2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
28 * Targeted preemption latency for CPU-bound tasks:
29 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
31 * NOTE: this latency value is not the same as the concept of
32 * 'timeslice length' - timeslices in CFS are of variable length
33 * and have no persistent notion like in traditional, time-slice
34 * based scheduling concepts.
36 * (to see the precise effective timeslice length of your workload,
37 * run vmstat and monitor the context-switches (cs) field)
39 unsigned int sysctl_sched_latency = 6000000ULL;
40 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
43 * The initial- and re-scaling of tunables is configurable
44 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
47 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
48 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
49 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 enum sched_tunable_scaling sysctl_sched_tunable_scaling
52 = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
56 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
58 unsigned int sysctl_sched_min_granularity = 750000ULL;
59 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
62 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
64 static unsigned int sched_nr_latency = 8;
67 * After fork, child runs first. If set to 0 (default) then
68 * parent will (try to) run first.
70 unsigned int sysctl_sched_child_runs_first __read_mostly;
73 * SCHED_OTHER wake-up granularity.
74 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
81 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
83 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
86 * The exponential sliding window over which load is averaged for shares
90 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
92 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) {
150 * Ensure we either appear before our parent (if already
151 * enqueued) or force our parent to appear after us when it is
152 * enqueued. The fact that we always enqueue bottom-up
153 * reduces this to two cases.
155 if (cfs_rq->tg->parent &&
156 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
157 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
158 &rq_of(cfs_rq)->leaf_cfs_rq_list);
160 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
161 &rq_of(cfs_rq)->leaf_cfs_rq_list);
168 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
170 if (cfs_rq->on_list) {
171 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
176 /* Iterate thr' all leaf cfs_rq's on a runqueue */
177 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
178 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
180 /* Do the two (enqueued) entities belong to the same group ? */
182 is_same_group(struct sched_entity *se, struct sched_entity *pse)
184 if (se->cfs_rq == pse->cfs_rq)
190 static inline struct sched_entity *parent_entity(struct sched_entity *se)
195 /* return depth at which a sched entity is present in the hierarchy */
196 static inline int depth_se(struct sched_entity *se)
200 for_each_sched_entity(se)
207 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
209 int se_depth, pse_depth;
212 * preemption test can be made between sibling entities who are in the
213 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
214 * both tasks until we find their ancestors who are siblings of common
218 /* First walk up until both entities are at same depth */
219 se_depth = depth_se(*se);
220 pse_depth = depth_se(*pse);
222 while (se_depth > pse_depth) {
224 *se = parent_entity(*se);
227 while (pse_depth > se_depth) {
229 *pse = parent_entity(*pse);
232 while (!is_same_group(*se, *pse)) {
233 *se = parent_entity(*se);
234 *pse = parent_entity(*pse);
238 #else /* !CONFIG_FAIR_GROUP_SCHED */
240 static inline struct task_struct *task_of(struct sched_entity *se)
242 return container_of(se, struct task_struct, se);
245 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
247 return container_of(cfs_rq, struct rq, cfs);
250 #define entity_is_task(se) 1
252 #define for_each_sched_entity(se) \
253 for (; se; se = NULL)
255 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
257 return &task_rq(p)->cfs;
260 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
262 struct task_struct *p = task_of(se);
263 struct rq *rq = task_rq(p);
268 /* runqueue "owned" by this group */
269 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
274 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
276 return &cpu_rq(this_cpu)->cfs;
279 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
283 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
288 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
291 is_same_group(struct sched_entity *se, struct sched_entity *pse)
296 static inline struct sched_entity *parent_entity(struct sched_entity *se)
302 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
306 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 /**************************************************************
310 * Scheduling class tree data structure manipulation methods:
313 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
315 s64 delta = (s64)(vruntime - min_vruntime);
317 min_vruntime = vruntime;
322 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
324 s64 delta = (s64)(vruntime - min_vruntime);
326 min_vruntime = vruntime;
331 static inline int entity_before(struct sched_entity *a,
332 struct sched_entity *b)
334 return (s64)(a->vruntime - b->vruntime) < 0;
337 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
339 return se->vruntime - cfs_rq->min_vruntime;
342 static void update_min_vruntime(struct cfs_rq *cfs_rq)
344 u64 vruntime = cfs_rq->min_vruntime;
347 vruntime = cfs_rq->curr->vruntime;
349 if (cfs_rq->rb_leftmost) {
350 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
355 vruntime = se->vruntime;
357 vruntime = min_vruntime(vruntime, se->vruntime);
360 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
363 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
368 * Enqueue an entity into the rb-tree:
370 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
372 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
373 struct rb_node *parent = NULL;
374 struct sched_entity *entry;
375 s64 key = entity_key(cfs_rq, se);
379 * Find the right place in the rbtree:
383 entry = rb_entry(parent, struct sched_entity, run_node);
385 * We dont care about collisions. Nodes with
386 * the same key stay together.
388 if (key < entity_key(cfs_rq, entry)) {
389 link = &parent->rb_left;
391 link = &parent->rb_right;
397 * Maintain a cache of leftmost tree entries (it is frequently
401 cfs_rq->rb_leftmost = &se->run_node;
403 rb_link_node(&se->run_node, parent, link);
404 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
407 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
409 if (cfs_rq->rb_leftmost == &se->run_node) {
410 struct rb_node *next_node;
412 next_node = rb_next(&se->run_node);
413 cfs_rq->rb_leftmost = next_node;
416 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
419 static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
421 struct rb_node *left = cfs_rq->rb_leftmost;
426 return rb_entry(left, struct sched_entity, run_node);
429 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
431 struct rb_node *next = rb_next(&se->run_node);
436 return rb_entry(next, struct sched_entity, run_node);
439 #ifdef CONFIG_SCHED_DEBUG
440 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
442 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
447 return rb_entry(last, struct sched_entity, run_node);
450 /**************************************************************
451 * Scheduling class statistics methods:
454 int sched_proc_update_handler(struct ctl_table *table, int write,
455 void __user *buffer, size_t *lenp,
458 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
459 int factor = get_update_sysctl_factor();
464 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
465 sysctl_sched_min_granularity);
467 #define WRT_SYSCTL(name) \
468 (normalized_sysctl_##name = sysctl_##name / (factor))
469 WRT_SYSCTL(sched_min_granularity);
470 WRT_SYSCTL(sched_latency);
471 WRT_SYSCTL(sched_wakeup_granularity);
481 static inline unsigned long
482 calc_delta_fair(unsigned long delta, struct sched_entity *se)
484 if (unlikely(se->load.weight != NICE_0_LOAD))
485 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
491 * The idea is to set a period in which each task runs once.
493 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
494 * this period because otherwise the slices get too small.
496 * p = (nr <= nl) ? l : l*nr/nl
498 static u64 __sched_period(unsigned long nr_running)
500 u64 period = sysctl_sched_latency;
501 unsigned long nr_latency = sched_nr_latency;
503 if (unlikely(nr_running > nr_latency)) {
504 period = sysctl_sched_min_granularity;
505 period *= nr_running;
512 * We calculate the wall-time slice from the period by taking a part
513 * proportional to the weight.
517 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
519 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
521 for_each_sched_entity(se) {
522 struct load_weight *load;
523 struct load_weight lw;
525 cfs_rq = cfs_rq_of(se);
526 load = &cfs_rq->load;
528 if (unlikely(!se->on_rq)) {
531 update_load_add(&lw, se->load.weight);
534 slice = calc_delta_mine(slice, se->load.weight, load);
540 * We calculate the vruntime slice of a to be inserted task
544 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
546 return calc_delta_fair(sched_slice(cfs_rq, se), se);
549 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
550 static void update_cfs_shares(struct cfs_rq *cfs_rq);
553 * Update the current task's runtime statistics. Skip current tasks that
554 * are not in our scheduling class.
557 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
558 unsigned long delta_exec)
560 unsigned long delta_exec_weighted;
562 schedstat_set(curr->statistics.exec_max,
563 max((u64)delta_exec, curr->statistics.exec_max));
565 curr->sum_exec_runtime += delta_exec;
566 schedstat_add(cfs_rq, exec_clock, delta_exec);
567 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
569 curr->vruntime += delta_exec_weighted;
570 update_min_vruntime(cfs_rq);
572 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
573 cfs_rq->load_unacc_exec_time += delta_exec;
577 static void update_curr(struct cfs_rq *cfs_rq)
579 struct sched_entity *curr = cfs_rq->curr;
580 u64 now = rq_of(cfs_rq)->clock_task;
581 unsigned long delta_exec;
587 * Get the amount of time the current task was running
588 * since the last time we changed load (this cannot
589 * overflow on 32 bits):
591 delta_exec = (unsigned long)(now - curr->exec_start);
595 __update_curr(cfs_rq, curr, delta_exec);
596 curr->exec_start = now;
598 if (entity_is_task(curr)) {
599 struct task_struct *curtask = task_of(curr);
601 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
602 cpuacct_charge(curtask, delta_exec);
603 account_group_exec_runtime(curtask, delta_exec);
608 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
610 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
614 * Task is being enqueued - update stats:
616 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
619 * Are we enqueueing a waiting task? (for current tasks
620 * a dequeue/enqueue event is a NOP)
622 if (se != cfs_rq->curr)
623 update_stats_wait_start(cfs_rq, se);
627 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
629 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
630 rq_of(cfs_rq)->clock - se->statistics.wait_start));
631 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
632 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
633 rq_of(cfs_rq)->clock - se->statistics.wait_start);
634 #ifdef CONFIG_SCHEDSTATS
635 if (entity_is_task(se)) {
636 trace_sched_stat_wait(task_of(se),
637 rq_of(cfs_rq)->clock - se->statistics.wait_start);
640 schedstat_set(se->statistics.wait_start, 0);
644 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
647 * Mark the end of the wait period if dequeueing a
650 if (se != cfs_rq->curr)
651 update_stats_wait_end(cfs_rq, se);
655 * We are picking a new current task - update its stats:
658 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
661 * We are starting a new run period:
663 se->exec_start = rq_of(cfs_rq)->clock_task;
666 /**************************************************
667 * Scheduling class queueing methods:
670 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
672 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
674 cfs_rq->task_weight += weight;
678 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
684 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 update_load_add(&cfs_rq->load, se->load.weight);
687 if (!parent_entity(se))
688 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
689 if (entity_is_task(se)) {
690 add_cfs_task_weight(cfs_rq, se->load.weight);
691 list_add(&se->group_node, &cfs_rq->tasks);
693 cfs_rq->nr_running++;
697 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
699 update_load_sub(&cfs_rq->load, se->load.weight);
700 if (!parent_entity(se))
701 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
702 if (entity_is_task(se)) {
703 add_cfs_task_weight(cfs_rq, -se->load.weight);
704 list_del_init(&se->group_node);
706 cfs_rq->nr_running--;
709 #ifdef CONFIG_FAIR_GROUP_SCHED
711 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
714 struct task_group *tg = cfs_rq->tg;
717 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
718 load_avg -= cfs_rq->load_contribution;
720 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
721 atomic_add(load_avg, &tg->load_weight);
722 cfs_rq->load_contribution += load_avg;
726 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
728 u64 period = sysctl_sched_shares_window;
730 unsigned long load = cfs_rq->load.weight;
732 if (cfs_rq->tg == &root_task_group)
735 now = rq_of(cfs_rq)->clock_task;
736 delta = now - cfs_rq->load_stamp;
738 /* truncate load history at 4 idle periods */
739 if (cfs_rq->load_stamp > cfs_rq->load_last &&
740 now - cfs_rq->load_last > 4 * period) {
741 cfs_rq->load_period = 0;
742 cfs_rq->load_avg = 0;
746 cfs_rq->load_stamp = now;
747 cfs_rq->load_unacc_exec_time = 0;
748 cfs_rq->load_period += delta;
750 cfs_rq->load_last = now;
751 cfs_rq->load_avg += delta * load;
754 /* consider updating load contribution on each fold or truncate */
755 if (global_update || cfs_rq->load_period > period
756 || !cfs_rq->load_period)
757 update_cfs_rq_load_contribution(cfs_rq, global_update);
759 while (cfs_rq->load_period > period) {
761 * Inline assembly required to prevent the compiler
762 * optimising this loop into a divmod call.
