2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
28 * Targeted preemption latency for CPU-bound tasks:
29 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
31 * NOTE: this latency value is not the same as the concept of
32 * 'timeslice length' - timeslices in CFS are of variable length
33 * and have no persistent notion like in traditional, time-slice
34 * based scheduling concepts.
36 * (to see the precise effective timeslice length of your workload,
37 * run vmstat and monitor the context-switches (cs) field)
39 unsigned int sysctl_sched_latency = 6000000ULL;
40 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
43 * The initial- and re-scaling of tunables is configurable
44 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
47 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
48 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
49 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 enum sched_tunable_scaling sysctl_sched_tunable_scaling
52 = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
56 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
58 unsigned int sysctl_sched_min_granularity = 750000ULL;
59 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
62 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
64 static unsigned int sched_nr_latency = 8;
67 * After fork, child runs first. If set to 0 (default) then
68 * parent will (try to) run first.
70 unsigned int sysctl_sched_child_runs_first __read_mostly;
73 * SCHED_OTHER wake-up granularity.
74 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
81 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
83 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
86 * The exponential sliding window over which load is averaged for shares
90 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
92 static const struct sched_class fair_sched_class;
94 /**************************************************************
95 * CFS operations on generic schedulable entities:
98 #ifdef CONFIG_FAIR_GROUP_SCHED
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
106 /* An entity is a task if it doesn't "own" a runqueue */
107 #define entity_is_task(se) (!se->my_q)
109 static inline struct task_struct *task_of(struct sched_entity *se)
111 #ifdef CONFIG_SCHED_DEBUG
112 WARN_ON_ONCE(!entity_is_task(se));
114 return container_of(se, struct task_struct, se);
117 /* Walk up scheduling entities hierarchy */
118 #define for_each_sched_entity(se) \
119 for (; se; se = se->parent)
121 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
126 /* runqueue on which this entity is (to be) queued */
127 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
132 /* runqueue "owned" by this group */
133 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
138 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
140 if (!cfs_rq->on_list) {
142 * Ensure we either appear before our parent (if already
143 * enqueued) or force our parent to appear after us when it is
144 * enqueued. The fact that we always enqueue bottom-up
145 * reduces this to two cases.
147 if (cfs_rq->tg->parent &&
148 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
149 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
150 &rq_of(cfs_rq)->leaf_cfs_rq_list);
152 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
153 &rq_of(cfs_rq)->leaf_cfs_rq_list);
160 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
162 if (cfs_rq->on_list) {
163 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
168 /* Iterate thr' all leaf cfs_rq's on a runqueue */
169 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
170 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
172 /* Do the two (enqueued) entities belong to the same group ? */
174 is_same_group(struct sched_entity *se, struct sched_entity *pse)
176 if (se->cfs_rq == pse->cfs_rq)
182 static inline struct sched_entity *parent_entity(struct sched_entity *se)
187 /* return depth at which a sched entity is present in the hierarchy */
188 static inline int depth_se(struct sched_entity *se)
192 for_each_sched_entity(se)
199 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
201 int se_depth, pse_depth;
204 * preemption test can be made between sibling entities who are in the
205 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
206 * both tasks until we find their ancestors who are siblings of common
210 /* First walk up until both entities are at same depth */
211 se_depth = depth_se(*se);
212 pse_depth = depth_se(*pse);
214 while (se_depth > pse_depth) {
216 *se = parent_entity(*se);
219 while (pse_depth > se_depth) {
221 *pse = parent_entity(*pse);
224 while (!is_same_group(*se, *pse)) {
225 *se = parent_entity(*se);
226 *pse = parent_entity(*pse);
230 #else /* !CONFIG_FAIR_GROUP_SCHED */
232 static inline struct task_struct *task_of(struct sched_entity *se)
234 return container_of(se, struct task_struct, se);
237 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
239 return container_of(cfs_rq, struct rq, cfs);
242 #define entity_is_task(se) 1
244 #define for_each_sched_entity(se) \
245 for (; se; se = NULL)
247 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
249 return &task_rq(p)->cfs;
252 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
254 struct task_struct *p = task_of(se);
255 struct rq *rq = task_rq(p);
260 /* runqueue "owned" by this group */
261 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
266 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
270 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
274 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
275 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
278 is_same_group(struct sched_entity *se, struct sched_entity *pse)
283 static inline struct sched_entity *parent_entity(struct sched_entity *se)
289 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
293 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 /**************************************************************
297 * Scheduling class tree data structure manipulation methods:
300 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
302 s64 delta = (s64)(vruntime - min_vruntime);
304 min_vruntime = vruntime;
309 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
311 s64 delta = (s64)(vruntime - min_vruntime);
313 min_vruntime = vruntime;
318 static inline int entity_before(struct sched_entity *a,
319 struct sched_entity *b)
321 return (s64)(a->vruntime - b->vruntime) < 0;
324 static void update_min_vruntime(struct cfs_rq *cfs_rq)
326 u64 vruntime = cfs_rq->min_vruntime;
329 vruntime = cfs_rq->curr->vruntime;
331 if (cfs_rq->rb_leftmost) {
332 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
337 vruntime = se->vruntime;
339 vruntime = min_vruntime(vruntime, se->vruntime);
342 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
345 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
350 * Enqueue an entity into the rb-tree:
352 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
354 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
355 struct rb_node *parent = NULL;
356 struct sched_entity *entry;
360 * Find the right place in the rbtree:
364 entry = rb_entry(parent, struct sched_entity, run_node);
366 * We dont care about collisions. Nodes with
367 * the same key stay together.
