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>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 8;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
92 static const struct sched_class fair_sched_class;
94 /**************************************************************
95 * CFS operations on generic schedulable entities:
98 #ifdef CONFIG_FAIR_GROUP_SCHED
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
106 /* An entity is a task if it doesn't "own" a runqueue */
107 #define entity_is_task(se) (!se->my_q)
109 static inline struct task_struct *task_of(struct sched_entity *se)
111 #ifdef CONFIG_SCHED_DEBUG
112 WARN_ON_ONCE(!entity_is_task(se));
114 return container_of(se, struct task_struct, se);
117 /* Walk up scheduling entities hierarchy */
118 #define for_each_sched_entity(se) \
119 for (; se; se = se->parent)
121 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
126 /* runqueue on which this entity is (to be) queued */
127 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
132 /* runqueue "owned" by this group */
133 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
138 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
139 * another cpu ('this_cpu')
141 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
143 return cfs_rq->tg->cfs_rq[this_cpu];
146 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
148 if (!cfs_rq->on_list) {
149 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
150 &rq_of(cfs_rq)->leaf_cfs_rq_list);
156 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
158 if (cfs_rq->on_list) {
159 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
164 /* Iterate thr' all leaf cfs_rq's on a runqueue */
165 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
166 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
168 /* Do the two (enqueued) entities belong to the same group ? */
170 is_same_group(struct sched_entity *se, struct sched_entity *pse)
172 if (se->cfs_rq == pse->cfs_rq)
178 static inline struct sched_entity *parent_entity(struct sched_entity *se)
183 /* return depth at which a sched entity is present in the hierarchy */
184 static inline int depth_se(struct sched_entity *se)
188 for_each_sched_entity(se)
195 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
197 int se_depth, pse_depth;
200 * preemption test can be made between sibling entities who are in the
201 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
202 * both tasks until we find their ancestors who are siblings of common
206 /* First walk up until both entities are at same depth */
207 se_depth = depth_se(*se);
208 pse_depth = depth_se(*pse);
210 while (se_depth > pse_depth) {
212 *se = parent_entity(*se);
215 while (pse_depth > se_depth) {
217 *pse = parent_entity(*pse);
220 while (!is_same_group(*se, *pse)) {
221 *se = parent_entity(*se);
222 *pse = parent_entity(*pse);
226 #else /* !CONFIG_FAIR_GROUP_SCHED */
228 static inline struct task_struct *task_of(struct sched_entity *se)
230 return container_of(se, struct task_struct, se);
233 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
235 return container_of(cfs_rq, struct rq, cfs);
238 #define entity_is_task(se) 1
240 #define for_each_sched_entity(se) \
241 for (; se; se = NULL)
243 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
245 return &task_rq(p)->cfs;
248 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
250 struct task_struct *p = task_of(se);
251 struct rq *rq = task_rq(p);
256 /* runqueue "owned" by this group */
257 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
262 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
264 return &cpu_rq(this_cpu)->cfs;
267 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
271 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
275 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
276 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
279 is_same_group(struct sched_entity *se, struct sched_entity *pse)
284 static inline struct sched_entity *parent_entity(struct sched_entity *se)
290 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
297 /**************************************************************
298 * Scheduling class tree data structure manipulation methods:
301 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
303 s64 delta = (s64)(vruntime - min_vruntime);
305 min_vruntime = vruntime;
310 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
312 s64 delta = (s64)(vruntime - min_vruntime);
314 min_vruntime = vruntime;
319 static inline int entity_before(struct sched_entity *a,
320 struct sched_entity *b)
322 return (s64)(a->vruntime - b->vruntime) < 0;
325 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
327 return se->vruntime - cfs_rq->min_vruntime;
330 static void update_min_vruntime(struct cfs_rq *cfs_rq)
332 u64 vruntime = cfs_rq->min_vruntime;
335 vruntime = cfs_rq->curr->vruntime;
337 if (cfs_rq->rb_leftmost) {
338 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
343 vruntime = se->vruntime;
345 vruntime = min_vruntime(vruntime, se->vruntime);
348 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
352 * Enqueue an entity into the rb-tree:
354 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
356 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
357 struct rb_node *parent = NULL;
358 struct sched_entity *entry;
359 s64 key = entity_key(cfs_rq, se);
363 * Find the right place in the rbtree:
367 entry = rb_entry(parent, struct sched_entity, run_node);
369 * We dont care about collisions. Nodes with
370 * the same key stay together.
372 if (key < entity_key(cfs_rq, entry)) {
373 link = &parent->rb_left;
375 link = &parent->rb_right;
381 * Maintain a cache of leftmost tree entries (it is frequently
385 cfs_rq->rb_leftmost = &se->run_node;
387 rb_link_node(&se->run_node, parent, link);
388 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
391 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
393 if (cfs_rq->rb_leftmost == &se->run_node) {
394 struct rb_node *next_node;
396 next_node = rb_next(&se->run_node);
397 cfs_rq->rb_leftmost = next_node;
400 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
403 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
405 struct rb_node *left = cfs_rq->rb_leftmost;
410 return rb_entry(left, struct sched_entity, run_node);
413 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
415 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
420 return rb_entry(last, struct sched_entity, run_node);
423 /**************************************************************
424 * Scheduling class statistics methods:
427 #ifdef CONFIG_SCHED_DEBUG
428 int sched_proc_update_handler(struct ctl_table *table, int write,
429 void __user *buffer, size_t *lenp,
432 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
433 int factor = get_update_sysctl_factor();
438 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
439 sysctl_sched_min_granularity);
441 #define WRT_SYSCTL(name) \
442 (normalized_sysctl_##name = sysctl_##name / (factor))
443 WRT_SYSCTL(sched_min_granularity);
444 WRT_SYSCTL(sched_latency);
445 WRT_SYSCTL(sched_wakeup_granularity);
455 static inline unsigned long
456 calc_delta_fair(unsigned long delta, struct sched_entity *se)
458 if (unlikely(se->load.weight != NICE_0_LOAD))
459 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
465 * The idea is to set a period in which each task runs once.
467 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
468 * this period because otherwise the slices get too small.
470 * p = (nr <= nl) ? l : l*nr/nl
472 static u64 __sched_period(unsigned long nr_running)
474 u64 period = sysctl_sched_latency;
475 unsigned long nr_latency = sched_nr_latency;
477 if (unlikely(nr_running > nr_latency)) {
478 period = sysctl_sched_min_granularity;
479 period *= nr_running;
486 * We calculate the wall-time slice from the period by taking a part
487 * proportional to the weight.
491 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
493 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
495 for_each_sched_entity(se) {
496 struct load_weight *load;
497 struct load_weight lw;
499 cfs_rq = cfs_rq_of(se);
500 load = &cfs_rq->load;
502 if (unlikely(!se->on_rq)) {
505 update_load_add(&lw, se->load.weight);
508 slice = calc_delta_mine(slice, se->load.weight, load);
514 * We calculate the vruntime slice of a to be inserted task
518 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
520 return calc_delta_fair(sched_slice(cfs_rq, se), se);
524 * Update the current task's runtime statistics. Skip current tasks that
525 * are not in our scheduling class.
528 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
529 unsigned long delta_exec)
531 unsigned long delta_exec_weighted;
533 schedstat_set(curr->statistics.exec_max,
534 max((u64)delta_exec, curr->statistics.exec_max));
536 curr->sum_exec_runtime += delta_exec;
537 schedstat_add(cfs_rq, exec_clock, delta_exec);
538 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
540 curr->vruntime += delta_exec_weighted;
541 update_min_vruntime(cfs_rq);
544 static void update_curr(struct cfs_rq *cfs_rq)
546 struct sched_entity *curr = cfs_rq->curr;
547 u64 now = rq_of(cfs_rq)->clock_task;
548 unsigned long delta_exec;
554 * Get the amount of time the current task was running
555 * since the last time we changed load (this cannot
556 * overflow on 32 bits):
558 delta_exec = (unsigned long)(now - curr->exec_start);
562 __update_curr(cfs_rq, curr, delta_exec);
563 curr->exec_start = now;
565 if (entity_is_task(curr)) {
566 struct task_struct *curtask = task_of(curr);
568 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
569 cpuacct_charge(curtask, delta_exec);
570 account_group_exec_runtime(curtask, delta_exec);
575 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
577 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
581 * Task is being enqueued - update stats:
583 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
586 * Are we enqueueing a waiting task? (for current tasks
587 * a dequeue/enqueue event is a NOP)
589 if (se != cfs_rq->curr)
590 update_stats_wait_start(cfs_rq, se);
594 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
596 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
597 rq_of(cfs_rq)->clock - se->statistics.wait_start));
598 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
599 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
600 rq_of(cfs_rq)->clock - se->statistics.wait_start);
601 #ifdef CONFIG_SCHEDSTATS
602 if (entity_is_task(se)) {
603 trace_sched_stat_wait(task_of(se),
604 rq_of(cfs_rq)->clock - se->statistics.wait_start);
607 schedstat_set(se->statistics.wait_start, 0);
611 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
614 * Mark the end of the wait period if dequeueing a
617 if (se != cfs_rq->curr)
618 update_stats_wait_end(cfs_rq, se);
622 * We are picking a new current task - update its stats:
625 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 * We are starting a new run period:
630 se->exec_start = rq_of(cfs_rq)->clock_task;
633 /**************************************************
634 * Scheduling class queueing methods:
637 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
639 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
641 cfs_rq->task_weight += weight;
645 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
651 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 update_load_add(&cfs_rq->load, se->load.weight);
654 if (!parent_entity(se))
655 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
656 if (entity_is_task(se)) {
657 add_cfs_task_weight(cfs_rq, se->load.weight);
658 list_add(&se->group_node, &cfs_rq->tasks);
660 cfs_rq->nr_running++;
664 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
666 update_load_sub(&cfs_rq->load, se->load.weight);
667 if (!parent_entity(se))
668 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
669 if (entity_is_task(se)) {
670 add_cfs_task_weight(cfs_rq, -se->load.weight);
671 list_del_init(&se->group_node);
673 cfs_rq->nr_running--;
676 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
677 static void update_cfs_load(struct cfs_rq *cfs_rq)
679 u64 period = sched_avg_period();
681 unsigned long load = cfs_rq->load.weight;
686 now = rq_of(cfs_rq)->clock;
687 delta = now - cfs_rq->load_stamp;
689 /* truncate load history at 4 idle periods */
690 if (cfs_rq->load_stamp > cfs_rq->load_last &&
691 now - cfs_rq->load_last > 4 * period) {
692 cfs_rq->load_period = 0;
693 cfs_rq->load_avg = 0;
696 cfs_rq->load_stamp = now;
697 cfs_rq->load_period += delta;
699 cfs_rq->load_last = now;
700 cfs_rq->load_avg += delta * load;
703 while (cfs_rq->load_period > period) {
705 * Inline assembly required to prevent the compiler
706 * optimising this loop into a divmod call.
707 * See __iter_div_u64_rem() for another example of this.