763 * See __iter_div_u64_rem() for another example of this.
765 asm("" : "+rm" (cfs_rq->load_period));
766 cfs_rq->load_period /= 2;
767 cfs_rq->load_avg /= 2;
770 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
771 list_del_leaf_cfs_rq(cfs_rq);
774 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
776 long load_weight, load, shares;
778 load = cfs_rq->load.weight;
780 load_weight = atomic_read(&tg->load_weight);
782 load_weight -= cfs_rq->load_contribution;
784 shares = (tg->shares * load);
786 shares /= load_weight;
788 if (shares < MIN_SHARES)
790 if (shares > tg->shares)
796 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
798 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
799 update_cfs_load(cfs_rq, 0);
800 update_cfs_shares(cfs_rq);
803 # else /* CONFIG_SMP */
804 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
808 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
813 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
816 # endif /* CONFIG_SMP */
817 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
818 unsigned long weight)
821 /* commit outstanding execution time */
822 if (cfs_rq->curr == se)
824 account_entity_dequeue(cfs_rq, se);
827 update_load_set(&se->load, weight);
830 account_entity_enqueue(cfs_rq, se);
833 static void update_cfs_shares(struct cfs_rq *cfs_rq)
835 struct task_group *tg;
836 struct sched_entity *se;
840 se = tg->se[cpu_of(rq_of(cfs_rq))];
844 if (likely(se->load.weight == tg->shares))
847 shares = calc_cfs_shares(cfs_rq, tg);
849 reweight_entity(cfs_rq_of(se), se, shares);
851 #else /* CONFIG_FAIR_GROUP_SCHED */
852 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
856 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
860 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
863 #endif /* CONFIG_FAIR_GROUP_SCHED */
865 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
867 #ifdef CONFIG_SCHEDSTATS
868 struct task_struct *tsk = NULL;
870 if (entity_is_task(se))
873 if (se->statistics.sleep_start) {
874 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
879 if (unlikely(delta > se->statistics.sleep_max))
880 se->statistics.sleep_max = delta;
882 se->statistics.sleep_start = 0;
883 se->statistics.sum_sleep_runtime += delta;
886 account_scheduler_latency(tsk, delta >> 10, 1);
887 trace_sched_stat_sleep(tsk, delta);
890 if (se->statistics.block_start) {
891 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
896 if (unlikely(delta > se->statistics.block_max))
897 se->statistics.block_max = delta;
899 se->statistics.block_start = 0;
900 se->statistics.sum_sleep_runtime += delta;
903 if (tsk->in_iowait) {
904 se->statistics.iowait_sum += delta;
905 se->statistics.iowait_count++;
906 trace_sched_stat_iowait(tsk, delta);
910 * Blocking time is in units of nanosecs, so shift by
911 * 20 to get a milliseconds-range estimation of the
912 * amount of time that the task spent sleeping:
914 if (unlikely(prof_on == SLEEP_PROFILING)) {
915 profile_hits(SLEEP_PROFILING,
916 (void *)get_wchan(tsk),
919 account_scheduler_latency(tsk, delta >> 10, 0);
925 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
927 #ifdef CONFIG_SCHED_DEBUG
928 s64 d = se->vruntime - cfs_rq->min_vruntime;
933 if (d > 3*sysctl_sched_latency)
934 schedstat_inc(cfs_rq, nr_spread_over);
939 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
941 u64 vruntime = cfs_rq->min_vruntime;
944 * The 'current' period is already promised to the current tasks,
945 * however the extra weight of the new task will slow them down a
946 * little, place the new task so that it fits in the slot that
947 * stays open at the end.
949 if (initial && sched_feat(START_DEBIT))
950 vruntime += sched_vslice(cfs_rq, se);
952 /* sleeps up to a single latency don't count. */
954 unsigned long thresh = sysctl_sched_latency;
957 * Halve their sleep time's effect, to allow
958 * for a gentler effect of sleepers:
960 if (sched_feat(GENTLE_FAIR_SLEEPERS))
966 /* ensure we never gain time by being placed backwards. */
967 vruntime = max_vruntime(se->vruntime, vruntime);
969 se->vruntime = vruntime;
973 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
976 * Update the normalized vruntime before updating min_vruntime
977 * through callig update_curr().
979 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
980 se->vruntime += cfs_rq->min_vruntime;
983 * Update run-time statistics of the 'current'.
986 update_cfs_load(cfs_rq, 0);
987 account_entity_enqueue(cfs_rq, se);
988 update_cfs_shares(cfs_rq);
990 if (flags & ENQUEUE_WAKEUP) {
991 place_entity(cfs_rq, se, 0);
992 enqueue_sleeper(cfs_rq, se);
995 update_stats_enqueue(cfs_rq, se);
996 check_spread(cfs_rq, se);
997 if (se != cfs_rq->curr)
998 __enqueue_entity(cfs_rq, se);
1001 if (cfs_rq->nr_running == 1)
1002 list_add_leaf_cfs_rq(cfs_rq);
1005 static void __clear_buddies_last(struct sched_entity *se)
1007 for_each_sched_entity(se) {
1008 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1009 if (cfs_rq->last == se)
1010 cfs_rq->last = NULL;
1016 static void __clear_buddies_next(struct sched_entity *se)
1018 for_each_sched_entity(se) {
1019 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1020 if (cfs_rq->next == se)
1021 cfs_rq->next = NULL;
1027 static void __clear_buddies_skip(struct sched_entity *se)
1029 for_each_sched_entity(se) {
1030 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1031 if (cfs_rq->skip == se)
1032 cfs_rq->skip = NULL;
1038 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1040 if (cfs_rq->last == se)
1041 __clear_buddies_last(se);
1043 if (cfs_rq->next == se)
1044 __clear_buddies_next(se);
1046 if (cfs_rq->skip == se)
1047 __clear_buddies_skip(se);
1051 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1054 * Update run-time statistics of the 'current'.
1056 update_curr(cfs_rq);
1058 update_stats_dequeue(cfs_rq, se);
1059 if (flags & DEQUEUE_SLEEP) {
1060 #ifdef CONFIG_SCHEDSTATS
1061 if (entity_is_task(se)) {
1062 struct task_struct *tsk = task_of(se);
1064 if (tsk->state & TASK_INTERRUPTIBLE)
1065 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1066 if (tsk->state & TASK_UNINTERRUPTIBLE)
1067 se->statistics.block_start = rq_of(cfs_rq)->clock;
1072 clear_buddies(cfs_rq, se);
1074 if (se != cfs_rq->curr)
1075 __dequeue_entity(cfs_rq, se);
1077 update_cfs_load(cfs_rq, 0);
1078 account_entity_dequeue(cfs_rq, se);
1079 update_min_vruntime(cfs_rq);
1080 update_cfs_shares(cfs_rq);
1083 * Normalize the entity after updating the min_vruntime because the
1084 * update can refer to the ->curr item and we need to reflect this
1085 * movement in our normalized position.
1087 if (!(flags & DEQUEUE_SLEEP))
1088 se->vruntime -= cfs_rq->min_vruntime;
1092 * Preempt the current task with a newly woken task if needed:
1095 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1097 unsigned long ideal_runtime, delta_exec;
1099 ideal_runtime = sched_slice(cfs_rq, curr);
1100 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1101 if (delta_exec > ideal_runtime) {
1102 resched_task(rq_of(cfs_rq)->curr);
1104 * The current task ran long enough, ensure it doesn't get
1105 * re-elected due to buddy favours.
1107 clear_buddies(cfs_rq, curr);
1112 * Ensure that a task that missed wakeup preemption by a
1113 * narrow margin doesn't have to wait for a full slice.
1114 * This also mitigates buddy induced latencies under load.
1116 if (!sched_feat(WAKEUP_PREEMPT))
1119 if (delta_exec < sysctl_sched_min_granularity)
1122 if (cfs_rq->nr_running > 1) {
1123 struct sched_entity *se = __pick_first_entity(cfs_rq);
1124 s64 delta = curr->vruntime - se->vruntime;
1129 if (delta > ideal_runtime)
1130 resched_task(rq_of(cfs_rq)->curr);
1135 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1137 /* 'current' is not kept within the tree. */
1140 * Any task has to be enqueued before it get to execute on
1141 * a CPU. So account for the time it spent waiting on the
1144 update_stats_wait_end(cfs_rq, se);
1145 __dequeue_entity(cfs_rq, se);
1148 update_stats_curr_start(cfs_rq, se);
1150 #ifdef CONFIG_SCHEDSTATS
1152 * Track our maximum slice length, if the CPU's load is at
1153 * least twice that of our own weight (i.e. dont track it
1154 * when there are only lesser-weight tasks around):
1156 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1157 se->statistics.slice_max = max(se->statistics.slice_max,
1158 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1161 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1165 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1168 * Pick the next process, keeping these things in mind, in this order:
1169 * 1) keep things fair between processes/task groups
1170 * 2) pick the "next" process, since someone really wants that to run
1171 * 3) pick the "last" process, for cache locality
1172 * 4) do not run the "skip" process, if something else is available
1174 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1176 struct sched_entity *se = __pick_first_entity(cfs_rq);
1177 struct sched_entity *left = se;
1180 * Avoid running the skip buddy, if running something else can
1181 * be done without getting too unfair.
1183 if (cfs_rq->skip == se) {
1184 struct sched_entity *second = __pick_next_entity(se);
1185 if (second && wakeup_preempt_entity(second, left) < 1)
1190 * Prefer last buddy, try to return the CPU to a preempted task.
1192 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1196 * Someone really wants this to run. If it's not unfair, run it.
1198 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1201 clear_buddies(cfs_rq, se);
1206 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1209 * If still on the runqueue then deactivate_task()
1210 * was not called and update_curr() has to be done:
1213 update_curr(cfs_rq);
1215 check_spread(cfs_rq, prev);
1217 update_stats_wait_start(cfs_rq, prev);
1218 /* Put 'current' back into the tree. */
1219 __enqueue_entity(cfs_rq, prev);
1221 cfs_rq->curr = NULL;
1225 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1228 * Update run-time statistics of the 'current'.
1230 update_curr(cfs_rq);
1233 * Update share accounting for long-running entities.
1235 update_entity_shares_tick(cfs_rq);
1237 #ifdef CONFIG_SCHED_HRTICK
1239 * queued ticks are scheduled to match the slice, so don't bother
1240 * validating it and just reschedule.
1243 resched_task(rq_of(cfs_rq)->curr);
1247 * don't let the period tick interfere with the hrtick preemption
1249 if (!sched_feat(DOUBLE_TICK) &&
1250 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1254 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1255 check_preempt_tick(cfs_rq, curr);
1258 /**************************************************
1259 * CFS operations on tasks:
1262 #ifdef CONFIG_SCHED_HRTICK
1263 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1265 struct sched_entity *se = &p->se;
1266 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1268 WARN_ON(task_rq(p) != rq);
1270 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1271 u64 slice = sched_slice(cfs_rq, se);
1272 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1273 s64 delta = slice - ran;
1282 * Don't schedule slices shorter than 10000ns, that just
1283 * doesn't make sense. Rely on vruntime for fairness.