369 if (entity_before(se, entry)) {
370 link = &parent->rb_left;
372 link = &parent->rb_right;
378 * Maintain a cache of leftmost tree entries (it is frequently
382 cfs_rq->rb_leftmost = &se->run_node;
384 rb_link_node(&se->run_node, parent, link);
385 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
388 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
390 if (cfs_rq->rb_leftmost == &se->run_node) {
391 struct rb_node *next_node;
393 next_node = rb_next(&se->run_node);
394 cfs_rq->rb_leftmost = next_node;
397 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
400 static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
402 struct rb_node *left = cfs_rq->rb_leftmost;
407 return rb_entry(left, struct sched_entity, run_node);
410 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
412 struct rb_node *next = rb_next(&se->run_node);
417 return rb_entry(next, struct sched_entity, run_node);
420 #ifdef CONFIG_SCHED_DEBUG
421 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
423 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
428 return rb_entry(last, struct sched_entity, run_node);
431 /**************************************************************
432 * Scheduling class statistics methods:
435 int sched_proc_update_handler(struct ctl_table *table, int write,
436 void __user *buffer, size_t *lenp,
439 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
440 int factor = get_update_sysctl_factor();
445 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
446 sysctl_sched_min_granularity);
448 #define WRT_SYSCTL(name) \
449 (normalized_sysctl_##name = sysctl_##name / (factor))
450 WRT_SYSCTL(sched_min_granularity);
451 WRT_SYSCTL(sched_latency);
452 WRT_SYSCTL(sched_wakeup_granularity);
462 static inline unsigned long
463 calc_delta_fair(unsigned long delta, struct sched_entity *se)
465 if (unlikely(se->load.weight != NICE_0_LOAD))
466 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
472 * The idea is to set a period in which each task runs once.
474 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
475 * this period because otherwise the slices get too small.
477 * p = (nr <= nl) ? l : l*nr/nl
479 static u64 __sched_period(unsigned long nr_running)
481 u64 period = sysctl_sched_latency;
482 unsigned long nr_latency = sched_nr_latency;
484 if (unlikely(nr_running > nr_latency)) {
485 period = sysctl_sched_min_granularity;
486 period *= nr_running;
493 * We calculate the wall-time slice from the period by taking a part
494 * proportional to the weight.
498 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
500 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
502 for_each_sched_entity(se) {
503 struct load_weight *load;
504 struct load_weight lw;
506 cfs_rq = cfs_rq_of(se);
507 load = &cfs_rq->load;
509 if (unlikely(!se->on_rq)) {
512 update_load_add(&lw, se->load.weight);
515 slice = calc_delta_mine(slice, se->load.weight, load);
521 * We calculate the vruntime slice of a to be inserted task
525 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
527 return calc_delta_fair(sched_slice(cfs_rq, se), se);
530 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
531 static void update_cfs_shares(struct cfs_rq *cfs_rq);
534 * Update the current task's runtime statistics. Skip current tasks that
535 * are not in our scheduling class.
