busiest_group ************************/
2746 * sd_lb_stats - Structure to store the statistics of a sched_domain
2747 * during load balancing.
2749 struct sd_lb_stats {
2750 struct sched_group *busiest; /* Busiest group in this sd */
2751 struct sched_group *this; /* Local group in this sd */
2752 unsigned long total_load; /* Total load of all groups in sd */
2753 unsigned long total_pwr; /* Total power of all groups in sd */
2754 unsigned long avg_load; /* Average load across all groups in sd */
2756 /** Statistics of this group */
2757 unsigned long this_load;
2758 unsigned long this_load_per_task;
2759 unsigned long this_nr_running;
2760 unsigned long this_has_capacity;
2761 unsigned int this_idle_cpus;
2763 /* Statistics of the busiest group */
2764 unsigned int busiest_idle_cpus;
2765 unsigned long max_load;
2766 unsigned long busiest_load_per_task;
2767 unsigned long busiest_nr_running;
2768 unsigned long busiest_group_capacity;
2769 unsigned long busiest_has_capacity;
2770 unsigned int busiest_group_weight;
2772 int group_imb; /* Is there imbalance in this sd */
2773 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2774 int power_savings_balance; /* Is powersave balance needed for this sd */
2775 struct sched_group *group_min; /* Least loaded group in sd */
2776 struct sched_group *group_leader; /* Group which relieves group_min */
2777 unsigned long min_load_per_task; /* load_per_task in group_min */
2778 unsigned long leader_nr_running; /* Nr running of group_leader */
2779 unsigned long min_nr_running; /* Nr running of group_min */
2784 * sg_lb_stats - stats of a sched_group required for load_balancing
2786 struct sg_lb_stats {
2787 unsigned long avg_load; /*Avg load across the CPUs of the group */
2788 unsigned long group_load; /* Total load over the CPUs of the group */
2789 unsigned long sum_nr_running; /* Nr tasks running in the group */
2790 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2791 unsigned long group_capacity;
2792 unsigned long idle_cpus;
2793 unsigned long group_weight;
2794 int group_imb; /* Is there an imbalance in the group ? */
2795 int group_has_capacity; /* Is there extra capacity in the group? */
2799 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2800 * @group: The group whose first cpu is to be returned.
2802 static inline unsigned int group_first_cpu(struct sched_group *group)
2804 return cpumask_first(sched_group_cpus(group));
2808 * get_sd_load_idx - Obtain the load index for a given sched domain.
2809 * @sd: The sched_domain whose load_idx is to be obtained.
2810 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2812 static inline int get_sd_load_idx(struct sched_domain *sd,
2813 enum cpu_idle_type idle)
2819 load_idx = sd->busy_idx;
2822 case CPU_NEWLY_IDLE:
2823 load_idx = sd->newidle_idx;
2826 load_idx = sd->idle_idx;
2834 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2836 * init_sd_power_savings_stats - Initialize power savings statistics for
2837 * the given sched_domain, during load balancing.
2839 * @sd: Sched domain whose power-savings statistics are to be initialized.
2840 * @sds: Variable containing the statistics for sd.
2841 * @idle: Idle status of the CPU at which we're performing load-balancing.
2843 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2844 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2847 * Busy processors will not participate in power savings
2850 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2851 sds->power_savings_balance = 0;
2853 sds->power_savings_balance = 1;
2854 sds->min_nr_running = ULONG_MAX;
2855 sds->leader_nr_running = 0;
2860 * update_sd_power_savings_stats - Update the power saving stats for a
2861 * sched_domain while performing load balancing.
2863 * @group: sched_group belonging to the sched_domain under consideration.
2864 * @sds: Variable containing the statistics of the sched_domain
2865 * @local_group: Does group contain the CPU for which we're performing
2867 * @sgs: Variable containing the statistics of the group.
2869 static inline void update_sd_power_savings_stats(struct sched_group *group,
2870 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2873 if (!sds->power_savings_balance)
2877 * If the local group is idle or completely loaded
2878 * no need to do power savings balance at this domain
2880 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2881 !sds->this_nr_running))
2882 sds->power_savings_balance = 0;
2885 * If a group is already running at full capacity or idle,
2886 * don't include that group in power savings calculations
2888 if (!sds->power_savings_balance ||
2889 sgs->sum_nr_running >= sgs->group_capacity ||
2890 !sgs->sum_nr_running)
2894 * Calculate the group which has the least non-idle load.
2895 * This is the group from where we need to pick up the load
2898 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2899 (sgs->sum_nr_running == sds->min_nr_running &&
2900 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2901 sds->group_min = group;
2902 sds->min_nr_running = sgs->sum_nr_running;
2903 sds->min_load_per_task = sgs->sum_weighted_load /
2904 sgs->sum_nr_running;
2908 * Calculate the group which is almost near its
2909 * capacity but still has some space to pick up some load
2910 * from other group and save more power
2912 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2915 if (sgs->sum_nr_running > sds->leader_nr_running ||
2916 (sgs->sum_nr_running == sds->leader_nr_running &&
2917 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2918 sds->group_leader = group;
2919 sds->leader_nr_running = sgs->sum_nr_running;
2924 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2925 * @sds: Variable containing the statistics of the sched_domain
2926 * under consideration.
2927 * @this_cpu: Cpu at which we're currently performing load-balancing.
2928 * @imbalance: Variable to store the imbalance.
2931 * Check if we have potential to perform some power-savings balance.
2932 * If yes, set the busiest group to be the least loaded group in the
2933 * sched_domain, so that it's CPUs can be put to idle.
2935 * Returns 1 if there is potential to perform power-savings balance.