709 asm("" : "+rm" (cfs_rq->load_period));
710 cfs_rq->load_period /= 2;
711 cfs_rq->load_avg /= 2;
714 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
715 list_del_leaf_cfs_rq(cfs_rq);
718 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
719 unsigned long weight)
722 account_entity_dequeue(cfs_rq, se);
724 update_load_set(&se->load, weight);
727 account_entity_enqueue(cfs_rq, se);
730 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
732 struct task_group *tg;
733 struct sched_entity *se;
734 long load_weight, load, shares;
740 se = tg->se[cpu_of(rq_of(cfs_rq))];
744 load = cfs_rq->load.weight + weight_delta;
746 load_weight = atomic_read(&tg->load_weight);
747 load_weight -= cfs_rq->load_contribution;
750 shares = (tg->shares * load);
752 shares /= load_weight;
754 if (shares < MIN_SHARES)
756 if (shares > tg->shares)
759 reweight_entity(cfs_rq_of(se), se, shares);
761 #else /* CONFIG_FAIR_GROUP_SCHED */
762 static inline void update_cfs_load(struct cfs_rq *cfs_rq)
766 static inline void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
769 #endif /* CONFIG_FAIR_GROUP_SCHED */
771 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 #ifdef CONFIG_SCHEDSTATS
774 struct task_struct *tsk = NULL;
776 if (entity_is_task(se))
779 if (se->statistics.sleep_start) {
780 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
785 if (unlikely(delta > se->statistics.sleep_max))
786 se->statistics.sleep_max = delta;
788 se->statistics.sleep_start = 0;
789 se->statistics.sum_sleep_runtime += delta;
792 account_scheduler_latency(tsk, delta >> 10, 1);
793 trace_sched_stat_sleep(tsk, delta);
796 if (se->statistics.block_start) {
797 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
802 if (unlikely(delta > se->statistics.block_max))
803 se->statistics.block_max = delta;
805 se->statistics.block_start = 0;
806 se->statistics.sum_sleep_runtime += delta;
809 if (tsk->in_iowait) {
810 se->statistics.iowait_sum += delta;
811 se->statistics.iowait_count++;
812 trace_sched_stat_iowait(tsk, delta);
816 * Blocking time is in units of nanosecs, so shift by
817 * 20 to get a milliseconds-range estimation of the
818 * amount of time that the task spent sleeping:
820 if (unlikely(prof_on == SLEEP_PROFILING)) {
821 profile_hits(SLEEP_PROFILING,
822 (void *)get_wchan(tsk),
825 account_scheduler_latency(tsk, delta >> 10, 0);
831 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
833 #ifdef CONFIG_SCHED_DEBUG
834 s64 d = se->vruntime - cfs_rq->min_vruntime;
839 if (d > 3*sysctl_sched_latency)
840 schedstat_inc(cfs_rq, nr_spread_over);
845 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
847 u64 vruntime = cfs_rq->min_vruntime;
850 * The 'current' period is already promised to the current tasks,
851 * however the extra weight of the new task will slow them down a
852 * little, place the new task so that it fits in the slot that
853 * stays open at the end.
855 if (initial && sched_feat(START_DEBIT))
856 vruntime += sched_vslice(cfs_rq, se);
858 /* sleeps up to a single latency don't count. */
860 unsigned long thresh = sysctl_sched_latency;
863 * Halve their sleep time's effect, to allow
864 * for a gentler effect of sleepers:
866 if (sched_feat(GENTLE_FAIR_SLEEPERS))
872 /* ensure we never gain time by being placed backwards. */
873 vruntime = max_vruntime(se->vruntime, vruntime);
875 se->vruntime = vruntime;
879 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
882 * Update the normalized vruntime before updating min_vruntime
883 * through callig update_curr().
885 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
886 se->vruntime += cfs_rq->min_vruntime;
889 * Update run-time statistics of the 'current'.
892 update_cfs_load(cfs_rq);
893 update_cfs_shares(cfs_rq, se->load.weight);
894 account_entity_enqueue(cfs_rq, se);
896 if (flags & ENQUEUE_WAKEUP) {
897 place_entity(cfs_rq, se, 0);
898 enqueue_sleeper(cfs_rq, se);
901 update_stats_enqueue(cfs_rq, se);
902 check_spread(cfs_rq, se);
903 if (se != cfs_rq->curr)
904 __enqueue_entity(cfs_rq, se);
907 if (cfs_rq->nr_running == 1)
908 list_add_leaf_cfs_rq(cfs_rq);
911 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
913 if (!se || cfs_rq->last == se)
916 if (!se || cfs_rq->next == se)
920 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
922 for_each_sched_entity(se)
923 __clear_buddies(cfs_rq_of(se), se);
927 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
930 * Update run-time statistics of the 'current'.
934 update_stats_dequeue(cfs_rq, se);
935 if (flags & DEQUEUE_SLEEP) {
936 #ifdef CONFIG_SCHEDSTATS
937 if (entity_is_task(se)) {
938 struct task_struct *tsk = task_of(se);
940 if (tsk->state & TASK_INTERRUPTIBLE)
941 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
942 if (tsk->state & TASK_UNINTERRUPTIBLE)
943 se->statistics.block_start = rq_of(cfs_rq)->clock;
948 clear_buddies(cfs_rq, se);
950 if (se != cfs_rq->curr)
951 __dequeue_entity(cfs_rq, se);
953 update_cfs_load(cfs_rq);
954 account_entity_dequeue(cfs_rq, se);
955 update_min_vruntime(cfs_rq);
956 update_cfs_shares(cfs_rq, 0);
959 * Normalize the entity after updating the min_vruntime because the
960 * update can refer to the ->curr item and we need to reflect this
961 * movement in our normalized position.
963 if (!(flags & DEQUEUE_SLEEP))
964 se->vruntime -= cfs_rq->min_vruntime;
968 * Preempt the current task with a newly woken task if needed:
971 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
973 unsigned long ideal_runtime, delta_exec;
975 ideal_runtime = sched_slice(cfs_rq, curr);
976 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
977 if (delta_exec > ideal_runtime) {
978 resched_task(rq_of(cfs_rq)->curr);
980 * The current task ran long enough, ensure it doesn't get
981 * re-elected due to buddy favours.
983 clear_buddies(cfs_rq, curr);
988 * Ensure that a task that missed wakeup preemption by a
989 * narrow margin doesn't have to wait for a full slice.
990 * This also mitigates buddy induced latencies under load.
992 if (!sched_feat(WAKEUP_PREEMPT))
995 if (delta_exec < sysctl_sched_min_granularity)
998 if (cfs_rq->nr_running > 1) {
999 struct sched_entity *se = __pick_next_entity(cfs_rq);
1000 s64 delta = curr->vruntime - se->vruntime;
1002 if (delta > ideal_runtime)
1003 resched_task(rq_of(cfs_rq)->curr);
1008 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1010 /* 'current' is not kept within the tree. */
1013 * Any task has to be enqueued before it get to execute on
1014 * a CPU. So account for the time it spent waiting on the
1017 update_stats_wait_end(cfs_rq, se);
1018 __dequeue_entity(cfs_rq, se);
1021 update_stats_curr_start(cfs_rq, se);
1023 #ifdef CONFIG_SCHEDSTATS
1025 * Track our maximum slice length, if the CPU's load is at
1026 * least twice that of our own weight (i.e. dont track it
1027 * when there are only lesser-weight tasks around):
1029 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1030 se->statistics.slice_max = max(se->statistics.slice_max,
1031 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1034 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1038 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1040 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1042 struct sched_entity *se = __pick_next_entity(cfs_rq);
1043 struct sched_entity *left = se;
1045 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1049 * Prefer last buddy, try to return the CPU to a preempted task.
1051 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1054 clear_buddies(cfs_rq, se);
1059 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1062 * If still on the runqueue then deactivate_task()
1063 * was not called and update_curr() has to be done:
1066 update_curr(cfs_rq);
1068 check_spread(cfs_rq, prev);
1070 update_stats_wait_start(cfs_rq, prev);
1071 /* Put 'current' back into the tree. */
1072 __enqueue_entity(cfs_rq, prev);
1074 cfs_rq->curr = NULL;
1078 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1081 * Update run-time statistics of the 'current'.
1083 update_curr(cfs_rq);
1085 #ifdef CONFIG_SCHED_HRTICK
1087 * queued ticks are scheduled to match the slice, so don't bother
1088 * validating it and just reschedule.
1091 resched_task(rq_of(cfs_rq)->curr);
1095 * don't let the period tick interfere with the hrtick preemption
1097 if (!sched_feat(DOUBLE_TICK) &&
1098 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1102 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1103 check_preempt_tick(cfs_rq, curr);
1106 /**************************************************
1107 * CFS operations on tasks:
1110 #ifdef CONFIG_SCHED_HRTICK
1111 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1113 struct sched_entity *se = &p->se;
1114 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1116 WARN_ON(task_rq(p) != rq);
1118 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1119 u64 slice = sched_slice(cfs_rq, se);
1120 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1121 s64 delta = slice - ran;
1130 * Don't schedule slices shorter than 10000ns, that just
1131 * doesn't make sense. Rely on vruntime for fairness.
1134 delta = max_t(s64, 10000LL, delta);
1136 hrtick_start(rq, delta);
1141 * called from enqueue/dequeue and updates the hrtick when the
1142 * current task is from our class and nr_running is low enough
1145 static void hrtick_update(struct rq *rq)
1147 struct task_struct *curr = rq->curr;
1149 if (curr->sched_class != &fair_sched_class)
1152 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1153 hrtick_start_fair(rq, curr);
1155 #else /* !CONFIG_SCHED_HRTICK */
1157 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1161 static inline void hrtick_update(struct rq *rq)
1167 * The enqueue_task method is called before nr_running is
1168 * increased. Here we update the fair scheduling stats and
1169 * then put the task into the rbtree:
1172 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1174 struct cfs_rq *cfs_rq;
1175 struct sched_entity *se = &p->se;
1177 for_each_sched_entity(se) {
1180 cfs_rq = cfs_rq_of(se);
1181 enqueue_entity(cfs_rq, se, flags);
1182 flags = ENQUEUE_WAKEUP;
1185 for_each_sched_entity(se) {
1186 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1188 update_cfs_load(cfs_rq);
1189 update_cfs_shares(cfs_rq, 0);
1196 * The dequeue_task method is called before nr_running is
1197 * decreased. We remove the task from the rbtree and
1198 * update the fair scheduling stats:
1200 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1202 struct cfs_rq *cfs_rq;
1203 struct sched_entity *se = &p->se;
1205 for_each_sched_entity(se) {
1206 cfs_rq = cfs_rq_of(se);
1207 dequeue_entity(cfs_rq, se, flags);
1209 /* Don't dequeue parent if it has other entities besides us */
1210 if (cfs_rq->load.weight)
1212 flags |= DEQUEUE_SLEEP;
1215 for_each_sched_entity(se) {
1216 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1218 update_cfs_load(cfs_rq);
1219 update_cfs_shares(cfs_rq, 0);
1226 * sched_yield() support is very simple - we dequeue and enqueue.