538 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
539 unsigned long delta_exec)
541 unsigned long delta_exec_weighted;
543 schedstat_set(curr->statistics.exec_max,
544 max((u64)delta_exec, curr->statistics.exec_max));
546 curr->sum_exec_runtime += delta_exec;
547 schedstat_add(cfs_rq, exec_clock, delta_exec);
548 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
550 curr->vruntime += delta_exec_weighted;
551 update_min_vruntime(cfs_rq);
553 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
554 cfs_rq->load_unacc_exec_time += delta_exec;
558 static void update_curr(struct cfs_rq *cfs_rq)
560 struct sched_entity *curr = cfs_rq->curr;
561 u64 now = rq_of(cfs_rq)->clock_task;
562 unsigned long delta_exec;
568 * Get the amount of time the current task was running
569 * since the last time we changed load (this cannot
570 * overflow on 32 bits):
572 delta_exec = (unsigned long)(now - curr->exec_start);
576 __update_curr(cfs_rq, curr, delta_exec);
577 curr->exec_start = now;
579 if (entity_is_task(curr)) {
580 struct task_struct *curtask = task_of(curr);
582 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
583 cpuacct_charge(curtask, delta_exec);
584 account_group_exec_runtime(curtask, delta_exec);
589 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
591 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
595 * Task is being enqueued - update stats:
597 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
600 * Are we enqueueing a waiting task? (for current tasks
601 * a dequeue/enqueue event is a NOP)
603 if (se != cfs_rq->curr)
604 update_stats_wait_start(cfs_rq, se);
608 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
610 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
611 rq_of(cfs_rq)->clock - se->statistics.wait_start));
612 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
613 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
614 rq_of(cfs_rq)->clock - se->statistics.wait_start);
615 #ifdef CONFIG_SCHEDSTATS
616 if (entity_is_task(se)) {
617 trace_sched_stat_wait(task_of(se),
618 rq_of(cfs_rq)->clock - se->statistics.wait_start);
621 schedstat_set(se->statistics.wait_start, 0);
625 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 * Mark the end of the wait period if dequeueing a
631 if (se != cfs_rq->curr)
632 update_stats_wait_end(cfs_rq, se);
636 * We are picking a new current task - update its stats:
639 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
642 * We are starting a new run period:
644 se->exec_start = rq_of(cfs_rq)->clock_task;
647 /**************************************************
648 * Scheduling class queueing methods:
651 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
653 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
655 cfs_rq->task_weight += weight;
659 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
665 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
667 update_load_add(&cfs_rq->load, se->load.weight);
668 if (!parent_entity(se))
669 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
670 if (entity_is_task(se)) {
671 add_cfs_task_weight(cfs_rq, se->load.weight);
672 list_add(&se->group_node, &cfs_rq->tasks);
674 cfs_rq->nr_running++;
678 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
680 update_load_sub(&cfs_rq->load, se->load.weight);
681 if (!parent_entity(se))
682 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
683 if (entity_is_task(se)) {
684 add_cfs_task_weight(cfs_rq, -se->load.weight);
685 list_del_init(&se->group_node);
687 cfs_rq->nr_running--;
690 #ifdef CONFIG_FAIR_GROUP_SCHED
692 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
695 struct task_group *tg = cfs_rq->tg;
698 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
699 load_avg -= cfs_rq->load_contribution;
701 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
702 atomic_add(load_avg, &tg->load_weight);
703 cfs_rq->load_contribution += load_avg;
707 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
709 u64 period = sysctl_sched_shares_window;
711 unsigned long load = cfs_rq->load.weight;
713 if (cfs_rq->tg == &root_task_group)
716 now = rq_of(cfs_rq)->clock_task;
717 delta = now - cfs_rq->load_stamp;
719 /* truncate load history at 4 idle periods */
720 if (cfs_rq->load_stamp > cfs_rq->load_last &&
721 now - cfs_rq->load_last > 4 * period) {
722 cfs_rq->load_period = 0;
723 cfs_rq->load_avg = 0;
727 cfs_rq->load_stamp = now;
728 cfs_rq->load_unacc_exec_time = 0;
729 cfs_rq->load_period += delta;
731 cfs_rq->load_last = now;
732 cfs_rq->load_avg += delta * load;
735 /* consider updating load contribution on each fold or truncate */
736 if (global_update || cfs_rq->load_period > period
737 || !cfs_rq->load_period)
738 update_cfs_rq_load_contribution(cfs_rq, global_update);
740 while (cfs_rq->load_period > period) {
742 * Inline assembly required to prevent the compiler
743 * optimising this loop into a divmod call.
744 * See __iter_div_u64_rem() for another example of this.
746 asm("" : "+rm" (cfs_rq->load_period));
747 cfs_rq->load_period /= 2;
748 cfs_rq->load_avg /= 2;
751 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
752 list_del_leaf_cfs_rq(cfs_rq);
755 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
757 long load_weight, load, shares;
759 load = cfs_rq->load.weight;
761 load_weight = atomic_read(&tg->load_weight);
763 load_weight -= cfs_rq->load_contribution;
765 shares = (tg->shares * load);
767 shares /= load_weight;
769 if (shares < MIN_SHARES)
771 if (shares > tg->shares)
777 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
779 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
780 update_cfs_load(cfs_rq, 0);
781 update_cfs_shares(cfs_rq);
784 # else /* CONFIG_SMP */
