2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 8;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
99 static const struct sched_class fair_sched_class;
101 /**************************************************************
102 * CFS operations on generic schedulable entities:
105 #ifdef CONFIG_FAIR_GROUP_SCHED
107 /* cpu runqueue to which this cfs_rq is attached */
108 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
113 /* An entity is a task if it doesn't "own" a runqueue */
114 #define entity_is_task(se) (!se->my_q)
116 static inline struct task_struct *task_of(struct sched_entity *se)
118 #ifdef CONFIG_SCHED_DEBUG
119 WARN_ON_ONCE(!entity_is_task(se));
121 return container_of(se, struct task_struct, se);
124 /* Walk up scheduling entities hierarchy */
125 #define for_each_sched_entity(se) \
126 for (; se; se = se->parent)
128 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
133 /* runqueue on which this entity is (to be) queued */
134 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
139 /* runqueue "owned" by this group */
140 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
145 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
146 * another cpu ('this_cpu')
148 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
150 return cfs_rq->tg->cfs_rq[this_cpu];
153 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
155 if (!cfs_rq->on_list) {
157 * Ensure we either appear before our parent (if already
158 * enqueued) or force our parent to appear after us when it is
159 * enqueued. The fact that we always enqueue bottom-up
160 * reduces this to two cases.
162 if (cfs_rq->tg->parent &&
163 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
164 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
165 &rq_of(cfs_rq)->leaf_cfs_rq_list);
167 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
168 &rq_of(cfs_rq)->leaf_cfs_rq_list);
175 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
177 if (cfs_rq->on_list) {
178 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
183 /* Iterate thr' all leaf cfs_rq's on a runqueue */
184 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
185 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
187 /* Do the two (enqueued) entities belong to the same group ? */
189 is_same_group(struct sched_entity *se, struct sched_entity *pse)
191 if (se->cfs_rq == pse->cfs_rq)
197 static inline struct sched_entity *parent_entity(struct sched_entity *se)
202 /* return depth at which a sched entity is present in the hierarchy */
203 static inline int depth_se(struct sched_entity *se)
207 for_each_sched_entity(se)
214 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
216 int se_depth, pse_depth;
219 * preemption test can be made between sibling entities who are in the
220 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
221 * both tasks until we find their ancestors who are siblings of common
225 /* First walk up until both entities are at same depth */
226 se_depth = depth_se(*se);
227 pse_depth = depth_se(*pse);
229 while (se_depth > pse_depth) {
231 *se = parent_entity(*se);
234 while (pse_depth > se_depth) {
236 *pse = parent_entity(*pse);
239 while (!is_same_group(*se, *pse)) {
240 *se = parent_entity(*se);
241 *pse = parent_entity(*pse);
245 #else /* !CONFIG_FAIR_GROUP_SCHED */
247 static inline struct task_struct *task_of(struct sched_entity *se)
249 return container_of(se, struct task_struct, se);
252 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 return container_of(cfs_rq, struct rq, cfs);
257 #define entity_is_task(se) 1
259 #define for_each_sched_entity(se) \
260 for (; se; se = NULL)
262 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
264 return &task_rq(p)->cfs;
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
269 struct task_struct *p = task_of(se);
270 struct rq *rq = task_rq(p);
275 /* runqueue "owned" by this group */
276 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
283 return &cpu_rq(this_cpu)->cfs;
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
295 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
298 is_same_group(struct sched_entity *se, struct sched_entity *pse)
303 static inline struct sched_entity *parent_entity(struct sched_entity *se)
309 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
313 #endif /* CONFIG_FAIR_GROUP_SCHED */
316 /**************************************************************
317 * Scheduling class tree data structure manipulation methods:
320 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
322 s64 delta = (s64)(vruntime - min_vruntime);
324 min_vruntime = vruntime;
329 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
331 s64 delta = (s64)(vruntime - min_vruntime);
333 min_vruntime = vruntime;
338 static inline int entity_before(struct sched_entity *a,
339 struct sched_entity *b)
341 return (s64)(a->vruntime - b->vruntime) < 0;
344 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
346 return se->vruntime - cfs_rq->min_vruntime;
349 static void update_min_vruntime(struct cfs_rq *cfs_rq)
351 u64 vruntime = cfs_rq->min_vruntime;
354 vruntime = cfs_rq->curr->vruntime;
356 if (cfs_rq->rb_leftmost) {
357 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
362 vruntime = se->vruntime;
364 vruntime = min_vruntime(vruntime, se->vruntime);
367 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
371 * Enqueue an entity into the rb-tree:
373 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
375 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
376 struct rb_node *parent = NULL;
377 struct sched_entity *entry;
378 s64 key = entity_key(cfs_rq, se);
382 * Find the right place in the rbtree:
386 entry = rb_entry(parent, struct sched_entity, run_node);
388 * We dont care about collisions. Nodes with
389 * the same key stay together.
391 if (key < entity_key(cfs_rq, entry)) {
392 link = &parent->rb_left;
394 link = &parent->rb_right;
400 * Maintain a cache of leftmost tree entries (it is frequently
404 cfs_rq->rb_leftmost = &se->run_node;
406 rb_link_node(&se->run_node, parent, link);
407 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
410 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
412 if (cfs_rq->rb_leftmost == &se->run_node) {
413 struct rb_node *next_node;
415 next_node = rb_next(&se->run_node);
416 cfs_rq->rb_leftmost = next_node;
419 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
422 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
424 struct rb_node *left = cfs_rq->rb_leftmost;
429 return rb_entry(left, struct sched_entity, run_node);
432 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
434 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
439 return rb_entry(last, struct sched_entity, run_node);
442 /**************************************************************
443 * Scheduling class statistics methods:
446 #ifdef CONFIG_SCHED_DEBUG
447 int sched_proc_update_handler(struct ctl_table *table, int write,
448 void __user *buffer, size_t *lenp,
451 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
452 int factor = get_update_sysctl_factor();
457 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
458 sysctl_sched_min_granularity);
460 #define WRT_SYSCTL(name) \
461 (normalized_sysctl_##name = sysctl_##name / (factor))
462 WRT_SYSCTL(sched_min_granularity);
463 WRT_SYSCTL(sched_latency);
464 WRT_SYSCTL(sched_wakeup_granularity);
474 static inline unsigned long
475 calc_delta_fair(unsigned long delta, struct sched_entity *se)
477 if (unlikely(se->load.weight != NICE_0_LOAD))
478 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
484 * The idea is to set a period in which each task runs once.
486 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
487 * this period because otherwise the slices get too small.
489 * p = (nr <= nl) ? l : l*nr/nl
491 static u64 __sched_period(unsigned long nr_running)
493 u64 period = sysctl_sched_latency;
494 unsigned long nr_latency = sched_nr_latency;
496 if (unlikely(nr_running > nr_latency)) {
497 period = sysctl_sched_min_granularity;
498 period *= nr_running;
505 * We calculate the wall-time slice from the period by taking a part
506 * proportional to the weight.
510 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
512 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
514 for_each_sched_entity(se) {
515 struct load_weight *load;
516 struct load_weight lw;
518 cfs_rq = cfs_rq_of(se);
519 load = &cfs_rq->load;
521 if (unlikely(!se->on_rq)) {
524 update_load_add(&lw, se->load.weight);
527 slice = calc_delta_mine(slice, se->load.weight, load);
533 * We calculate the vruntime slice of a to be inserted task
537 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
539 return calc_delta_fair(sched_slice(cfs_rq, se), se);
542 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
543 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta);
546 * Update the current task's runtime statistics. Skip current tasks that
547 * are not in our scheduling class.
550 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
551 unsigned long delta_exec)
553 unsigned long delta_exec_weighted;
555 schedstat_set(curr->statistics.exec_max,
556 max((u64)delta_exec, curr->statistics.exec_max));
558 curr->sum_exec_runtime += delta_exec;
559 schedstat_add(cfs_rq, exec_clock, delta_exec);
560 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
562 curr->vruntime += delta_exec_weighted;
563 update_min_vruntime(cfs_rq);
565 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
566 cfs_rq->load_unacc_exec_time += delta_exec;
567 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
568 update_cfs_load(cfs_rq, 0);
569 update_cfs_shares(cfs_rq, 0);
574 static void update_curr(struct cfs_rq *cfs_rq)
576 struct sched_entity *curr = cfs_rq->curr;
577 u64 now = rq_of(cfs_rq)->clock_task;
578 unsigned long delta_exec;
584 * Get the amount of time the current task was running
585 * since the last time we changed load (this cannot
586 * overflow on 32 bits):
588 delta_exec = (unsigned long)(now - curr->exec_start);
592 __update_curr(cfs_rq, curr, delta_exec);
593 curr->exec_start = now;
595 if (entity_is_task(curr)) {
596 struct task_struct *curtask = task_of(curr);
598 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
599 cpuacct_charge(curtask, delta_exec);
600 account_group_exec_runtime(curtask, delta_exec);
605 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
607 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
611 * Task is being enqueued - update stats:
613 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
616 * Are we enqueueing a waiting task? (for current tasks
617 * a dequeue/enqueue event is a NOP)
619 if (se != cfs_rq->curr)
620 update_stats_wait_start(cfs_rq, se);
624 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
626 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
627 rq_of(cfs_rq)->clock - se->statistics.wait_start));
628 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
629 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
630 rq_of(cfs_rq)->clock - se->statistics.wait_start);
631 #ifdef CONFIG_SCHEDSTATS
632 if (entity_is_task(se)) {
633 trace_sched_stat_wait(task_of(se),
634 rq_of(cfs_rq)->clock - se->statistics.wait_start);
637 schedstat_set(se->statistics.wait_start, 0);
641 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
644 * Mark the end of the wait period if dequeueing a
647 if (se != cfs_rq->curr)
648 update_stats_wait_end(cfs_rq, se);
652 * We are picking a new current task - update its stats:
655 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
658 * We are starting a new run period:
660 se->exec_start = rq_of(cfs_rq)->clock_task;
663 /**************************************************
664 * Scheduling class queueing methods:
667 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
669 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
671 cfs_rq->task_weight += weight;
675 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
681 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
683 update_load_add(&cfs_rq->load, se->load.weight);
684 if (!parent_entity(se))
685 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
686 if (entity_is_task(se)) {
687 add_cfs_task_weight(cfs_rq, se->load.weight);
688 list_add(&se->group_node, &cfs_rq->tasks);
690 cfs_rq->nr_running++;
694 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
696 update_load_sub(&cfs_rq->load, se->load.weight);
697 if (!parent_entity(se))
698 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
699 if (entity_is_task(se)) {
700 add_cfs_task_weight(cfs_rq, -se->load.weight);
701 list_del_init(&se->group_node);
703 cfs_rq->nr_running--;
706 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
707 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
710 struct task_group *tg = cfs_rq->tg;
713 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
714 load_avg -= cfs_rq->load_contribution;
716 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
717 atomic_add(load_avg, &tg->load_weight);
718 cfs_rq->load_contribution += load_avg;
722 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
724 u64 period = sysctl_sched_shares_window;
726 unsigned long load = cfs_rq->load.weight;
731 now = rq_of(cfs_rq)->clock;
732 delta = now - cfs_rq->load_stamp;
734 /* truncate load history at 4 idle periods */
735 if (cfs_rq->load_stamp > cfs_rq->load_last &&
736 now - cfs_rq->load_last > 4 * period) {
737 cfs_rq->load_period = 0;
738 cfs_rq->load_avg = 0;
741 cfs_rq->load_stamp = now;
742 cfs_rq->load_unacc_exec_time = 0;
743 cfs_rq->load_period += delta;
745 cfs_rq->load_last = now;
746 cfs_rq->load_avg += delta * load;
749 /* consider updating load contribution on each fold or truncate */
750 if (global_update || cfs_rq->load_period > period
751 || !cfs_rq->load_period)
752 update_cfs_rq_load_contribution(cfs_rq, global_update);
754 while (cfs_rq->load_period > period) {
756 * Inline assembly required to prevent the compiler
757 * optimising this loop into a divmod call.