1286 delta = max_t(s64, 10000LL, delta);
1288 hrtick_start(rq, delta);
1293 * called from enqueue/dequeue and updates the hrtick when the
1294 * current task is from our class and nr_running is low enough
1297 static void hrtick_update(struct rq *rq)
1299 struct task_struct *curr = rq->curr;
1301 if (curr->sched_class != &fair_sched_class)
1304 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1305 hrtick_start_fair(rq, curr);
1307 #else /* !CONFIG_SCHED_HRTICK */
1309 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1313 static inline void hrtick_update(struct rq *rq)
1319 * The enqueue_task method is called before nr_running is
1320 * increased. Here we update the fair scheduling stats and
1321 * then put the task into the rbtree:
1324 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1326 struct cfs_rq *cfs_rq;
1327 struct sched_entity *se = &p->se;
1329 for_each_sched_entity(se) {
1332 cfs_rq = cfs_rq_of(se);
1333 enqueue_entity(cfs_rq, se, flags);
1334 flags = ENQUEUE_WAKEUP;
1337 for_each_sched_entity(se) {
1338 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1340 update_cfs_load(cfs_rq, 0);
1341 update_cfs_shares(cfs_rq);
1347 static void set_next_buddy(struct sched_entity *se);
1350 * The dequeue_task method is called before nr_running is
1351 * decreased. We remove the task from the rbtree and
1352 * update the fair scheduling stats:
1354 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1356 struct cfs_rq *cfs_rq;
1357 struct sched_entity *se = &p->se;
1358 int task_sleep = flags & DEQUEUE_SLEEP;
1360 for_each_sched_entity(se) {
1361 cfs_rq = cfs_rq_of(se);
1362 dequeue_entity(cfs_rq, se, flags);
1364 /* Don't dequeue parent if it has other entities besides us */
1365 if (cfs_rq->load.weight) {
1367 * Bias pick_next to pick a task from this cfs_rq, as
1368 * p is sleeping when it is within its sched_slice.
1370 if (task_sleep && parent_entity(se))
1371 set_next_buddy(parent_entity(se));
1374 flags |= DEQUEUE_SLEEP;
1377 for_each_sched_entity(se) {
1378 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1380 update_cfs_load(cfs_rq, 0);
1381 update_cfs_shares(cfs_rq);
1389 static void task_waking_fair(struct task_struct *p)
1391 struct sched_entity *se = &p->se;
1392 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1395 #ifndef CONFIG_64BIT
1396 u64 min_vruntime_copy;
1399 min_vruntime_copy = cfs_rq->min_vruntime_copy;
1401 min_vruntime = cfs_rq->min_vruntime;
1402 } while (min_vruntime != min_vruntime_copy);
1404 min_vruntime = cfs_rq->min_vruntime;
1407 se->vruntime -= min_vruntime;
1410 #ifdef CONFIG_FAIR_GROUP_SCHED
1412 * effective_load() calculates the load change as seen from the root_task_group
1414 * Adding load to a group doesn't make a group heavier, but can cause movement
1415 * of group shares between cpus. Assuming the shares were perfectly aligned one
1416 * can calculate the shift in shares.
1418 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1420 struct sched_entity *se = tg->se[cpu];
1425 for_each_sched_entity(se) {
1429 w = se->my_q->load.weight;
1431 /* use this cpu's instantaneous contribution */
1432 lw = atomic_read(&tg->load_weight);
1433 lw -= se->my_q->load_contribution;
1438 if (lw > 0 && wl < lw)
1439 wl = (wl * tg->shares) / lw;
1443 /* zero point is MIN_SHARES */
1444 if (wl < MIN_SHARES)
1446 wl -= se->load.weight;
1455 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1456 unsigned long wl, unsigned long wg)
1463 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1465 s64 this_load, load;
1466 int idx, this_cpu, prev_cpu;
1467 unsigned long tl_per_task;
1468 struct task_group *tg;
1469 unsigned long weight;
1473 this_cpu = smp_processor_id();
1474 prev_cpu = task_cpu(p);
1475 load = source_load(prev_cpu, idx);
1476 this_load = target_load(this_cpu, idx);
1479 * If sync wakeup then subtract the (maximum possible)
1480 * effect of the currently running task from the load
1481 * of the current CPU:
1485 tg = task_group(current);
1486 weight = current->se.load.weight;
1488 this_load += effective_load(tg, this_cpu, -weight, -weight);
1489 load += effective_load(tg, prev_cpu, 0, -weight);
1493 weight = p->se.load.weight;
1496 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1497 * due to the sync cause above having dropped this_load to 0, we'll
1498 * always have an imbalance, but there's really nothing you can do
1499 * about that, so that's good too.
1501 * Otherwise check if either cpus are near enough in load to allow this
1502 * task to be woken on this_cpu.
1504 if (this_load > 0) {
1505 s64 this_eff_load, prev_eff_load;
1507 this_eff_load = 100;
1508 this_eff_load *= power_of(prev_cpu);
1509 this_eff_load *= this_load +
1510 effective_load(tg, this_cpu, weight, weight);
1512 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1513 prev_eff_load *= power_of(this_cpu);
1514 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1516 balanced = this_eff_load <= prev_eff_load;
1522 * If the currently running task will sleep within
1523 * a reasonable amount of time then attract this newly
1526 if (sync && balanced)
1529 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1530 tl_per_task = cpu_avg_load_per_task(this_cpu);
1533 (this_load <= load &&
1534 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1536 * This domain has SD_WAKE_AFFINE and
1537 * p is cache cold in this domain, and
1538 * there is no bad imbalance.
1540 schedstat_inc(sd, ttwu_move_affine);
1541 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1549 * find_idlest_group finds and returns the least busy CPU group within the
1552 static struct sched_group *
1553 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1554 int this_cpu, int load_idx)
1556 struct sched_group *idlest = NULL, *group = sd->groups;
1557 unsigned long min_load = ULONG_MAX, this_load = 0;
1558 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1561 unsigned long load, avg_load;
1565 /* Skip over this group if it has no CPUs allowed */
1566 if (!cpumask_intersects(sched_group_cpus(group),
1570 local_group = cpumask_test_cpu(this_cpu,
1571 sched_group_cpus(group));
1573 /* Tally up the load of all CPUs in the group */
1576 for_each_cpu(i, sched_group_cpus(group)) {
1577 /* Bias balancing toward cpus of our domain */
1579 load = source_load(i, load_idx);
1581 load = target_load(i, load_idx);
1586 /* Adjust by relative CPU power of the group */
1587 avg_load = (avg_load * SCHED_POWER_SCALE) / group->cpu_power;
1590 this_load = avg_load;
1591 } else if (avg_load < min_load) {
1592 min_load = avg_load;
1595 } while (group = group->next, group != sd->groups);
1597 if (!idlest || 100*this_load < imbalance*min_load)
1603 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1606 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1608 unsigned long load, min_load = ULONG_MAX;
1612 /* Traverse only the allowed CPUs */
1613 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1614 load = weighted_cpuload(i);
1616 if (load < min_load || (load == min_load && i == this_cpu)) {
1626 * Try and locate an idle CPU in the sched_domain.
1628 static int select_idle_sibling(struct task_struct *p, int target)
1630 int cpu = smp_processor_id();
1631 int prev_cpu = task_cpu(p);
1632 struct sched_domain *sd;
1636 * If the task is going to be woken-up on this cpu and if it is
1637 * already idle, then it is the right target.
1639 if (target == cpu && idle_cpu(cpu))
1643 * If the task is going to be woken-up on the cpu where it previously
1644 * ran and if it is currently idle, then it the right target.
1646 if (target == prev_cpu && idle_cpu(prev_cpu))
1650 * Otherwise, iterate the domains and find an elegible idle cpu.
1653 for_each_domain(target, sd) {
1654 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1657 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1665 * Lets stop looking for an idle sibling when we reached
1666 * the domain that spans the current cpu and prev_cpu.
1668 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1669 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1678 * sched_balance_self: balance the current task (running on cpu) in domains
1679 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1682 * Balance, ie. select the least loaded group.
1684 * Returns the target CPU number, or the same CPU if no balancing is needed.
1686 * preempt must be disabled.
1689 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1691 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1692 int cpu = smp_processor_id();
1693 int prev_cpu = task_cpu(p);
1695 int want_affine = 0;
1697 int sync = wake_flags & WF_SYNC;
1699 if (sd_flag & SD_BALANCE_WAKE) {
1700 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1706 for_each_domain(cpu, tmp) {
1707 if (!(tmp->flags & SD_LOAD_BALANCE))
1711 * If power savings logic is enabled for a domain, see if we
1712 * are not overloaded, if so, don't balance wider.
1714 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1715 unsigned long power = 0;
1716 unsigned long nr_running = 0;
1717 unsigned long capacity;
1720 for_each_cpu(i, sched_domain_span(tmp)) {
1721 power += power_of(i);
1722 nr_running += cpu_rq(i)->cfs.nr_running;
1725 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
1727 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1730 if (nr_running < capacity)
1735 * If both cpu and prev_cpu are part of this domain,
1736 * cpu is a valid SD_WAKE_AFFINE target.
1738 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1739 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1744 if (!want_sd && !want_affine)
1747 if (!(tmp->flags & sd_flag))
1755 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1758 new_cpu = select_idle_sibling(p, prev_cpu);
1763 int load_idx = sd->forkexec_idx;
1764 struct sched_group *group;
1767 if (!(sd->flags & sd_flag)) {
1772 if (sd_flag & SD_BALANCE_WAKE)
1773 load_idx = sd->wake_idx;
1775 group = find_idlest_group(sd, p, cpu, load_idx);
1781 new_cpu = find_idlest_cpu(group, p, cpu);
1782 if (new_cpu == -1 || new_cpu == cpu) {
1783 /* Now try balancing at a lower domain level of cpu */
1788 /* Now try balancing at a lower domain level of new_cpu */
1790 weight = sd->span_weight;
1792 for_each_domain(cpu, tmp) {
1793 if (weight <= tmp->span_weight)
1795 if (tmp->flags & sd_flag)
1798 /* while loop will break here if sd == NULL */
1805 #endif /* CONFIG_SMP */
1807 static unsigned long
1808 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1810 unsigned long gran = sysctl_sched_wakeup_granularity;
1813 * Since its curr running now, convert the gran from real-time
1814 * to virtual-time in his units.
1816 * By using 'se' instead of 'curr' we penalize light tasks, so
1817 * they get preempted easier. That is, if 'se' < 'curr' then
1818 * the resulting gran will be larger, therefore penalizing the
1819 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1820 * be smaller, again penalizing the lighter task.
1822 * This is especially important for buddies when the leftmost
1823 * task is higher priority than the buddy.
1825 return calc_delta_fair(gran, se);
1829 * Should 'se' preempt 'curr'.