2938 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2939 int this_cpu, unsigned long *imbalance)
2941 if (!sds->power_savings_balance)
2944 if (sds->this != sds->group_leader ||
2945 sds->group_leader == sds->group_min)
2948 *imbalance = sds->min_load_per_task;
2949 sds->busiest = sds->group_min;
2954 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2955 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2956 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2961 static inline void update_sd_power_savings_stats(struct sched_group *group,
2962 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2967 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2968 int this_cpu, unsigned long *imbalance)
2972 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2975 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2977 return SCHED_POWER_SCALE;
2980 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2982 return default_scale_freq_power(sd, cpu);
2985 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2987 unsigned long weight = sd->span_weight;
2988 unsigned long smt_gain = sd->smt_gain;
2995 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2997 return default_scale_smt_power(sd, cpu);
3000 unsigned long scale_rt_power(int cpu)
3002 struct rq *rq = cpu_rq(cpu);
3003 u64 total, available;
3005 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3007 if (unlikely(total < rq->rt_avg)) {
3008 /* Ensures that power won't end up being negative */
3011 available = total - rq->rt_avg;
3014 if (unlikely((s64)total < SCHED_POWER_SCALE))
3015 total = SCHED_POWER_SCALE;
3017 total >>= SCHED_POWER_SHIFT;
3019 return div_u64(available, total);
3022 static void update_cpu_power(struct sched_domain *sd, int cpu)
3024 unsigned long weight = sd->span_weight;
3025 unsigned long power = SCHED_POWER_SCALE;
3026 struct sched_group *sdg = sd->groups;
3028 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3029 if (sched_feat(ARCH_POWER))
3030 power *= arch_scale_smt_power(sd, cpu);
3032 power *= default_scale_smt_power(sd, cpu);
3034 power >>= SCHED_POWER_SHIFT;
3037 sdg->sgp->power_orig = power;
3039 if (sched_feat(ARCH_POWER))
3040 power *= arch_scale_freq_power(sd, cpu);
3042 power *= default_scale_freq_power(sd, cpu);
3044 power >>= SCHED_POWER_SHIFT;
3046 power *= scale_rt_power(cpu);
3047 power >>= SCHED_POWER_SHIFT;
3052 cpu_rq(cpu)->cpu_power = power;
3053 sdg->sgp->power = power;
3056 static void update_group_power(struct sched_domain *sd, int cpu)
3058 struct sched_domain *child = sd->child;
3059 struct sched_group *group, *sdg = sd->groups;
3060 unsigned long power;
3063 update_cpu_power(sd, cpu);
3069 group = child->groups;
3071 power += group->sgp->power;
3072 group = group->next;
3073 } while (group != child->groups);
3075 sdg->sgp->power = power;
3079 * Try and fix up capacity for tiny siblings, this is needed when
3080 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3081 * which on its own isn't powerful enough.
3083 * See update_sd_pick_busiest() and check_asym_packing().
3086 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3089 * Only siblings can have significantly less than SCHED_POWER_SCALE
3091 if (!(sd->flags & SD_SHARE_CPUPOWER))
3095 * If ~90% of the cpu_power is still there, we're good.
3097 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3104 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3105 * @sd: The sched_domain whose statistics are to be updated.
3106 * @group: sched_group whose statistics are to be updated.
3107 * @this_cpu: Cpu for which load balance is currently performed.
3108 * @idle: Idle status of this_cpu
3109 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3110 * @local_group: Does group contain this_cpu.
3111 * @cpus: Set of cpus considered for load balancing.
3112 * @balance: Should we balance.
3113 * @sgs: variable to hold the statistics for this group.
3115 static inline void update_sg_lb_stats(struct sched_domain *sd,
3116 struct sched_group *group, int this_cpu,
3117 enum cpu_idle_type idle, int load_idx,
3118 int local_group, const struct cpumask *cpus,
3119 int *balance, struct sg_lb_stats *sgs)
3121 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3123 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3124 unsigned long avg_load_per_task = 0;
3127 balance_cpu = group_first_cpu(group);
3129 /* Tally up the load of all CPUs in the group */
3131 min_cpu_load = ~0UL;
3134 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3135 struct rq *rq = cpu_rq(i);
3137 /* Bias balancing toward cpus of our domain */
3139 if (idle_cpu(i) && !first_idle_cpu) {
3144 load = target_load(i, load_idx);
3146 load = source_load(i, load_idx);
3147 if (load > max_cpu_load) {
3148 max_cpu_load = load;
3149 max_nr_running = rq->nr_running;
3151 if (min_cpu_load > load)
3152 min_cpu_load = load;
3155 sgs->group_load += load;
3156 sgs->sum_nr_running += rq->nr_running;
3157 sgs->sum_weighted_load += weighted_cpuload(i);
3163 * First idle cpu or the first cpu(busiest) in this sched group
3164 * is eligible for doing load balancing at this and above
3165 * domains. In the newly idle case, we will allow all the cpu's
3166 * to do the newly idle load balance.
3168 if (idle != CPU_NEWLY_IDLE && local_group) {
3169 if (balance_cpu != this_cpu) {
3173 update_group_power(sd, this_cpu);
3176 /* Adjust by relative CPU power of the group */
3177 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3180 * Consider the group unbalanced when the imbalance is larger
3181 * than the average weight of a task.
3183 * APZ: with cgroup the avg task weight can vary wildly and
3184 * might not be a suitable number - should we keep a
3185 * normalized nr_running number somewhere that negates
3188 if (sgs->sum_nr_running)
3189 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3191 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3194 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3196 if (!sgs->group_capacity)
3197 sgs->group_capacity = fix_small_capacity(sd, group);
3198 sgs->group_weight = group->group_weight;
3200 if (sgs->group_capacity > sgs->sum_nr_running)
3201 sgs->group_has_capacity = 1;
3205 * update_sd_pick_busiest - return 1 on busiest group
3206 * @sd: sched_domain whose statistics are to be checked
3207 * @sds: sched_domain statistics
3208 * @sg: sched_group candidate to be checked for being the busiest
3209 * @sgs: sched_group statistics
3210 * @this_cpu: the current cpu
3212 * Determine if @sg is a busier group than the previously selected
3215 static bool update_sd_pick_busiest(struct sched_domain *sd,
3216 struct sd_lb_stats *sds,
3217 struct sched_group *sg,
3218 struct sg_lb_stats *sgs,
3221 if (sgs->avg_load <= sds->max_load)
3224 if (sgs->sum_nr_running > sgs->group_capacity)
3231 * ASYM_PACKING needs to move all the work to the lowest
3232 * numbered CPUs in the group, therefore mark all groups
3233 * higher than ourself as busy.
3235 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3236 this_cpu < group_first_cpu(sg)) {
3240 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3248 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3249 * @sd: sched_domain whose statistics are to be updated.