1228 * If compat_yield is turned on then we requeue to the end of the tree.
1230 static void yield_task_fair(struct rq *rq)
1232 struct task_struct *curr = rq->curr;
1233 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1234 struct sched_entity *rightmost, *se = &curr->se;
1237 * Are we the only task in the tree?
1239 if (unlikely(cfs_rq->nr_running == 1))
1242 clear_buddies(cfs_rq, se);
1244 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1245 update_rq_clock(rq);
1247 * Update run-time statistics of the 'current'.
1249 update_curr(cfs_rq);
1254 * Find the rightmost entry in the rbtree:
1256 rightmost = __pick_last_entity(cfs_rq);
1258 * Already in the rightmost position?
1260 if (unlikely(!rightmost || entity_before(rightmost, se)))
1264 * Minimally necessary key value to be last in the tree:
1265 * Upon rescheduling, sched_class::put_prev_task() will place
1266 * 'current' within the tree based on its new key value.
1268 se->vruntime = rightmost->vruntime + 1;
1273 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1275 struct sched_entity *se = &p->se;
1276 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1278 se->vruntime -= cfs_rq->min_vruntime;
1281 #ifdef CONFIG_FAIR_GROUP_SCHED
1283 * effective_load() calculates the load change as seen from the root_task_group
1285 * Adding load to a group doesn't make a group heavier, but can cause movement
1286 * of group shares between cpus. Assuming the shares were perfectly aligned one
1287 * can calculate the shift in shares.
1289 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1291 struct sched_entity *se = tg->se[cpu];
1296 for_each_sched_entity(se) {
1297 long S, rw, s, a, b;
1299 S = se->my_q->tg->shares;
1300 s = se->load.weight;
1301 rw = se->my_q->load.weight;
1312 * Assume the group is already running and will
1313 * thus already be accounted for in the weight.
1315 * That is, moving shares between CPUs, does not
1316 * alter the group weight.
1326 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1327 unsigned long wl, unsigned long wg)
1334 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1336 unsigned long this_load, load;
1337 int idx, this_cpu, prev_cpu;
1338 unsigned long tl_per_task;
1339 struct task_group *tg;
1340 unsigned long weight;
1344 this_cpu = smp_processor_id();
1345 prev_cpu = task_cpu(p);
1346 load = source_load(prev_cpu, idx);
1347 this_load = target_load(this_cpu, idx);
1350 * If sync wakeup then subtract the (maximum possible)
1351 * effect of the currently running task from the load
1352 * of the current CPU:
1356 tg = task_group(current);
1357 weight = current->se.load.weight;
1359 this_load += effective_load(tg, this_cpu, -weight, -weight);
1360 load += effective_load(tg, prev_cpu, 0, -weight);
1364 weight = p->se.load.weight;
1367 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1368 * due to the sync cause above having dropped this_load to 0, we'll
1369 * always have an imbalance, but there's really nothing you can do
1370 * about that, so that's good too.
1372 * Otherwise check if either cpus are near enough in load to allow this
1373 * task to be woken on this_cpu.
1376 unsigned long this_eff_load, prev_eff_load;
1378 this_eff_load = 100;
1379 this_eff_load *= power_of(prev_cpu);
1380 this_eff_load *= this_load +
1381 effective_load(tg, this_cpu, weight, weight);
1383 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1384 prev_eff_load *= power_of(this_cpu);
1385 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1387 balanced = this_eff_load <= prev_eff_load;
1393 * If the currently running task will sleep within
1394 * a reasonable amount of time then attract this newly
1397 if (sync && balanced)
1400 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1401 tl_per_task = cpu_avg_load_per_task(this_cpu);
1404 (this_load <= load &&
1405 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1407 * This domain has SD_WAKE_AFFINE and
1408 * p is cache cold in this domain, and
1409 * there is no bad imbalance.
1411 schedstat_inc(sd, ttwu_move_affine);
1412 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1420 * find_idlest_group finds and returns the least busy CPU group within the
1423 static struct sched_group *
1424 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1425 int this_cpu, int load_idx)
1427 struct sched_group *idlest = NULL, *group = sd->groups;
1428 unsigned long min_load = ULONG_MAX, this_load = 0;
1429 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1432 unsigned long load, avg_load;
1436 /* Skip over this group if it has no CPUs allowed */
1437 if (!cpumask_intersects(sched_group_cpus(group),
1441 local_group = cpumask_test_cpu(this_cpu,
1442 sched_group_cpus(group));
1444 /* Tally up the load of all CPUs in the group */
1447 for_each_cpu(i, sched_group_cpus(group)) {
1448 /* Bias balancing toward cpus of our domain */
1450 load = source_load(i, load_idx);
1452 load = target_load(i, load_idx);
1457 /* Adjust by relative CPU power of the group */
1458 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1461 this_load = avg_load;
1462 } else if (avg_load < min_load) {
1463 min_load = avg_load;
1466 } while (group = group->next, group != sd->groups);
1468 if (!idlest || 100*this_load < imbalance*min_load)
1474 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1477 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1479 unsigned long load, min_load = ULONG_MAX;
1483 /* Traverse only the allowed CPUs */
1484 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1485 load = weighted_cpuload(i);
1487 if (load < min_load || (load == min_load && i == this_cpu)) {
1497 * Try and locate an idle CPU in the sched_domain.
1499 static int select_idle_sibling(struct task_struct *p, int target)
1501 int cpu = smp_processor_id();
1502 int prev_cpu = task_cpu(p);
1503 struct sched_domain *sd;
1507 * If the task is going to be woken-up on this cpu and if it is
1508 * already idle, then it is the right target.
1510 if (target == cpu && idle_cpu(cpu))
1514 * If the task is going to be woken-up on the cpu where it previously
1515 * ran and if it is currently idle, then it the right target.
1517 if (target == prev_cpu && idle_cpu(prev_cpu))
1521 * Otherwise, iterate the domains and find an elegible idle cpu.
1523 for_each_domain(target, sd) {
1524 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1527 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1535 * Lets stop looking for an idle sibling when we reached
1536 * the domain that spans the current cpu and prev_cpu.
1538 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1539 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1547 * sched_balance_self: balance the current task (running on cpu) in domains
1548 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1551 * Balance, ie. select the least loaded group.
1553 * Returns the target CPU number, or the same CPU if no balancing is needed.
1555 * preempt must be disabled.
1558 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1560 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1561 int cpu = smp_processor_id();
1562 int prev_cpu = task_cpu(p);
1564 int want_affine = 0;
1566 int sync = wake_flags & WF_SYNC;
1568 if (sd_flag & SD_BALANCE_WAKE) {
1569 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1574 for_each_domain(cpu, tmp) {
1575 if (!(tmp->flags & SD_LOAD_BALANCE))
1579 * If power savings logic is enabled for a domain, see if we
1580 * are not overloaded, if so, don't balance wider.
1582 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1583 unsigned long power = 0;
1584 unsigned long nr_running = 0;
1585 unsigned long capacity;
1588 for_each_cpu(i, sched_domain_span(tmp)) {
1589 power += power_of(i);
1590 nr_running += cpu_rq(i)->cfs.nr_running;
1593 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1595 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1598 if (nr_running < capacity)
1603 * If both cpu and prev_cpu are part of this domain,
1604 * cpu is a valid SD_WAKE_AFFINE target.
1606 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1607 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1612 if (!want_sd && !want_affine)
1615 if (!(tmp->flags & sd_flag))
1623 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1624 return select_idle_sibling(p, cpu);
1626 return select_idle_sibling(p, prev_cpu);
1630 int load_idx = sd->forkexec_idx;
1631 struct sched_group *group;
1634 if (!(sd->flags & sd_flag)) {
1639 if (sd_flag & SD_BALANCE_WAKE)
1640 load_idx = sd->wake_idx;
1642 group = find_idlest_group(sd, p, cpu, load_idx);
1648 new_cpu = find_idlest_cpu(group, p, cpu);
1649 if (new_cpu == -1 || new_cpu == cpu) {
1650 /* Now try balancing at a lower domain level of cpu */
1655 /* Now try balancing at a lower domain level of new_cpu */
1657 weight = sd->span_weight;
1659 for_each_domain(cpu, tmp) {
1660 if (weight <= tmp->span_weight)
1662 if (tmp->flags & sd_flag)
1665 /* while loop will break here if sd == NULL */
1670 #endif /* CONFIG_SMP */
1672 static unsigned long
1673 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1675 unsigned long gran = sysctl_sched_wakeup_granularity;
1678 * Since its curr running now, convert the gran from real-time
1679 * to virtual-time in his units.
1681 * By using 'se' instead of 'curr' we penalize light tasks, so
1682 * they get preempted easier. That is, if 'se' < 'curr' then
1683 * the resulting gran will be larger, therefore penalizing the
1684 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1685 * be smaller, again penalizing the lighter task.
1687 * This is especially important for buddies when the leftmost
1688 * task is higher priority than the buddy.
1690 if (unlikely(se->load.weight != NICE_0_LOAD))
1691 gran = calc_delta_fair(gran, se);
1697 * Should 'se' preempt 'curr'.
1711 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1713 s64 gran, vdiff = curr->vruntime - se->vruntime;
1718 gran = wakeup_gran(curr, se);
1725 static void set_last_buddy(struct sched_entity *se)
1727 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1728 for_each_sched_entity(se)
1729 cfs_rq_of(se)->last = se;
1733 static void set_next_buddy(struct sched_entity *se)
1735 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1736 for_each_sched_entity(se)
1737 cfs_rq_of(se)->next = se;
1742 * Preempt the current task with a newly woken task if needed:
1744 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1746 struct task_struct *curr = rq->curr;
1747 struct sched_entity *se = &curr->se, *pse = &p->se;
1748 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1749 int scale = cfs_rq->nr_running >= sched_nr_latency;
1751 if (unlikely(rt_prio(p->prio)))
1754 if (unlikely(p->sched_class != &fair_sched_class))
1757 if (unlikely(se == pse))
1760 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1761 set_next_buddy(pse);
1764 * We can come here with TIF_NEED_RESCHED already set from new task
1767 if (test_tsk_need_resched(curr))
1771 * Batch and idle tasks do not preempt (their preemption is driven by
1774 if (unlikely(p->policy != SCHED_NORMAL))
1777 /* Idle tasks are by definition preempted by everybody. */
1778 if (unlikely(curr->policy == SCHED_IDLE))
1781 if (!sched_feat(WAKEUP_PREEMPT))
1784 update_curr(cfs_rq);
1785 find_matching_se(&se, &pse);
1787 if (wakeup_preempt_entity(se, pse) == 1)
1795 * Only set the backward buddy when the current task is still
1796 * on the rq. This can happen when a wakeup gets interleaved
1797 * with schedule on the ->pre_schedule() or idle_balance()
1798 * point, either of which can * drop the rq lock.
1800 * Also, during early boot the idle thread is in the fair class,
1801 * for obvious reasons its a bad idea to schedule back to it.