1843 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1845 s64 gran, vdiff = curr->vruntime - se->vruntime;
1850 gran = wakeup_gran(curr, se);
1857 static void set_last_buddy(struct sched_entity *se)
1859 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
1862 for_each_sched_entity(se)
1863 cfs_rq_of(se)->last = se;
1866 static void set_next_buddy(struct sched_entity *se)
1868 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
1871 for_each_sched_entity(se)
1872 cfs_rq_of(se)->next = se;
1875 static void set_skip_buddy(struct sched_entity *se)
1877 for_each_sched_entity(se)
1878 cfs_rq_of(se)->skip = se;
1882 * Preempt the current task with a newly woken task if needed:
1884 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1886 struct task_struct *curr = rq->curr;
1887 struct sched_entity *se = &curr->se, *pse = &p->se;
1888 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1889 int scale = cfs_rq->nr_running >= sched_nr_latency;
1890 int next_buddy_marked = 0;
1892 if (unlikely(se == pse))
1895 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
1896 set_next_buddy(pse);
1897 next_buddy_marked = 1;
1901 * We can come here with TIF_NEED_RESCHED already set from new task
1904 if (test_tsk_need_resched(curr))
1907 /* Idle tasks are by definition preempted by non-idle tasks. */
1908 if (unlikely(curr->policy == SCHED_IDLE) &&
1909 likely(p->policy != SCHED_IDLE))
1913 * Batch and idle tasks do not preempt non-idle tasks (their preemption
1914 * is driven by the tick):
1916 if (unlikely(p->policy != SCHED_NORMAL))
1920 if (!sched_feat(WAKEUP_PREEMPT))
1923 update_curr(cfs_rq);
1924 find_matching_se(&se, &pse);
1926 if (wakeup_preempt_entity(se, pse) == 1) {
1928 * Bias pick_next to pick the sched entity that is
1929 * triggering this preemption.
1931 if (!next_buddy_marked)
1932 set_next_buddy(pse);
1941 * Only set the backward buddy when the current task is still
1942 * on the rq. This can happen when a wakeup gets interleaved
1943 * with schedule on the ->pre_schedule() or idle_balance()
1944 * point, either of which can * drop the rq lock.
1946 * Also, during early boot the idle thread is in the fair class,
1947 * for obvious reasons its a bad idea to schedule back to it.
1949 if (unlikely(!se->on_rq || curr == rq->idle))
1952 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1956 static struct task_struct *pick_next_task_fair(struct rq *rq)
1958 struct task_struct *p;
1959 struct cfs_rq *cfs_rq = &rq->cfs;
1960 struct sched_entity *se;
1962 if (!cfs_rq->nr_running)
1966 se = pick_next_entity(cfs_rq);
1967 set_next_entity(cfs_rq, se);
1968 cfs_rq = group_cfs_rq(se);
1972 hrtick_start_fair(rq, p);
1978 * Account for a descheduled task:
1980 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1982 struct sched_entity *se = &prev->se;
1983 struct cfs_rq *cfs_rq;
1985 for_each_sched_entity(se) {
1986 cfs_rq = cfs_rq_of(se);
1987 put_prev_entity(cfs_rq, se);
1992 * sched_yield() is very simple
1994 * The magic of dealing with the ->skip buddy is in pick_next_entity.
1996 static void yield_task_fair(struct rq *rq)
1998 struct task_struct *curr = rq->curr;
1999 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2000 struct sched_entity *se = &curr->se;
2003 * Are we the only task in the tree?
2005 if (unlikely(rq->nr_running == 1))
2008 clear_buddies(cfs_rq, se);
2010 if (curr->policy != SCHED_BATCH) {
2011 update_rq_clock(rq);
2013 * Update run-time statistics of the 'current'.
2015 update_curr(cfs_rq);
2021 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2023 struct sched_entity *se = &p->se;
2028 /* Tell the scheduler that we'd really like pse to run next. */
2031 yield_task_fair(rq);
2037 /**************************************************
2038 * Fair scheduling class load-balancing methods:
2042 * pull_task - move a task from a remote runqueue to the local runqueue.
2043 * Both runqueues must be locked.
2045 static void pull_task(struct rq *src_rq, struct task_struct *p,
2046 struct rq *this_rq, int this_cpu)
2048 deactivate_task(src_rq, p, 0);
2049 set_task_cpu(p, this_cpu);
2050 activate_task(this_rq, p, 0);
2051 check_preempt_curr(this_rq, p, 0);
2055 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2058 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2059 struct sched_domain *sd, enum cpu_idle_type idle,
2062 int tsk_cache_hot = 0;
2064 * We do not migrate tasks that are:
2065 * 1) running (obviously), or
2066 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2067 * 3) are cache-hot on their current CPU.
2069 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2070 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2075 if (task_running(rq, p)) {
2076 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2081 * Aggressive migration if:
2082 * 1) task is cache cold, or
2083 * 2) too many balance attempts have failed.
2086 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2087 if (!tsk_cache_hot ||
2088 sd->nr_balance_failed > sd->cache_nice_tries) {
2089 #ifdef CONFIG_SCHEDSTATS
2090 if (tsk_cache_hot) {
2091 schedstat_inc(sd, lb_hot_gained[idle]);
2092 schedstat_inc(p, se.statistics.nr_forced_migrations);
2098 if (tsk_cache_hot) {
2099 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2106 * move_one_task tries to move exactly one task from busiest to this_rq, as
2107 * part of active balancing operations within "domain".
2108 * Returns 1 if successful and 0 otherwise.
2110 * Called with both runqueues locked.
2113 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2114 struct sched_domain *sd, enum cpu_idle_type idle)
2116 struct task_struct *p, *n;
2117 struct cfs_rq *cfs_rq;
2120 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2121 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2123 if (!can_migrate_task(p, busiest, this_cpu,
2127 pull_task(busiest, p, this_rq, this_cpu);
2129 * Right now, this is only the second place pull_task()
2130 * is called, so we can safely collect pull_task()
2131 * stats here rather than inside pull_task().
2133 schedstat_inc(sd, lb_gained[idle]);
2141 static unsigned long
2142 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2143 unsigned long max_load_move, struct sched_domain *sd,
2144 enum cpu_idle_type idle, int *all_pinned,
2145 struct cfs_rq *busiest_cfs_rq)
2147 int loops = 0, pulled = 0;
2148 long rem_load_move = max_load_move;
2149 struct task_struct *p, *n;
2151 if (max_load_move == 0)
2154 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2155 if (loops++ > sysctl_sched_nr_migrate)
2158 if ((p->se.load.weight >> 1) > rem_load_move ||
2159 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2163 pull_task(busiest, p, this_rq, this_cpu);
2165 rem_load_move -= p->se.load.weight;
2167 #ifdef CONFIG_PREEMPT
2169 * NEWIDLE balancing is a source of latency, so preemptible
2170 * kernels will stop after the first task is pulled to minimize
2171 * the critical section.
2173 if (idle == CPU_NEWLY_IDLE)
2178 * We only want to steal up to the prescribed amount of
2181 if (rem_load_move <= 0)
2186 * Right now, this is one of only two places pull_task() is called,
2187 * so we can safely collect pull_task() stats here rather than
2188 * inside pull_task().
2190 schedstat_add(sd, lb_gained[idle], pulled);
2192 return max_load_move - rem_load_move;
2195 #ifdef CONFIG_FAIR_GROUP_SCHED
2197 * update tg->load_weight by folding this cpu's load_avg
2199 static int update_shares_cpu(struct task_group *tg, int cpu)
2201 struct cfs_rq *cfs_rq;
2202 unsigned long flags;
2209 cfs_rq = tg->cfs_rq[cpu];
2211 raw_spin_lock_irqsave(&rq->lock, flags);
2213 update_rq_clock(rq);
2214 update_cfs_load(cfs_rq, 1);
2217 * We need to update shares after updating tg->load_weight in
2218 * order to adjust the weight of groups with long running tasks.
2220 update_cfs_shares(cfs_rq);
2222 raw_spin_unlock_irqrestore(&rq->lock, flags);
2227 static void update_shares(int cpu)
2229 struct cfs_rq *cfs_rq;
2230 struct rq *rq = cpu_rq(cpu);
2233 for_each_leaf_cfs_rq(rq, cfs_rq)
2234 update_shares_cpu(cfs_rq->tg, cpu);
2238 static unsigned long
2239 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2240 unsigned long max_load_move,
2241 struct sched_domain *sd, enum cpu_idle_type idle,
2244 long rem_load_move = max_load_move;
2245 int busiest_cpu = cpu_of(busiest);
2246 struct task_group *tg;
2249 update_h_load(busiest_cpu);
2251 list_for_each_entry_rcu(tg, &task_groups, list) {
2252 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2253 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2254 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2255 u64 rem_load, moved_load;
2260 if (!busiest_cfs_rq->task_weight)
2263 rem_load = (u64)rem_load_move * busiest_weight;
2264 rem_load = div_u64(rem_load, busiest_h_load + 1);
2266 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2267 rem_load, sd, idle, all_pinned,
2273 moved_load *= busiest_h_load;
2274 moved_load = div_u64(moved_load, busiest_weight + 1);
2276 rem_load_move -= moved_load;
2277 if (rem_load_move < 0)
2282 return max_load_move - rem_load_move;
2285 static inline void update_shares(int cpu)
2289 static unsigned long
2290 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2291 unsigned long max_load_move,
2292 struct sched_domain *sd, enum cpu_idle_type idle,
2295 return balance_tasks(this_rq, this_cpu, busiest,
2296 max_load_move, sd, idle, all_pinned,
2302 * move_tasks tries to move up to max_load_move weighted load from busiest to
2303 * this_rq, as part of a balancing operation within domain "sd".
2304 * Returns 1 if successful and 0 otherwise.
2306 * Called with both runqueues locked.
2308 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2309 unsigned long max_load_move,
2310 struct sched_domain *sd, enum cpu_idle_type idle,
2313 unsigned long total_load_moved = 0, load_moved;
2316 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2317 max_load_move - total_load_moved,
2318 sd, idle, all_pinned);
2320 total_load_moved += load_moved;
2322 #ifdef CONFIG_PREEMPT
2324 * NEWIDLE balancing is a source of latency, so preemptible
2325 * kernels will stop after the first task is pulled to minimize
2326 * the critical section.
2328 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2331 if (raw_spin_is_contended(&this_rq->lock) ||
2332 raw_spin_is_contended(&busiest->lock))
2335 } while (load_moved && max_load_move > total_load_moved);
2337 return total_load_moved > 0;
2340 /********** Helpers for find_busiest_group ************************/
2342 * sd_lb_stats - Structure to store the statistics of a sched_domain
2343 * during load balancing.
2345 struct sd_lb_stats {
2346 struct sched_group *busiest; /* Busiest group in this sd */
2347 struct sched_group *this; /* Local group in this sd */
2348 unsigned long total_load; /* Total load of all groups in sd */
2349 unsigned long total_pwr; /* Total power of all groups in sd */
2350 unsigned long avg_load; /* Average load across all groups in sd */
2352 /** Statistics of this group */
2353 unsigned long this_load;
2354 unsigned long this_load_per_task;
2355 unsigned long this_nr_running;
2356 unsigned long this_has_capacity;
2357 unsigned int this_idle_cpus;
2359 /* Statistics of the busiest group */
2360 unsigned int busiest_idle_cpus;
2361 unsigned long max_load;
2362 unsigned long busiest_load_per_task;
2363 unsigned long busiest_nr_running;
2364 unsigned long busiest_group_capacity;
2365 unsigned long busiest_has_capacity;
2366 unsigned int busiest_group_weight;
2368 int group_imb; /* Is there imbalance in this sd */
2369 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2370 int power_savings_balance; /* Is powersave balance needed for this sd */
2371 struct sched_group *group_min; /* Least loaded group in sd */
2372 struct sched_group *group_leader; /* Group which relieves group_min */
2373 unsigned long min_load_per_task; /* load_per_task in group_min */
2374 unsigned long leader_nr_running; /* Nr running of group_leader */
2375 unsigned long min_nr_running; /* Nr running of group_min */
2380 * sg_lb_stats - stats of a sched_group required for load_balancing
2382 struct sg_lb_stats {
2383 unsigned long avg_load; /*Avg load across the CPUs of the group */
2384 unsigned long group_load; /* Total load over the CPUs of the group */
2385 unsigned long sum_nr_running; /* Nr tasks running in the group */
2386 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2387 unsigned long group_capacity;
2388 unsigned long idle_cpus;
2389 unsigned long group_weight;
2390 int group_imb; /* Is there an imbalance in the group ? */
2391 int group_has_capacity; /* Is there extra capacity in the group? */
2395 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2396 * @group: The group whose first cpu is to be returned.