758 * See __iter_div_u64_rem() for another example of this.
760 asm("" : "+rm" (cfs_rq->load_period));
761 cfs_rq->load_period /= 2;
762 cfs_rq->load_avg /= 2;
765 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
766 list_del_leaf_cfs_rq(cfs_rq);
769 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
770 unsigned long weight)
773 account_entity_dequeue(cfs_rq, se);
775 update_load_set(&se->load, weight);
778 account_entity_enqueue(cfs_rq, se);
781 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
783 struct task_group *tg;
784 struct sched_entity *se;
785 long load_weight, load, shares;
791 se = tg->se[cpu_of(rq_of(cfs_rq))];
795 load = cfs_rq->load.weight + weight_delta;
797 load_weight = atomic_read(&tg->load_weight);
798 load_weight -= cfs_rq->load_contribution;
801 shares = (tg->shares * load);
803 shares /= load_weight;
805 if (shares < MIN_SHARES)
807 if (shares > tg->shares)
810 reweight_entity(cfs_rq_of(se), se, shares);
812 #else /* CONFIG_FAIR_GROUP_SCHED */
813 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
817 static inline void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
820 #endif /* CONFIG_FAIR_GROUP_SCHED */
822 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
824 #ifdef CONFIG_SCHEDSTATS
825 struct task_struct *tsk = NULL;
827 if (entity_is_task(se))
830 if (se->statistics.sleep_start) {
831 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
836 if (unlikely(delta > se->statistics.sleep_max))
837 se->statistics.sleep_max = delta;
839 se->statistics.sleep_start = 0;
840 se->statistics.sum_sleep_runtime += delta;
843 account_scheduler_latency(tsk, delta >> 10, 1);
844 trace_sched_stat_sleep(tsk, delta);
847 if (se->statistics.block_start) {
848 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
853 if (unlikely(delta > se->statistics.block_max))
854 se->statistics.block_max = delta;
856 se->statistics.block_start = 0;
857 se->statistics.sum_sleep_runtime += delta;
860 if (tsk->in_iowait) {
861 se->statistics.iowait_sum += delta;
862 se->statistics.iowait_count++;
863 trace_sched_stat_iowait(tsk, delta);
867 * Blocking time is in units of nanosecs, so shift by
868 * 20 to get a milliseconds-range estimation of the
869 * amount of time that the task spent sleeping:
871 if (unlikely(prof_on == SLEEP_PROFILING)) {
872 profile_hits(SLEEP_PROFILING,
873 (void *)get_wchan(tsk),
876 account_scheduler_latency(tsk, delta >> 10, 0);
882 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
884 #ifdef CONFIG_SCHED_DEBUG
885 s64 d = se->vruntime - cfs_rq->min_vruntime;
890 if (d > 3*sysctl_sched_latency)
891 schedstat_inc(cfs_rq, nr_spread_over);
896 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
898 u64 vruntime = cfs_rq->min_vruntime;
901 * The 'current' period is already promised to the current tasks,
902 * however the extra weight of the new task will slow them down a
903 * little, place the new task so that it fits in the slot that
904 * stays open at the end.
906 if (initial && sched_feat(START_DEBIT))
907 vruntime += sched_vslice(cfs_rq, se);
909 /* sleeps up to a single latency don't count. */
911 unsigned long thresh = sysctl_sched_latency;
914 * Halve their sleep time's effect, to allow
915 * for a gentler effect of sleepers:
917 if (sched_feat(GENTLE_FAIR_SLEEPERS))
923 /* ensure we never gain time by being placed backwards. */
924 vruntime = max_vruntime(se->vruntime, vruntime);
926 se->vruntime = vruntime;
930 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
933 * Update the normalized vruntime before updating min_vruntime
934 * through callig update_curr().
936 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
937 se->vruntime += cfs_rq->min_vruntime;
940 * Update run-time statistics of the 'current'.
943 update_cfs_load(cfs_rq, 0);
944 update_cfs_shares(cfs_rq, se->load.weight);
945 account_entity_enqueue(cfs_rq, se);
947 if (flags & ENQUEUE_WAKEUP) {
948 place_entity(cfs_rq, se, 0);
949 enqueue_sleeper(cfs_rq, se);
952 update_stats_enqueue(cfs_rq, se);
953 check_spread(cfs_rq, se);
954 if (se != cfs_rq->curr)
955 __enqueue_entity(cfs_rq, se);
958 if (cfs_rq->nr_running == 1)
959 list_add_leaf_cfs_rq(cfs_rq);
962 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
964 if (!se || cfs_rq->last == se)
967 if (!se || cfs_rq->next == se)
971 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
973 for_each_sched_entity(se)
974 __clear_buddies(cfs_rq_of(se), se);
978 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
981 * Update run-time statistics of the 'current'.
985 update_stats_dequeue(cfs_rq, se);
986 if (flags & DEQUEUE_SLEEP) {
987 #ifdef CONFIG_SCHEDSTATS
988 if (entity_is_task(se)) {
989 struct task_struct *tsk = task_of(se);
991 if (tsk->state & TASK_INTERRUPTIBLE)
992 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
993 if (tsk->state & TASK_UNINTERRUPTIBLE)
994 se->statistics.block_start = rq_of(cfs_rq)->clock;
999 clear_buddies(cfs_rq, se);
1001 if (se != cfs_rq->curr)
1002 __dequeue_entity(cfs_rq, se);
1004 update_cfs_load(cfs_rq, 0);
1005 account_entity_dequeue(cfs_rq, se);
1006 update_min_vruntime(cfs_rq);
1007 update_cfs_shares(cfs_rq, 0);
1010 * Normalize the entity after updating the min_vruntime because the
1011 * update can refer to the ->curr item and we need to reflect this
1012 * movement in our normalized position.
1014 if (!(flags & DEQUEUE_SLEEP))
1015 se->vruntime -= cfs_rq->min_vruntime;
1019 * Preempt the current task with a newly woken task if needed:
1022 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1024 unsigned long ideal_runtime, delta_exec;
1026 ideal_runtime = sched_slice(cfs_rq, curr);
1027 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1028 if (delta_exec > ideal_runtime) {
1029 resched_task(rq_of(cfs_rq)->curr);
1031 * The current task ran long enough, ensure it doesn't get
1032 * re-elected due to buddy favours.
1034 clear_buddies(cfs_rq, curr);
1039 * Ensure that a task that missed wakeup preemption by a
1040 * narrow margin doesn't have to wait for a full slice.
1041 * This also mitigates buddy induced latencies under load.
1043 if (!sched_feat(WAKEUP_PREEMPT))
1046 if (delta_exec < sysctl_sched_min_granularity)
1049 if (cfs_rq->nr_running > 1) {
1050 struct sched_entity *se = __pick_next_entity(cfs_rq);
1051 s64 delta = curr->vruntime - se->vruntime;
1053 if (delta > ideal_runtime)
1054 resched_task(rq_of(cfs_rq)->curr);
1059 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1061 /* 'current' is not kept within the tree. */
1064 * Any task has to be enqueued before it get to execute on
1065 * a CPU. So account for the time it spent waiting on the
1068 update_stats_wait_end(cfs_rq, se);
1069 __dequeue_entity(cfs_rq, se);
1072 update_stats_curr_start(cfs_rq, se);
1074 #ifdef CONFIG_SCHEDSTATS
1076 * Track our maximum slice length, if the CPU's load is at
1077 * least twice that of our own weight (i.e. dont track it
1078 * when there are only lesser-weight tasks around):
1080 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1081 se->statistics.slice_max = max(se->statistics.slice_max,
1082 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1085 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1089 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1091 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1093 struct sched_entity *se = __pick_next_entity(cfs_rq);
1094 struct sched_entity *left = se;
1096 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1100 * Prefer last buddy, try to return the CPU to a preempted task.
1102 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1105 clear_buddies(cfs_rq, se);
1110 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1113 * If still on the runqueue then deactivate_task()
1114 * was not called and update_curr() has to be done:
1117 update_curr(cfs_rq);
1119 check_spread(cfs_rq, prev);
1121 update_stats_wait_start(cfs_rq, prev);
1122 /* Put 'current' back into the tree. */
1123 __enqueue_entity(cfs_rq, prev);
1125 cfs_rq->curr = NULL;
1129 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1132 * Update run-time statistics of the 'current'.
1134 update_curr(cfs_rq);
1136 #ifdef CONFIG_SCHED_HRTICK
1138 * queued ticks are scheduled to match the slice, so don't bother
1139 * validating it and just reschedule.
1142 resched_task(rq_of(cfs_rq)->curr);
1146 * don't let the period tick interfere with the hrtick preemption
1148 if (!sched_feat(DOUBLE_TICK) &&
1149 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1153 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1154 check_preempt_tick(cfs_rq, curr);
1157 /**************************************************
1158 * CFS operations on tasks:
1161 #ifdef CONFIG_SCHED_HRTICK
1162 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1164 struct sched_entity *se = &p->se;
1165 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1167 WARN_ON(task_rq(p) != rq);
1169 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1170 u64 slice = sched_slice(cfs_rq, se);
1171 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1172 s64 delta = slice - ran;
1181 * Don't schedule slices shorter than 10000ns, that just
1182 * doesn't make sense. Rely on vruntime for fairness.
1185 delta = max_t(s64, 10000LL, delta);
1187 hrtick_start(rq, delta);
1192 * called from enqueue/dequeue and updates the hrtick when the
1193 * current task is from our class and nr_running is low enough
1196 static void hrtick_update(struct rq *rq)
1198 struct task_struct *curr = rq->curr;
1200 if (curr->sched_class != &fair_sched_class)
1203 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1204 hrtick_start_fair(rq, curr);
1206 #else /* !CONFIG_SCHED_HRTICK */
1208 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1212 static inline void hrtick_update(struct rq *rq)
1218 * The enqueue_task method is called before nr_running is
1219 * increased. Here we update the fair scheduling stats and
1220 * then put the task into the rbtree:
1223 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1225 struct cfs_rq *cfs_rq;
1226 struct sched_entity *se = &p->se;
1228 for_each_sched_entity(se) {
1231 cfs_rq = cfs_rq_of(se);
1232 enqueue_entity(cfs_rq, se, flags);
1233 flags = ENQUEUE_WAKEUP;
1236 for_each_sched_entity(se) {
1237 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1239 update_cfs_load(cfs_rq, 0);
1240 update_cfs_shares(cfs_rq, 0);
1247 * The dequeue_task method is called before nr_running is
1248 * decreased. We remove the task from the rbtree and
1249 * update the fair scheduling stats:
1251 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1253 struct cfs_rq *cfs_rq;
1254 struct sched_entity *se = &p->se;
1256 for_each_sched_entity(se) {
1257 cfs_rq = cfs_rq_of(se);
1258 dequeue_entity(cfs_rq, se, flags);
1260 /* Don't dequeue parent if it has other entities besides us */
1261 if (cfs_rq->load.weight)
1263 flags |= DEQUEUE_SLEEP;
1266 for_each_sched_entity(se) {
1267 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1269 update_cfs_load(cfs_rq, 0);
1270 update_cfs_shares(cfs_rq, 0);
1277 * sched_yield() support is very simple - we dequeue and enqueue.