1803 if (unlikely(!se->on_rq || curr == rq->idle))
1806 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1810 static struct task_struct *pick_next_task_fair(struct rq *rq)
1812 struct task_struct *p;
1813 struct cfs_rq *cfs_rq = &rq->cfs;
1814 struct sched_entity *se;
1816 if (!cfs_rq->nr_running)
1820 se = pick_next_entity(cfs_rq);
1821 set_next_entity(cfs_rq, se);
1822 cfs_rq = group_cfs_rq(se);
1826 hrtick_start_fair(rq, p);
1832 * Account for a descheduled task:
1834 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1836 struct sched_entity *se = &prev->se;
1837 struct cfs_rq *cfs_rq;
1839 for_each_sched_entity(se) {
1840 cfs_rq = cfs_rq_of(se);
1841 put_prev_entity(cfs_rq, se);
1846 /**************************************************
1847 * Fair scheduling class load-balancing methods:
1851 * pull_task - move a task from a remote runqueue to the local runqueue.
1852 * Both runqueues must be locked.
1854 static void pull_task(struct rq *src_rq, struct task_struct *p,
1855 struct rq *this_rq, int this_cpu)
1857 deactivate_task(src_rq, p, 0);
1858 set_task_cpu(p, this_cpu);
1859 activate_task(this_rq, p, 0);
1860 check_preempt_curr(this_rq, p, 0);
1862 /* re-arm NEWIDLE balancing when moving tasks */
1863 src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost;
1864 this_rq->idle_stamp = 0;
1868 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1871 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1872 struct sched_domain *sd, enum cpu_idle_type idle,
1875 int tsk_cache_hot = 0;
1877 * We do not migrate tasks that are:
1878 * 1) running (obviously), or
1879 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1880 * 3) are cache-hot on their current CPU.
1882 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1883 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1888 if (task_running(rq, p)) {
1889 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1894 * Aggressive migration if:
1895 * 1) task is cache cold, or
1896 * 2) too many balance attempts have failed.
1899 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
1900 if (!tsk_cache_hot ||
1901 sd->nr_balance_failed > sd->cache_nice_tries) {
1902 #ifdef CONFIG_SCHEDSTATS
1903 if (tsk_cache_hot) {
1904 schedstat_inc(sd, lb_hot_gained[idle]);
1905 schedstat_inc(p, se.statistics.nr_forced_migrations);
1911 if (tsk_cache_hot) {
1912 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
1919 * move_one_task tries to move exactly one task from busiest to this_rq, as
1920 * part of active balancing operations within "domain".
1921 * Returns 1 if successful and 0 otherwise.
1923 * Called with both runqueues locked.
1926 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1927 struct sched_domain *sd, enum cpu_idle_type idle)
1929 struct task_struct *p, *n;
1930 struct cfs_rq *cfs_rq;
1933 for_each_leaf_cfs_rq(busiest, cfs_rq) {
1934 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1936 if (!can_migrate_task(p, busiest, this_cpu,
1940 pull_task(busiest, p, this_rq, this_cpu);
1942 * Right now, this is only the second place pull_task()
1943 * is called, so we can safely collect pull_task()
1944 * stats here rather than inside pull_task().
1946 schedstat_inc(sd, lb_gained[idle]);
1954 static unsigned long
1955 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1956 unsigned long max_load_move, struct sched_domain *sd,
1957 enum cpu_idle_type idle, int *all_pinned,
1958 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
1960 int loops = 0, pulled = 0, pinned = 0;
1961 long rem_load_move = max_load_move;
1962 struct task_struct *p, *n;
1964 if (max_load_move == 0)
1969 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
1970 if (loops++ > sysctl_sched_nr_migrate)
1973 if ((p->se.load.weight >> 1) > rem_load_move ||
1974 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
1977 pull_task(busiest, p, this_rq, this_cpu);
1979 rem_load_move -= p->se.load.weight;
1981 #ifdef CONFIG_PREEMPT
1983 * NEWIDLE balancing is a source of latency, so preemptible
1984 * kernels will stop after the first task is pulled to minimize
1985 * the critical section.
1987 if (idle == CPU_NEWLY_IDLE)
1992 * We only want to steal up to the prescribed amount of
1995 if (rem_load_move <= 0)
1998 if (p->prio < *this_best_prio)
1999 *this_best_prio = p->prio;
2003 * Right now, this is one of only two places pull_task() is called,
2004 * so we can safely collect pull_task() stats here rather than
2005 * inside pull_task().
2007 schedstat_add(sd, lb_gained[idle], pulled);
2010 *all_pinned = pinned;
2012 return max_load_move - rem_load_move;
2015 #ifdef CONFIG_FAIR_GROUP_SCHED
2017 * update tg->load_weight by folding this cpu's load_avg
2019 static int tg_shares_up(struct task_group *tg, int cpu)
2021 struct cfs_rq *cfs_rq;
2022 unsigned long flags;
2030 cfs_rq = tg->cfs_rq[cpu];
2032 raw_spin_lock_irqsave(&rq->lock, flags);
2034 update_rq_clock(rq);
2035 update_cfs_load(cfs_rq);
2037 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
2038 load_avg -= cfs_rq->load_contribution;
2039 atomic_add(load_avg, &tg->load_weight);
2040 cfs_rq->load_contribution += load_avg;
2043 * We need to update shares after updating tg->load_weight in
2044 * order to adjust the weight of groups with long running tasks.
2046 update_cfs_shares(cfs_rq, 0);
2048 raw_spin_unlock_irqrestore(&rq->lock, flags);
2053 static void update_shares(int cpu)
2055 struct cfs_rq *cfs_rq;
2056 struct rq *rq = cpu_rq(cpu);
2059 for_each_leaf_cfs_rq(rq, cfs_rq) {
2060 struct task_group *tg = cfs_rq->tg;
2063 tg_shares_up(tg, cpu);
2070 static unsigned long
2071 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2072 unsigned long max_load_move,
2073 struct sched_domain *sd, enum cpu_idle_type idle,
2074 int *all_pinned, int *this_best_prio)
2076 long rem_load_move = max_load_move;
2077 int busiest_cpu = cpu_of(busiest);
2078 struct task_group *tg;
2081 update_h_load(busiest_cpu);
2083 list_for_each_entry_rcu(tg, &task_groups, list) {
2084 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2085 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2086 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2087 u64 rem_load, moved_load;
2092 if (!busiest_cfs_rq->task_weight)
2095 rem_load = (u64)rem_load_move * busiest_weight;
2096 rem_load = div_u64(rem_load, busiest_h_load + 1);
2098 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2099 rem_load, sd, idle, all_pinned, this_best_prio,
2105 moved_load *= busiest_h_load;
2106 moved_load = div_u64(moved_load, busiest_weight + 1);
2108 rem_load_move -= moved_load;
2109 if (rem_load_move < 0)
2114 return max_load_move - rem_load_move;
2117 static inline void update_shares(int cpu)
2121 static unsigned long
2122 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2123 unsigned long max_load_move,
2124 struct sched_domain *sd, enum cpu_idle_type idle,
2125 int *all_pinned, int *this_best_prio)
2127 return balance_tasks(this_rq, this_cpu, busiest,
2128 max_load_move, sd, idle, all_pinned,
2129 this_best_prio, &busiest->cfs);
2134 * move_tasks tries to move up to max_load_move weighted load from busiest to
2135 * this_rq, as part of a balancing operation within domain "sd".
2136 * Returns 1 if successful and 0 otherwise.
2138 * Called with both runqueues locked.
2140 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2141 unsigned long max_load_move,
2142 struct sched_domain *sd, enum cpu_idle_type idle,
2145 unsigned long total_load_moved = 0, load_moved;
2146 int this_best_prio = this_rq->curr->prio;
2149 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2150 max_load_move - total_load_moved,
2151 sd, idle, all_pinned, &this_best_prio);
2153 total_load_moved += load_moved;
2155 #ifdef CONFIG_PREEMPT
2157 * NEWIDLE balancing is a source of latency, so preemptible
2158 * kernels will stop after the first task is pulled to minimize
2159 * the critical section.
2161 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2164 if (raw_spin_is_contended(&this_rq->lock) ||
2165 raw_spin_is_contended(&busiest->lock))
2168 } while (load_moved && max_load_move > total_load_moved);
2170 return total_load_moved > 0;
2173 /********** Helpers for find_busiest_group ************************/
2175 * sd_lb_stats - Structure to store the statistics of a sched_domain
2176 * during load balancing.
2178 struct sd_lb_stats {
2179 struct sched_group *busiest; /* Busiest group in this sd */
2180 struct sched_group *this; /* Local group in this sd */
2181 unsigned long total_load; /* Total load of all groups in sd */
2182 unsigned long total_pwr; /* Total power of all groups in sd */
2183 unsigned long avg_load; /* Average load across all groups in sd */
2185 /** Statistics of this group */
2186 unsigned long this_load;
2187 unsigned long this_load_per_task;
2188 unsigned long this_nr_running;
2189 unsigned long this_has_capacity;
2191 /* Statistics of the busiest group */
2192 unsigned long max_load;
2193 unsigned long busiest_load_per_task;
2194 unsigned long busiest_nr_running;
2195 unsigned long busiest_group_capacity;
2196 unsigned long busiest_has_capacity;
2198 int group_imb; /* Is there imbalance in this sd */
2199 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2200 int power_savings_balance; /* Is powersave balance needed for this sd */
2201 struct sched_group *group_min; /* Least loaded group in sd */
2202 struct sched_group *group_leader; /* Group which relieves group_min */
2203 unsigned long min_load_per_task; /* load_per_task in group_min */
2204 unsigned long leader_nr_running; /* Nr running of group_leader */
2205 unsigned long min_nr_running; /* Nr running of group_min */
2210 * sg_lb_stats - stats of a sched_group required for load_balancing
2212 struct sg_lb_stats {
2213 unsigned long avg_load; /*Avg load across the CPUs of the group */
2214 unsigned long group_load; /* Total load over the CPUs of the group */
2215 unsigned long sum_nr_running; /* Nr tasks running in the group */
2216 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2217 unsigned long group_capacity;
2218 int group_imb; /* Is there an imbalance in the group ? */
2219 int group_has_capacity; /* Is there extra capacity in the group? */
2223 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2224 * @group: The group whose first cpu is to be returned.
2226 static inline unsigned int group_first_cpu(struct sched_group *group)
2228 return cpumask_first(sched_group_cpus(group));
2232 * get_sd_load_idx - Obtain the load index for a given sched domain.
2233 * @sd: The sched_domain whose load_idx is to be obtained.
2234 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2236 static inline int get_sd_load_idx(struct sched_domain *sd,
2237 enum cpu_idle_type idle)
2243 load_idx = sd->busy_idx;
2246 case CPU_NEWLY_IDLE:
2247 load_idx = sd->newidle_idx;
2250 load_idx = sd->idle_idx;
2258 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2260 * init_sd_power_savings_stats - Initialize power savings statistics for
2261 * the given sched_domain, during load balancing.
2263 * @sd: Sched domain whose power-savings statistics are to be initialized.
2264 * @sds: Variable containing the statistics for sd.
2265 * @idle: Idle status of the CPU at which we're performing load-balancing.