2398 static inline unsigned int group_first_cpu(struct sched_group *group)
2400 return cpumask_first(sched_group_cpus(group));
2404 * get_sd_load_idx - Obtain the load index for a given sched domain.
2405 * @sd: The sched_domain whose load_idx is to be obtained.
2406 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2408 static inline int get_sd_load_idx(struct sched_domain *sd,
2409 enum cpu_idle_type idle)
2415 load_idx = sd->busy_idx;
2418 case CPU_NEWLY_IDLE:
2419 load_idx = sd->newidle_idx;
2422 load_idx = sd->idle_idx;
2430 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2432 * init_sd_power_savings_stats - Initialize power savings statistics for
2433 * the given sched_domain, during load balancing.
2435 * @sd: Sched domain whose power-savings statistics are to be initialized.
2436 * @sds: Variable containing the statistics for sd.
2437 * @idle: Idle status of the CPU at which we're performing load-balancing.
2439 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2440 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2443 * Busy processors will not participate in power savings
2446 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2447 sds->power_savings_balance = 0;
2449 sds->power_savings_balance = 1;
2450 sds->min_nr_running = ULONG_MAX;
2451 sds->leader_nr_running = 0;
2456 * update_sd_power_savings_stats - Update the power saving stats for a
2457 * sched_domain while performing load balancing.
2459 * @group: sched_group belonging to the sched_domain under consideration.
2460 * @sds: Variable containing the statistics of the sched_domain
2461 * @local_group: Does group contain the CPU for which we're performing
2463 * @sgs: Variable containing the statistics of the group.
2465 static inline void update_sd_power_savings_stats(struct sched_group *group,
2466 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2469 if (!sds->power_savings_balance)
2473 * If the local group is idle or completely loaded
2474 * no need to do power savings balance at this domain
2476 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2477 !sds->this_nr_running))
2478 sds->power_savings_balance = 0;
2481 * If a group is already running at full capacity or idle,
2482 * don't include that group in power savings calculations
2484 if (!sds->power_savings_balance ||
2485 sgs->sum_nr_running >= sgs->group_capacity ||
2486 !sgs->sum_nr_running)
2490 * Calculate the group which has the least non-idle load.
2491 * This is the group from where we need to pick up the load
2494 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2495 (sgs->sum_nr_running == sds->min_nr_running &&
2496 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2497 sds->group_min = group;
2498 sds->min_nr_running = sgs->sum_nr_running;
2499 sds->min_load_per_task = sgs->sum_weighted_load /
2500 sgs->sum_nr_running;
2504 * Calculate the group which is almost near its
2505 * capacity but still has some space to pick up some load
2506 * from other group and save more power
2508 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2511 if (sgs->sum_nr_running > sds->leader_nr_running ||
2512 (sgs->sum_nr_running == sds->leader_nr_running &&
2513 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2514 sds->group_leader = group;
2515 sds->leader_nr_running = sgs->sum_nr_running;
2520 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2521 * @sds: Variable containing the statistics of the sched_domain
2522 * under consideration.
2523 * @this_cpu: Cpu at which we're currently performing load-balancing.
2524 * @imbalance: Variable to store the imbalance.
2527 * Check if we have potential to perform some power-savings balance.
2528 * If yes, set the busiest group to be the least loaded group in the
2529 * sched_domain, so that it's CPUs can be put to idle.
2531 * Returns 1 if there is potential to perform power-savings balance.
2534 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2535 int this_cpu, unsigned long *imbalance)
2537 if (!sds->power_savings_balance)
2540 if (sds->this != sds->group_leader ||
2541 sds->group_leader == sds->group_min)
2544 *imbalance = sds->min_load_per_task;
2545 sds->busiest = sds->group_min;
2550 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2551 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2552 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2557 static inline void update_sd_power_savings_stats(struct sched_group *group,
2558 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2563 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2564 int this_cpu, unsigned long *imbalance)
2568 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2571 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2573 return SCHED_POWER_SCALE;
2576 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2578 return default_scale_freq_power(sd, cpu);
2581 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2583 unsigned long weight = sd->span_weight;
2584 unsigned long smt_gain = sd->smt_gain;
2591 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2593 return default_scale_smt_power(sd, cpu);
2596 unsigned long scale_rt_power(int cpu)
2598 struct rq *rq = cpu_rq(cpu);
2599 u64 total, available;
2601 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2603 if (unlikely(total < rq->rt_avg)) {
2604 /* Ensures that power won't end up being negative */
2607 available = total - rq->rt_avg;
2610 if (unlikely((s64)total < SCHED_POWER_SCALE))
2611 total = SCHED_POWER_SCALE;
2613 total >>= SCHED_POWER_SHIFT;
2615 return div_u64(available, total);
2618 static void update_cpu_power(struct sched_domain *sd, int cpu)
2620 unsigned long weight = sd->span_weight;
2621 unsigned long power = SCHED_POWER_SCALE;
2622 struct sched_group *sdg = sd->groups;
2624 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2625 if (sched_feat(ARCH_POWER))
2626 power *= arch_scale_smt_power(sd, cpu);
2628 power *= default_scale_smt_power(sd, cpu);
2630 power >>= SCHED_POWER_SHIFT;
2633 sdg->cpu_power_orig = power;
2635 if (sched_feat(ARCH_POWER))
2636 power *= arch_scale_freq_power(sd, cpu);
2638 power *= default_scale_freq_power(sd, cpu);
2640 power >>= SCHED_POWER_SHIFT;
2642 power *= scale_rt_power(cpu);
2643 power >>= SCHED_POWER_SHIFT;
2648 cpu_rq(cpu)->cpu_power = power;
2649 sdg->cpu_power = power;
2652 static void update_group_power(struct sched_domain *sd, int cpu)
2654 struct sched_domain *child = sd->child;
2655 struct sched_group *group, *sdg = sd->groups;
2656 unsigned long power;
2659 update_cpu_power(sd, cpu);
2665 group = child->groups;
2667 power += group->cpu_power;
2668 group = group->next;
2669 } while (group != child->groups);
2671 sdg->cpu_power = power;
2675 * Try and fix up capacity for tiny siblings, this is needed when
2676 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2677 * which on its own isn't powerful enough.
2679 * See update_sd_pick_busiest() and check_asym_packing().
2682 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2685 * Only siblings can have significantly less than SCHED_POWER_SCALE
2687 if (!(sd->flags & SD_SHARE_CPUPOWER))
2691 * If ~90% of the cpu_power is still there, we're good.
2693 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2700 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2701 * @sd: The sched_domain whose statistics are to be updated.
2702 * @group: sched_group whose statistics are to be updated.
2703 * @this_cpu: Cpu for which load balance is currently performed.
2704 * @idle: Idle status of this_cpu
2705 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2706 * @local_group: Does group contain this_cpu.
2707 * @cpus: Set of cpus considered for load balancing.
2708 * @balance: Should we balance.
2709 * @sgs: variable to hold the statistics for this group.
2711 static inline void update_sg_lb_stats(struct sched_domain *sd,
2712 struct sched_group *group, int this_cpu,
2713 enum cpu_idle_type idle, int load_idx,
2714 int local_group, const struct cpumask *cpus,
2715 int *balance, struct sg_lb_stats *sgs)
2717 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2719 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2720 unsigned long avg_load_per_task = 0;
2723 balance_cpu = group_first_cpu(group);
2725 /* Tally up the load of all CPUs in the group */
2727 min_cpu_load = ~0UL;
2730 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2731 struct rq *rq = cpu_rq(i);
2733 /* Bias balancing toward cpus of our domain */
2735 if (idle_cpu(i) && !first_idle_cpu) {
2740 load = target_load(i, load_idx);
2742 load = source_load(i, load_idx);
2743 if (load > max_cpu_load) {
2744 max_cpu_load = load;
2745 max_nr_running = rq->nr_running;
2747 if (min_cpu_load > load)
2748 min_cpu_load = load;
2751 sgs->group_load += load;
2752 sgs->sum_nr_running += rq->nr_running;
2753 sgs->sum_weighted_load += weighted_cpuload(i);
2759 * First idle cpu or the first cpu(busiest) in this sched group
2760 * is eligible for doing load balancing at this and above
2761 * domains. In the newly idle case, we will allow all the cpu's
2762 * to do the newly idle load balance.
2764 if (idle != CPU_NEWLY_IDLE && local_group) {
2765 if (balance_cpu != this_cpu) {
2769 update_group_power(sd, this_cpu);
2772 /* Adjust by relative CPU power of the group */
2773 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->cpu_power;
2776 * Consider the group unbalanced when the imbalance is larger
2777 * than the average weight of a task.
2779 * APZ: with cgroup the avg task weight can vary wildly and
2780 * might not be a suitable number - should we keep a
2781 * normalized nr_running number somewhere that negates
2784 if (sgs->sum_nr_running)
2785 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2787 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
2790 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power,
2792 if (!sgs->group_capacity)
2793 sgs->group_capacity = fix_small_capacity(sd, group);
2794 sgs->group_weight = group->group_weight;
2796 if (sgs->group_capacity > sgs->sum_nr_running)
2797 sgs->group_has_capacity = 1;
2801 * update_sd_pick_busiest - return 1 on busiest group
2802 * @sd: sched_domain whose statistics are to be checked
2803 * @sds: sched_domain statistics
2804 * @sg: sched_group candidate to be checked for being the busiest
2805 * @sgs: sched_group statistics
2806 * @this_cpu: the current cpu
2808 * Determine if @sg is a busier group than the previously selected
2811 static bool update_sd_pick_busiest(struct sched_domain *sd,
2812 struct sd_lb_stats *sds,
2813 struct sched_group *sg,
2814 struct sg_lb_stats *sgs,
2817 if (sgs->avg_load <= sds->max_load)
2820 if (sgs->sum_nr_running > sgs->group_capacity)
2827 * ASYM_PACKING needs to move all the work to the lowest
2828 * numbered CPUs in the group, therefore mark all groups
2829 * higher than ourself as busy.
2831 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2832 this_cpu < group_first_cpu(sg)) {
2836 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2844 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2845 * @sd: sched_domain whose statistics are to be updated.
2846 * @this_cpu: Cpu for which load balance is currently performed.
2847 * @idle: Idle status of this_cpu
2848 * @cpus: Set of cpus considered for load balancing.
2849 * @balance: Should we balance.
2850 * @sds: variable to hold the statistics for this sched_domain.
2852 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2853 enum cpu_idle_type idle, const struct cpumask *cpus,
2854 int *balance, struct sd_lb_stats *sds)
2856 struct sched_domain *child = sd->child;
2857 struct sched_group *sg = sd->groups;
2858 struct sg_lb_stats sgs;
2859 int load_idx, prefer_sibling = 0;
2861 if (child && child->flags & SD_PREFER_SIBLING)
2864 init_sd_power_savings_stats(sd, sds, idle);
2865 load_idx = get_sd_load_idx(sd, idle);
2870 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2871 memset(&sgs, 0, sizeof(sgs));
2872 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
2873 local_group, cpus, balance, &sgs);
2875 if (local_group && !(*balance))
2878 sds->total_load += sgs.group_load;
2879 sds->total_pwr += sg->cpu_power;
2882 * In case the child domain prefers tasks go to siblings
2883 * first, lower the sg capacity to one so that we'll try
2884 * and move all the excess tasks away. We lower the capacity
2885 * of a group only if the local group has the capacity to fit
2886 * these excess tasks, i.e. nr_running < group_capacity. The
2887 * extra check prevents the case where you always pull from the
2888 * heaviest group when it is already under-utilized (possible
2889 * with a large weight task outweighs the tasks on the system).