3250 * @this_cpu: Cpu for which load balance is currently performed.
3251 * @idle: Idle status of this_cpu
3252 * @cpus: Set of cpus considered for load balancing.
3253 * @balance: Should we balance.
3254 * @sds: variable to hold the statistics for this sched_domain.
3256 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3257 enum cpu_idle_type idle, const struct cpumask *cpus,
3258 int *balance, struct sd_lb_stats *sds)
3260 struct sched_domain *child = sd->child;
3261 struct sched_group *sg = sd->groups;
3262 struct sg_lb_stats sgs;
3263 int load_idx, prefer_sibling = 0;
3265 if (child && child->flags & SD_PREFER_SIBLING)
3268 init_sd_power_savings_stats(sd, sds, idle);
3269 load_idx = get_sd_load_idx(sd, idle);
3274 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3275 memset(&sgs, 0, sizeof(sgs));
3276 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3277 local_group, cpus, balance, &sgs);
3279 if (local_group && !(*balance))
3282 sds->total_load += sgs.group_load;
3283 sds->total_pwr += sg->sgp->power;
3286 * In case the child domain prefers tasks go to siblings
3287 * first, lower the sg capacity to one so that we'll try
3288 * and move all the excess tasks away. We lower the capacity
3289 * of a group only if the local group has the capacity to fit
3290 * these excess tasks, i.e. nr_running < group_capacity. The
3291 * extra check prevents the case where you always pull from the
3292 * heaviest group when it is already under-utilized (possible
3293 * with a large weight task outweighs the tasks on the system).
3295 if (prefer_sibling && !local_group && sds->this_has_capacity)
3296 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3299 sds->this_load = sgs.avg_load;
3301 sds->this_nr_running = sgs.sum_nr_running;
3302 sds->this_load_per_task = sgs.sum_weighted_load;
3303 sds->this_has_capacity = sgs.group_has_capacity;
3304 sds->this_idle_cpus = sgs.idle_cpus;
3305 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
3306 sds->max_load = sgs.avg_load;
3308 sds->busiest_nr_running = sgs.sum_nr_running;
3309 sds->busiest_idle_cpus = sgs.idle_cpus;
3310 sds->busiest_group_capacity = sgs.group_capacity;
3311 sds->busiest_load_per_task = sgs.sum_weighted_load;
3312 sds->busiest_has_capacity = sgs.group_has_capacity;
3313 sds->busiest_group_weight = sgs.group_weight;
3314 sds->group_imb = sgs.group_imb;
3317 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
3319 } while (sg != sd->groups);
3322 int __weak arch_sd_sibling_asym_packing(void)
3324 return 0*SD_ASYM_PACKING;
3328 * check_asym_packing - Check to see if the group is packed into the
3331 * This is primarily intended to used at the sibling level. Some
3332 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3333 * case of POWER7, it can move to lower SMT modes only when higher
3334 * threads are idle. When in lower SMT modes, the threads will
3335 * perform better since they share less core resources. Hence when we
3336 * have idle threads, we want them to be the higher ones.
3338 * This packing function is run on idle threads. It checks to see if
3339 * the busiest CPU in this domain (core in the P7 case) has a higher
3340 * CPU number than the packing function is being run on. Here we are
3341 * assuming lower CPU number will be equivalent to lower a SMT thread
3344 * Returns 1 when packing is required and a task should be moved to
3345 * this CPU. The amount of the imbalance is returned in *imbalance.
3347 * @sd: The sched_domain whose packing is to be checked.
3348 * @sds: Statistics of the sched_domain which is to be packed
3349 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3350 * @imbalance: returns amount of imbalanced due to packing.
3352 static int check_asym_packing(struct sched_domain *sd,
3353 struct sd_lb_stats *sds,
3354 int this_cpu, unsigned long *imbalance)
3358 if (!(sd->flags & SD_ASYM_PACKING))
3364 busiest_cpu = group_first_cpu(sds->busiest);
3365 if (this_cpu > busiest_cpu)
3368 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
3374 * fix_small_imbalance - Calculate the minor imbalance that exists
3375 * amongst the groups of a sched_domain, during
3377 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3378 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3379 * @imbalance: Variable to store the imbalance.
3381 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3382 int this_cpu, unsigned long *imbalance)
3384 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3385 unsigned int imbn = 2;
3386 unsigned long scaled_busy_load_per_task;
3388 if (sds->this_nr_running) {
3389 sds->this_load_per_task /= sds->this_nr_running;
3390 if (sds->busiest_load_per_task >
3391 sds->this_load_per_task)
3394 sds->this_load_per_task =
3395 cpu_avg_load_per_task(this_cpu);
3397 scaled_busy_load_per_task = sds->busiest_load_per_task
3398 * SCHED_POWER_SCALE;
3399 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3401 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3402 (scaled_busy_load_per_task * imbn)) {
3403 *imbalance = sds->busiest_load_per_task;
3408 * OK, we don't have enough imbalance to justify moving tasks,
3409 * however we may be able to increase total CPU power used by
3413 pwr_now += sds->busiest->sgp->power *
3414 min(sds->busiest_load_per_task, sds->max_load);
3415 pwr_now += sds->this->sgp->power *
3416 min(sds->this_load_per_task, sds->this_load);
3417 pwr_now /= SCHED_POWER_SCALE;
3419 /* Amount of load we'd subtract */
3420 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3421 sds->busiest->sgp->power;
3422 if (sds->max_load > tmp)
3423 pwr_move += sds->busiest->sgp->power *
3424 min(sds->busiest_load_per_task, sds->max_load - tmp);
3426 /* Amount of load we'd add */
3427 if (sds->max_load * sds->busiest->sgp->power <
3428 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3429 tmp = (sds->max_load * sds->busiest->sgp->power) /
3430 sds->this->sgp->power;
3432 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3433 sds->this->sgp->power;
3434 pwr_move += sds->this->sgp->power *
3435 min(sds->this_load_per_task, sds->this_load + tmp);
3436 pwr_move /= SCHED_POWER_SCALE;
3438 /* Move if we gain throughput */
3439 if (pwr_move > pwr_now)
3440 *imbalance = sds->busiest_load_per_task;
3444 * calculate_imbalance - Calculate the amount of imbalance present within the
3445 * groups of a given sched_domain during load balance.