2267 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2268 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2271 * Busy processors will not participate in power savings
2274 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2275 sds->power_savings_balance = 0;
2277 sds->power_savings_balance = 1;
2278 sds->min_nr_running = ULONG_MAX;
2279 sds->leader_nr_running = 0;
2284 * update_sd_power_savings_stats - Update the power saving stats for a
2285 * sched_domain while performing load balancing.
2287 * @group: sched_group belonging to the sched_domain under consideration.
2288 * @sds: Variable containing the statistics of the sched_domain
2289 * @local_group: Does group contain the CPU for which we're performing
2291 * @sgs: Variable containing the statistics of the group.
2293 static inline void update_sd_power_savings_stats(struct sched_group *group,
2294 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2297 if (!sds->power_savings_balance)
2301 * If the local group is idle or completely loaded
2302 * no need to do power savings balance at this domain
2304 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2305 !sds->this_nr_running))
2306 sds->power_savings_balance = 0;
2309 * If a group is already running at full capacity or idle,
2310 * don't include that group in power savings calculations
2312 if (!sds->power_savings_balance ||
2313 sgs->sum_nr_running >= sgs->group_capacity ||
2314 !sgs->sum_nr_running)
2318 * Calculate the group which has the least non-idle load.
2319 * This is the group from where we need to pick up the load
2322 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2323 (sgs->sum_nr_running == sds->min_nr_running &&
2324 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2325 sds->group_min = group;
2326 sds->min_nr_running = sgs->sum_nr_running;
2327 sds->min_load_per_task = sgs->sum_weighted_load /
2328 sgs->sum_nr_running;
2332 * Calculate the group which is almost near its
2333 * capacity but still has some space to pick up some load
2334 * from other group and save more power
2336 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2339 if (sgs->sum_nr_running > sds->leader_nr_running ||
2340 (sgs->sum_nr_running == sds->leader_nr_running &&
2341 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2342 sds->group_leader = group;
2343 sds->leader_nr_running = sgs->sum_nr_running;
2348 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2349 * @sds: Variable containing the statistics of the sched_domain
2350 * under consideration.
2351 * @this_cpu: Cpu at which we're currently performing load-balancing.
2352 * @imbalance: Variable to store the imbalance.
2355 * Check if we have potential to perform some power-savings balance.
2356 * If yes, set the busiest group to be the least loaded group in the
2357 * sched_domain, so that it's CPUs can be put to idle.
2359 * Returns 1 if there is potential to perform power-savings balance.
2362 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2363 int this_cpu, unsigned long *imbalance)
2365 if (!sds->power_savings_balance)
2368 if (sds->this != sds->group_leader ||
2369 sds->group_leader == sds->group_min)
2372 *imbalance = sds->min_load_per_task;
2373 sds->busiest = sds->group_min;
2378 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2379 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2380 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2385 static inline void update_sd_power_savings_stats(struct sched_group *group,
2386 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2391 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2392 int this_cpu, unsigned long *imbalance)
2396 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2399 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2401 return SCHED_LOAD_SCALE;
2404 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2406 return default_scale_freq_power(sd, cpu);
2409 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2411 unsigned long weight = sd->span_weight;
2412 unsigned long smt_gain = sd->smt_gain;
2419 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2421 return default_scale_smt_power(sd, cpu);
2424 unsigned long scale_rt_power(int cpu)
2426 struct rq *rq = cpu_rq(cpu);
2427 u64 total, available;
2429 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2431 if (unlikely(total < rq->rt_avg)) {
2432 /* Ensures that power won't end up being negative */
2435 available = total - rq->rt_avg;
2438 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2439 total = SCHED_LOAD_SCALE;
2441 total >>= SCHED_LOAD_SHIFT;
2443 return div_u64(available, total);
2446 static void update_cpu_power(struct sched_domain *sd, int cpu)
2448 unsigned long weight = sd->span_weight;
2449 unsigned long power = SCHED_LOAD_SCALE;
2450 struct sched_group *sdg = sd->groups;
2452 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2453 if (sched_feat(ARCH_POWER))
2454 power *= arch_scale_smt_power(sd, cpu);
2456 power *= default_scale_smt_power(sd, cpu);
2458 power >>= SCHED_LOAD_SHIFT;
2461 sdg->cpu_power_orig = power;
2463 if (sched_feat(ARCH_POWER))
2464 power *= arch_scale_freq_power(sd, cpu);
2466 power *= default_scale_freq_power(sd, cpu);
2468 power >>= SCHED_LOAD_SHIFT;
2470 power *= scale_rt_power(cpu);
2471 power >>= SCHED_LOAD_SHIFT;
2476 cpu_rq(cpu)->cpu_power = power;
2477 sdg->cpu_power = power;
2480 static void update_group_power(struct sched_domain *sd, int cpu)
2482 struct sched_domain *child = sd->child;
2483 struct sched_group *group, *sdg = sd->groups;
2484 unsigned long power;
2487 update_cpu_power(sd, cpu);
2493 group = child->groups;
2495 power += group->cpu_power;
2496 group = group->next;
2497 } while (group != child->groups);
2499 sdg->cpu_power = power;
2503 * Try and fix up capacity for tiny siblings, this is needed when
2504 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2505 * which on its own isn't powerful enough.
2507 * See update_sd_pick_busiest() and check_asym_packing().
2510 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2513 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2515 if (sd->level != SD_LV_SIBLING)
2519 * If ~90% of the cpu_power is still there, we're good.
2521 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2528 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2529 * @sd: The sched_domain whose statistics are to be updated.
2530 * @group: sched_group whose statistics are to be updated.
2531 * @this_cpu: Cpu for which load balance is currently performed.
2532 * @idle: Idle status of this_cpu
2533 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2534 * @sd_idle: Idle status of the sched_domain containing group.
2535 * @local_group: Does group contain this_cpu.
2536 * @cpus: Set of cpus considered for load balancing.
2537 * @balance: Should we balance.
2538 * @sgs: variable to hold the statistics for this group.
2540 static inline void update_sg_lb_stats(struct sched_domain *sd,
2541 struct sched_group *group, int this_cpu,
2542 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2543 int local_group, const struct cpumask *cpus,
2544 int *balance, struct sg_lb_stats *sgs)
2546 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2548 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2549 unsigned long avg_load_per_task = 0;
2552 balance_cpu = group_first_cpu(group);
2554 /* Tally up the load of all CPUs in the group */
2556 min_cpu_load = ~0UL;
2559 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2560 struct rq *rq = cpu_rq(i);
2562 if (*sd_idle && rq->nr_running)
2565 /* Bias balancing toward cpus of our domain */
2567 if (idle_cpu(i) && !first_idle_cpu) {
2572 load = target_load(i, load_idx);
2574 load = source_load(i, load_idx);
2575 if (load > max_cpu_load) {
2576 max_cpu_load = load;
2577 max_nr_running = rq->nr_running;
2579 if (min_cpu_load > load)
2580 min_cpu_load = load;
2583 sgs->group_load += load;
2584 sgs->sum_nr_running += rq->nr_running;
2585 sgs->sum_weighted_load += weighted_cpuload(i);
2590 * First idle cpu or the first cpu(busiest) in this sched group
2591 * is eligible for doing load balancing at this and above
2592 * domains. In the newly idle case, we will allow all the cpu's
2593 * to do the newly idle load balance.
2595 if (idle != CPU_NEWLY_IDLE && local_group) {
2596 if (balance_cpu != this_cpu) {
2600 update_group_power(sd, this_cpu);
2603 /* Adjust by relative CPU power of the group */
2604 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2607 * Consider the group unbalanced when the imbalance is larger
2608 * than the average weight of two tasks.
2610 * APZ: with cgroup the avg task weight can vary wildly and
2611 * might not be a suitable number - should we keep a
2612 * normalized nr_running number somewhere that negates
2615 if (sgs->sum_nr_running)
2616 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2618 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2621 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2622 if (!sgs->group_capacity)
2623 sgs->group_capacity = fix_small_capacity(sd, group);
2625 if (sgs->group_capacity > sgs->sum_nr_running)
2626 sgs->group_has_capacity = 1;
2630 * update_sd_pick_busiest - return 1 on busiest group
2631 * @sd: sched_domain whose statistics are to be checked
2632 * @sds: sched_domain statistics
2633 * @sg: sched_group candidate to be checked for being the busiest
2634 * @sgs: sched_group statistics
2635 * @this_cpu: the current cpu
2637 * Determine if @sg is a busier group than the previously selected
2640 static bool update_sd_pick_busiest(struct sched_domain *sd,
2641 struct sd_lb_stats *sds,
2642 struct sched_group *sg,
2643 struct sg_lb_stats *sgs,
2646 if (sgs->avg_load <= sds->max_load)
2649 if (sgs->sum_nr_running > sgs->group_capacity)
2656 * ASYM_PACKING needs to move all the work to the lowest
2657 * numbered CPUs in the group, therefore mark all groups
2658 * higher than ourself as busy.
2660 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2661 this_cpu < group_first_cpu(sg)) {
2665 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2673 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2674 * @sd: sched_domain whose statistics are to be updated.
2675 * @this_cpu: Cpu for which load balance is currently performed.
2676 * @idle: Idle status of this_cpu
2677 * @sd_idle: Idle status of the sched_domain containing sg.
2678 * @cpus: Set of cpus considered for load balancing.
2679 * @balance: Should we balance.
2680 * @sds: variable to hold the statistics for this sched_domain.
2682 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2683 enum cpu_idle_type idle, int *sd_idle,
2684 const struct cpumask *cpus, int *balance,
2685 struct sd_lb_stats *sds)
2687 struct sched_domain *child = sd->child;
2688 struct sched_group *sg = sd->groups;
2689 struct sg_lb_stats sgs;
2690 int load_idx, prefer_sibling = 0;
2692 if (child && child->flags & SD_PREFER_SIBLING)
2695 init_sd_power_savings_stats(sd, sds, idle);
2696 load_idx = get_sd_load_idx(sd, idle);
2701 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2702 memset(&sgs, 0, sizeof(sgs));
2703 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2704 local_group, cpus, balance, &sgs);
2706 if (local_group && !(*balance))
2709 sds->total_load += sgs.group_load;
2710 sds->total_pwr += sg->cpu_power;
2713 * In case the child domain prefers tasks go to siblings
2714 * first, lower the sg capacity to one so that we'll try
2715 * and move all the excess tasks away. We lower the capacity
2716 * of a group only if the local group has the capacity to fit
2717 * these excess tasks, i.e. nr_running < group_capacity. The
2718 * extra check prevents the case where you always pull from the
2719 * heaviest group when it is already under-utilized (possible
2720 * with a large weight task outweighs the tasks on the system).