2891 if (prefer_sibling && !local_group && sds->this_has_capacity)
2892 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2895 sds->this_load = sgs.avg_load;
2897 sds->this_nr_running = sgs.sum_nr_running;
2898 sds->this_load_per_task = sgs.sum_weighted_load;
2899 sds->this_has_capacity = sgs.group_has_capacity;
2900 sds->this_idle_cpus = sgs.idle_cpus;
2901 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2902 sds->max_load = sgs.avg_load;
2904 sds->busiest_nr_running = sgs.sum_nr_running;
2905 sds->busiest_idle_cpus = sgs.idle_cpus;
2906 sds->busiest_group_capacity = sgs.group_capacity;
2907 sds->busiest_load_per_task = sgs.sum_weighted_load;
2908 sds->busiest_has_capacity = sgs.group_has_capacity;
2909 sds->busiest_group_weight = sgs.group_weight;
2910 sds->group_imb = sgs.group_imb;
2913 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2915 } while (sg != sd->groups);
2918 int __weak arch_sd_sibling_asym_packing(void)
2920 return 0*SD_ASYM_PACKING;
2924 * check_asym_packing - Check to see if the group is packed into the
2927 * This is primarily intended to used at the sibling level. Some
2928 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2929 * case of POWER7, it can move to lower SMT modes only when higher
2930 * threads are idle. When in lower SMT modes, the threads will
2931 * perform better since they share less core resources. Hence when we
2932 * have idle threads, we want them to be the higher ones.
2934 * This packing function is run on idle threads. It checks to see if
2935 * the busiest CPU in this domain (core in the P7 case) has a higher
2936 * CPU number than the packing function is being run on. Here we are
2937 * assuming lower CPU number will be equivalent to lower a SMT thread
2940 * Returns 1 when packing is required and a task should be moved to
2941 * this CPU. The amount of the imbalance is returned in *imbalance.
2943 * @sd: The sched_domain whose packing is to be checked.
2944 * @sds: Statistics of the sched_domain which is to be packed
2945 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2946 * @imbalance: returns amount of imbalanced due to packing.
2948 static int check_asym_packing(struct sched_domain *sd,
2949 struct sd_lb_stats *sds,
2950 int this_cpu, unsigned long *imbalance)
2954 if (!(sd->flags & SD_ASYM_PACKING))
2960 busiest_cpu = group_first_cpu(sds->busiest);
2961 if (this_cpu > busiest_cpu)
2964 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2970 * fix_small_imbalance - Calculate the minor imbalance that exists
2971 * amongst the groups of a sched_domain, during
2973 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2974 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2975 * @imbalance: Variable to store the imbalance.
2977 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2978 int this_cpu, unsigned long *imbalance)
2980 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2981 unsigned int imbn = 2;
2982 unsigned long scaled_busy_load_per_task;
2984 if (sds->this_nr_running) {
2985 sds->this_load_per_task /= sds->this_nr_running;
2986 if (sds->busiest_load_per_task >
2987 sds->this_load_per_task)
2990 sds->this_load_per_task =
2991 cpu_avg_load_per_task(this_cpu);
2993 scaled_busy_load_per_task = sds->busiest_load_per_task
2994 * SCHED_POWER_SCALE;
2995 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2997 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2998 (scaled_busy_load_per_task * imbn)) {
2999 *imbalance = sds->busiest_load_per_task;
3004 * OK, we don't have enough imbalance to justify moving tasks,
3005 * however we may be able to increase total CPU power used by
3009 pwr_now += sds->busiest->cpu_power *
3010 min(sds->busiest_load_per_task, sds->max_load);
3011 pwr_now += sds->this->cpu_power *
3012 min(sds->this_load_per_task, sds->this_load);
3013 pwr_now /= SCHED_POWER_SCALE;
3015 /* Amount of load we'd subtract */
3016 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3017 sds->busiest->cpu_power;
3018 if (sds->max_load > tmp)
3019 pwr_move += sds->busiest->cpu_power *
3020 min(sds->busiest_load_per_task, sds->max_load - tmp);
3022 /* Amount of load we'd add */
3023 if (sds->max_load * sds->busiest->cpu_power <
3024 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3025 tmp = (sds->max_load * sds->busiest->cpu_power) /
3026 sds->this->cpu_power;
3028 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3029 sds->this->cpu_power;
3030 pwr_move += sds->this->cpu_power *
3031 min(sds->this_load_per_task, sds->this_load + tmp);
3032 pwr_move /= SCHED_POWER_SCALE;
3034 /* Move if we gain throughput */
3035 if (pwr_move > pwr_now)
3036 *imbalance = sds->busiest_load_per_task;
3040 * calculate_imbalance - Calculate the amount of imbalance present within the
3041 * groups of a given sched_domain during load balance.
3042 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3043 * @this_cpu: Cpu for which currently load balance is being performed.
3044 * @imbalance: The variable to store the imbalance.
3046 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3047 unsigned long *imbalance)
3049 unsigned long max_pull, load_above_capacity = ~0UL;
3051 sds->busiest_load_per_task /= sds->busiest_nr_running;
3052 if (sds->group_imb) {
3053 sds->busiest_load_per_task =
3054 min(sds->busiest_load_per_task, sds->avg_load);
3058 * In the presence of smp nice balancing, certain scenarios can have
3059 * max load less than avg load(as we skip the groups at or below
3060 * its cpu_power, while calculating max_load..)
3062 if (sds->max_load < sds->avg_load) {
3064 return fix_small_imbalance(sds, this_cpu, imbalance);
3067 if (!sds->group_imb) {
3069 * Don't want to pull so many tasks that a group would go idle.
3071 load_above_capacity = (sds->busiest_nr_running -
3072 sds->busiest_group_capacity);
3074 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3076 load_above_capacity /= sds->busiest->cpu_power;
3080 * We're trying to get all the cpus to the average_load, so we don't
3081 * want to push ourselves above the average load, nor do we wish to
3082 * reduce the max loaded cpu below the average load. At the same time,
3083 * we also don't want to reduce the group load below the group capacity
3084 * (so that we can implement power-savings policies etc). Thus we look
3085 * for the minimum possible imbalance.
3086 * Be careful of negative numbers as they'll appear as very large values
3087 * with unsigned longs.
3089 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3091 /* How much load to actually move to equalise the imbalance */
3092 *imbalance = min(max_pull * sds->busiest->cpu_power,
3093 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3094 / SCHED_POWER_SCALE;
3097 * if *imbalance is less than the average load per runnable task
3098 * there is no guarantee that any tasks will be moved so we'll have
3099 * a think about bumping its value to force at least one task to be
3102 if (*imbalance < sds->busiest_load_per_task)
3103 return fix_small_imbalance(sds, this_cpu, imbalance);
3107 /******* find_busiest_group() helpers end here *********************/
3110 * find_busiest_group - Returns the busiest group within the sched_domain
3111 * if there is an imbalance. If there isn't an imbalance, and
3112 * the user has opted for power-savings, it returns a group whose
3113 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3114 * such a group exists.
3116 * Also calculates the amount of weighted load which should be moved
3117 * to restore balance.
3119 * @sd: The sched_domain whose busiest group is to be returned.
3120 * @this_cpu: The cpu for which load balancing is currently being performed.
3121 * @imbalance: Variable which stores amount of weighted load which should
3122 * be moved to restore balance/put a group to idle.
3123 * @idle: The idle status of this_cpu.
3124 * @cpus: The set of CPUs under consideration for load-balancing.
3125 * @balance: Pointer to a variable indicating if this_cpu
3126 * is the appropriate cpu to perform load balancing at this_level.
3128 * Returns: - the busiest group if imbalance exists.
3129 * - If no imbalance and user has opted for power-savings balance,
3130 * return the least loaded group whose CPUs can be
3131 * put to idle by rebalancing its tasks onto our group.
3133 static struct sched_group *
3134 find_busiest_group(struct sched_domain *sd, int this_cpu,
3135 unsigned long *imbalance, enum cpu_idle_type idle,
3136 const struct cpumask *cpus, int *balance)
3138 struct sd_lb_stats sds;
3140 memset(&sds, 0, sizeof(sds));
3143 * Compute the various statistics relavent for load balancing at
3146 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3149 * this_cpu is not the appropriate cpu to perform load balancing at
3155 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3156 check_asym_packing(sd, &sds, this_cpu, imbalance))
3159 /* There is no busy sibling group to pull tasks from */
3160 if (!sds.busiest || sds.busiest_nr_running == 0)
3163 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3166 * If the busiest group is imbalanced the below checks don't
3167 * work because they assumes all things are equal, which typically
3168 * isn't true due to cpus_allowed constraints and the like.
3173 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3174 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3175 !sds.busiest_has_capacity)
3179 * If the local group is more busy than the selected busiest group
3180 * don't try and pull any tasks.
3182 if (sds.this_load >= sds.max_load)
3186 * Don't pull any tasks if this group is already above the domain
3189 if (sds.this_load >= sds.avg_load)
3192 if (idle == CPU_IDLE) {
3194 * This cpu is idle. If the busiest group load doesn't
3195 * have more tasks than the number of available cpu's and
3196 * there is no imbalance between this and busiest group
3197 * wrt to idle cpu's, it is balanced.
3199 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3200 sds.busiest_nr_running <= sds.busiest_group_weight)
3204 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3205 * imbalance_pct to be conservative.
3207 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3212 /* Looks like there is an imbalance. Compute it */
3213 calculate_imbalance(&sds, this_cpu, imbalance);
3218 * There is no obvious imbalance. But check if we can do some balancing
3221 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3229 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3232 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3233 enum cpu_idle_type idle, unsigned long imbalance,
3234 const struct cpumask *cpus)
3236 struct rq *busiest = NULL, *rq;
3237 unsigned long max_load = 0;
3240 for_each_cpu(i, sched_group_cpus(group)) {
3241 unsigned long power = power_of(i);
3242 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3247 capacity = fix_small_capacity(sd, group);
3249 if (!cpumask_test_cpu(i, cpus))
3253 wl = weighted_cpuload(i);
3256 * When comparing with imbalance, use weighted_cpuload()
3257 * which is not scaled with the cpu power.
3259 if (capacity && rq->nr_running == 1 && wl > imbalance)
3263 * For the load comparisons with the other cpu's, consider
3264 * the weighted_cpuload() scaled with the cpu power, so that
3265 * the load can be moved away from the cpu that is potentially
3266 * running at a lower capacity.
3268 wl = (wl * SCHED_POWER_SCALE) / power;
3270 if (wl > max_load) {
3280 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3281 * so long as it is large enough.
3283 #define MAX_PINNED_INTERVAL 512
3285 /* Working cpumask for load_balance and load_balance_newidle. */
3286 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3288 static int need_active_balance(struct sched_domain *sd, int idle,
3289 int busiest_cpu, int this_cpu)
3291 if (idle == CPU_NEWLY_IDLE) {
3294 * ASYM_PACKING needs to force migrate tasks from busy but
3295 * higher numbered CPUs in order to pack all tasks in the
3296 * lowest numbered CPUs.
3298 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3302 * The only task running in a non-idle cpu can be moved to this
3303 * cpu in an attempt to completely freeup the other CPU
3306 * The package power saving logic comes from
3307 * find_busiest_group(). If there are no imbalance, then
3308 * f_b_g() will return NULL. However when sched_mc={1,2} then
3309 * f_b_g() will select a group from which a running task may be
3310 * pulled to this cpu in order to make the other package idle.