785 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
789 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
794 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
797 # endif /* CONFIG_SMP */
798 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
799 unsigned long weight)
802 /* commit outstanding execution time */
803 if (cfs_rq->curr == se)
805 account_entity_dequeue(cfs_rq, se);
808 update_load_set(&se->load, weight);
811 account_entity_enqueue(cfs_rq, se);
814 static void update_cfs_shares(struct cfs_rq *cfs_rq)
816 struct task_group *tg;
817 struct sched_entity *se;
821 se = tg->se[cpu_of(rq_of(cfs_rq))];
825 if (likely(se->load.weight == tg->shares))
828 shares = calc_cfs_shares(cfs_rq, tg);
830 reweight_entity(cfs_rq_of(se), se, shares);
832 #else /* CONFIG_FAIR_GROUP_SCHED */
833 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
837 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
841 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
844 #endif /* CONFIG_FAIR_GROUP_SCHED */
846 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
848 #ifdef CONFIG_SCHEDSTATS
849 struct task_struct *tsk = NULL;
851 if (entity_is_task(se))
854 if (se->statistics.sleep_start) {
855 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
860 if (unlikely(delta > se->statistics.sleep_max))
861 se->statistics.sleep_max = delta;
863 se->statistics.sleep_start = 0;
864 se->statistics.sum_sleep_runtime += delta;
867 account_scheduler_latency(tsk, delta >> 10, 1);
868 trace_sched_stat_sleep(tsk, delta);
871 if (se->statistics.block_start) {
872 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
877 if (unlikely(delta > se->statistics.block_max))
878 se->statistics.block_max = delta;
880 se->statistics.block_start = 0;
881 se->statistics.sum_sleep_runtime += delta;
884 if (tsk->in_iowait) {
885 se->statistics.iowait_sum += delta;
886 se->statistics.iowait_count++;
887 trace_sched_stat_iowait(tsk, delta);
891 * Blocking time is in units of nanosecs, so shift by
892 * 20 to get a milliseconds-range estimation of the
893 * amount of time that the task spent sleeping:
895 if (unlikely(prof_on == SLEEP_PROFILING)) {
896 profile_hits(SLEEP_PROFILING,
897 (void *)get_wchan(tsk),
900 account_scheduler_latency(tsk, delta >> 10, 0);
906 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
908 #ifdef CONFIG_SCHED_DEBUG
909 s64 d = se->vruntime - cfs_rq->min_vruntime;
914 if (d > 3*sysctl_sched_latency)
915 schedstat_inc(cfs_rq, nr_spread_over);
920 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
922 u64 vruntime = cfs_rq->min_vruntime;
925 * The 'current' period is already promised to the current tasks,
926 * however the extra weight of the new task will slow them down a
927 * little, place the new task so that it fits in the slot that
928 * stays open at the end.
930 if (initial && sched_feat(START_DEBIT))
931 vruntime += sched_vslice(cfs_rq, se);
933 /* sleeps up to a single latency don't count. */
935 unsigned long thresh = sysctl_sched_latency;
938 * Halve their sleep time's effect, to allow
939 * for a gentler effect of sleepers:
941 if (sched_feat(GENTLE_FAIR_SLEEPERS))
947 /* ensure we never gain time by being placed backwards. */
948 vruntime = max_vruntime(se->vruntime, vruntime);
950 se->vruntime = vruntime;
954 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
957 * Update the normalized vruntime before updating min_vruntime
958 * through callig update_curr().
960 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
961 se->vruntime += cfs_rq->min_vruntime;
964 * Update run-time statistics of the 'current'.
967 update_cfs_load(cfs_rq, 0);
968 account_entity_enqueue(cfs_rq, se);
969 update_cfs_shares(cfs_rq);
971 if (flags & ENQUEUE_WAKEUP) {
972 place_entity(cfs_rq, se, 0);
973 enqueue_sleeper(cfs_rq, se);
976 update_stats_enqueue(cfs_rq, se);
977 check_spread(cfs_rq, se);
978 if (se != cfs_rq->curr)
979 __enqueue_entity(cfs_rq, se);
982 if (cfs_rq->nr_running == 1)
983 list_add_leaf_cfs_rq(cfs_rq);
986 static void __clear_buddies_last(struct sched_entity *se)
988 for_each_sched_entity(se) {
989 struct cfs_rq *cfs_rq = cfs_rq_of(se);
990 if (cfs_rq->last == se)
997 static void __clear_buddies_next(struct sched_entity *se)
999 for_each_sched_entity(se) {
1000 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1001 if (cfs_rq->next == se)
1002 cfs_rq->next = NULL;
1008 static void __clear_buddies_skip(struct sched_entity *se)
1010 for_each_sched_entity(se) {
1011 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1012 if (cfs_rq->skip == se)
1013 cfs_rq->skip = NULL;
1019 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1021 if (cfs_rq->last == se)
1022 __clear_buddies_last(se);
1024 if (cfs_rq->next == se)
1025 __clear_buddies_next(se);
1027 if (cfs_rq->skip == se)
1028 __clear_buddies_skip(se);
1032 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1035 * Update run-time statistics of the 'current'.
1037 update_curr(cfs_rq);
1039 update_stats_dequeue(cfs_rq, se);
1040 if (flags & DEQUEUE_SLEEP) {
1041 #ifdef CONFIG_SCHEDSTATS
1042 if (entity_is_task(se)) {
1043 struct task_struct *tsk = task_of(se);
1045 if (tsk->state & TASK_INTERRUPTIBLE)
1046 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1047 if (tsk->state & TASK_UNINTERRUPTIBLE)
1048 se->statistics.block_start = rq_of(cfs_rq)->clock;
1053 clear_buddies(cfs_rq, se);
1055 if (se != cfs_rq->curr)
1056 __dequeue_entity(cfs_rq, se);
1058 update_cfs_load(cfs_rq, 0);
1059 account_entity_dequeue(cfs_rq, se);
1062 * Normalize the entity after updating the min_vruntime because the
1063 * update can refer to the ->curr item and we need to reflect this
1064 * movement in our normalized position.