3446 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3447 * @this_cpu: Cpu for which currently load balance is being performed.
3448 * @imbalance: The variable to store the imbalance.
3450 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3451 unsigned long *imbalance)
3453 unsigned long max_pull, load_above_capacity = ~0UL;
3455 sds->busiest_load_per_task /= sds->busiest_nr_running;
3456 if (sds->group_imb) {
3457 sds->busiest_load_per_task =
3458 min(sds->busiest_load_per_task, sds->avg_load);
3462 * In the presence of smp nice balancing, certain scenarios can have
3463 * max load less than avg load(as we skip the groups at or below
3464 * its cpu_power, while calculating max_load..)
3466 if (sds->max_load < sds->avg_load) {
3468 return fix_small_imbalance(sds, this_cpu, imbalance);
3471 if (!sds->group_imb) {
3473 * Don't want to pull so many tasks that a group would go idle.
3475 load_above_capacity = (sds->busiest_nr_running -
3476 sds->busiest_group_capacity);
3478 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3480 load_above_capacity /= sds->busiest->sgp->power;
3484 * We're trying to get all the cpus to the average_load, so we don't
3485 * want to push ourselves above the average load, nor do we wish to
3486 * reduce the max loaded cpu below the average load. At the same time,
3487 * we also don't want to reduce the group load below the group capacity
3488 * (so that we can implement power-savings policies etc). Thus we look
3489 * for the minimum possible imbalance.
3490 * Be careful of negative numbers as they'll appear as very large values
3491 * with unsigned longs.
3493 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3495 /* How much load to actually move to equalise the imbalance */
3496 *imbalance = min(max_pull * sds->busiest->sgp->power,
3497 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3498 / SCHED_POWER_SCALE;
3501 * if *imbalance is less than the average load per runnable task
3502 * there is no guarantee that any tasks will be moved so we'll have
3503 * a think about bumping its value to force at least one task to be
3506 if (*imbalance < sds->busiest_load_per_task)
3507 return fix_small_imbalance(sds, this_cpu, imbalance);
3511 /******* find_busiest_group() helpers end here *********************/
3514 * find_busiest_group - Returns the busiest group within the sched_domain
3515 * if there is an imbalance. If there isn't an imbalance, and
3516 * the user has opted for power-savings, it returns a group whose
3517 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3518 * such a group exists.
3520 * Also calculates the amount of weighted load which should be moved
3521 * to restore balance.
3523 * @sd: The sched_domain whose busiest group is to be returned.
3524 * @this_cpu: The cpu for which load balancing is currently being performed.
3525 * @imbalance: Variable which stores amount of weighted load which should
3526 * be moved to restore balance/put a group to idle.
3527 * @idle: The idle status of this_cpu.
3528 * @cpus: The set of CPUs under consideration for load-balancing.
3529 * @balance: Pointer to a variable indicating if this_cpu
3530 * is the appropriate cpu to perform load balancing at this_level.
3532 * Returns: - the busiest group if imbalance exists.
3533 * - If no imbalance and user has opted for power-savings balance,
3534 * return the least loaded group whose CPUs can be
3535 * put to idle by rebalancing its tasks onto our group.
3537 static struct sched_group *
3538 find_busiest_group(struct sched_domain *sd, int this_cpu,
3539 unsigned long *imbalance, enum cpu_idle_type idle,
3540 const struct cpumask *cpus, int *balance)
3542 struct sd_lb_stats sds;
3544 memset(&sds, 0, sizeof(sds));
3547 * Compute the various statistics relavent for load balancing at
3550 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3553 * this_cpu is not the appropriate cpu to perform load balancing at
3559 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3560 check_asym_packing(sd, &sds, this_cpu, imbalance))
3563 /* There is no busy sibling group to pull tasks from */
3564 if (!sds.busiest || sds.busiest_nr_running == 0)
3567 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3570 * If the busiest group is imbalanced the below checks don't
3571 * work because they assumes all things are equal, which typically
3572 * isn't true due to cpus_allowed constraints and the like.
3577 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3578 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3579 !sds.busiest_has_capacity)
3583 * If the local group is more busy than the selected busiest group
3584 * don't try and pull any tasks.
3586 if (sds.this_load >= sds.max_load)
3590 * Don't pull any tasks if this group is already above the domain
3593 if (sds.this_load >= sds.avg_load)
3596 if (idle == CPU_IDLE) {
3598 * This cpu is idle. If the busiest group load doesn't
3599 * have more tasks than the number of available cpu's and
3600 * there is no imbalance between this and busiest group
3601 * wrt to idle cpu's, it is balanced.
3603 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3604 sds.busiest_nr_running <= sds.busiest_group_weight)
3608 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3609 * imbalance_pct to be conservative.
3611 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3616 /* Looks like there is an imbalance. Compute it */
3617 calculate_imbalance(&sds, this_cpu, imbalance);
3622 * There is no obvious imbalance. But check if we can do some balancing
3625 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3633 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3636 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3637 enum cpu_idle_type idle, unsigned long imbalance,
3638 const struct cpumask *cpus)
3640 struct rq *busiest = NULL, *rq;
3641 unsigned long max_load = 0;
3644 for_each_cpu(i, sched_group_cpus(group)) {
3645 unsigned long power = power_of(i);
3646 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3651 capacity = fix_small_capacity(sd, group);
3653 if (!cpumask_test_cpu(i, cpus))
3657 wl = weighted_cpuload(i);
3660 * When comparing with imbalance, use weighted_cpuload()
3661 * which is not scaled with the cpu power.
3663 if (capacity && rq->nr_running == 1 && wl > imbalance)
3667 * For the load comparisons with the other cpu's, consider
3668 * the weighted_cpuload() scaled with the cpu power, so that
3669 * the load can be moved away from the cpu that is potentially
3670 * running at a lower capacity.
3672 wl = (wl * SCHED_POWER_SCALE) / power;
3674 if (wl > max_load) {
3684 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3685 * so long as it is large enough.