1279 * If compat_yield is turned on then we requeue to the end of the tree.
1281 static void yield_task_fair(struct rq *rq)
1283 struct task_struct *curr = rq->curr;
1284 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1285 struct sched_entity *rightmost, *se = &curr->se;
1288 * Are we the only task in the tree?
1290 if (unlikely(cfs_rq->nr_running == 1))
1293 clear_buddies(cfs_rq, se);
1295 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1296 update_rq_clock(rq);
1298 * Update run-time statistics of the 'current'.
1300 update_curr(cfs_rq);
1305 * Find the rightmost entry in the rbtree:
1307 rightmost = __pick_last_entity(cfs_rq);
1309 * Already in the rightmost position?
1311 if (unlikely(!rightmost || entity_before(rightmost, se)))
1315 * Minimally necessary key value to be last in the tree:
1316 * Upon rescheduling, sched_class::put_prev_task() will place
1317 * 'current' within the tree based on its new key value.
1319 se->vruntime = rightmost->vruntime + 1;
1324 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1326 struct sched_entity *se = &p->se;
1327 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1329 se->vruntime -= cfs_rq->min_vruntime;
1332 #ifdef CONFIG_FAIR_GROUP_SCHED
1334 * effective_load() calculates the load change as seen from the root_task_group
1336 * Adding load to a group doesn't make a group heavier, but can cause movement
1337 * of group shares between cpus. Assuming the shares were perfectly aligned one
1338 * can calculate the shift in shares.
1340 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1342 struct sched_entity *se = tg->se[cpu];
1347 for_each_sched_entity(se) {
1348 long S, rw, s, a, b;
1350 S = se->my_q->tg->shares;
1351 s = se->load.weight;
1352 rw = se->my_q->load.weight;
1363 * Assume the group is already running and will
1364 * thus already be accounted for in the weight.
1366 * That is, moving shares between CPUs, does not
1367 * alter the group weight.
1377 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1378 unsigned long wl, unsigned long wg)
1385 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1387 unsigned long this_load, load;
1388 int idx, this_cpu, prev_cpu;
1389 unsigned long tl_per_task;
1390 struct task_group *tg;
1391 unsigned long weight;
1395 this_cpu = smp_processor_id();
1396 prev_cpu = task_cpu(p);
1397 load = source_load(prev_cpu, idx);
1398 this_load = target_load(this_cpu, idx);
1401 * If sync wakeup then subtract the (maximum possible)
1402 * effect of the currently running task from the load
1403 * of the current CPU:
1407 tg = task_group(current);
1408 weight = current->se.load.weight;
1410 this_load += effective_load(tg, this_cpu, -weight, -weight);
1411 load += effective_load(tg, prev_cpu, 0, -weight);
1415 weight = p->se.load.weight;
1418 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1419 * due to the sync cause above having dropped this_load to 0, we'll
1420 * always have an imbalance, but there's really nothing you can do
1421 * about that, so that's good too.
1423 * Otherwise check if either cpus are near enough in load to allow this
1424 * task to be woken on this_cpu.
1427 unsigned long this_eff_load, prev_eff_load;
1429 this_eff_load = 100;
1430 this_eff_load *= power_of(prev_cpu);
1431 this_eff_load *= this_load +
1432 effective_load(tg, this_cpu, weight, weight);
1434 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1435 prev_eff_load *= power_of(this_cpu);
1436 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1438 balanced = this_eff_load <= prev_eff_load;
1444 * If the currently running task will sleep within
1445 * a reasonable amount of time then attract this newly
1448 if (sync && balanced)
1451 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1452 tl_per_task = cpu_avg_load_per_task(this_cpu);
1455 (this_load <= load &&
1456 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1458 * This domain has SD_WAKE_AFFINE and
1459 * p is cache cold in this domain, and
1460 * there is no bad imbalance.
1462 schedstat_inc(sd, ttwu_move_affine);
1463 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1471 * find_idlest_group finds and returns the least busy CPU group within the
1474 static struct sched_group *
1475 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1476 int this_cpu, int load_idx)
1478 struct sched_group *idlest = NULL, *group = sd->groups;
1479 unsigned long min_load = ULONG_MAX, this_load = 0;
1480 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1483 unsigned long load, avg_load;
1487 /* Skip over this group if it has no CPUs allowed */
1488 if (!cpumask_intersects(sched_group_cpus(group),
1492 local_group = cpumask_test_cpu(this_cpu,
1493 sched_group_cpus(group));
1495 /* Tally up the load of all CPUs in the group */
1498 for_each_cpu(i, sched_group_cpus(group)) {
1499 /* Bias balancing toward cpus of our domain */
1501 load = source_load(i, load_idx);
1503 load = target_load(i, load_idx);
1508 /* Adjust by relative CPU power of the group */
1509 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1512 this_load = avg_load;
1513 } else if (avg_load < min_load) {
1514 min_load = avg_load;
1517 } while (group = group->next, group != sd->groups);
1519 if (!idlest || 100*this_load < imbalance*min_load)
1525 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1528 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1530 unsigned long load, min_load = ULONG_MAX;
1534 /* Traverse only the allowed CPUs */
1535 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1536 load = weighted_cpuload(i);
1538 if (load < min_load || (load == min_load && i == this_cpu)) {
1548 * Try and locate an idle CPU in the sched_domain.
1550 static int select_idle_sibling(struct task_struct *p, int target)
1552 int cpu = smp_processor_id();
1553 int prev_cpu = task_cpu(p);
1554 struct sched_domain *sd;
1558 * If the task is going to be woken-up on this cpu and if it is
1559 * already idle, then it is the right target.
1561 if (target == cpu && idle_cpu(cpu))
1565 * If the task is going to be woken-up on the cpu where it previously
1566 * ran and if it is currently idle, then it the right target.
1568 if (target == prev_cpu && idle_cpu(prev_cpu))
1572 * Otherwise, iterate the domains and find an elegible idle cpu.
1574 for_each_domain(target, sd) {
1575 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1578 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1586 * Lets stop looking for an idle sibling when we reached
1587 * the domain that spans the current cpu and prev_cpu.
1589 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1590 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1598 * sched_balance_self: balance the current task (running on cpu) in domains
1599 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1602 * Balance, ie. select the least loaded group.
1604 * Returns the target CPU number, or the same CPU if no balancing is needed.
1606 * preempt must be disabled.
1609 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1611 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1612 int cpu = smp_processor_id();
1613 int prev_cpu = task_cpu(p);
1615 int want_affine = 0;
1617 int sync = wake_flags & WF_SYNC;
1619 if (sd_flag & SD_BALANCE_WAKE) {
1620 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1625 for_each_domain(cpu, tmp) {
1626 if (!(tmp->flags & SD_LOAD_BALANCE))
1630 * If power savings logic is enabled for a domain, see if we
1631 * are not overloaded, if so, don't balance wider.
1633 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1634 unsigned long power = 0;
1635 unsigned long nr_running = 0;
1636 unsigned long capacity;
1639 for_each_cpu(i, sched_domain_span(tmp)) {
1640 power += power_of(i);
1641 nr_running += cpu_rq(i)->cfs.nr_running;
1644 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1646 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1649 if (nr_running < capacity)
1654 * If both cpu and prev_cpu are part of this domain,
1655 * cpu is a valid SD_WAKE_AFFINE target.
1657 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1658 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1663 if (!want_sd && !want_affine)
1666 if (!(tmp->flags & sd_flag))
1674 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1675 return select_idle_sibling(p, cpu);
1677 return select_idle_sibling(p, prev_cpu);
1681 int load_idx = sd->forkexec_idx;
1682 struct sched_group *group;
1685 if (!(sd->flags & sd_flag)) {
1690 if (sd_flag & SD_BALANCE_WAKE)
1691 load_idx = sd->wake_idx;
1693 group = find_idlest_group(sd, p, cpu, load_idx);
1699 new_cpu = find_idlest_cpu(group, p, cpu);
1700 if (new_cpu == -1 || new_cpu == cpu) {
1701 /* Now try balancing at a lower domain level of cpu */
1706 /* Now try balancing at a lower domain level of new_cpu */
1708 weight = sd->span_weight;
1710 for_each_domain(cpu, tmp) {
1711 if (weight <= tmp->span_weight)
1713 if (tmp->flags & sd_flag)
1716 /* while loop will break here if sd == NULL */
1721 #endif /* CONFIG_SMP */
1723 static unsigned long
1724 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1726 unsigned long gran = sysctl_sched_wakeup_granularity;
1729 * Since its curr running now, convert the gran from real-time
1730 * to virtual-time in his units.
1732 * By using 'se' instead of 'curr' we penalize light tasks, so
1733 * they get preempted easier. That is, if 'se' < 'curr' then
1734 * the resulting gran will be larger, therefore penalizing the
1735 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1736 * be smaller, again penalizing the lighter task.
1738 * This is especially important for buddies when the leftmost
1739 * task is higher priority than the buddy.
1741 if (unlikely(se->load.weight != NICE_0_LOAD))
1742 gran = calc_delta_fair(gran, se);
1748 * Should 'se' preempt 'curr'.
1762 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1764 s64 gran, vdiff = curr->vruntime - se->vruntime;
1769 gran = wakeup_gran(curr, se);
1776 static void set_last_buddy(struct sched_entity *se)
1778 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1779 for_each_sched_entity(se)
1780 cfs_rq_of(se)->last = se;
1784 static void set_next_buddy(struct sched_entity *se)
1786 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1787 for_each_sched_entity(se)
1788 cfs_rq_of(se)->next = se;
1793 * Preempt the current task with a newly woken task if needed:
1795 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1797 struct task_struct *curr = rq->curr;
1798 struct sched_entity *se = &curr->se, *pse = &p->se;
1799 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1800 int scale = cfs_rq->nr_running >= sched_nr_latency;
1802 if (unlikely(rt_prio(p->prio)))
1805 if (unlikely(p->sched_class != &fair_sched_class))
1808 if (unlikely(se == pse))
1811 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1812 set_next_buddy(pse);
1815 * We can come here with TIF_NEED_RESCHED already set from new task
1818 if (test_tsk_need_resched(curr))
1822 * Batch and idle tasks do not preempt (their preemption is driven by
1825 if (unlikely(p->policy != SCHED_NORMAL))
1828 /* Idle tasks are by definition preempted by everybody. */
1829 if (unlikely(curr->policy == SCHED_IDLE))
1832 if (!sched_feat(WAKEUP_PREEMPT))
1835 update_curr(cfs_rq);
1836 find_matching_se(&se, &pse);
1838 if (wakeup_preempt_entity(se, pse) == 1)
1846 * Only set the backward buddy when the current task is still
1847 * on the rq. This can happen when a wakeup gets interleaved
1848 * with schedule on the ->pre_schedule() or idle_balance()
1849 * point, either of which can * drop the rq lock.