2722 if (prefer_sibling && !local_group && sds->this_has_capacity)
2723 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2726 sds->this_load = sgs.avg_load;
2728 sds->this_nr_running = sgs.sum_nr_running;
2729 sds->this_load_per_task = sgs.sum_weighted_load;
2730 sds->this_has_capacity = sgs.group_has_capacity;
2731 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2732 sds->max_load = sgs.avg_load;
2734 sds->busiest_nr_running = sgs.sum_nr_running;
2735 sds->busiest_group_capacity = sgs.group_capacity;
2736 sds->busiest_load_per_task = sgs.sum_weighted_load;
2737 sds->busiest_has_capacity = sgs.group_has_capacity;
2738 sds->group_imb = sgs.group_imb;
2741 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2743 } while (sg != sd->groups);
2746 int __weak arch_sd_sibling_asym_packing(void)
2748 return 0*SD_ASYM_PACKING;
2752 * check_asym_packing - Check to see if the group is packed into the
2755 * This is primarily intended to used at the sibling level. Some
2756 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2757 * case of POWER7, it can move to lower SMT modes only when higher
2758 * threads are idle. When in lower SMT modes, the threads will
2759 * perform better since they share less core resources. Hence when we
2760 * have idle threads, we want them to be the higher ones.
2762 * This packing function is run on idle threads. It checks to see if
2763 * the busiest CPU in this domain (core in the P7 case) has a higher
2764 * CPU number than the packing function is being run on. Here we are
2765 * assuming lower CPU number will be equivalent to lower a SMT thread
2768 * Returns 1 when packing is required and a task should be moved to
2769 * this CPU. The amount of the imbalance is returned in *imbalance.
2771 * @sd: The sched_domain whose packing is to be checked.
2772 * @sds: Statistics of the sched_domain which is to be packed
2773 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2774 * @imbalance: returns amount of imbalanced due to packing.
2776 static int check_asym_packing(struct sched_domain *sd,
2777 struct sd_lb_stats *sds,
2778 int this_cpu, unsigned long *imbalance)
2782 if (!(sd->flags & SD_ASYM_PACKING))
2788 busiest_cpu = group_first_cpu(sds->busiest);
2789 if (this_cpu > busiest_cpu)
2792 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2798 * fix_small_imbalance - Calculate the minor imbalance that exists
2799 * amongst the groups of a sched_domain, during
2801 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2802 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2803 * @imbalance: Variable to store the imbalance.
2805 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2806 int this_cpu, unsigned long *imbalance)
2808 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2809 unsigned int imbn = 2;
2810 unsigned long scaled_busy_load_per_task;
2812 if (sds->this_nr_running) {
2813 sds->this_load_per_task /= sds->this_nr_running;
2814 if (sds->busiest_load_per_task >
2815 sds->this_load_per_task)
2818 sds->this_load_per_task =
2819 cpu_avg_load_per_task(this_cpu);
2821 scaled_busy_load_per_task = sds->busiest_load_per_task
2823 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2825 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2826 (scaled_busy_load_per_task * imbn)) {
2827 *imbalance = sds->busiest_load_per_task;
2832 * OK, we don't have enough imbalance to justify moving tasks,
2833 * however we may be able to increase total CPU power used by
2837 pwr_now += sds->busiest->cpu_power *
2838 min(sds->busiest_load_per_task, sds->max_load);
2839 pwr_now += sds->this->cpu_power *
2840 min(sds->this_load_per_task, sds->this_load);
2841 pwr_now /= SCHED_LOAD_SCALE;
2843 /* Amount of load we'd subtract */
2844 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2845 sds->busiest->cpu_power;
2846 if (sds->max_load > tmp)
2847 pwr_move += sds->busiest->cpu_power *
2848 min(sds->busiest_load_per_task, sds->max_load - tmp);
2850 /* Amount of load we'd add */
2851 if (sds->max_load * sds->busiest->cpu_power <
2852 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2853 tmp = (sds->max_load * sds->busiest->cpu_power) /
2854 sds->this->cpu_power;
2856 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2857 sds->this->cpu_power;
2858 pwr_move += sds->this->cpu_power *
2859 min(sds->this_load_per_task, sds->this_load + tmp);
2860 pwr_move /= SCHED_LOAD_SCALE;
2862 /* Move if we gain throughput */
2863 if (pwr_move > pwr_now)
2864 *imbalance = sds->busiest_load_per_task;
2868 * calculate_imbalance - Calculate the amount of imbalance present within the
2869 * groups of a given sched_domain during load balance.
2870 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2871 * @this_cpu: Cpu for which currently load balance is being performed.
2872 * @imbalance: The variable to store the imbalance.
2874 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2875 unsigned long *imbalance)
2877 unsigned long max_pull, load_above_capacity = ~0UL;
2879 sds->busiest_load_per_task /= sds->busiest_nr_running;
2880 if (sds->group_imb) {
2881 sds->busiest_load_per_task =
2882 min(sds->busiest_load_per_task, sds->avg_load);
2886 * In the presence of smp nice balancing, certain scenarios can have
2887 * max load less than avg load(as we skip the groups at or below
2888 * its cpu_power, while calculating max_load..)
2890 if (sds->max_load < sds->avg_load) {
2892 return fix_small_imbalance(sds, this_cpu, imbalance);
2895 if (!sds->group_imb) {
2897 * Don't want to pull so many tasks that a group would go idle.
2899 load_above_capacity = (sds->busiest_nr_running -
2900 sds->busiest_group_capacity);
2902 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2904 load_above_capacity /= sds->busiest->cpu_power;
2908 * We're trying to get all the cpus to the average_load, so we don't
2909 * want to push ourselves above the average load, nor do we wish to
2910 * reduce the max loaded cpu below the average load. At the same time,
2911 * we also don't want to reduce the group load below the group capacity
2912 * (so that we can implement power-savings policies etc). Thus we look
2913 * for the minimum possible imbalance.
2914 * Be careful of negative numbers as they'll appear as very large values
2915 * with unsigned longs.
2917 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2919 /* How much load to actually move to equalise the imbalance */
2920 *imbalance = min(max_pull * sds->busiest->cpu_power,
2921 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
2925 * if *imbalance is less than the average load per runnable task
2926 * there is no gaurantee that any tasks will be moved so we'll have
2927 * a think about bumping its value to force at least one task to be
2930 if (*imbalance < sds->busiest_load_per_task)
2931 return fix_small_imbalance(sds, this_cpu, imbalance);
2935 /******* find_busiest_group() helpers end here *********************/
2938 * find_busiest_group - Returns the busiest group within the sched_domain
2939 * if there is an imbalance. If there isn't an imbalance, and
2940 * the user has opted for power-savings, it returns a group whose
2941 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2942 * such a group exists.
2944 * Also calculates the amount of weighted load which should be moved
2945 * to restore balance.
2947 * @sd: The sched_domain whose busiest group is to be returned.
2948 * @this_cpu: The cpu for which load balancing is currently being performed.
2949 * @imbalance: Variable which stores amount of weighted load which should
2950 * be moved to restore balance/put a group to idle.
2951 * @idle: The idle status of this_cpu.
2952 * @sd_idle: The idleness of sd
2953 * @cpus: The set of CPUs under consideration for load-balancing.
2954 * @balance: Pointer to a variable indicating if this_cpu
2955 * is the appropriate cpu to perform load balancing at this_level.
2957 * Returns: - the busiest group if imbalance exists.
2958 * - If no imbalance and user has opted for power-savings balance,
2959 * return the least loaded group whose CPUs can be
2960 * put to idle by rebalancing its tasks onto our group.
2962 static struct sched_group *
2963 find_busiest_group(struct sched_domain *sd, int this_cpu,
2964 unsigned long *imbalance, enum cpu_idle_type idle,
2965 int *sd_idle, const struct cpumask *cpus, int *balance)
2967 struct sd_lb_stats sds;
2969 memset(&sds, 0, sizeof(sds));
2972 * Compute the various statistics relavent for load balancing at
2975 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
2978 /* Cases where imbalance does not exist from POV of this_cpu */
2979 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2981 * 2) There is no busy sibling group to pull from.
2982 * 3) This group is the busiest group.
2983 * 4) This group is more busy than the avg busieness at this
2985 * 5) The imbalance is within the specified limit.
2987 * Note: when doing newidle balance, if the local group has excess
2988 * capacity (i.e. nr_running < group_capacity) and the busiest group
2989 * does not have any capacity, we force a load balance to pull tasks
2990 * to the local group. In this case, we skip past checks 3, 4 and 5.
2995 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
2996 check_asym_packing(sd, &sds, this_cpu, imbalance))
2999 if (!sds.busiest || sds.busiest_nr_running == 0)
3002 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3003 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3004 !sds.busiest_has_capacity)
3007 if (sds.this_load >= sds.max_load)
3010 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3012 if (sds.this_load >= sds.avg_load)
3015 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3019 /* Looks like there is an imbalance. Compute it */
3020 calculate_imbalance(&sds, this_cpu, imbalance);
3025 * There is no obvious imbalance. But check if we can do some balancing
3028 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3036 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3039 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3040 enum cpu_idle_type idle, unsigned long imbalance,
3041 const struct cpumask *cpus)
3043 struct rq *busiest = NULL, *rq;
3044 unsigned long max_load = 0;
3047 for_each_cpu(i, sched_group_cpus(group)) {
3048 unsigned long power = power_of(i);
3049 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3053 capacity = fix_small_capacity(sd, group);
3055 if (!cpumask_test_cpu(i, cpus))
3059 wl = weighted_cpuload(i);
3062 * When comparing with imbalance, use weighted_cpuload()
3063 * which is not scaled with the cpu power.
3065 if (capacity && rq->nr_running == 1 && wl > imbalance)
3069 * For the load comparisons with the other cpu's, consider
3070 * the weighted_cpuload() scaled with the cpu power, so that
3071 * the load can be moved away from the cpu that is potentially
3072 * running at a lower capacity.
3074 wl = (wl * SCHED_LOAD_SCALE) / power;
3076 if (wl > max_load) {
3086 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3087 * so long as it is large enough.
3089 #define MAX_PINNED_INTERVAL 512
3091 /* Working cpumask for load_balance and load_balance_newidle. */
3092 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3094 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
3095 int busiest_cpu, int this_cpu)
3097 if (idle == CPU_NEWLY_IDLE) {
3100 * ASYM_PACKING needs to force migrate tasks from busy but
3101 * higher numbered CPUs in order to pack all tasks in the
3102 * lowest numbered CPUs.
3104 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3108 * The only task running in a non-idle cpu can be moved to this
3109 * cpu in an attempt to completely freeup the other CPU
3112 * The package power saving logic comes from
3113 * find_busiest_group(). If there are no imbalance, then
3114 * f_b_g() will return NULL. However when sched_mc={1,2} then
3115 * f_b_g() will select a group from which a running task may be
3116 * pulled to this cpu in order to make the other package idle.
3117 * If there is no opportunity to make a package idle and if
3118 * there are no imbalance, then f_b_g() will return NULL and no
3119 * action will be taken in load_balance_newidle().
3121 * Under normal task pull operation due to imbalance, there
3122 * will be more than one task in the source run queue and
3123 * move_tasks() will succeed. ld_moved will be true and this
3124 * active balance code will not be triggered.
3126 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3127 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3130 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3134 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3137 static int active_load_balance_cpu_stop(void *data);
3140 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3141 * tasks if there is an imbalance.