3687 #define MAX_PINNED_INTERVAL 512
3689 /* Working cpumask for load_balance and load_balance_newidle. */
3690 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3692 static int need_active_balance(struct sched_domain *sd, int idle,
3693 int busiest_cpu, int this_cpu)
3695 if (idle == CPU_NEWLY_IDLE) {
3698 * ASYM_PACKING needs to force migrate tasks from busy but
3699 * higher numbered CPUs in order to pack all tasks in the
3700 * lowest numbered CPUs.
3702 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3706 * The only task running in a non-idle cpu can be moved to this
3707 * cpu in an attempt to completely freeup the other CPU
3710 * The package power saving logic comes from
3711 * find_busiest_group(). If there are no imbalance, then
3712 * f_b_g() will return NULL. However when sched_mc={1,2} then
3713 * f_b_g() will select a group from which a running task may be
3714 * pulled to this cpu in order to make the other package idle.
3715 * If there is no opportunity to make a package idle and if
3716 * there are no imbalance, then f_b_g() will return NULL and no
3717 * action will be taken in load_balance_newidle().
3719 * Under normal task pull operation due to imbalance, there
3720 * will be more than one task in the source run queue and
3721 * move_tasks() will succeed. ld_moved will be true and this
3722 * active balance code will not be triggered.
3724 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3728 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3731 static int active_load_balance_cpu_stop(void *data);
3734 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3735 * tasks if there is an imbalance.
3737 static int load_balance(int this_cpu, struct rq *this_rq,
3738 struct sched_domain *sd, enum cpu_idle_type idle,
3741 int ld_moved, all_pinned = 0, active_balance = 0;
3742 struct sched_group *group;
3743 unsigned long imbalance;
3745 unsigned long flags;
3746 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3748 cpumask_copy(cpus, cpu_active_mask);
3750 schedstat_inc(sd, lb_count[idle]);
3753 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
3760 schedstat_inc(sd, lb_nobusyg[idle]);
3764 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3766 schedstat_inc(sd, lb_nobusyq[idle]);
3770 BUG_ON(busiest == this_rq);
3772 schedstat_add(sd, lb_imbalance[idle], imbalance);
3775 if (busiest->nr_running > 1) {
3777 * Attempt to move tasks. If find_busiest_group has found
3778 * an imbalance but busiest->nr_running <= 1, the group is
3779 * still unbalanced. ld_moved simply stays zero, so it is
3780 * correctly treated as an imbalance.
3783 local_irq_save(flags);
3784 double_rq_lock(this_rq, busiest);
3785 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3786 imbalance, sd, idle, &all_pinned);
3787 double_rq_unlock(this_rq, busiest);
3788 local_irq_restore(flags);
3791 * some other cpu did the load balance for us.
3793 if (ld_moved && this_cpu != smp_processor_id())
3794 resched_cpu(this_cpu);
3796 /* All tasks on this runqueue were pinned by CPU affinity */
3797 if (unlikely(all_pinned)) {
3798 cpumask_clear_cpu(cpu_of(busiest), cpus);
3799 if (!cpumask_empty(cpus))
3806 schedstat_inc(sd, lb_failed[idle]);
3808 * Increment the failure counter only on periodic balance.
3809 * We do not want newidle balance, which can be very
3810 * frequent, pollute the failure counter causing
3811 * excessive cache_hot migrations and active balances.
3813 if (idle != CPU_NEWLY_IDLE)
3814 sd->nr_balance_failed++;
3816 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
3817 raw_spin_lock_irqsave(&busiest->lock, flags);
3819 /* don't kick the active_load_balance_cpu_stop,
3820 * if the curr task on busiest cpu can't be
3823 if (!cpumask_test_cpu(this_cpu,
3824 &busiest->curr->cpus_allowed)) {
3825 raw_spin_unlock_irqrestore(&busiest->lock,
3828 goto out_one_pinned;
3832 * ->active_balance synchronizes accesses to
3833 * ->active_balance_work. Once set, it's cleared
3834 * only after active load balance is finished.
3836 if (!busiest->active_balance) {
3837 busiest->active_balance = 1;
3838 busiest->push_cpu = this_cpu;
3841 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3844 stop_one_cpu_nowait(cpu_of(busiest),
3845 active_load_balance_cpu_stop, busiest,
3846 &busiest->active_balance_work);
3849 * We've kicked active balancing, reset the failure
3852 sd->nr_balance_failed = sd->cache_nice_tries+1;
3855 sd->nr_balance_failed = 0;
3857 if (likely(!active_balance)) {
3858 /* We were unbalanced, so reset the balancing interval */
3859 sd->balance_interval = sd->min_interval;
3862 * If we've begun active balancing, start to back off. This
3863 * case may not be covered by the all_pinned logic if there
3864 * is only 1 task on the busy runqueue (because we don't call
3867 if (sd->balance_interval < sd->max_interval)
3868 sd->balance_interval *= 2;
3874 schedstat_inc(sd, lb_balanced[idle]);
3876 sd->nr_balance_failed = 0;
3879 /* tune up the balancing interval */
3880 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3881 (sd->balance_interval < sd->max_interval))
3882 sd->balance_interval *= 2;
3890 * idle_balance is called by schedule() if this_cpu is about to become
3891 * idle. Attempts to pull tasks from other CPUs.
3893 static void idle_balance(int this_cpu, struct rq *this_rq)
3895 struct sched_domain *sd;
3896 int pulled_task = 0;
3897 unsigned long next_balance = jiffies + HZ;
3899 this_rq->idle_stamp = this_rq->clock;
3901 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3905 * Drop the rq->lock, but keep IRQ/preempt disabled.
3907 raw_spin_unlock(&this_rq->lock);
3909 update_shares(this_cpu);
3911 for_each_domain(this_cpu, sd) {
3912 unsigned long interval;
3915 if (!(sd->flags & SD_LOAD_BALANCE))
3918 if (sd->flags & SD_BALANCE_NEWIDLE) {
3919 /* If we've pulled tasks over stop searching: */
3920 pulled_task = load_balance(this_cpu, this_rq,
3921 sd, CPU_NEWLY_IDLE, &balance);
3924 interval = msecs_to_jiffies(sd->balance_interval);
3925 if (time_after(next_balance, sd->last_balance + interval))
3926 next_balance = sd->last_balance + interval;
3928 this_rq->idle_stamp = 0;
3934 raw_spin_lock(&this_rq->lock);
3936 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3938 * We are going idle. next_balance may be set based on
3939 * a busy processor. So reset next_balance.