1066 if (!(flags & DEQUEUE_SLEEP))
1067 se->vruntime -= cfs_rq->min_vruntime;
1069 update_min_vruntime(cfs_rq);
1070 update_cfs_shares(cfs_rq);
1074 * Preempt the current task with a newly woken task if needed:
1077 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1079 unsigned long ideal_runtime, delta_exec;
1081 ideal_runtime = sched_slice(cfs_rq, curr);
1082 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1083 if (delta_exec > ideal_runtime) {
1084 resched_task(rq_of(cfs_rq)->curr);
1086 * The current task ran long enough, ensure it doesn't get
1087 * re-elected due to buddy favours.
1089 clear_buddies(cfs_rq, curr);
1094 * Ensure that a task that missed wakeup preemption by a
1095 * narrow margin doesn't have to wait for a full slice.
1096 * This also mitigates buddy induced latencies under load.
1098 if (delta_exec < sysctl_sched_min_granularity)
1101 if (cfs_rq->nr_running > 1) {
1102 struct sched_entity *se = __pick_first_entity(cfs_rq);
1103 s64 delta = curr->vruntime - se->vruntime;
1108 if (delta > ideal_runtime)
1109 resched_task(rq_of(cfs_rq)->curr);
1114 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1116 /* 'current' is not kept within the tree. */
1119 * Any task has to be enqueued before it get to execute on
1120 * a CPU. So account for the time it spent waiting on the
1123 update_stats_wait_end(cfs_rq, se);
1124 __dequeue_entity(cfs_rq, se);
1127 update_stats_curr_start(cfs_rq, se);
1129 #ifdef CONFIG_SCHEDSTATS
1131 * Track our maximum slice length, if the CPU's load is at
1132 * least twice that of our own weight (i.e. dont track it
1133 * when there are only lesser-weight tasks around):
1135 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1136 se->statistics.slice_max = max(se->statistics.slice_max,
1137 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1140 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1144 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1147 * Pick the next process, keeping these things in mind, in this order:
1148 * 1) keep things fair between processes/task groups
1149 * 2) pick the "next" process, since someone really wants that to run
1150 * 3) pick the "last" process, for cache locality
1151 * 4) do not run the "skip" process, if something else is available
1153 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1155 struct sched_entity *se = __pick_first_entity(cfs_rq);
1156 struct sched_entity *left = se;
1159 * Avoid running the skip buddy, if running something else can
1160 * be done without getting too unfair.
1162 if (cfs_rq->skip == se) {
1163 struct sched_entity *second = __pick_next_entity(se);
1164 if (second && wakeup_preempt_entity(second, left) < 1)
1169 * Prefer last buddy, try to return the CPU to a preempted task.
1171 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1175 * Someone really wants this to run. If it's not unfair, run it.
1177 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1180 clear_buddies(cfs_rq, se);
1185 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1188 * If still on the runqueue then deactivate_task()
1189 * was not called and update_curr() has to be done:
1192 update_curr(cfs_rq);
1194 check_spread(cfs_rq, prev);
1196 update_stats_wait_start(cfs_rq, prev);
1197 /* Put 'current' back into the tree. */
1198 __enqueue_entity(cfs_rq, prev);
1200 cfs_rq->curr = NULL;
1204 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1207 * Update run-time statistics of the 'current'.
1209 update_curr(cfs_rq);
1212 * Update share accounting for long-running entities.
1214 update_entity_shares_tick(cfs_rq);
1216 #ifdef CONFIG_SCHED_HRTICK
1218 * queued ticks are scheduled to match the slice, so don't bother
1219 * validating it and just reschedule.
1222 resched_task(rq_of(cfs_rq)->curr);
1226 * don't let the period tick interfere with the hrtick preemption
1228 if (!sched_feat(DOUBLE_TICK) &&
1229 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1233 if (cfs_rq->nr_running > 1)
1234 check_preempt_tick(cfs_rq, curr);
1238 /**************************************************
1239 * CFS bandwidth control machinery
1242 #ifdef CONFIG_CFS_BANDWIDTH
1244 * default period for cfs group bandwidth.
1245 * default: 0.1s, units: nanoseconds
1247 static inline u64 default_cfs_period(void)
1249 return 100000000ULL;
1253 /**************************************************
1254 * CFS operations on tasks:
1257 #ifdef CONFIG_SCHED_HRTICK
1258 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1260 struct sched_entity *se = &p->se;
1261 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1263 WARN_ON(task_rq(p) != rq);
1265 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1266 u64 slice = sched_slice(cfs_rq, se);
1267 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1268 s64 delta = slice - ran;
1277 * Don't schedule slices shorter than 10000ns, that just
1278 * doesn't make sense. Rely on vruntime for fairness.