1851 * Also, during early boot the idle thread is in the fair class,
1852 * for obvious reasons its a bad idea to schedule back to it.
1854 if (unlikely(!se->on_rq || curr == rq->idle))
1857 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1861 static struct task_struct *pick_next_task_fair(struct rq *rq)
1863 struct task_struct *p;
1864 struct cfs_rq *cfs_rq = &rq->cfs;
1865 struct sched_entity *se;
1867 if (!cfs_rq->nr_running)
1871 se = pick_next_entity(cfs_rq);
1872 set_next_entity(cfs_rq, se);
1873 cfs_rq = group_cfs_rq(se);
1877 hrtick_start_fair(rq, p);
1883 * Account for a descheduled task:
1885 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1887 struct sched_entity *se = &prev->se;
1888 struct cfs_rq *cfs_rq;
1890 for_each_sched_entity(se) {
1891 cfs_rq = cfs_rq_of(se);
1892 put_prev_entity(cfs_rq, se);
1897 /**************************************************
1898 * Fair scheduling class load-balancing methods:
1902 * pull_task - move a task from a remote runqueue to the local runqueue.
1903 * Both runqueues must be locked.
1905 static void pull_task(struct rq *src_rq, struct task_struct *p,
1906 struct rq *this_rq, int this_cpu)
1908 deactivate_task(src_rq, p, 0);
1909 set_task_cpu(p, this_cpu);
1910 activate_task(this_rq, p, 0);
1911 check_preempt_curr(this_rq, p, 0);
1913 /* re-arm NEWIDLE balancing when moving tasks */
1914 src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost;
1915 this_rq->idle_stamp = 0;
1919 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1922 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1923 struct sched_domain *sd, enum cpu_idle_type idle,
1926 int tsk_cache_hot = 0;
1928 * We do not migrate tasks that are:
1929 * 1) running (obviously), or
1930 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1931 * 3) are cache-hot on their current CPU.
1933 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1934 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1939 if (task_running(rq, p)) {
1940 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1945 * Aggressive migration if:
1946 * 1) task is cache cold, or
1947 * 2) too many balance attempts have failed.
1950 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
1951 if (!tsk_cache_hot ||
1952 sd->nr_balance_failed > sd->cache_nice_tries) {
1953 #ifdef CONFIG_SCHEDSTATS
1954 if (tsk_cache_hot) {
1955 schedstat_inc(sd, lb_hot_gained[idle]);
1956 schedstat_inc(p, se.statistics.nr_forced_migrations);
1962 if (tsk_cache_hot) {
1963 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
1970 * move_one_task tries to move exactly one task from busiest to this_rq, as
1971 * part of active balancing operations within "domain".
1972 * Returns 1 if successful and 0 otherwise.
1974 * Called with both runqueues locked.
1977 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1978 struct sched_domain *sd, enum cpu_idle_type idle)
1980 struct task_struct *p, *n;
1981 struct cfs_rq *cfs_rq;
1984 for_each_leaf_cfs_rq(busiest, cfs_rq) {
1985 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1987 if (!can_migrate_task(p, busiest, this_cpu,
1991 pull_task(busiest, p, this_rq, this_cpu);
1993 * Right now, this is only the second place pull_task()
1994 * is called, so we can safely collect pull_task()
1995 * stats here rather than inside pull_task().
1997 schedstat_inc(sd, lb_gained[idle]);
2005 static unsigned long
2006 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2007 unsigned long max_load_move, struct sched_domain *sd,
2008 enum cpu_idle_type idle, int *all_pinned,
2009 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
2011 int loops = 0, pulled = 0, pinned = 0;
2012 long rem_load_move = max_load_move;
2013 struct task_struct *p, *n;
2015 if (max_load_move == 0)
2020 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2021 if (loops++ > sysctl_sched_nr_migrate)
2024 if ((p->se.load.weight >> 1) > rem_load_move ||
2025 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
2028 pull_task(busiest, p, this_rq, this_cpu);
2030 rem_load_move -= p->se.load.weight;
2032 #ifdef CONFIG_PREEMPT
2034 * NEWIDLE balancing is a source of latency, so preemptible
2035 * kernels will stop after the first task is pulled to minimize
2036 * the critical section.
2038 if (idle == CPU_NEWLY_IDLE)
2043 * We only want to steal up to the prescribed amount of
2046 if (rem_load_move <= 0)
2049 if (p->prio < *this_best_prio)
2050 *this_best_prio = p->prio;
2054 * Right now, this is one of only two places pull_task() is called,
2055 * so we can safely collect pull_task() stats here rather than
2056 * inside pull_task().
2058 schedstat_add(sd, lb_gained[idle], pulled);
2061 *all_pinned = pinned;
2063 return max_load_move - rem_load_move;
2066 #ifdef CONFIG_FAIR_GROUP_SCHED
2068 * update tg->load_weight by folding this cpu's load_avg
2070 static int update_shares_cpu(struct task_group *tg, int cpu)
2072 struct cfs_rq *cfs_rq;
2073 unsigned long flags;
2080 cfs_rq = tg->cfs_rq[cpu];
2082 raw_spin_lock_irqsave(&rq->lock, flags);
2084 update_rq_clock(rq);
2085 update_cfs_load(cfs_rq, 1);
2088 * We need to update shares after updating tg->load_weight in
2089 * order to adjust the weight of groups with long running tasks.
2091 update_cfs_shares(cfs_rq, 0);
2093 raw_spin_unlock_irqrestore(&rq->lock, flags);
2098 static void update_shares(int cpu)
2100 struct cfs_rq *cfs_rq;
2101 struct rq *rq = cpu_rq(cpu);
2104 for_each_leaf_cfs_rq(rq, cfs_rq)
2105 update_shares_cpu(cfs_rq->tg, cpu);
2109 static unsigned long
2110 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2111 unsigned long max_load_move,
2112 struct sched_domain *sd, enum cpu_idle_type idle,
2113 int *all_pinned, int *this_best_prio)
2115 long rem_load_move = max_load_move;
2116 int busiest_cpu = cpu_of(busiest);
2117 struct task_group *tg;
2120 update_h_load(busiest_cpu);
2122 list_for_each_entry_rcu(tg, &task_groups, list) {
2123 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2124 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2125 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2126 u64 rem_load, moved_load;
2131 if (!busiest_cfs_rq->task_weight)
2134 rem_load = (u64)rem_load_move * busiest_weight;
2135 rem_load = div_u64(rem_load, busiest_h_load + 1);
2137 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2138 rem_load, sd, idle, all_pinned, this_best_prio,
2144 moved_load *= busiest_h_load;
2145 moved_load = div_u64(moved_load, busiest_weight + 1);
2147 rem_load_move -= moved_load;
2148 if (rem_load_move < 0)
2153 return max_load_move - rem_load_move;
2156 static inline void update_shares(int cpu)
2160 static unsigned long
2161 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2162 unsigned long max_load_move,
2163 struct sched_domain *sd, enum cpu_idle_type idle,
2164 int *all_pinned, int *this_best_prio)
2166 return balance_tasks(this_rq, this_cpu, busiest,
2167 max_load_move, sd, idle, all_pinned,
2168 this_best_prio, &busiest->cfs);
2173 * move_tasks tries to move up to max_load_move weighted load from busiest to
2174 * this_rq, as part of a balancing operation within domain "sd".
2175 * Returns 1 if successful and 0 otherwise.
2177 * Called with both runqueues locked.
2179 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2180 unsigned long max_load_move,
2181 struct sched_domain *sd, enum cpu_idle_type idle,
2184 unsigned long total_load_moved = 0, load_moved;
2185 int this_best_prio = this_rq->curr->prio;
2188 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2189 max_load_move - total_load_moved,
2190 sd, idle, all_pinned, &this_best_prio);
2192 total_load_moved += load_moved;
2194 #ifdef CONFIG_PREEMPT
2196 * NEWIDLE balancing is a source of latency, so preemptible
2197 * kernels will stop after the first task is pulled to minimize
2198 * the critical section.
2200 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2203 if (raw_spin_is_contended(&this_rq->lock) ||
2204 raw_spin_is_contended(&busiest->lock))
2207 } while (load_moved && max_load_move > total_load_moved);
2209 return total_load_moved > 0;
2212 /********** Helpers for find_busiest_group ************************/
2214 * sd_lb_stats - Structure to store the statistics of a sched_domain
2215 * during load balancing.
2217 struct sd_lb_stats {
2218 struct sched_group *busiest; /* Busiest group in this sd */
2219 struct sched_group *this; /* Local group in this sd */
2220 unsigned long total_load; /* Total load of all groups in sd */
2221 unsigned long total_pwr; /* Total power of all groups in sd */
2222 unsigned long avg_load; /* Average load across all groups in sd */
2224 /** Statistics of this group */
2225 unsigned long this_load;
2226 unsigned long this_load_per_task;
2227 unsigned long this_nr_running;
2228 unsigned long this_has_capacity;
2230 /* Statistics of the busiest group */
2231 unsigned long max_load;
2232 unsigned long busiest_load_per_task;
2233 unsigned long busiest_nr_running;
2234 unsigned long busiest_group_capacity;
2235 unsigned long busiest_has_capacity;
2237 int group_imb; /* Is there imbalance in this sd */
2238 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2239 int power_savings_balance; /* Is powersave balance needed for this sd */
2240 struct sched_group *group_min; /* Least loaded group in sd */
2241 struct sched_group *group_leader; /* Group which relieves group_min */
2242 unsigned long min_load_per_task; /* load_per_task in group_min */
2243 unsigned long leader_nr_running; /* Nr running of group_leader */
2244 unsigned long min_nr_running; /* Nr running of group_min */
2249 * sg_lb_stats - stats of a sched_group required for load_balancing
2251 struct sg_lb_stats {
2252 unsigned long avg_load; /*Avg load across the CPUs of the group */
2253 unsigned long group_load; /* Total load over the CPUs of the group */
2254 unsigned long sum_nr_running; /* Nr tasks running in the group */
2255 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2256 unsigned long group_capacity;
2257 int group_imb; /* Is there an imbalance in the group ? */
2258 int group_has_capacity; /* Is there extra capacity in the group? */
2262 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2263 * @group: The group whose first cpu is to be returned.
2265 static inline unsigned int group_first_cpu(struct sched_group *group)
2267 return cpumask_first(sched_group_cpus(group));
2271 * get_sd_load_idx - Obtain the load index for a given sched domain.
2272 * @sd: The sched_domain whose load_idx is to be obtained.