3941 this_rq->next_balance = next_balance;
3946 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3947 * running tasks off the busiest CPU onto idle CPUs. It requires at
3948 * least 1 task to be running on each physical CPU where possible, and
3949 * avoids physical / logical imbalances.
3951 static int active_load_balance_cpu_stop(void *data)
3953 struct rq *busiest_rq = data;
3954 int busiest_cpu = cpu_of(busiest_rq);
3955 int target_cpu = busiest_rq->push_cpu;
3956 struct rq *target_rq = cpu_rq(target_cpu);
3957 struct sched_domain *sd;
3959 raw_spin_lock_irq(&busiest_rq->lock);
3961 /* make sure the requested cpu hasn't gone down in the meantime */
3962 if (unlikely(busiest_cpu != smp_processor_id() ||
3963 !busiest_rq->active_balance))
3966 /* Is there any task to move? */
3967 if (busiest_rq->nr_running <= 1)
3971 * This condition is "impossible", if it occurs
3972 * we need to fix it. Originally reported by
3973 * Bjorn Helgaas on a 128-cpu setup.
3975 BUG_ON(busiest_rq == target_rq);
3977 /* move a task from busiest_rq to target_rq */
3978 double_lock_balance(busiest_rq, target_rq);
3980 /* Search for an sd spanning us and the target CPU. */
3982 for_each_domain(target_cpu, sd) {
3983 if ((sd->flags & SD_LOAD_BALANCE) &&
3984 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3989 schedstat_inc(sd, alb_count);
3991 if (move_one_task(target_rq, target_cpu, busiest_rq,
3993 schedstat_inc(sd, alb_pushed);
3995 schedstat_inc(sd, alb_failed);
3998 double_unlock_balance(busiest_rq, target_rq);
4000 busiest_rq->active_balance = 0;
4001 raw_spin_unlock_irq(&busiest_rq->lock);
4007 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
4009 static void trigger_sched_softirq(void *data)
4011 raise_softirq_irqoff(SCHED_SOFTIRQ);
4014 static inline void init_sched_softirq_csd(struct call_single_data *csd)
4016 csd->func = trigger_sched_softirq;
4023 * idle load balancing details
4024 * - One of the idle CPUs nominates itself as idle load_balancer, while
4026 * - This idle load balancer CPU will also go into tickless mode when
4027 * it is idle, just like all other idle CPUs
4028 * - When one of the busy CPUs notice that there may be an idle rebalancing
4029 * needed, they will kick the idle load balancer, which then does idle
4030 * load balancing for all the idle CPUs.
4033 atomic_t load_balancer;
4034 atomic_t first_pick_cpu;
4035 atomic_t second_pick_cpu;
4036 cpumask_var_t idle_cpus_mask;
4037 cpumask_var_t grp_idle_mask;
4038 unsigned long next_balance; /* in jiffy units */
4039 } nohz ____cacheline_aligned;
4041 int get_nohz_load_balancer(void)
4043 return atomic_read(&nohz.load_balancer);
4046 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4048 * lowest_flag_domain - Return lowest sched_domain containing flag.
4049 * @cpu: The cpu whose lowest level of sched domain is to
4051 * @flag: The flag to check for the lowest sched_domain
4052 * for the given cpu.
4054 * Returns the lowest sched_domain of a cpu which contains the given flag.
4056 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4058 struct sched_domain *sd;
4060 for_each_domain(cpu, sd)
4061 if (sd->flags & flag)
4068 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4069 * @cpu: The cpu whose domains we're iterating over.
4070 * @sd: variable holding the value of the power_savings_sd
4072 * @flag: The flag to filter the sched_domains to be iterated.
4074 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4075 * set, starting from the lowest sched_domain to the highest.
4077 #define for_each_flag_domain(cpu, sd, flag) \
4078 for (sd = lowest_flag_domain(cpu, flag); \
4079 (sd && (sd->flags & flag)); sd = sd->parent)
4082 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4083 * @ilb_group: group to be checked for semi-idleness
4085 * Returns: 1 if the group is semi-idle. 0 otherwise.
4087 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4088 * and atleast one non-idle CPU. This helper function checks if the given
4089 * sched_group is semi-idle or not.
4091 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4093 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
4094 sched_group_cpus(ilb_group));
4097 * A sched_group is semi-idle when it has atleast one busy cpu
4098 * and atleast one idle cpu.
4100 if (cpumask_empty(nohz.grp_idle_mask))
4103 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
4109 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4110 * @cpu: The cpu which is nominating a new idle_load_balancer.
4112 * Returns: Returns the id of the idle load balancer if it exists,
4113 * Else, returns >= nr_cpu_ids.
4115 * This algorithm picks the idle load balancer such that it belongs to a
4116 * semi-idle powersavings sched_domain. The idea is to try and avoid
4117 * completely idle packages/cores just for the purpose of idle load balancing
4118 * when there are other idle cpu's which are better suited for that job.
4120 static int find_new_ilb(int cpu)
4122 struct sched_domain *sd;
4123 struct sched_group *ilb_group;
4124 int ilb = nr_cpu_ids;
4127 * Have idle load balancer selection from semi-idle packages only
4128 * when power-aware load balancing is enabled
4130 if (!(sched_smt_power_savings || sched_mc_power_savings))
4134 * Optimize for the case when we have no idle CPUs or only one
4135 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4137 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4141 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4142 ilb_group = sd->groups;
4145 if (is_semi_idle_group(ilb_group)) {
4146 ilb = cpumask_first(nohz.grp_idle_mask);
4150 ilb_group = ilb_group->next;
4152 } while (ilb_group != sd->groups);
4160 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4161 static inline int find_new_ilb(int call_cpu)
4168 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4169 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4170 * CPU (if there is one).