3143 static int load_balance(int this_cpu, struct rq *this_rq,
3144 struct sched_domain *sd, enum cpu_idle_type idle,
3147 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3148 struct sched_group *group;
3149 unsigned long imbalance;
3151 unsigned long flags;
3152 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3154 cpumask_copy(cpus, cpu_active_mask);
3157 * When power savings policy is enabled for the parent domain, idle
3158 * sibling can pick up load irrespective of busy siblings. In this case,
3159 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3160 * portraying it as CPU_NOT_IDLE.
3162 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3163 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3166 schedstat_inc(sd, lb_count[idle]);
3169 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3176 schedstat_inc(sd, lb_nobusyg[idle]);
3180 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3182 schedstat_inc(sd, lb_nobusyq[idle]);
3186 BUG_ON(busiest == this_rq);
3188 schedstat_add(sd, lb_imbalance[idle], imbalance);
3191 if (busiest->nr_running > 1) {
3193 * Attempt to move tasks. If find_busiest_group has found
3194 * an imbalance but busiest->nr_running <= 1, the group is
3195 * still unbalanced. ld_moved simply stays zero, so it is
3196 * correctly treated as an imbalance.
3198 local_irq_save(flags);
3199 double_rq_lock(this_rq, busiest);
3200 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3201 imbalance, sd, idle, &all_pinned);
3202 double_rq_unlock(this_rq, busiest);
3203 local_irq_restore(flags);
3206 * some other cpu did the load balance for us.
3208 if (ld_moved && this_cpu != smp_processor_id())
3209 resched_cpu(this_cpu);
3211 /* All tasks on this runqueue were pinned by CPU affinity */
3212 if (unlikely(all_pinned)) {
3213 cpumask_clear_cpu(cpu_of(busiest), cpus);
3214 if (!cpumask_empty(cpus))
3221 schedstat_inc(sd, lb_failed[idle]);
3223 * Increment the failure counter only on periodic balance.
3224 * We do not want newidle balance, which can be very
3225 * frequent, pollute the failure counter causing
3226 * excessive cache_hot migrations and active balances.
3228 if (idle != CPU_NEWLY_IDLE)
3229 sd->nr_balance_failed++;
3231 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3233 raw_spin_lock_irqsave(&busiest->lock, flags);
3235 /* don't kick the active_load_balance_cpu_stop,
3236 * if the curr task on busiest cpu can't be
3239 if (!cpumask_test_cpu(this_cpu,
3240 &busiest->curr->cpus_allowed)) {
3241 raw_spin_unlock_irqrestore(&busiest->lock,
3244 goto out_one_pinned;
3248 * ->active_balance synchronizes accesses to
3249 * ->active_balance_work. Once set, it's cleared
3250 * only after active load balance is finished.
3252 if (!busiest->active_balance) {
3253 busiest->active_balance = 1;
3254 busiest->push_cpu = this_cpu;
3257 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3260 stop_one_cpu_nowait(cpu_of(busiest),
3261 active_load_balance_cpu_stop, busiest,
3262 &busiest->active_balance_work);
3265 * We've kicked active balancing, reset the failure
3268 sd->nr_balance_failed = sd->cache_nice_tries+1;
3271 sd->nr_balance_failed = 0;
3273 if (likely(!active_balance)) {
3274 /* We were unbalanced, so reset the balancing interval */
3275 sd->balance_interval = sd->min_interval;
3278 * If we've begun active balancing, start to back off. This
3279 * case may not be covered by the all_pinned logic if there
3280 * is only 1 task on the busy runqueue (because we don't call
3283 if (sd->balance_interval < sd->max_interval)
3284 sd->balance_interval *= 2;
3287 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3288 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3294 schedstat_inc(sd, lb_balanced[idle]);
3296 sd->nr_balance_failed = 0;
3299 /* tune up the balancing interval */
3300 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3301 (sd->balance_interval < sd->max_interval))
3302 sd->balance_interval *= 2;
3304 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3305 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3314 * idle_balance is called by schedule() if this_cpu is about to become
3315 * idle. Attempts to pull tasks from other CPUs.
3317 static void idle_balance(int this_cpu, struct rq *this_rq)
3319 struct sched_domain *sd;
3320 int pulled_task = 0;
3321 unsigned long next_balance = jiffies + HZ;
3323 this_rq->idle_stamp = this_rq->clock;
3325 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3329 * Drop the rq->lock, but keep IRQ/preempt disabled.
3331 raw_spin_unlock(&this_rq->lock);
3333 for_each_domain(this_cpu, sd) {
3334 unsigned long interval;
3337 if (!(sd->flags & SD_LOAD_BALANCE))
3340 if (sd->flags & SD_BALANCE_NEWIDLE) {
3341 /* If we've pulled tasks over stop searching: */
3342 pulled_task = load_balance(this_cpu, this_rq,
3343 sd, CPU_NEWLY_IDLE, &balance);
3346 interval = msecs_to_jiffies(sd->balance_interval);
3347 if (time_after(next_balance, sd->last_balance + interval))
3348 next_balance = sd->last_balance + interval;
3353 raw_spin_lock(&this_rq->lock);
3355 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3357 * We are going idle. next_balance may be set based on
3358 * a busy processor. So reset next_balance.
3360 this_rq->next_balance = next_balance;
3365 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3366 * running tasks off the busiest CPU onto idle CPUs. It requires at
3367 * least 1 task to be running on each physical CPU where possible, and
3368 * avoids physical / logical imbalances.
3370 static int active_load_balance_cpu_stop(void *data)
3372 struct rq *busiest_rq = data;
3373 int busiest_cpu = cpu_of(busiest_rq);
3374 int target_cpu = busiest_rq->push_cpu;
3375 struct rq *target_rq = cpu_rq(target_cpu);
3376 struct sched_domain *sd;
3378 raw_spin_lock_irq(&busiest_rq->lock);
3380 /* make sure the requested cpu hasn't gone down in the meantime */
3381 if (unlikely(busiest_cpu != smp_processor_id() ||
3382 !busiest_rq->active_balance))
3385 /* Is there any task to move? */
3386 if (busiest_rq->nr_running <= 1)
3390 * This condition is "impossible", if it occurs
3391 * we need to fix it. Originally reported by
3392 * Bjorn Helgaas on a 128-cpu setup.
3394 BUG_ON(busiest_rq == target_rq);
3396 /* move a task from busiest_rq to target_rq */
3397 double_lock_balance(busiest_rq, target_rq);
3399 /* Search for an sd spanning us and the target CPU. */
3400 for_each_domain(target_cpu, sd) {
3401 if ((sd->flags & SD_LOAD_BALANCE) &&
3402 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3407 schedstat_inc(sd, alb_count);
3409 if (move_one_task(target_rq, target_cpu, busiest_rq,
3411 schedstat_inc(sd, alb_pushed);
3413 schedstat_inc(sd, alb_failed);
3415 double_unlock_balance(busiest_rq, target_rq);
3417 busiest_rq->active_balance = 0;
3418 raw_spin_unlock_irq(&busiest_rq->lock);
3424 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3426 static void trigger_sched_softirq(void *data)
3428 raise_softirq_irqoff(SCHED_SOFTIRQ);
3431 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3433 csd->func = trigger_sched_softirq;
3440 * idle load balancing details
3441 * - One of the idle CPUs nominates itself as idle load_balancer, while
3443 * - This idle load balancer CPU will also go into tickless mode when
3444 * it is idle, just like all other idle CPUs
3445 * - When one of the busy CPUs notice that there may be an idle rebalancing
3446 * needed, they will kick the idle load balancer, which then does idle
3447 * load balancing for all the idle CPUs.
3450 atomic_t load_balancer;
3451 atomic_t first_pick_cpu;
3452 atomic_t second_pick_cpu;
3453 cpumask_var_t idle_cpus_mask;
3454 cpumask_var_t grp_idle_mask;
3455 unsigned long next_balance; /* in jiffy units */
3456 } nohz ____cacheline_aligned;
3458 int get_nohz_load_balancer(void)
3460 return atomic_read(&nohz.load_balancer);
3463 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3465 * lowest_flag_domain - Return lowest sched_domain containing flag.
3466 * @cpu: The cpu whose lowest level of sched domain is to
3468 * @flag: The flag to check for the lowest sched_domain
3469 * for the given cpu.
3471 * Returns the lowest sched_domain of a cpu which contains the given flag.
3473 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3475 struct sched_domain *sd;
3477 for_each_domain(cpu, sd)
3478 if (sd && (sd->flags & flag))
3485 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3486 * @cpu: The cpu whose domains we're iterating over.
3487 * @sd: variable holding the value of the power_savings_sd
3489 * @flag: The flag to filter the sched_domains to be iterated.
3491 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3492 * set, starting from the lowest sched_domain to the highest.
3494 #define for_each_flag_domain(cpu, sd, flag) \
3495 for (sd = lowest_flag_domain(cpu, flag); \
3496 (sd && (sd->flags & flag)); sd = sd->parent)
3499 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3500 * @ilb_group: group to be checked for semi-idleness
3502 * Returns: 1 if the group is semi-idle. 0 otherwise.
3504 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3505 * and atleast one non-idle CPU. This helper function checks if the given
3506 * sched_group is semi-idle or not.
3508 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3510 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3511 sched_group_cpus(ilb_group));
3514 * A sched_group is semi-idle when it has atleast one busy cpu
3515 * and atleast one idle cpu.
3517 if (cpumask_empty(nohz.grp_idle_mask))
3520 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3526 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3527 * @cpu: The cpu which is nominating a new idle_load_balancer.
3529 * Returns: Returns the id of the idle load balancer if it exists,
3530 * Else, returns >= nr_cpu_ids.
3532 * This algorithm picks the idle load balancer such that it belongs to a
3533 * semi-idle powersavings sched_domain. The idea is to try and avoid
3534 * completely idle packages/cores just for the purpose of idle load balancing
3535 * when there are other idle cpu's which are better suited for that job.
3537 static int find_new_ilb(int cpu)
3539 struct sched_domain *sd;
3540 struct sched_group *ilb_group;
3543 * Have idle load balancer selection from semi-idle packages only
3544 * when power-aware load balancing is enabled
3546 if (!(sched_smt_power_savings || sched_mc_power_savings))
3550 * Optimize for the case when we have no idle CPUs or only one
3551 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3553 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3556 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3557 ilb_group = sd->groups;
3560 if (is_semi_idle_group(ilb_group))
3561 return cpumask_first(nohz.grp_idle_mask);
3563 ilb_group = ilb_group->next;
3565 } while (ilb_group != sd->groups);
3571 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3572 static inline int find_new_ilb(int call_cpu)
3579 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3580 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3581 * CPU (if there is one).
3583 static void nohz_balancer_kick(int cpu)
3587 nohz.next_balance++;
3589 ilb_cpu = get_nohz_load_balancer();
3591 if (ilb_cpu >= nr_cpu_ids) {
3592 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3593 if (ilb_cpu >= nr_cpu_ids)
3597 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3598 struct call_single_data *cp;
3600 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3601 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3602 __smp_call_function_single(ilb_cpu, cp, 0);
3608 * This routine will try to nominate the ilb (idle load balancing)
3609 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3610 * load balancing on behalf of all those cpus.