3311 * If there is no opportunity to make a package idle and if
3312 * there are no imbalance, then f_b_g() will return NULL and no
3313 * action will be taken in load_balance_newidle().
3315 * Under normal task pull operation due to imbalance, there
3316 * will be more than one task in the source run queue and
3317 * move_tasks() will succeed. ld_moved will be true and this
3318 * active balance code will not be triggered.
3320 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3324 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3327 static int active_load_balance_cpu_stop(void *data);
3330 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3331 * tasks if there is an imbalance.
3333 static int load_balance(int this_cpu, struct rq *this_rq,
3334 struct sched_domain *sd, enum cpu_idle_type idle,
3337 int ld_moved, all_pinned = 0, active_balance = 0;
3338 struct sched_group *group;
3339 unsigned long imbalance;
3341 unsigned long flags;
3342 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3344 cpumask_copy(cpus, cpu_active_mask);
3346 schedstat_inc(sd, lb_count[idle]);
3349 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
3356 schedstat_inc(sd, lb_nobusyg[idle]);
3360 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3362 schedstat_inc(sd, lb_nobusyq[idle]);
3366 BUG_ON(busiest == this_rq);
3368 schedstat_add(sd, lb_imbalance[idle], imbalance);
3371 if (busiest->nr_running > 1) {
3373 * Attempt to move tasks. If find_busiest_group has found
3374 * an imbalance but busiest->nr_running <= 1, the group is
3375 * still unbalanced. ld_moved simply stays zero, so it is
3376 * correctly treated as an imbalance.
3379 local_irq_save(flags);
3380 double_rq_lock(this_rq, busiest);
3381 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3382 imbalance, sd, idle, &all_pinned);
3383 double_rq_unlock(this_rq, busiest);
3384 local_irq_restore(flags);
3387 * some other cpu did the load balance for us.
3389 if (ld_moved && this_cpu != smp_processor_id())
3390 resched_cpu(this_cpu);
3392 /* All tasks on this runqueue were pinned by CPU affinity */
3393 if (unlikely(all_pinned)) {
3394 cpumask_clear_cpu(cpu_of(busiest), cpus);
3395 if (!cpumask_empty(cpus))
3402 schedstat_inc(sd, lb_failed[idle]);
3404 * Increment the failure counter only on periodic balance.
3405 * We do not want newidle balance, which can be very
3406 * frequent, pollute the failure counter causing
3407 * excessive cache_hot migrations and active balances.
3409 if (idle != CPU_NEWLY_IDLE)
3410 sd->nr_balance_failed++;
3412 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
3413 raw_spin_lock_irqsave(&busiest->lock, flags);
3415 /* don't kick the active_load_balance_cpu_stop,
3416 * if the curr task on busiest cpu can't be
3419 if (!cpumask_test_cpu(this_cpu,
3420 &busiest->curr->cpus_allowed)) {
3421 raw_spin_unlock_irqrestore(&busiest->lock,
3424 goto out_one_pinned;
3428 * ->active_balance synchronizes accesses to
3429 * ->active_balance_work. Once set, it's cleared
3430 * only after active load balance is finished.
3432 if (!busiest->active_balance) {
3433 busiest->active_balance = 1;
3434 busiest->push_cpu = this_cpu;
3437 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3440 stop_one_cpu_nowait(cpu_of(busiest),
3441 active_load_balance_cpu_stop, busiest,
3442 &busiest->active_balance_work);
3445 * We've kicked active balancing, reset the failure
3448 sd->nr_balance_failed = sd->cache_nice_tries+1;
3451 sd->nr_balance_failed = 0;
3453 if (likely(!active_balance)) {
3454 /* We were unbalanced, so reset the balancing interval */
3455 sd->balance_interval = sd->min_interval;
3458 * If we've begun active balancing, start to back off. This
3459 * case may not be covered by the all_pinned logic if there
3460 * is only 1 task on the busy runqueue (because we don't call
3463 if (sd->balance_interval < sd->max_interval)
3464 sd->balance_interval *= 2;
3470 schedstat_inc(sd, lb_balanced[idle]);
3472 sd->nr_balance_failed = 0;
3475 /* tune up the balancing interval */
3476 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3477 (sd->balance_interval < sd->max_interval))
3478 sd->balance_interval *= 2;
3486 * idle_balance is called by schedule() if this_cpu is about to become
3487 * idle. Attempts to pull tasks from other CPUs.
3489 static void idle_balance(int this_cpu, struct rq *this_rq)
3491 struct sched_domain *sd;
3492 int pulled_task = 0;
3493 unsigned long next_balance = jiffies + HZ;
3495 this_rq->idle_stamp = this_rq->clock;
3497 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3501 * Drop the rq->lock, but keep IRQ/preempt disabled.
3503 raw_spin_unlock(&this_rq->lock);
3505 update_shares(this_cpu);
3507 for_each_domain(this_cpu, sd) {
3508 unsigned long interval;
3511 if (!(sd->flags & SD_LOAD_BALANCE))
3514 if (sd->flags & SD_BALANCE_NEWIDLE) {
3515 /* If we've pulled tasks over stop searching: */
3516 pulled_task = load_balance(this_cpu, this_rq,
3517 sd, CPU_NEWLY_IDLE, &balance);
3520 interval = msecs_to_jiffies(sd->balance_interval);
3521 if (time_after(next_balance, sd->last_balance + interval))
3522 next_balance = sd->last_balance + interval;
3524 this_rq->idle_stamp = 0;
3530 raw_spin_lock(&this_rq->lock);
3532 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3534 * We are going idle. next_balance may be set based on
3535 * a busy processor. So reset next_balance.
3537 this_rq->next_balance = next_balance;
3542 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3543 * running tasks off the busiest CPU onto idle CPUs. It requires at
3544 * least 1 task to be running on each physical CPU where possible, and
3545 * avoids physical / logical imbalances.
3547 static int active_load_balance_cpu_stop(void *data)
3549 struct rq *busiest_rq = data;
3550 int busiest_cpu = cpu_of(busiest_rq);
3551 int target_cpu = busiest_rq->push_cpu;
3552 struct rq *target_rq = cpu_rq(target_cpu);
3553 struct sched_domain *sd;
3555 raw_spin_lock_irq(&busiest_rq->lock);
3557 /* make sure the requested cpu hasn't gone down in the meantime */
3558 if (unlikely(busiest_cpu != smp_processor_id() ||
3559 !busiest_rq->active_balance))
3562 /* Is there any task to move? */
3563 if (busiest_rq->nr_running <= 1)
3567 * This condition is "impossible", if it occurs
3568 * we need to fix it. Originally reported by
3569 * Bjorn Helgaas on a 128-cpu setup.
3571 BUG_ON(busiest_rq == target_rq);
3573 /* move a task from busiest_rq to target_rq */
3574 double_lock_balance(busiest_rq, target_rq);
3576 /* Search for an sd spanning us and the target CPU. */
3578 for_each_domain(target_cpu, sd) {
3579 if ((sd->flags & SD_LOAD_BALANCE) &&
3580 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3585 schedstat_inc(sd, alb_count);
3587 if (move_one_task(target_rq, target_cpu, busiest_rq,
3589 schedstat_inc(sd, alb_pushed);
3591 schedstat_inc(sd, alb_failed);
3594 double_unlock_balance(busiest_rq, target_rq);
3596 busiest_rq->active_balance = 0;
3597 raw_spin_unlock_irq(&busiest_rq->lock);
3603 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3605 static void trigger_sched_softirq(void *data)
3607 raise_softirq_irqoff(SCHED_SOFTIRQ);
3610 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3612 csd->func = trigger_sched_softirq;
3619 * idle load balancing details
3620 * - One of the idle CPUs nominates itself as idle load_balancer, while
3622 * - This idle load balancer CPU will also go into tickless mode when
3623 * it is idle, just like all other idle CPUs
3624 * - When one of the busy CPUs notice that there may be an idle rebalancing
3625 * needed, they will kick the idle load balancer, which then does idle
3626 * load balancing for all the idle CPUs.
3629 atomic_t load_balancer;
3630 atomic_t first_pick_cpu;
3631 atomic_t second_pick_cpu;
3632 cpumask_var_t idle_cpus_mask;
3633 cpumask_var_t grp_idle_mask;
3634 unsigned long next_balance; /* in jiffy units */
3635 } nohz ____cacheline_aligned;
3637 int get_nohz_load_balancer(void)
3639 return atomic_read(&nohz.load_balancer);
3642 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3644 * lowest_flag_domain - Return lowest sched_domain containing flag.
3645 * @cpu: The cpu whose lowest level of sched domain is to
3647 * @flag: The flag to check for the lowest sched_domain
3648 * for the given cpu.
3650 * Returns the lowest sched_domain of a cpu which contains the given flag.
3652 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3654 struct sched_domain *sd;
3656 for_each_domain(cpu, sd)
3657 if (sd && (sd->flags & flag))
3664 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3665 * @cpu: The cpu whose domains we're iterating over.
3666 * @sd: variable holding the value of the power_savings_sd
3668 * @flag: The flag to filter the sched_domains to be iterated.
3670 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3671 * set, starting from the lowest sched_domain to the highest.
3673 #define for_each_flag_domain(cpu, sd, flag) \
3674 for (sd = lowest_flag_domain(cpu, flag); \
3675 (sd && (sd->flags & flag)); sd = sd->parent)
3678 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3679 * @ilb_group: group to be checked for semi-idleness
3681 * Returns: 1 if the group is semi-idle. 0 otherwise.
3683 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3684 * and atleast one non-idle CPU. This helper function checks if the given
3685 * sched_group is semi-idle or not.
3687 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3689 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3690 sched_group_cpus(ilb_group));
3693 * A sched_group is semi-idle when it has atleast one busy cpu
3694 * and atleast one idle cpu.
3696 if (cpumask_empty(nohz.grp_idle_mask))
3699 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3705 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3706 * @cpu: The cpu which is nominating a new idle_load_balancer.
3708 * Returns: Returns the id of the idle load balancer if it exists,
3709 * Else, returns >= nr_cpu_ids.
3711 * This algorithm picks the idle load balancer such that it belongs to a
3712 * semi-idle powersavings sched_domain. The idea is to try and avoid
3713 * completely idle packages/cores just for the purpose of idle load balancing
3714 * when there are other idle cpu's which are better suited for that job.
3716 static int find_new_ilb(int cpu)
3718 struct sched_domain *sd;
3719 struct sched_group *ilb_group;
3720 int ilb = nr_cpu_ids;
3723 * Have idle load balancer selection from semi-idle packages only
3724 * when power-aware load balancing is enabled
3726 if (!(sched_smt_power_savings || sched_mc_power_savings))
3730 * Optimize for the case when we have no idle CPUs or only one
3731 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3733 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3737 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3738 ilb_group = sd->groups;
3741 if (is_semi_idle_group(ilb_group)) {
3742 ilb = cpumask_first(nohz.grp_idle_mask);
3746 ilb_group = ilb_group->next;
3748 } while (ilb_group != sd->groups);
3756 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3757 static inline int find_new_ilb(int call_cpu)
3764 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3765 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3766 * CPU (if there is one).
3768 static void nohz_balancer_kick(int cpu)
3772 nohz.next_balance++;
3774 ilb_cpu = get_nohz_load_balancer();
3776 if (ilb_cpu >= nr_cpu_ids) {
3777 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3778 if (ilb_cpu >= nr_cpu_ids)
3782 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3783 struct call_single_data *cp;
3785 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3786 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3787 __smp_call_function_single(ilb_cpu, cp, 0);
3793 * This routine will try to nominate the ilb (idle load balancing)
3794 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3795 * load balancing on behalf of all those cpus.