2273 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2275 static inline int get_sd_load_idx(struct sched_domain *sd,
2276 enum cpu_idle_type idle)
2282 load_idx = sd->busy_idx;
2285 case CPU_NEWLY_IDLE:
2286 load_idx = sd->newidle_idx;
2289 load_idx = sd->idle_idx;
2297 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2299 * init_sd_power_savings_stats - Initialize power savings statistics for
2300 * the given sched_domain, during load balancing.
2302 * @sd: Sched domain whose power-savings statistics are to be initialized.
2303 * @sds: Variable containing the statistics for sd.
2304 * @idle: Idle status of the CPU at which we're performing load-balancing.
2306 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2307 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2310 * Busy processors will not participate in power savings
2313 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2314 sds->power_savings_balance = 0;
2316 sds->power_savings_balance = 1;
2317 sds->min_nr_running = ULONG_MAX;
2318 sds->leader_nr_running = 0;
2323 * update_sd_power_savings_stats - Update the power saving stats for a
2324 * sched_domain while performing load balancing.
2326 * @group: sched_group belonging to the sched_domain under consideration.
2327 * @sds: Variable containing the statistics of the sched_domain
2328 * @local_group: Does group contain the CPU for which we're performing
2330 * @sgs: Variable containing the statistics of the group.
2332 static inline void update_sd_power_savings_stats(struct sched_group *group,
2333 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2336 if (!sds->power_savings_balance)
2340 * If the local group is idle or completely loaded
2341 * no need to do power savings balance at this domain
2343 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2344 !sds->this_nr_running))
2345 sds->power_savings_balance = 0;
2348 * If a group is already running at full capacity or idle,
2349 * don't include that group in power savings calculations
2351 if (!sds->power_savings_balance ||
2352 sgs->sum_nr_running >= sgs->group_capacity ||
2353 !sgs->sum_nr_running)
2357 * Calculate the group which has the least non-idle load.
2358 * This is the group from where we need to pick up the load
2361 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2362 (sgs->sum_nr_running == sds->min_nr_running &&
2363 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2364 sds->group_min = group;
2365 sds->min_nr_running = sgs->sum_nr_running;
2366 sds->min_load_per_task = sgs->sum_weighted_load /
2367 sgs->sum_nr_running;
2371 * Calculate the group which is almost near its
2372 * capacity but still has some space to pick up some load
2373 * from other group and save more power
2375 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2378 if (sgs->sum_nr_running > sds->leader_nr_running ||
2379 (sgs->sum_nr_running == sds->leader_nr_running &&
2380 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2381 sds->group_leader = group;
2382 sds->leader_nr_running = sgs->sum_nr_running;
2387 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2388 * @sds: Variable containing the statistics of the sched_domain
2389 * under consideration.
2390 * @this_cpu: Cpu at which we're currently performing load-balancing.
2391 * @imbalance: Variable to store the imbalance.
2394 * Check if we have potential to perform some power-savings balance.
2395 * If yes, set the busiest group to be the least loaded group in the
2396 * sched_domain, so that it's CPUs can be put to idle.
2398 * Returns 1 if there is potential to perform power-savings balance.
2401 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2402 int this_cpu, unsigned long *imbalance)
2404 if (!sds->power_savings_balance)
2407 if (sds->this != sds->group_leader ||
2408 sds->group_leader == sds->group_min)
2411 *imbalance = sds->min_load_per_task;
2412 sds->busiest = sds->group_min;
2417 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2418 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2419 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2424 static inline void update_sd_power_savings_stats(struct sched_group *group,
2425 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2430 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2431 int this_cpu, unsigned long *imbalance)
2435 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2438 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2440 return SCHED_LOAD_SCALE;
2443 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2445 return default_scale_freq_power(sd, cpu);
2448 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2450 unsigned long weight = sd->span_weight;
2451 unsigned long smt_gain = sd->smt_gain;
2458 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2460 return default_scale_smt_power(sd, cpu);
2463 unsigned long scale_rt_power(int cpu)
2465 struct rq *rq = cpu_rq(cpu);
2466 u64 total, available;
2468 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2470 if (unlikely(total < rq->rt_avg)) {
2471 /* Ensures that power won't end up being negative */
2474 available = total - rq->rt_avg;
2477 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2478 total = SCHED_LOAD_SCALE;
2480 total >>= SCHED_LOAD_SHIFT;
2482 return div_u64(available, total);
2485 static void update_cpu_power(struct sched_domain *sd, int cpu)
2487 unsigned long weight = sd->span_weight;
2488 unsigned long power = SCHED_LOAD_SCALE;
2489 struct sched_group *sdg = sd->groups;
2491 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2492 if (sched_feat(ARCH_POWER))
2493 power *= arch_scale_smt_power(sd, cpu);
2495 power *= default_scale_smt_power(sd, cpu);
2497 power >>= SCHED_LOAD_SHIFT;
2500 sdg->cpu_power_orig = power;
2502 if (sched_feat(ARCH_POWER))
2503 power *= arch_scale_freq_power(sd, cpu);
2505 power *= default_scale_freq_power(sd, cpu);
2507 power >>= SCHED_LOAD_SHIFT;
2509 power *= scale_rt_power(cpu);
2510 power >>= SCHED_LOAD_SHIFT;
2515 cpu_rq(cpu)->cpu_power = power;
2516 sdg->cpu_power = power;
2519 static void update_group_power(struct sched_domain *sd, int cpu)
2521 struct sched_domain *child = sd->child;
2522 struct sched_group *group, *sdg = sd->groups;
2523 unsigned long power;
2526 update_cpu_power(sd, cpu);
2532 group = child->groups;
2534 power += group->cpu_power;
2535 group = group->next;
2536 } while (group != child->groups);
2538 sdg->cpu_power = power;
2542 * Try and fix up capacity for tiny siblings, this is needed when
2543 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2544 * which on its own isn't powerful enough.
2546 * See update_sd_pick_busiest() and check_asym_packing().
2549 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2552 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2554 if (sd->level != SD_LV_SIBLING)
2558 * If ~90% of the cpu_power is still there, we're good.
2560 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2567 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2568 * @sd: The sched_domain whose statistics are to be updated.
2569 * @group: sched_group whose statistics are to be updated.
2570 * @this_cpu: Cpu for which load balance is currently performed.
2571 * @idle: Idle status of this_cpu
2572 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2573 * @sd_idle: Idle status of the sched_domain containing group.
2574 * @local_group: Does group contain this_cpu.
2575 * @cpus: Set of cpus considered for load balancing.
2576 * @balance: Should we balance.
2577 * @sgs: variable to hold the statistics for this group.
2579 static inline void update_sg_lb_stats(struct sched_domain *sd,
2580 struct sched_group *group, int this_cpu,
2581 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2582 int local_group, const struct cpumask *cpus,
2583 int *balance, struct sg_lb_stats *sgs)
2585 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2587 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2588 unsigned long avg_load_per_task = 0;
2591 balance_cpu = group_first_cpu(group);
2593 /* Tally up the load of all CPUs in the group */
2595 min_cpu_load = ~0UL;
2598 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2599 struct rq *rq = cpu_rq(i);
2601 if (*sd_idle && rq->nr_running)
2604 /* Bias balancing toward cpus of our domain */
2606 if (idle_cpu(i) && !first_idle_cpu) {
2611 load = target_load(i, load_idx);
2613 load = source_load(i, load_idx);
2614 if (load > max_cpu_load) {
2615 max_cpu_load = load;
2616 max_nr_running = rq->nr_running;
2618 if (min_cpu_load > load)
2619 min_cpu_load = load;
2622 sgs->group_load += load;
2623 sgs->sum_nr_running += rq->nr_running;
2624 sgs->sum_weighted_load += weighted_cpuload(i);
2629 * First idle cpu or the first cpu(busiest) in this sched group
2630 * is eligible for doing load balancing at this and above
2631 * domains. In the newly idle case, we will allow all the cpu's
2632 * to do the newly idle load balance.
2634 if (idle != CPU_NEWLY_IDLE && local_group) {
2635 if (balance_cpu != this_cpu) {
2639 update_group_power(sd, this_cpu);
2642 /* Adjust by relative CPU power of the group */
2643 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2646 * Consider the group unbalanced when the imbalance is larger
2647 * than the average weight of two tasks.
2649 * APZ: with cgroup the avg task weight can vary wildly and
2650 * might not be a suitable number - should we keep a
2651 * normalized nr_running number somewhere that negates
2654 if (sgs->sum_nr_running)
2655 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2657 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2660 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2661 if (!sgs->group_capacity)
2662 sgs->group_capacity = fix_small_capacity(sd, group);
2664 if (sgs->group_capacity > sgs->sum_nr_running)
2665 sgs->group_has_capacity = 1;
2669 * update_sd_pick_busiest - return 1 on busiest group
2670 * @sd: sched_domain whose statistics are to be checked
2671 * @sds: sched_domain statistics
2672 * @sg: sched_group candidate to be checked for being the busiest
2673 * @sgs: sched_group statistics
2674 * @this_cpu: the current cpu
2676 * Determine if @sg is a busier group than the previously selected
2679 static bool update_sd_pick_busiest(struct sched_domain *sd,
2680 struct sd_lb_stats *sds,
2681 struct sched_group *sg,
2682 struct sg_lb_stats *sgs,
2685 if (sgs->avg_load <= sds->max_load)
2688 if (sgs->sum_nr_running > sgs->group_capacity)
2695 * ASYM_PACKING needs to move all the work to the lowest
2696 * numbered CPUs in the group, therefore mark all groups
2697 * higher than ourself as busy.
2699 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2700 this_cpu < group_first_cpu(sg)) {
2704 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2712 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2713 * @sd: sched_domain whose statistics are to be updated.
2714 * @this_cpu: Cpu for which load balance is currently performed.
2715 * @idle: Idle status of this_cpu
2716 * @sd_idle: Idle status of the sched_domain containing sg.
2717 * @cpus: Set of cpus considered for load balancing.
2718 * @balance: Should we balance.
2719 * @sds: variable to hold the statistics for this sched_domain.
2721 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2722 enum cpu_idle_type idle, int *sd_idle,
2723 const struct cpumask *cpus, int *balance,
2724 struct sd_lb_stats *sds)
2726 struct sched_domain *child = sd->child;
2727 struct sched_group *sg = sd->groups;
2728 struct sg_lb_stats sgs;
2729 int load_idx, prefer_sibling = 0;
2731 if (child && child->flags & SD_PREFER_SIBLING)
2734 init_sd_power_savings_stats(sd, sds, idle);
2735 load_idx = get_sd_load_idx(sd, idle);
2740 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2741 memset(&sgs, 0, sizeof(sgs));
2742 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2743 local_group, cpus, balance, &sgs);
2745 if (local_group && !(*balance))
2748 sds->total_load += sgs.group_load;
2749 sds->total_pwr += sg->cpu_power;
2752 * In case the child domain prefers tasks go to siblings
2753 * first, lower the sg capacity to one so that we'll try
2754 * and move all the excess tasks away. We lower the capacity
2755 * of a group only if the local group has the capacity to fit
2756 * these excess tasks, i.e. nr_running < group_capacity. The
2757 * extra check prevents the case where you always pull from the
2758 * heaviest group when it is already under-utilized (possible
2759 * with a large weight task outweighs the tasks on the system).