1281 delta = max_t(s64, 10000LL, delta);
1283 hrtick_start(rq, delta);
1288 * called from enqueue/dequeue and updates the hrtick when the
1289 * current task is from our class and nr_running is low enough
1292 static void hrtick_update(struct rq *rq)
1294 struct task_struct *curr = rq->curr;
1296 if (curr->sched_class != &fair_sched_class)
1299 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1300 hrtick_start_fair(rq, curr);
1302 #else /* !CONFIG_SCHED_HRTICK */
1304 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1308 static inline void hrtick_update(struct rq *rq)
1314 * The enqueue_task method is called before nr_running is
1315 * increased. Here we update the fair scheduling stats and
1316 * then put the task into the rbtree:
1319 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1321 struct cfs_rq *cfs_rq;
1322 struct sched_entity *se = &p->se;
1324 for_each_sched_entity(se) {
1327 cfs_rq = cfs_rq_of(se);
1328 enqueue_entity(cfs_rq, se, flags);
1329 cfs_rq->h_nr_running++;
1330 flags = ENQUEUE_WAKEUP;
1333 for_each_sched_entity(se) {
1334 cfs_rq = cfs_rq_of(se);
1335 cfs_rq->h_nr_running++;
1337 update_cfs_load(cfs_rq, 0);
1338 update_cfs_shares(cfs_rq);
1345 static void set_next_buddy(struct sched_entity *se);
1348 * The dequeue_task method is called before nr_running is
1349 * decreased. We remove the task from the rbtree and
1350 * update the fair scheduling stats:
1352 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1354 struct cfs_rq *cfs_rq;
1355 struct sched_entity *se = &p->se;
1356 int task_sleep = flags & DEQUEUE_SLEEP;
1358 for_each_sched_entity(se) {
1359 cfs_rq = cfs_rq_of(se);
1360 dequeue_entity(cfs_rq, se, flags);
1361 cfs_rq->h_nr_running--;
1363 /* Don't dequeue parent if it has other entities besides us */
1364 if (cfs_rq->load.weight) {
1366 * Bias pick_next to pick a task from this cfs_rq, as
1367 * p is sleeping when it is within its sched_slice.
1369 if (task_sleep && parent_entity(se))
1370 set_next_buddy(parent_entity(se));
1372 /* avoid re-evaluating load for this entity */
1373 se = parent_entity(se);
1376 flags |= DEQUEUE_SLEEP;
1379 for_each_sched_entity(se) {
1380 cfs_rq = cfs_rq_of(se);
1381 cfs_rq->h_nr_running--;
1383 update_cfs_load(cfs_rq, 0);
1384 update_cfs_shares(cfs_rq);
1393 static void task_waking_fair(struct task_struct *p)
1395 struct sched_entity *se = &p->se;
1396 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1399 #ifndef CONFIG_64BIT
1400 u64 min_vruntime_copy;
1403 min_vruntime_copy = cfs_rq->min_vruntime_copy;
1405 min_vruntime = cfs_rq->min_vruntime;
1406 } while (min_vruntime != min_vruntime_copy);
1408 min_vruntime = cfs_rq->min_vruntime;
1411 se->vruntime -= min_vruntime;
1414 #ifdef CONFIG_FAIR_GROUP_SCHED
1416 * effective_load() calculates the load change as seen from the root_task_group
1418 * Adding load to a group doesn't make a group heavier, but can cause movement
1419 * of group shares between cpus. Assuming the shares were perfectly aligned one
1420 * can calculate the shift in shares.
1422 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1424 struct sched_entity *se = tg->se[cpu];
1429 for_each_sched_entity(se) {
1433 w = se->my_q->load.weight;
1435 /* use this cpu's instantaneous contribution */
1436 lw = atomic_read(&tg->load_weight);
1437 lw -= se->my_q->load_contribution;
1442 if (lw > 0 && wl < lw)
1443 wl = (wl * tg->shares) / lw;
1447 /* zero point is MIN_SHARES */
1448 if (wl < MIN_SHARES)
1450 wl -= se->load.weight;
1459 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1460 unsigned long wl, unsigned long wg)
1467 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1469 s64 this_load, load;
1470 int idx, this_cpu, prev_cpu;
1471 unsigned long tl_per_task;
1472 struct task_group *tg;
1473 unsigned long weight;
1477 this_cpu = smp_processor_id();
1478 prev_cpu = task_cpu(p);
1479 load = source_load(prev_cpu, idx);
1480 this_load = target_load(this_cpu, idx);
1483 * If sync wakeup then subtract the (maximum possible)
1484 * effect of the currently running task from the load
1485 * of the current CPU:
1488 tg = task_group(current);
1489 weight = current->se.load.weight;
1491 this_load += effective_load(tg, this_cpu, -weight, -weight);
1492 load += effective_load(tg, prev_cpu, 0, -weight);
1496 weight = p->se.load.weight;
1499 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1500 * due to the sync cause above having dropped this_load to 0, we'll
1501 * always have an imbalance, but there's really nothing you can do
1502 * about that, so that's good too.
1504 * Otherwise check if either cpus are near enough in load to allow this
1505 * task to be woken on this_cpu.