3797 * When the ilb owner becomes busy, we will not have new ilb owner until some
3798 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3799 * idle load balancing by kicking one of the idle CPUs.
3801 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3802 * ilb owner CPU in future (when there is a need for idle load balancing on
3803 * behalf of all idle CPUs).
3805 void select_nohz_load_balancer(int stop_tick)
3807 int cpu = smp_processor_id();
3810 if (!cpu_active(cpu)) {
3811 if (atomic_read(&nohz.load_balancer) != cpu)
3815 * If we are going offline and still the leader,
3818 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3825 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3827 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3828 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3829 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3830 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3832 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3835 /* make me the ilb owner */
3836 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3841 * Check to see if there is a more power-efficient
3844 new_ilb = find_new_ilb(cpu);
3845 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3846 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3847 resched_cpu(new_ilb);
3853 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3856 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3858 if (atomic_read(&nohz.load_balancer) == cpu)
3859 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3867 static DEFINE_SPINLOCK(balancing);
3869 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3872 * Scale the max load_balance interval with the number of CPUs in the system.
3873 * This trades load-balance latency on larger machines for less cross talk.
3875 static void update_max_interval(void)
3877 max_load_balance_interval = HZ*num_online_cpus()/10;
3881 * It checks each scheduling domain to see if it is due to be balanced,
3882 * and initiates a balancing operation if so.
3884 * Balancing parameters are set up in arch_init_sched_domains.
3886 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3889 struct rq *rq = cpu_rq(cpu);
3890 unsigned long interval;
3891 struct sched_domain *sd;
3892 /* Earliest time when we have to do rebalance again */
3893 unsigned long next_balance = jiffies + 60*HZ;
3894 int update_next_balance = 0;
3900 for_each_domain(cpu, sd) {
3901 if (!(sd->flags & SD_LOAD_BALANCE))
3904 interval = sd->balance_interval;
3905 if (idle != CPU_IDLE)
3906 interval *= sd->busy_factor;
3908 /* scale ms to jiffies */
3909 interval = msecs_to_jiffies(interval);
3910 interval = clamp(interval, 1UL, max_load_balance_interval);
3912 need_serialize = sd->flags & SD_SERIALIZE;
3914 if (need_serialize) {
3915 if (!spin_trylock(&balancing))
3919 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3920 if (load_balance(cpu, rq, sd, idle, &balance)) {
3922 * We've pulled tasks over so either we're no
3925 idle = CPU_NOT_IDLE;
3927 sd->last_balance = jiffies;
3930 spin_unlock(&balancing);
3932 if (time_after(next_balance, sd->last_balance + interval)) {
3933 next_balance = sd->last_balance + interval;
3934 update_next_balance = 1;
3938 * Stop the load balance at this level. There is another
3939 * CPU in our sched group which is doing load balancing more
3948 * next_balance will be updated only when there is a need.
3949 * When the cpu is attached to null domain for ex, it will not be
3952 if (likely(update_next_balance))
3953 rq->next_balance = next_balance;
3958 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3959 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3961 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3963 struct rq *this_rq = cpu_rq(this_cpu);
3967 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3970 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3971 if (balance_cpu == this_cpu)
3975 * If this cpu gets work to do, stop the load balancing
3976 * work being done for other cpus. Next load
3977 * balancing owner will pick it up.
3979 if (need_resched()) {
3980 this_rq->nohz_balance_kick = 0;
3984 raw_spin_lock_irq(&this_rq->lock);
3985 update_rq_clock(this_rq);
3986 update_cpu_load(this_rq);
3987 raw_spin_unlock_irq(&this_rq->lock);
3989 rebalance_domains(balance_cpu, CPU_IDLE);
3991 rq = cpu_rq(balance_cpu);
3992 if (time_after(this_rq->next_balance, rq->next_balance))
3993 this_rq->next_balance = rq->next_balance;
3995 nohz.next_balance = this_rq->next_balance;
3996 this_rq->nohz_balance_kick = 0;
4000 * Current heuristic for kicking the idle load balancer
4001 * - first_pick_cpu is the one of the busy CPUs. It will kick
4002 * idle load balancer when it has more than one process active. This
4003 * eliminates the need for idle load balancing altogether when we have
4004 * only one running process in the system (common case).
4005 * - If there are more than one busy CPU, idle load balancer may have
4006 * to run for active_load_balance to happen (i.e., two busy CPUs are
4007 * SMT or core siblings and can run better if they move to different
4008 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4009 * which will kick idle load balancer as soon as it has any load.
4011 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4013 unsigned long now = jiffies;
4015 int first_pick_cpu, second_pick_cpu;
4017 if (time_before(now, nohz.next_balance))
4020 if (rq->idle_at_tick)
4023 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4024 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4026 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4027 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4030 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4031 if (ret == nr_cpu_ids || ret == cpu) {
4032 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4033 if (rq->nr_running > 1)
4036 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4037 if (ret == nr_cpu_ids || ret == cpu) {
4045 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4049 * run_rebalance_domains is triggered when needed from the scheduler tick.
4050 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4052 static void run_rebalance_domains(struct softirq_action *h)
4054 int this_cpu = smp_processor_id();
4055 struct rq *this_rq = cpu_rq(this_cpu);
4056 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4057 CPU_IDLE : CPU_NOT_IDLE;
4059 rebalance_domains(this_cpu, idle);
4062 * If this cpu has a pending nohz_balance_kick, then do the
4063 * balancing on behalf of the other idle cpus whose ticks are
4066 nohz_idle_balance(this_cpu, idle);
4069 static inline int on_null_domain(int cpu)
4071 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4075 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4077 static inline void trigger_load_balance(struct rq *rq, int cpu)
4079 /* Don't need to rebalance while attached to NULL domain */
4080 if (time_after_eq(jiffies, rq->next_balance) &&
4081 likely(!on_null_domain(cpu)))
4082 raise_softirq(SCHED_SOFTIRQ);
4084 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4085 nohz_balancer_kick(cpu);
4089 static void rq_online_fair(struct rq *rq)
4094 static void rq_offline_fair(struct rq *rq)
4099 #else /* CONFIG_SMP */
4102 * on UP we do not need to balance between CPUs:
4104 static inline void idle_balance(int cpu, struct rq *rq)
4108 #endif /* CONFIG_SMP */
4111 * scheduler tick hitting a task of our scheduling class:
4113 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4115 struct cfs_rq *cfs_rq;
4116 struct sched_entity *se = &curr->se;
4118 for_each_sched_entity(se) {
4119 cfs_rq = cfs_rq_of(se);
4120 entity_tick(cfs_rq, se, queued);
4125 * called on fork with the child task as argument from the parent's context
4126 * - child not yet on the tasklist
4127 * - preemption disabled
4129 static void task_fork_fair(struct task_struct *p)
4131 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4132 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4133 int this_cpu = smp_processor_id();
4134 struct rq *rq = this_rq();
4135 unsigned long flags;
4137 raw_spin_lock_irqsave(&rq->lock, flags);
4139 update_rq_clock(rq);
4141 if (unlikely(task_cpu(p) != this_cpu)) {
4143 __set_task_cpu(p, this_cpu);
4147 update_curr(cfs_rq);
4150 se->vruntime = curr->vruntime;
4151 place_entity(cfs_rq, se, 1);
4153 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4155 * Upon rescheduling, sched_class::put_prev_task() will place
4156 * 'current' within the tree based on its new key value.
4158 swap(curr->vruntime, se->vruntime);
4159 resched_task(rq->curr);
4162 se->vruntime -= cfs_rq->min_vruntime;
4164 raw_spin_unlock_irqrestore(&rq->lock, flags);
4168 * Priority of the task has changed. Check to see if we preempt
4172 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4178 * Reschedule if we are currently running on this runqueue and
4179 * our priority decreased, or if we are not currently running on
4180 * this runqueue and our priority is higher than the current's
4182 if (rq->curr == p) {
4183 if (p->prio > oldprio)
4184 resched_task(rq->curr);
4186 check_preempt_curr(rq, p, 0);
4189 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4191 struct sched_entity *se = &p->se;
4192 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4195 * Ensure the task's vruntime is normalized, so that when its
4196 * switched back to the fair class the enqueue_entity(.flags=0) will
4197 * do the right thing.
4199 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4200 * have normalized the vruntime, if it was !on_rq, then only when
4201 * the task is sleeping will it still have non-normalized vruntime.
4203 if (!se->on_rq && p->state != TASK_RUNNING) {
4205 * Fix up our vruntime so that the current sleep doesn't
4206 * cause 'unlimited' sleep bonus.
4208 place_entity(cfs_rq, se, 0);
4209 se->vruntime -= cfs_rq->min_vruntime;
4214 * We switched to the sched_fair class.
4216 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4222 * We were most likely switched from sched_rt, so
4223 * kick off the schedule if running, otherwise just see
4224 * if we can still preempt the current task.
4227 resched_task(rq->curr);
4229 check_preempt_curr(rq, p, 0);
4232 /* Account for a task changing its policy or group.
4234 * This routine is mostly called to set cfs_rq->curr field when a task
4235 * migrates between groups/classes.
4237 static void set_curr_task_fair(struct rq *rq)
4239 struct sched_entity *se = &rq->curr->se;
4241 for_each_sched_entity(se)
4242 set_next_entity(cfs_rq_of(se), se);
4245 #ifdef CONFIG_FAIR_GROUP_SCHED
4246 static void task_move_group_fair(struct task_struct *p, int on_rq)
4249 * If the task was not on the rq at the time of this cgroup movement
4250 * it must have been asleep, sleeping tasks keep their ->vruntime
4251 * absolute on their old rq until wakeup (needed for the fair sleeper
4252 * bonus in place_entity()).
4254 * If it was on the rq, we've just 'preempted' it, which does convert
4255 * ->vruntime to a relative base.
4257 * Make sure both cases convert their relative position when migrating
4258 * to another cgroup's rq. This does somewhat interfere with the
4259 * fair sleeper stuff for the first placement, but who cares.
4262 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4263 set_task_rq(p, task_cpu(p));
4265 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4269 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4271 struct sched_entity *se = &task->se;
4272 unsigned int rr_interval = 0;
4275 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4278 if (rq->cfs.load.weight)
4279 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4285 * All the scheduling class methods:
4287 static const struct sched_class fair_sched_class = {
4288 .next = &idle_sched_class,
4289 .enqueue_task = enqueue_task_fair,
4290 .dequeue_task = dequeue_task_fair,
4291 .yield_task = yield_task_fair,
4292 .yield_to_task = yield_to_task_fair,
4294 .check_preempt_curr = check_preempt_wakeup,
4296 .pick_next_task = pick_next_task_fair,
4297 .put_prev_task = put_prev_task_fair,
4300 .select_task_rq = select_task_rq_fair,
4302 .rq_online = rq_online_fair,
4303 .rq_offline = rq_offline_fair,
4305 .task_waking = task_waking_fair,
4308 .set_curr_task = set_curr_task_fair,
4309 .task_tick = task_tick_fair,
4310 .task_fork = task_fork_fair,
4312 .prio_changed = prio_changed_fair,
4313 .switched_from = switched_from_fair,
4314 .switched_to = switched_to_fair,
4316 .get_rr_interval = get_rr_interval_fair,
4318 #ifdef CONFIG_FAIR_GROUP_SCHED
4319 .task_move_group = task_move_group_fair,
4323 #ifdef CONFIG_SCHED_DEBUG
4324 static void print_cfs_stats(struct seq_file *m, int cpu)
4326 struct cfs_rq *cfs_rq;
4329 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4330 print_cfs_rq(m, cpu, cfs_rq);