2761 if (prefer_sibling && !local_group && sds->this_has_capacity)
2762 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2765 sds->this_load = sgs.avg_load;
2767 sds->this_nr_running = sgs.sum_nr_running;
2768 sds->this_load_per_task = sgs.sum_weighted_load;
2769 sds->this_has_capacity = sgs.group_has_capacity;
2770 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2771 sds->max_load = sgs.avg_load;
2773 sds->busiest_nr_running = sgs.sum_nr_running;
2774 sds->busiest_group_capacity = sgs.group_capacity;
2775 sds->busiest_load_per_task = sgs.sum_weighted_load;
2776 sds->busiest_has_capacity = sgs.group_has_capacity;
2777 sds->group_imb = sgs.group_imb;
2780 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2782 } while (sg != sd->groups);
2785 int __weak arch_sd_sibling_asym_packing(void)
2787 return 0*SD_ASYM_PACKING;
2791 * check_asym_packing - Check to see if the group is packed into the
2794 * This is primarily intended to used at the sibling level. Some
2795 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2796 * case of POWER7, it can move to lower SMT modes only when higher
2797 * threads are idle. When in lower SMT modes, the threads will
2798 * perform better since they share less core resources. Hence when we
2799 * have idle threads, we want them to be the higher ones.
2801 * This packing function is run on idle threads. It checks to see if
2802 * the busiest CPU in this domain (core in the P7 case) has a higher
2803 * CPU number than the packing function is being run on. Here we are
2804 * assuming lower CPU number will be equivalent to lower a SMT thread
2807 * Returns 1 when packing is required and a task should be moved to
2808 * this CPU. The amount of the imbalance is returned in *imbalance.
2810 * @sd: The sched_domain whose packing is to be checked.
2811 * @sds: Statistics of the sched_domain which is to be packed
2812 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2813 * @imbalance: returns amount of imbalanced due to packing.
2815 static int check_asym_packing(struct sched_domain *sd,
2816 struct sd_lb_stats *sds,
2817 int this_cpu, unsigned long *imbalance)
2821 if (!(sd->flags & SD_ASYM_PACKING))
2827 busiest_cpu = group_first_cpu(sds->busiest);
2828 if (this_cpu > busiest_cpu)
2831 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2837 * fix_small_imbalance - Calculate the minor imbalance that exists
2838 * amongst the groups of a sched_domain, during
2840 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2841 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2842 * @imbalance: Variable to store the imbalance.
2844 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2845 int this_cpu, unsigned long *imbalance)
2847 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2848 unsigned int imbn = 2;
2849 unsigned long scaled_busy_load_per_task;
2851 if (sds->this_nr_running) {
2852 sds->this_load_per_task /= sds->this_nr_running;
2853 if (sds->busiest_load_per_task >
2854 sds->this_load_per_task)
2857 sds->this_load_per_task =
2858 cpu_avg_load_per_task(this_cpu);
2860 scaled_busy_load_per_task = sds->busiest_load_per_task
2862 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2864 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2865 (scaled_busy_load_per_task * imbn)) {
2866 *imbalance = sds->busiest_load_per_task;
2871 * OK, we don't have enough imbalance to justify moving tasks,
2872 * however we may be able to increase total CPU power used by
2876 pwr_now += sds->busiest->cpu_power *
2877 min(sds->busiest_load_per_task, sds->max_load);
2878 pwr_now += sds->this->cpu_power *
2879 min(sds->this_load_per_task, sds->this_load);
2880 pwr_now /= SCHED_LOAD_SCALE;
2882 /* Amount of load we'd subtract */
2883 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2884 sds->busiest->cpu_power;
2885 if (sds->max_load > tmp)
2886 pwr_move += sds->busiest->cpu_power *
2887 min(sds->busiest_load_per_task, sds->max_load - tmp);
2889 /* Amount of load we'd add */
2890 if (sds->max_load * sds->busiest->cpu_power <
2891 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2892 tmp = (sds->max_load * sds->busiest->cpu_power) /
2893 sds->this->cpu_power;
2895 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2896 sds->this->cpu_power;
2897 pwr_move += sds->this->cpu_power *
2898 min(sds->this_load_per_task, sds->this_load + tmp);
2899 pwr_move /= SCHED_LOAD_SCALE;
2901 /* Move if we gain throughput */
2902 if (pwr_move > pwr_now)
2903 *imbalance = sds->busiest_load_per_task;
2907 * calculate_imbalance - Calculate the amount of imbalance present within the
2908 * groups of a given sched_domain during load balance.
2909 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2910 * @this_cpu: Cpu for which currently load balance is being performed.
2911 * @imbalance: The variable to store the imbalance.
2913 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2914 unsigned long *imbalance)
2916 unsigned long max_pull, load_above_capacity = ~0UL;
2918 sds->busiest_load_per_task /= sds->busiest_nr_running;
2919 if (sds->group_imb) {
2920 sds->busiest_load_per_task =
2921 min(sds->busiest_load_per_task, sds->avg_load);
2925 * In the presence of smp nice balancing, certain scenarios can have
2926 * max load less than avg load(as we skip the groups at or below
2927 * its cpu_power, while calculating max_load..)
2929 if (sds->max_load < sds->avg_load) {
2931 return fix_small_imbalance(sds, this_cpu, imbalance);
2934 if (!sds->group_imb) {
2936 * Don't want to pull so many tasks that a group would go idle.
2938 load_above_capacity = (sds->busiest_nr_running -
2939 sds->busiest_group_capacity);
2941 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2943 load_above_capacity /= sds->busiest->cpu_power;
2947 * We're trying to get all the cpus to the average_load, so we don't
2948 * want to push ourselves above the average load, nor do we wish to
2949 * reduce the max loaded cpu below the average load. At the same time,
2950 * we also don't want to reduce the group load below the group capacity
2951 * (so that we can implement power-savings policies etc). Thus we look
2952 * for the minimum possible imbalance.
2953 * Be careful of negative numbers as they'll appear as very large values
2954 * with unsigned longs.
2956 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2958 /* How much load to actually move to equalise the imbalance */
2959 *imbalance = min(max_pull * sds->busiest->cpu_power,
2960 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
2964 * if *imbalance is less than the average load per runnable task
2965 * there is no gaurantee that any tasks will be moved so we'll have
2966 * a think about bumping its value to force at least one task to be
2969 if (*imbalance < sds->busiest_load_per_task)
2970 return fix_small_imbalance(sds, this_cpu, imbalance);
2974 /******* find_busiest_group() helpers end here *********************/
2977 * find_busiest_group - Returns the busiest group within the sched_domain
2978 * if there is an imbalance. If there isn't an imbalance, and
2979 * the user has opted for power-savings, it returns a group whose
2980 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2981 * such a group exists.
2983 * Also calculates the amount of weighted load which should be moved
2984 * to restore balance.
2986 * @sd: The sched_domain whose busiest group is to be returned.
2987 * @this_cpu: The cpu for which load balancing is currently being performed.
2988 * @imbalance: Variable which stores amount of weighted load which should
2989 * be moved to restore balance/put a group to idle.
2990 * @idle: The idle status of this_cpu.
2991 * @sd_idle: The idleness of sd
2992 * @cpus: The set of CPUs under consideration for load-balancing.
2993 * @balance: Pointer to a variable indicating if this_cpu
2994 * is the appropriate cpu to perform load balancing at this_level.
2996 * Returns: - the busiest group if imbalance exists.
2997 * - If no imbalance and user has opted for power-savings balance,
2998 * return the least loaded group whose CPUs can be
2999 * put to idle by rebalancing its tasks onto our group.
3001 static struct sched_group *
3002 find_busiest_group(struct sched_domain *sd, int this_cpu,
3003 unsigned long *imbalance, enum cpu_idle_type idle,
3004 int *sd_idle, const struct cpumask *cpus, int *balance)
3006 struct sd_lb_stats sds;
3008 memset(&sds, 0, sizeof(sds));
3011 * Compute the various statistics relavent for load balancing at
3014 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3017 /* Cases where imbalance does not exist from POV of this_cpu */
3018 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3020 * 2) There is no busy sibling group to pull from.
3021 * 3) This group is the busiest group.
3022 * 4) This group is more busy than the avg busieness at this
3024 * 5) The imbalance is within the specified limit.
3026 * Note: when doing newidle balance, if the local group has excess
3027 * capacity (i.e. nr_running < group_capacity) and the busiest group
3028 * does not have any capacity, we force a load balance to pull tasks
3029 * to the local group. In this case, we skip past checks 3, 4 and 5.
3034 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3035 check_asym_packing(sd, &sds, this_cpu, imbalance))
3038 if (!sds.busiest || sds.busiest_nr_running == 0)
3041 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3042 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3043 !sds.busiest_has_capacity)
3046 if (sds.this_load >= sds.max_load)
3049 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3051 if (sds.this_load >= sds.avg_load)
3054 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3058 /* Looks like there is an imbalance. Compute it */
3059 calculate_imbalance(&sds, this_cpu, imbalance);
3064 * There is no obvious imbalance. But check if we can do some balancing
3067 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3075 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3078 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3079 enum cpu_idle_type idle, unsigned long imbalance,
3080 const struct cpumask *cpus)
3082 struct rq *busiest = NULL, *rq;
3083 unsigned long max_load = 0;
3086 for_each_cpu(i, sched_group_cpus(group)) {
3087 unsigned long power = power_of(i);
3088 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3092 capacity = fix_small_capacity(sd, group);
3094 if (!cpumask_test_cpu(i, cpus))
3098 wl = weighted_cpuload(i);
3101 * When comparing with imbalance, use weighted_cpuload()
3102 * which is not scaled with the cpu power.
3104 if (capacity && rq->nr_running == 1 && wl > imbalance)
3108 * For the load comparisons with the other cpu's, consider
3109 * the weighted_cpuload() scaled with the cpu power, so that
3110 * the load can be moved away from the cpu that is potentially
3111 * running at a lower capacity.
3113 wl = (wl * SCHED_LOAD_SCALE) / power;
3115 if (wl > max_load) {
3125 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3126 * so long as it is large enough.
3128 #define MAX_PINNED_INTERVAL 512
3130 /* Working cpumask for load_balance and load_balance_newidle. */
3131 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3133 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
3134 int busiest_cpu, int this_cpu)
3136 if (idle == CPU_NEWLY_IDLE) {
3139 * ASYM_PACKING needs to force migrate tasks from busy but
3140 * higher numbered CPUs in order to pack all tasks in the
3141 * lowest numbered CPUs.
3143 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3147 * The only task running in a non-idle cpu can be moved to this
3148 * cpu in an attempt to completely freeup the other CPU
3151 * The package power saving logic comes from
3152 * find_busiest_group(). If there are no imbalance, then
3153 * f_b_g() will return NULL. However when sched_mc={1,2} then
3154 * f_b_g() will select a group from which a running task may be
3155 * pulled to this cpu in order to make the other package idle.
3156 * If there is no opportunity to make a package idle and if
3157 * there are no imbalance, then f_b_g() will return NULL and no
3158 * action will be taken in load_balance_newidle().
3160 * Under normal task pull operation due to imbalance, there
3161 * will be more than one task in the source run queue and
3162 * move_tasks() will succeed. ld_moved will be true and this
3163 * active balance code will not be triggered.