4172 static void nohz_balancer_kick(int cpu)
4176 nohz.next_balance++;
4178 ilb_cpu = get_nohz_load_balancer();
4180 if (ilb_cpu >= nr_cpu_ids) {
4181 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4182 if (ilb_cpu >= nr_cpu_ids)
4186 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4187 struct call_single_data *cp;
4189 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4190 cp = &per_cpu(remote_sched_softirq_cb, cpu);
4191 __smp_call_function_single(ilb_cpu, cp, 0);
4197 * This routine will try to nominate the ilb (idle load balancing)
4198 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4199 * load balancing on behalf of all those cpus.
4201 * When the ilb owner becomes busy, we will not have new ilb owner until some
4202 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4203 * idle load balancing by kicking one of the idle CPUs.
4205 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4206 * ilb owner CPU in future (when there is a need for idle load balancing on
4207 * behalf of all idle CPUs).
4209 void select_nohz_load_balancer(int stop_tick)
4211 int cpu = smp_processor_id();
4214 if (!cpu_active(cpu)) {
4215 if (atomic_read(&nohz.load_balancer) != cpu)
4219 * If we are going offline and still the leader,
4222 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4229 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4231 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4232 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4233 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4234 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4236 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4239 /* make me the ilb owner */
4240 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4245 * Check to see if there is a more power-efficient
4248 new_ilb = find_new_ilb(cpu);
4249 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4250 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4251 resched_cpu(new_ilb);
4257 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4260 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4262 if (atomic_read(&nohz.load_balancer) == cpu)
4263 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4271 static DEFINE_SPINLOCK(balancing);
4273 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4276 * Scale the max load_balance interval with the number of CPUs in the system.
4277 * This trades load-balance latency on larger machines for less cross talk.
4279 static void update_max_interval(void)
4281 max_load_balance_interval = HZ*num_online_cpus()/10;
4285 * It checks each scheduling domain to see if it is due to be balanced,
4286 * and initiates a balancing operation if so.
4288 * Balancing parameters are set up in arch_init_sched_domains.
4290 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4293 struct rq *rq = cpu_rq(cpu);
4294 unsigned long interval;
4295 struct sched_domain *sd;
4296 /* Earliest time when we have to do rebalance again */
4297 unsigned long next_balance = jiffies + 60*HZ;
4298 int update_next_balance = 0;
4304 for_each_domain(cpu, sd) {
4305 if (!(sd->flags & SD_LOAD_BALANCE))
4308 interval = sd->balance_interval;
4309 if (idle != CPU_IDLE)
4310 interval *= sd->busy_factor;
4312 /* scale ms to jiffies */
4313 interval = msecs_to_jiffies(interval);
4314 interval = clamp(interval, 1UL, max_load_balance_interval);
4316 need_serialize = sd->flags & SD_SERIALIZE;
4318 if (need_serialize) {
4319 if (!spin_trylock(&balancing))
4323 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4324 if (load_balance(cpu, rq, sd, idle, &balance)) {
4326 * We've pulled tasks over so either we're no
4329 idle = CPU_NOT_IDLE;
4331 sd->last_balance = jiffies;
4334 spin_unlock(&balancing);
4336 if (time_after(next_balance, sd->last_balance + interval)) {
4337 next_balance = sd->last_balance + interval;
4338 update_next_balance = 1;
4342 * Stop the load balance at this level. There is another
4343 * CPU in our sched group which is doing load balancing more
4352 * next_balance will be updated only when there is a need.
4353 * When the cpu is attached to null domain for ex, it will not be
4356 if (likely(update_next_balance))
4357 rq->next_balance = next_balance;
4362 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4363 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4365 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4367 struct rq *this_rq = cpu_rq(this_cpu);
4371 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
4374 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4375 if (balance_cpu == this_cpu)
4379 * If this cpu gets work to do, stop the load balancing
4380 * work being done for other cpus. Next load
4381 * balancing owner will pick it up.
4383 if (need_resched()) {
4384 this_rq->nohz_balance_kick = 0;
4388 raw_spin_lock_irq(&this_rq->lock);
4389 update_rq_clock(this_rq);
4390 update_cpu_load(this_rq);
4391 raw_spin_unlock_irq(&this_rq->lock);
4393 rebalance_domains(balance_cpu, CPU_IDLE);
4395 rq = cpu_rq(balance_cpu);
4396 if (time_after(this_rq->next_balance, rq->next_balance))
4397 this_rq->next_balance = rq->next_balance;
4399 nohz.next_balance = this_rq->next_balance;
4400 this_rq->nohz_balance_kick = 0;
4404 * Current heuristic for kicking the idle load balancer
4405 * - first_pick_cpu is the one of the busy CPUs. It will kick
4406 * idle load balancer when it has more than one process active. This
4407 * eliminates the need for idle load balancing altogether when we have
4408 * only one running process in the system (common case).
4409 * - If there are more than one busy CPU, idle load balancer may have
4410 * to run for active_load_balance to happen (i.e., two busy CPUs are
4411 * SMT or core siblings and can run better if they move to different
4412 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4413 * which will kick idle load balancer as soon as it has any load.
4415 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4417 unsigned long now = jiffies;
4419 int first_pick_cpu, second_pick_cpu;
4421 if (time_before(now, nohz.next_balance))
4424 if (rq->idle_at_tick)
4427 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4428 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4430 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4431 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4434 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4435 if (ret == nr_cpu_ids || ret == cpu) {
4436 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4437 if (rq->nr_running > 1)
4440 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4441 if (ret == nr_cpu_ids || ret == cpu) {
4449 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4453 * run_rebalance_domains is triggered when needed from the scheduler tick.