1507 if (this_load > 0) {
1508 s64 this_eff_load, prev_eff_load;
1510 this_eff_load = 100;
1511 this_eff_load *= power_of(prev_cpu);
1512 this_eff_load *= this_load +
1513 effective_load(tg, this_cpu, weight, weight);
1515 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1516 prev_eff_load *= power_of(this_cpu);
1517 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1519 balanced = this_eff_load <= prev_eff_load;
1524 * If the currently running task will sleep within
1525 * a reasonable amount of time then attract this newly
1528 if (sync && balanced)
1531 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1532 tl_per_task = cpu_avg_load_per_task(this_cpu);
1535 (this_load <= load &&
1536 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1538 * This domain has SD_WAKE_AFFINE and
1539 * p is cache cold in this domain, and
1540 * there is no bad imbalance.
1542 schedstat_inc(sd, ttwu_move_affine);
1543 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1551 * find_idlest_group finds and returns the least busy CPU group within the
1554 static struct sched_group *
1555 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1556 int this_cpu, int load_idx)
1558 struct sched_group *idlest = NULL, *group = sd->groups;
1559 unsigned long min_load = ULONG_MAX, this_load = 0;
1560 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1563 unsigned long load, avg_load;
1567 /* Skip over this group if it has no CPUs allowed */
1568 if (!cpumask_intersects(sched_group_cpus(group),
1572 local_group = cpumask_test_cpu(this_cpu,
1573 sched_group_cpus(group));
1575 /* Tally up the load of all CPUs in the group */
1578 for_each_cpu(i, sched_group_cpus(group)) {
1579 /* Bias balancing toward cpus of our domain */
1581 load = source_load(i, load_idx);
1583 load = target_load(i, load_idx);
1588 /* Adjust by relative CPU power of the group */
1589 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
1592 this_load = avg_load;
1593 } else if (avg_load < min_load) {
1594 min_load = avg_load;
1597 } while (group = group->next, group != sd->groups);
1599 if (!idlest || 100*this_load < imbalance*min_load)
1605 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1608 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1610 unsigned long load, min_load = ULONG_MAX;
1614 /* Traverse only the allowed CPUs */
1615 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1616 load = weighted_cpuload(i);
1618 if (load < min_load || (load == min_load && i == this_cpu)) {
1628 * Try and locate an idle CPU in the sched_domain.
1630 static int select_idle_sibling(struct task_struct *p, int target)
1632 int cpu = smp_processor_id();
1633 int prev_cpu = task_cpu(p);
1634 struct sched_domain *sd;
1638 * If the task is going to be woken-up on this cpu and if it is
1639 * already idle, then it is the right target.
1641 if (target == cpu && idle_cpu(cpu))
1645 * If the task is going to be woken-up on the cpu where it previously
1646 * ran and if it is currently idle, then it the right target.
1648 if (target == prev_cpu && idle_cpu(prev_cpu))
1652 * Otherwise, iterate the domains and find an elegible idle cpu.
1655 for_each_domain(target, sd) {
1656 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1659 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1667 * Lets stop looking for an idle sibling when we reached
1668 * the domain that spans the current cpu and prev_cpu.
1670 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1671 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1680 * sched_balance_self: balance the current task (running on cpu) in domains
1681 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1684 * Balance, ie. select the least loaded group.
1686 * Returns the target CPU number, or the same CPU if no balancing is needed.
1688 * preempt must be disabled.
1691 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
1693 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1694 int cpu = smp_processor_id();
1695 int prev_cpu = task_cpu(p);
1697 int want_affine = 0;
1699 int sync = wake_flags & WF_SYNC;
1701 if (sd_flag & SD_BALANCE_WAKE) {
1702 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1708 for_each_domain(cpu, tmp) {
1709 if (!(tmp->flags & SD_LOAD_BALANCE))
1713 * If power savings logic is enabled for a domain, see if we
1714 * are not overloaded, if so, don't balance wider.
1716 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1717 unsigned long power = 0;
1718 unsigned long nr_running = 0;
1719 unsigned long capacity;
1722 for_each_cpu(i, sched_domain_span(tmp)) {
1723 power += power_of(i);
1724 nr_running += cpu_rq(i)->cfs.nr_running;
1727 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
1729 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1732 if (nr_running < capacity)
1737 * If both cpu and prev_cpu are part of this domain,
1738 * cpu is a valid SD_WAKE_AFFINE target.