3165 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3166 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3169 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3173 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3176 static int active_load_balance_cpu_stop(void *data);
3179 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3180 * tasks if there is an imbalance.
3182 static int load_balance(int this_cpu, struct rq *this_rq,
3183 struct sched_domain *sd, enum cpu_idle_type idle,
3186 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3187 struct sched_group *group;
3188 unsigned long imbalance;
3190 unsigned long flags;
3191 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3193 cpumask_copy(cpus, cpu_active_mask);
3196 * When power savings policy is enabled for the parent domain, idle
3197 * sibling can pick up load irrespective of busy siblings. In this case,
3198 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3199 * portraying it as CPU_NOT_IDLE.
3201 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3202 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3205 schedstat_inc(sd, lb_count[idle]);
3208 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3215 schedstat_inc(sd, lb_nobusyg[idle]);
3219 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3221 schedstat_inc(sd, lb_nobusyq[idle]);
3225 BUG_ON(busiest == this_rq);
3227 schedstat_add(sd, lb_imbalance[idle], imbalance);
3230 if (busiest->nr_running > 1) {
3232 * Attempt to move tasks. If find_busiest_group has found
3233 * an imbalance but busiest->nr_running <= 1, the group is
3234 * still unbalanced. ld_moved simply stays zero, so it is
3235 * correctly treated as an imbalance.
3237 local_irq_save(flags);
3238 double_rq_lock(this_rq, busiest);
3239 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3240 imbalance, sd, idle, &all_pinned);
3241 double_rq_unlock(this_rq, busiest);
3242 local_irq_restore(flags);
3245 * some other cpu did the load balance for us.
3247 if (ld_moved && this_cpu != smp_processor_id())
3248 resched_cpu(this_cpu);
3250 /* All tasks on this runqueue were pinned by CPU affinity */
3251 if (unlikely(all_pinned)) {
3252 cpumask_clear_cpu(cpu_of(busiest), cpus);
3253 if (!cpumask_empty(cpus))
3260 schedstat_inc(sd, lb_failed[idle]);
3262 * Increment the failure counter only on periodic balance.
3263 * We do not want newidle balance, which can be very
3264 * frequent, pollute the failure counter causing
3265 * excessive cache_hot migrations and active balances.
3267 if (idle != CPU_NEWLY_IDLE)
3268 sd->nr_balance_failed++;
3270 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3272 raw_spin_lock_irqsave(&busiest->lock, flags);
3274 /* don't kick the active_load_balance_cpu_stop,
3275 * if the curr task on busiest cpu can't be
3278 if (!cpumask_test_cpu(this_cpu,
3279 &busiest->curr->cpus_allowed)) {
3280 raw_spin_unlock_irqrestore(&busiest->lock,
3283 goto out_one_pinned;
3287 * ->active_balance synchronizes accesses to
3288 * ->active_balance_work. Once set, it's cleared
3289 * only after active load balance is finished.
3291 if (!busiest->active_balance) {
3292 busiest->active_balance = 1;
3293 busiest->push_cpu = this_cpu;
3296 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3299 stop_one_cpu_nowait(cpu_of(busiest),
3300 active_load_balance_cpu_stop, busiest,
3301 &busiest->active_balance_work);
3304 * We've kicked active balancing, reset the failure
3307 sd->nr_balance_failed = sd->cache_nice_tries+1;
3310 sd->nr_balance_failed = 0;
3312 if (likely(!active_balance)) {
3313 /* We were unbalanced, so reset the balancing interval */
3314 sd->balance_interval = sd->min_interval;
3317 * If we've begun active balancing, start to back off. This
3318 * case may not be covered by the all_pinned logic if there
3319 * is only 1 task on the busy runqueue (because we don't call
3322 if (sd->balance_interval < sd->max_interval)
3323 sd->balance_interval *= 2;
3326 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3327 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3333 schedstat_inc(sd, lb_balanced[idle]);
3335 sd->nr_balance_failed = 0;
3338 /* tune up the balancing interval */
3339 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3340 (sd->balance_interval < sd->max_interval))
3341 sd->balance_interval *= 2;
3343 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3344 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3353 * idle_balance is called by schedule() if this_cpu is about to become
3354 * idle. Attempts to pull tasks from other CPUs.
3356 static void idle_balance(int this_cpu, struct rq *this_rq)
3358 struct sched_domain *sd;
3359 int pulled_task = 0;
3360 unsigned long next_balance = jiffies + HZ;
3362 this_rq->idle_stamp = this_rq->clock;
3364 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3368 * Drop the rq->lock, but keep IRQ/preempt disabled.
3370 raw_spin_unlock(&this_rq->lock);
3372 update_shares(this_cpu);
3373 for_each_domain(this_cpu, sd) {
3374 unsigned long interval;
3377 if (!(sd->flags & SD_LOAD_BALANCE))
3380 if (sd->flags & SD_BALANCE_NEWIDLE) {
3381 /* If we've pulled tasks over stop searching: */
3382 pulled_task = load_balance(this_cpu, this_rq,
3383 sd, CPU_NEWLY_IDLE, &balance);
3386 interval = msecs_to_jiffies(sd->balance_interval);
3387 if (time_after(next_balance, sd->last_balance + interval))
3388 next_balance = sd->last_balance + interval;
3393 raw_spin_lock(&this_rq->lock);
3395 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3397 * We are going idle. next_balance may be set based on
3398 * a busy processor. So reset next_balance.
3400 this_rq->next_balance = next_balance;
3405 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3406 * running tasks off the busiest CPU onto idle CPUs. It requires at
3407 * least 1 task to be running on each physical CPU where possible, and
3408 * avoids physical / logical imbalances.
3410 static int active_load_balance_cpu_stop(void *data)
3412 struct rq *busiest_rq = data;
3413 int busiest_cpu = cpu_of(busiest_rq);
3414 int target_cpu = busiest_rq->push_cpu;
3415 struct rq *target_rq = cpu_rq(target_cpu);
3416 struct sched_domain *sd;
3418 raw_spin_lock_irq(&busiest_rq->lock);
3420 /* make sure the requested cpu hasn't gone down in the meantime */
3421 if (unlikely(busiest_cpu != smp_processor_id() ||
3422 !busiest_rq->active_balance))
3425 /* Is there any task to move? */
3426 if (busiest_rq->nr_running <= 1)
3430 * This condition is "impossible", if it occurs
3431 * we need to fix it. Originally reported by
3432 * Bjorn Helgaas on a 128-cpu setup.
3434 BUG_ON(busiest_rq == target_rq);
3436 /* move a task from busiest_rq to target_rq */
3437 double_lock_balance(busiest_rq, target_rq);
3439 /* Search for an sd spanning us and the target CPU. */
3440 for_each_domain(target_cpu, sd) {
3441 if ((sd->flags & SD_LOAD_BALANCE) &&
3442 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3447 schedstat_inc(sd, alb_count);
3449 if (move_one_task(target_rq, target_cpu, busiest_rq,
3451 schedstat_inc(sd, alb_pushed);
3453 schedstat_inc(sd, alb_failed);
3455 double_unlock_balance(busiest_rq, target_rq);
3457 busiest_rq->active_balance = 0;
3458 raw_spin_unlock_irq(&busiest_rq->lock);
3464 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3466 static void trigger_sched_softirq(void *data)
3468 raise_softirq_irqoff(SCHED_SOFTIRQ);
3471 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3473 csd->func = trigger_sched_softirq;
3480 * idle load balancing details
3481 * - One of the idle CPUs nominates itself as idle load_balancer, while
3483 * - This idle load balancer CPU will also go into tickless mode when
3484 * it is idle, just like all other idle CPUs
3485 * - When one of the busy CPUs notice that there may be an idle rebalancing
3486 * needed, they will kick the idle load balancer, which then does idle
3487 * load balancing for all the idle CPUs.
3490 atomic_t load_balancer;
3491 atomic_t first_pick_cpu;
3492 atomic_t second_pick_cpu;
3493 cpumask_var_t idle_cpus_mask;
3494 cpumask_var_t grp_idle_mask;
3495 unsigned long next_balance; /* in jiffy units */
3496 } nohz ____cacheline_aligned;
3498 int get_nohz_load_balancer(void)
3500 return atomic_read(&nohz.load_balancer);
3503 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3505 * lowest_flag_domain - Return lowest sched_domain containing flag.
3506 * @cpu: The cpu whose lowest level of sched domain is to
3508 * @flag: The flag to check for the lowest sched_domain
3509 * for the given cpu.
3511 * Returns the lowest sched_domain of a cpu which contains the given flag.
3513 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3515 struct sched_domain *sd;
3517 for_each_domain(cpu, sd)
3518 if (sd && (sd->flags & flag))
3525 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3526 * @cpu: The cpu whose domains we're iterating over.
3527 * @sd: variable holding the value of the power_savings_sd
3529 * @flag: The flag to filter the sched_domains to be iterated.
3531 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3532 * set, starting from the lowest sched_domain to the highest.
3534 #define for_each_flag_domain(cpu, sd, flag) \
3535 for (sd = lowest_flag_domain(cpu, flag); \
3536 (sd && (sd->flags & flag)); sd = sd->parent)
3539 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3540 * @ilb_group: group to be checked for semi-idleness
3542 * Returns: 1 if the group is semi-idle. 0 otherwise.
3544 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3545 * and atleast one non-idle CPU. This helper function checks if the given
3546 * sched_group is semi-idle or not.
3548 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3550 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3551 sched_group_cpus(ilb_group));
3554 * A sched_group is semi-idle when it has atleast one busy cpu
3555 * and atleast one idle cpu.
3557 if (cpumask_empty(nohz.grp_idle_mask))
3560 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3566 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3567 * @cpu: The cpu which is nominating a new idle_load_balancer.
3569 * Returns: Returns the id of the idle load balancer if it exists,
3570 * Else, returns >= nr_cpu_ids.
3572 * This algorithm picks the idle load balancer such that it belongs to a
3573 * semi-idle powersavings sched_domain. The idea is to try and avoid
3574 * completely idle packages/cores just for the purpose of idle load balancing
3575 * when there are other idle cpu's which are better suited for that job.
3577 static int find_new_ilb(int cpu)
3579 struct sched_domain *sd;
3580 struct sched_group *ilb_group;
3583 * Have idle load balancer selection from semi-idle packages only
3584 * when power-aware load balancing is enabled
3586 if (!(sched_smt_power_savings || sched_mc_power_savings))
3590 * Optimize for the case when we have no idle CPUs or only one
3591 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3593 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3596 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3597 ilb_group = sd->groups;
3600 if (is_semi_idle_group(ilb_group))
3601 return cpumask_first(nohz.grp_idle_mask);
3603 ilb_group = ilb_group->next;
3605 } while (ilb_group != sd->groups);
3611 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3612 static inline int find_new_ilb(int call_cpu)
3619 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3620 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3621 * CPU (if there is one).