3612 * When the ilb owner becomes busy, we will not have new ilb owner until some
3613 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3614 * idle load balancing by kicking one of the idle CPUs.
3616 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3617 * ilb owner CPU in future (when there is a need for idle load balancing on
3618 * behalf of all idle CPUs).
3620 void select_nohz_load_balancer(int stop_tick)
3622 int cpu = smp_processor_id();
3625 if (!cpu_active(cpu)) {
3626 if (atomic_read(&nohz.load_balancer) != cpu)
3630 * If we are going offline and still the leader,
3633 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3640 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3642 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3643 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3644 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3645 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3647 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3650 /* make me the ilb owner */
3651 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3656 * Check to see if there is a more power-efficient
3659 new_ilb = find_new_ilb(cpu);
3660 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3661 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3662 resched_cpu(new_ilb);
3668 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3671 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3673 if (atomic_read(&nohz.load_balancer) == cpu)
3674 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3682 static DEFINE_SPINLOCK(balancing);
3685 * It checks each scheduling domain to see if it is due to be balanced,
3686 * and initiates a balancing operation if so.
3688 * Balancing parameters are set up in arch_init_sched_domains.
3690 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3693 struct rq *rq = cpu_rq(cpu);
3694 unsigned long interval;
3695 struct sched_domain *sd;
3696 /* Earliest time when we have to do rebalance again */
3697 unsigned long next_balance = jiffies + 60*HZ;
3698 int update_next_balance = 0;
3703 for_each_domain(cpu, sd) {
3704 if (!(sd->flags & SD_LOAD_BALANCE))
3707 interval = sd->balance_interval;
3708 if (idle != CPU_IDLE)
3709 interval *= sd->busy_factor;
3711 /* scale ms to jiffies */
3712 interval = msecs_to_jiffies(interval);
3713 if (unlikely(!interval))
3715 if (interval > HZ*NR_CPUS/10)
3716 interval = HZ*NR_CPUS/10;
3718 need_serialize = sd->flags & SD_SERIALIZE;
3720 if (need_serialize) {
3721 if (!spin_trylock(&balancing))
3725 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3726 if (load_balance(cpu, rq, sd, idle, &balance)) {
3728 * We've pulled tasks over so either we're no
3729 * longer idle, or one of our SMT siblings is
3732 idle = CPU_NOT_IDLE;
3734 sd->last_balance = jiffies;
3737 spin_unlock(&balancing);
3739 if (time_after(next_balance, sd->last_balance + interval)) {
3740 next_balance = sd->last_balance + interval;
3741 update_next_balance = 1;
3745 * Stop the load balance at this level. There is another
3746 * CPU in our sched group which is doing load balancing more
3754 * next_balance will be updated only when there is a need.
3755 * When the cpu is attached to null domain for ex, it will not be
3758 if (likely(update_next_balance))
3759 rq->next_balance = next_balance;
3764 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3765 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3767 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3769 struct rq *this_rq = cpu_rq(this_cpu);
3773 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3776 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3777 if (balance_cpu == this_cpu)
3781 * If this cpu gets work to do, stop the load balancing
3782 * work being done for other cpus. Next load
3783 * balancing owner will pick it up.
3785 if (need_resched()) {
3786 this_rq->nohz_balance_kick = 0;
3790 raw_spin_lock_irq(&this_rq->lock);
3791 update_rq_clock(this_rq);
3792 update_cpu_load(this_rq);
3793 raw_spin_unlock_irq(&this_rq->lock);
3795 rebalance_domains(balance_cpu, CPU_IDLE);
3797 rq = cpu_rq(balance_cpu);
3798 if (time_after(this_rq->next_balance, rq->next_balance))
3799 this_rq->next_balance = rq->next_balance;
3801 nohz.next_balance = this_rq->next_balance;
3802 this_rq->nohz_balance_kick = 0;
3806 * Current heuristic for kicking the idle load balancer
3807 * - first_pick_cpu is the one of the busy CPUs. It will kick
3808 * idle load balancer when it has more than one process active. This
3809 * eliminates the need for idle load balancing altogether when we have
3810 * only one running process in the system (common case).
3811 * - If there are more than one busy CPU, idle load balancer may have
3812 * to run for active_load_balance to happen (i.e., two busy CPUs are
3813 * SMT or core siblings and can run better if they move to different
3814 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3815 * which will kick idle load balancer as soon as it has any load.
3817 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3819 unsigned long now = jiffies;
3821 int first_pick_cpu, second_pick_cpu;
3823 if (time_before(now, nohz.next_balance))
3826 if (rq->idle_at_tick)
3829 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3830 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3832 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3833 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3836 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3837 if (ret == nr_cpu_ids || ret == cpu) {
3838 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3839 if (rq->nr_running > 1)
3842 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3843 if (ret == nr_cpu_ids || ret == cpu) {
3851 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3855 * run_rebalance_domains is triggered when needed from the scheduler tick.
3856 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3858 static void run_rebalance_domains(struct softirq_action *h)
3860 int this_cpu = smp_processor_id();
3861 struct rq *this_rq = cpu_rq(this_cpu);
3862 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3863 CPU_IDLE : CPU_NOT_IDLE;
3865 rebalance_domains(this_cpu, idle);
3868 * If this cpu has a pending nohz_balance_kick, then do the
3869 * balancing on behalf of the other idle cpus whose ticks are
3872 nohz_idle_balance(this_cpu, idle);
3875 static inline int on_null_domain(int cpu)
3877 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3881 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3883 static inline void trigger_load_balance(struct rq *rq, int cpu)
3885 /* Don't need to rebalance while attached to NULL domain */
3886 if (time_after_eq(jiffies, rq->next_balance) &&
3887 likely(!on_null_domain(cpu)))
3888 raise_softirq(SCHED_SOFTIRQ);
3890 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
3891 nohz_balancer_kick(cpu);
3895 static void rq_online_fair(struct rq *rq)
3900 static void rq_offline_fair(struct rq *rq)
3905 #else /* CONFIG_SMP */
3908 * on UP we do not need to balance between CPUs:
3910 static inline void idle_balance(int cpu, struct rq *rq)
3914 #endif /* CONFIG_SMP */
3917 * scheduler tick hitting a task of our scheduling class:
3919 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
3921 struct cfs_rq *cfs_rq;
3922 struct sched_entity *se = &curr->se;
3924 for_each_sched_entity(se) {
3925 cfs_rq = cfs_rq_of(se);
3926 entity_tick(cfs_rq, se, queued);
3931 * called on fork with the child task as argument from the parent's context
3932 * - child not yet on the tasklist
3933 * - preemption disabled
3935 static void task_fork_fair(struct task_struct *p)
3937 struct cfs_rq *cfs_rq = task_cfs_rq(current);
3938 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
3939 int this_cpu = smp_processor_id();
3940 struct rq *rq = this_rq();
3941 unsigned long flags;
3943 raw_spin_lock_irqsave(&rq->lock, flags);
3945 update_rq_clock(rq);
3947 if (unlikely(task_cpu(p) != this_cpu)) {
3949 __set_task_cpu(p, this_cpu);
3953 update_curr(cfs_rq);
3956 se->vruntime = curr->vruntime;
3957 place_entity(cfs_rq, se, 1);
3959 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
3961 * Upon rescheduling, sched_class::put_prev_task() will place
3962 * 'current' within the tree based on its new key value.
3964 swap(curr->vruntime, se->vruntime);
3965 resched_task(rq->curr);
3968 se->vruntime -= cfs_rq->min_vruntime;
3970 raw_spin_unlock_irqrestore(&rq->lock, flags);
3974 * Priority of the task has changed. Check to see if we preempt
3977 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
3978 int oldprio, int running)
3981 * Reschedule if we are currently running on this runqueue and
3982 * our priority decreased, or if we are not currently running on
3983 * this runqueue and our priority is higher than the current's
3986 if (p->prio > oldprio)
3987 resched_task(rq->curr);
3989 check_preempt_curr(rq, p, 0);
3993 * We switched to the sched_fair class.
3995 static void switched_to_fair(struct rq *rq, struct task_struct *p,
3999 * We were most likely switched from sched_rt, so
4000 * kick off the schedule if running, otherwise just see
4001 * if we can still preempt the current task.
4004 resched_task(rq->curr);
4006 check_preempt_curr(rq, p, 0);
4009 /* Account for a task changing its policy or group.
4011 * This routine is mostly called to set cfs_rq->curr field when a task
4012 * migrates between groups/classes.
4014 static void set_curr_task_fair(struct rq *rq)
4016 struct sched_entity *se = &rq->curr->se;
4018 for_each_sched_entity(se)
4019 set_next_entity(cfs_rq_of(se), se);
4022 #ifdef CONFIG_FAIR_GROUP_SCHED
4023 static void task_move_group_fair(struct task_struct *p, int on_rq)
4026 * If the task was not on the rq at the time of this cgroup movement
4027 * it must have been asleep, sleeping tasks keep their ->vruntime
4028 * absolute on their old rq until wakeup (needed for the fair sleeper
4029 * bonus in place_entity()).
4031 * If it was on the rq, we've just 'preempted' it, which does convert
4032 * ->vruntime to a relative base.
4034 * Make sure both cases convert their relative position when migrating
4035 * to another cgroup's rq. This does somewhat interfere with the
4036 * fair sleeper stuff for the first placement, but who cares.
4039 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4040 set_task_rq(p, task_cpu(p));
4042 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4046 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4048 struct sched_entity *se = &task->se;
4049 unsigned int rr_interval = 0;
4052 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4055 if (rq->cfs.load.weight)
4056 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4062 * All the scheduling class methods:
4064 static const struct sched_class fair_sched_class = {
4065 .next = &idle_sched_class,
4066 .enqueue_task = enqueue_task_fair,
4067 .dequeue_task = dequeue_task_fair,
4068 .yield_task = yield_task_fair,
4070 .check_preempt_curr = check_preempt_wakeup,
4072 .pick_next_task = pick_next_task_fair,
4073 .put_prev_task = put_prev_task_fair,
4076 .select_task_rq = select_task_rq_fair,
4078 .rq_online = rq_online_fair,
4079 .rq_offline = rq_offline_fair,
4081 .task_waking = task_waking_fair,
4084 .set_curr_task = set_curr_task_fair,
4085 .task_tick = task_tick_fair,
4086 .task_fork = task_fork_fair,
4088 .prio_changed = prio_changed_fair,
4089 .switched_to = switched_to_fair,
4091 .get_rr_interval = get_rr_interval_fair,
4093 #ifdef CONFIG_FAIR_GROUP_SCHED
4094 .task_move_group = task_move_group_fair,
4098 #ifdef CONFIG_SCHED_DEBUG
4099 static void print_cfs_stats(struct seq_file *m, int cpu)
4101 struct cfs_rq *cfs_rq;
4104 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4105 print_cfs_rq(m, cpu, cfs_rq);