4454 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4456 static void run_rebalance_domains(struct softirq_action *h)
4458 int this_cpu = smp_processor_id();
4459 struct rq *this_rq = cpu_rq(this_cpu);
4460 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4461 CPU_IDLE : CPU_NOT_IDLE;
4463 rebalance_domains(this_cpu, idle);
4466 * If this cpu has a pending nohz_balance_kick, then do the
4467 * balancing on behalf of the other idle cpus whose ticks are
4470 nohz_idle_balance(this_cpu, idle);
4473 static inline int on_null_domain(int cpu)
4475 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4479 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4481 static inline void trigger_load_balance(struct rq *rq, int cpu)
4483 /* Don't need to rebalance while attached to NULL domain */
4484 if (time_after_eq(jiffies, rq->next_balance) &&
4485 likely(!on_null_domain(cpu)))
4486 raise_softirq(SCHED_SOFTIRQ);
4488 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4489 nohz_balancer_kick(cpu);
4493 static void rq_online_fair(struct rq *rq)
4498 static void rq_offline_fair(struct rq *rq)
4503 #else /* CONFIG_SMP */
4506 * on UP we do not need to balance between CPUs:
4508 static inline void idle_balance(int cpu, struct rq *rq)
4512 #endif /* CONFIG_SMP */
4515 * scheduler tick hitting a task of our scheduling class:
4517 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4519 struct cfs_rq *cfs_rq;
4520 struct sched_entity *se = &curr->se;
4522 for_each_sched_entity(se) {
4523 cfs_rq = cfs_rq_of(se);
4524 entity_tick(cfs_rq, se, queued);
4529 * called on fork with the child task as argument from the parent's context
4530 * - child not yet on the tasklist
4531 * - preemption disabled
4533 static void task_fork_fair(struct task_struct *p)
4535 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4536 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4537 int this_cpu = smp_processor_id();
4538 struct rq *rq = this_rq();
4539 unsigned long flags;
4541 raw_spin_lock_irqsave(&rq->lock, flags);
4543 update_rq_clock(rq);
4545 if (unlikely(task_cpu(p) != this_cpu)) {
4547 __set_task_cpu(p, this_cpu);
4551 update_curr(cfs_rq);
4554 se->vruntime = curr->vruntime;
4555 place_entity(cfs_rq, se, 1);
4557 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4559 * Upon rescheduling, sched_class::put_prev_task() will place
4560 * 'current' within the tree based on its new key value.
4562 swap(curr->vruntime, se->vruntime);
4563 resched_task(rq->curr);
4566 se->vruntime -= cfs_rq->min_vruntime;
4568 raw_spin_unlock_irqrestore(&rq->lock, flags);
4572 * Priority of the task has changed. Check to see if we preempt
4576 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4582 * Reschedule if we are currently running on this runqueue and
4583 * our priority decreased, or if we are not currently running on
4584 * this runqueue and our priority is higher than the current's
4586 if (rq->curr == p) {
4587 if (p->prio > oldprio)
4588 resched_task(rq->curr);
4590 check_preempt_curr(rq, p, 0);
4593 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4595 struct sched_entity *se = &p->se;
4596 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4599 * Ensure the task's vruntime is normalized, so that when its
4600 * switched back to the fair class the enqueue_entity(.flags=0) will
4601 * do the right thing.
4603 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4604 * have normalized the vruntime, if it was !on_rq, then only when
4605 * the task is sleeping will it still have non-normalized vruntime.
4607 if (!se->on_rq && p->state != TASK_RUNNING) {
4609 * Fix up our vruntime so that the current sleep doesn't
4610 * cause 'unlimited' sleep bonus.
4612 place_entity(cfs_rq, se, 0);
4613 se->vruntime -= cfs_rq->min_vruntime;
4618 * We switched to the sched_fair class.
4620 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4626 * We were most likely switched from sched_rt, so
4627 * kick off the schedule if running, otherwise just see
4628 * if we can still preempt the current task.
4631 resched_task(rq->curr);
4633 check_preempt_curr(rq, p, 0);
4636 /* Account for a task changing its policy or group.
4638 * This routine is mostly called to set cfs_rq->curr field when a task
4639 * migrates between groups/classes.
4641 static void set_curr_task_fair(struct rq *rq)
4643 struct sched_entity *se = &rq->curr->se;
4645 for_each_sched_entity(se) {
4646 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4648 set_next_entity(cfs_rq, se);
4649 /* ensure bandwidth has been allocated on our new cfs_rq */
4650 account_cfs_rq_runtime(cfs_rq, 0);
4654 #ifdef CONFIG_FAIR_GROUP_SCHED
4655 static void task_move_group_fair(struct task_struct *p, int on_rq)
4658 * If the task was not on the rq at the time of this cgroup movement
4659 * it must have been asleep, sleeping tasks keep their ->vruntime
4660 * absolute on their old rq until wakeup (needed for the fair sleeper
4661 * bonus in place_entity()).
4663 * If it was on the rq, we've just 'preempted' it, which does convert
4664 * ->vruntime to a relative base.
4666 * Make sure both cases convert their relative position when migrating
4667 * to another cgroup's rq. This does somewhat interfere with the
4668 * fair sleeper stuff for the first placement, but who cares.
4671 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4672 set_task_rq(p, task_cpu(p));
4674 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4678 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4680 struct sched_entity *se = &task->se;
4681 unsigned int rr_interval = 0;
4684 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4687 if (rq->cfs.load.weight)
4688 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4694 * All the scheduling class methods:
4696 static const struct sched_class fair_sched_class = {
4697 .next = &idle_sched_class,
4698 .enqueue_task = enqueue_task_fair,
4699 .dequeue_task = dequeue_task_fair,
4700 .yield_task = yield_task_fair,
4701 .yield_to_task = yield_to_task_fair,
4703 .check_preempt_curr = check_preempt_wakeup,
4705 .pick_next_task = pick_next_task_fair,
4706 .put_prev_task = put_prev_task_fair,
4709 .select_task_rq = select_task_rq_fair,
4711 .rq_online = rq_online_fair,
4712 .rq_offline = rq_offline_fair,
4714 .task_waking = task_waking_fair,
4717 .set_curr_task = set_curr_task_fair,
4718 .task_tick = task_tick_fair,
4719 .task_fork = task_fork_fair,
4721 .prio_changed = prio_changed_fair,
4722 .switched_from = switched_from_fair,
4723 .switched_to = switched_to_fair,
4725 .get_rr_interval = get_rr_interval_fair,
4727 #ifdef CONFIG_FAIR_GROUP_SCHED
4728 .task_move_group = task_move_group_fair,
4732 #ifdef CONFIG_SCHED_DEBUG
4733 static void print_cfs_stats(struct seq_file *m, int cpu)
4735 struct cfs_rq *cfs_rq;
4738 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4739 print_cfs_rq(m, cpu, cfs_rq);