3623 static void nohz_balancer_kick(int cpu)
3627 nohz.next_balance++;
3629 ilb_cpu = get_nohz_load_balancer();
3631 if (ilb_cpu >= nr_cpu_ids) {
3632 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3633 if (ilb_cpu >= nr_cpu_ids)
3637 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3638 struct call_single_data *cp;
3640 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3641 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3642 __smp_call_function_single(ilb_cpu, cp, 0);
3648 * This routine will try to nominate the ilb (idle load balancing)
3649 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3650 * load balancing on behalf of all those cpus.
3652 * When the ilb owner becomes busy, we will not have new ilb owner until some
3653 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3654 * idle load balancing by kicking one of the idle CPUs.
3656 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3657 * ilb owner CPU in future (when there is a need for idle load balancing on
3658 * behalf of all idle CPUs).
3660 void select_nohz_load_balancer(int stop_tick)
3662 int cpu = smp_processor_id();
3665 if (!cpu_active(cpu)) {
3666 if (atomic_read(&nohz.load_balancer) != cpu)
3670 * If we are going offline and still the leader,
3673 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3680 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3682 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3683 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3684 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3685 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3687 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3690 /* make me the ilb owner */
3691 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3696 * Check to see if there is a more power-efficient
3699 new_ilb = find_new_ilb(cpu);
3700 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3701 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3702 resched_cpu(new_ilb);
3708 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3711 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3713 if (atomic_read(&nohz.load_balancer) == cpu)
3714 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3722 static DEFINE_SPINLOCK(balancing);
3725 * It checks each scheduling domain to see if it is due to be balanced,
3726 * and initiates a balancing operation if so.
3728 * Balancing parameters are set up in arch_init_sched_domains.
3730 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3733 struct rq *rq = cpu_rq(cpu);
3734 unsigned long interval;
3735 struct sched_domain *sd;
3736 /* Earliest time when we have to do rebalance again */
3737 unsigned long next_balance = jiffies + 60*HZ;
3738 int update_next_balance = 0;
3743 for_each_domain(cpu, sd) {
3744 if (!(sd->flags & SD_LOAD_BALANCE))
3747 interval = sd->balance_interval;
3748 if (idle != CPU_IDLE)
3749 interval *= sd->busy_factor;
3751 /* scale ms to jiffies */
3752 interval = msecs_to_jiffies(interval);
3753 if (unlikely(!interval))
3755 if (interval > HZ*NR_CPUS/10)
3756 interval = HZ*NR_CPUS/10;
3758 need_serialize = sd->flags & SD_SERIALIZE;
3760 if (need_serialize) {
3761 if (!spin_trylock(&balancing))
3765 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3766 if (load_balance(cpu, rq, sd, idle, &balance)) {
3768 * We've pulled tasks over so either we're no
3769 * longer idle, or one of our SMT siblings is
3772 idle = CPU_NOT_IDLE;
3774 sd->last_balance = jiffies;
3777 spin_unlock(&balancing);
3779 if (time_after(next_balance, sd->last_balance + interval)) {
3780 next_balance = sd->last_balance + interval;
3781 update_next_balance = 1;
3785 * Stop the load balance at this level. There is another
3786 * CPU in our sched group which is doing load balancing more
3794 * next_balance will be updated only when there is a need.
3795 * When the cpu is attached to null domain for ex, it will not be
3798 if (likely(update_next_balance))
3799 rq->next_balance = next_balance;
3804 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3805 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3807 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3809 struct rq *this_rq = cpu_rq(this_cpu);
3813 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3816 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3817 if (balance_cpu == this_cpu)
3821 * If this cpu gets work to do, stop the load balancing
3822 * work being done for other cpus. Next load
3823 * balancing owner will pick it up.
3825 if (need_resched()) {
3826 this_rq->nohz_balance_kick = 0;
3830 raw_spin_lock_irq(&this_rq->lock);
3831 update_rq_clock(this_rq);
3832 update_cpu_load(this_rq);
3833 raw_spin_unlock_irq(&this_rq->lock);
3835 rebalance_domains(balance_cpu, CPU_IDLE);
3837 rq = cpu_rq(balance_cpu);
3838 if (time_after(this_rq->next_balance, rq->next_balance))
3839 this_rq->next_balance = rq->next_balance;
3841 nohz.next_balance = this_rq->next_balance;
3842 this_rq->nohz_balance_kick = 0;
3846 * Current heuristic for kicking the idle load balancer
3847 * - first_pick_cpu is the one of the busy CPUs. It will kick
3848 * idle load balancer when it has more than one process active. This
3849 * eliminates the need for idle load balancing altogether when we have
3850 * only one running process in the system (common case).
3851 * - If there are more than one busy CPU, idle load balancer may have
3852 * to run for active_load_balance to happen (i.e., two busy CPUs are
3853 * SMT or core siblings and can run better if they move to different
3854 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3855 * which will kick idle load balancer as soon as it has any load.
3857 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3859 unsigned long now = jiffies;
3861 int first_pick_cpu, second_pick_cpu;
3863 if (time_before(now, nohz.next_balance))
3866 if (rq->idle_at_tick)
3869 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3870 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3872 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3873 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3876 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3877 if (ret == nr_cpu_ids || ret == cpu) {
3878 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3879 if (rq->nr_running > 1)
3882 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3883 if (ret == nr_cpu_ids || ret == cpu) {
3891 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3895 * run_rebalance_domains is triggered when needed from the scheduler tick.
3896 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3898 static void run_rebalance_domains(struct softirq_action *h)
3900 int this_cpu = smp_processor_id();
3901 struct rq *this_rq = cpu_rq(this_cpu);
3902 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3903 CPU_IDLE : CPU_NOT_IDLE;
3905 rebalance_domains(this_cpu, idle);
3908 * If this cpu has a pending nohz_balance_kick, then do the
3909 * balancing on behalf of the other idle cpus whose ticks are
3912 nohz_idle_balance(this_cpu, idle);
3915 static inline int on_null_domain(int cpu)
3917 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3921 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3923 static inline void trigger_load_balance(struct rq *rq, int cpu)
3925 /* Don't need to rebalance while attached to NULL domain */
3926 if (time_after_eq(jiffies, rq->next_balance) &&
3927 likely(!on_null_domain(cpu)))
3928 raise_softirq(SCHED_SOFTIRQ);
3930 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
3931 nohz_balancer_kick(cpu);
3935 static void rq_online_fair(struct rq *rq)
3940 static void rq_offline_fair(struct rq *rq)
3945 #else /* CONFIG_SMP */
3948 * on UP we do not need to balance between CPUs:
3950 static inline void idle_balance(int cpu, struct rq *rq)
3954 #endif /* CONFIG_SMP */
3957 * scheduler tick hitting a task of our scheduling class:
3959 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
3961 struct cfs_rq *cfs_rq;
3962 struct sched_entity *se = &curr->se;
3964 for_each_sched_entity(se) {
3965 cfs_rq = cfs_rq_of(se);
3966 entity_tick(cfs_rq, se, queued);
3971 * called on fork with the child task as argument from the parent's context
3972 * - child not yet on the tasklist
3973 * - preemption disabled
3975 static void task_fork_fair(struct task_struct *p)
3977 struct cfs_rq *cfs_rq = task_cfs_rq(current);
3978 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
3979 int this_cpu = smp_processor_id();
3980 struct rq *rq = this_rq();
3981 unsigned long flags;
3983 raw_spin_lock_irqsave(&rq->lock, flags);
3985 update_rq_clock(rq);
3987 if (unlikely(task_cpu(p) != this_cpu)) {
3989 __set_task_cpu(p, this_cpu);
3993 update_curr(cfs_rq);
3996 se->vruntime = curr->vruntime;
3997 place_entity(cfs_rq, se, 1);
3999 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4001 * Upon rescheduling, sched_class::put_prev_task() will place
4002 * 'current' within the tree based on its new key value.
4004 swap(curr->vruntime, se->vruntime);
4005 resched_task(rq->curr);
4008 se->vruntime -= cfs_rq->min_vruntime;
4010 raw_spin_unlock_irqrestore(&rq->lock, flags);
4014 * Priority of the task has changed. Check to see if we preempt
4017 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
4018 int oldprio, int running)
4021 * Reschedule if we are currently running on this runqueue and
4022 * our priority decreased, or if we are not currently running on
4023 * this runqueue and our priority is higher than the current's
4026 if (p->prio > oldprio)
4027 resched_task(rq->curr);
4029 check_preempt_curr(rq, p, 0);
4033 * We switched to the sched_fair class.
4035 static void switched_to_fair(struct rq *rq, struct task_struct *p,
4039 * We were most likely switched from sched_rt, so
4040 * kick off the schedule if running, otherwise just see
4041 * if we can still preempt the current task.
4044 resched_task(rq->curr);
4046 check_preempt_curr(rq, p, 0);
4049 /* Account for a task changing its policy or group.
4051 * This routine is mostly called to set cfs_rq->curr field when a task
4052 * migrates between groups/classes.
4054 static void set_curr_task_fair(struct rq *rq)
4056 struct sched_entity *se = &rq->curr->se;
4058 for_each_sched_entity(se)
4059 set_next_entity(cfs_rq_of(se), se);
4062 #ifdef CONFIG_FAIR_GROUP_SCHED
4063 static void task_move_group_fair(struct task_struct *p, int on_rq)
4066 * If the task was not on the rq at the time of this cgroup movement
4067 * it must have been asleep, sleeping tasks keep their ->vruntime
4068 * absolute on their old rq until wakeup (needed for the fair sleeper
4069 * bonus in place_entity()).
4071 * If it was on the rq, we've just 'preempted' it, which does convert
4072 * ->vruntime to a relative base.
4074 * Make sure both cases convert their relative position when migrating
4075 * to another cgroup's rq. This does somewhat interfere with the
4076 * fair sleeper stuff for the first placement, but who cares.
4079 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4080 set_task_rq(p, task_cpu(p));
4082 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4086 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4088 struct sched_entity *se = &task->se;
4089 unsigned int rr_interval = 0;
4092 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4095 if (rq->cfs.load.weight)
4096 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4102 * All the scheduling class methods:
4104 static const struct sched_class fair_sched_class = {
4105 .next = &idle_sched_class,
4106 .enqueue_task = enqueue_task_fair,
4107 .dequeue_task = dequeue_task_fair,
4108 .yield_task = yield_task_fair,
4110 .check_preempt_curr = check_preempt_wakeup,
4112 .pick_next_task = pick_next_task_fair,
4113 .put_prev_task = put_prev_task_fair,
4116 .select_task_rq = select_task_rq_fair,
4118 .rq_online = rq_online_fair,
4119 .rq_offline = rq_offline_fair,
4121 .task_waking = task_waking_fair,
4124 .set_curr_task = set_curr_task_fair,
4125 .task_tick = task_tick_fair,
4126 .task_fork = task_fork_fair,
4128 .prio_changed = prio_changed_fair,
4129 .switched_to = switched_to_fair,
4131 .get_rr_interval = get_rr_interval_fair,
4133 #ifdef CONFIG_FAIR_GROUP_SCHED
4134 .task_move_group = task_move_group_fair,
4138 #ifdef CONFIG_SCHED_DEBUG
4139 static void print_cfs_stats(struct seq_file *m, int cpu)
4141 struct cfs_rq *cfs_rq;
4144 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4145 print_cfs_rq(m, cpu, cfs_rq);