1740 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1741 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1746 if (!want_sd && !want_affine)
1749 if (!(tmp->flags & sd_flag))
1757 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1760 new_cpu = select_idle_sibling(p, prev_cpu);
1765 int load_idx = sd->forkexec_idx;
1766 struct sched_group *group;
1769 if (!(sd->flags & sd_flag)) {
1774 if (sd_flag & SD_BALANCE_WAKE)
1775 load_idx = sd->wake_idx;
1777 group = find_idlest_group(sd, p, cpu, load_idx);
1783 new_cpu = find_idlest_cpu(group, p, cpu);
1784 if (new_cpu == -1 || new_cpu == cpu) {
1785 /* Now try balancing at a lower domain level of cpu */
1790 /* Now try balancing at a lower domain level of new_cpu */
1792 weight = sd->span_weight;
1794 for_each_domain(cpu, tmp) {
1795 if (weight <= tmp->span_weight)
1797 if (tmp->flags & sd_flag)
1800 /* while loop will break here if sd == NULL */
1807 #endif /* CONFIG_SMP */
1809 static unsigned long
1810 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1812 unsigned long gran = sysctl_sched_wakeup_granularity;
1815 * Since its curr running now, convert the gran from real-time
1816 * to virtual-time in his units.
1818 * By using 'se' instead of 'curr' we penalize light tasks, so
1819 * they get preempted easier. That is, if 'se' < 'curr' then
1820 * the resulting gran will be larger, therefore penalizing the
1821 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1822 * be smaller, again penalizing the lighter task.
1824 * This is especially important for buddies when the leftmost
1825 * task is higher priority than the buddy.
1827 return calc_delta_fair(gran, se);
1831 * Should 'se' preempt 'curr'.
1845 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1847 s64 gran, vdiff = curr->vruntime - se->vruntime;
1852 gran = wakeup_gran(curr, se);
1859 static void set_last_buddy(struct sched_entity *se)
1861 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
1864 for_each_sched_entity(se)
1865 cfs_rq_of(se)->last = se;
1868 static void set_next_buddy(struct sched_entity *se)
1870 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
1873 for_each_sched_entity(se)
1874 cfs_rq_of(se)->next = se;
1877 static void set_skip_buddy(struct sched_entity *se)
1879 for_each_sched_entity(se)
1880 cfs_rq_of(se)->skip = se;
1884 * Preempt the current task with a newly woken task if needed:
1886 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1888 struct task_struct *curr = rq->curr;
1889 struct sched_entity *se = &curr->se, *pse = &p->se;
1890 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1891 int scale = cfs_rq->nr_running >= sched_nr_latency;
1892 int next_buddy_marked = 0;
1894 if (unlikely(se == pse))
1897 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
1898 set_next_buddy(pse);
1899 next_buddy_marked = 1;
1903 * We can come here with TIF_NEED_RESCHED already set from new task
1906 if (test_tsk_need_resched(curr))
1909 /* Idle tasks are by definition preempted by non-idle tasks. */
1910 if (unlikely(curr->policy == SCHED_IDLE) &&
1911 likely(p->policy != SCHED_IDLE))
1915 * Batch and idle tasks do not preempt non-idle tasks (their preemption
1916 * is driven by the tick):
1918 if (unlikely(p->policy != SCHED_NORMAL))
1921 find_matching_se(&se, &pse);
1922 update_curr(cfs_rq_of(se));
1924 if (wakeup_preempt_entity(se, pse) == 1) {
1926 * Bias pick_next to pick the sched entity that is
1927 * triggering this preemption.
1929 if (!next_buddy_marked)
1930 set_next_buddy(pse);
1939 * Only set the backward buddy when the current task is still
1940 * on the rq. This can happen when a wakeup gets interleaved
1941 * with schedule on the ->pre_schedule() or idle_balance()
1942 * point, either of which can * drop the rq lock.
1944 * Also, during early boot the idle thread is in the fair class,
1945 * for obvious reasons its a bad idea to schedule back to it.
1947 if (unlikely(!se->on_rq || curr == rq->idle))
1950 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1954 static struct task_struct *pick_next_task_fair(struct rq *rq)
1956 struct task_struct *p;
1957 struct cfs_rq *cfs_rq = &rq->cfs;
1958 struct sched_entity *se;
1960 if (!cfs_rq->nr_running)
1964 se = pick_next_entity(cfs_rq);
1965 set_next_entity(cfs_rq, se);
1966 cfs_rq = group_cfs_rq(se);
1970 hrtick_start_fair(rq, p);
1976 * Account for a descheduled task:
1978 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1980 struct sched_entity *se = &prev->se;
1981 struct cfs_rq *cfs_rq;
1983 for_each_sched_entity(se) {
1984 cfs_rq = cfs_rq_of(se);
1985 put_prev_entity(cfs_rq, se);
1990 * sched_yield() is very simple
1992 * The magic of dealing with the ->skip buddy is in pick_next_entity.
1994 static void yield_task_fair(struct rq *rq)
1996 struct task_struct *curr = rq->curr;
1997 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1998 struct sched_entity *se = &curr->se;
2001 * Are we the only task in the tree?
2003 if (unlikely(rq->nr_running == 1))
2006 clear_buddies(cfs_rq, se);
2008 if (curr->policy != SCHED_BATCH) {
2009 update_rq_clock(rq);
2011 * Update run-time statistics of the 'current'.
2013 update_curr(cfs_rq);
2019 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2021 struct sched_entity *se = &p->se;
2026 /* Tell the scheduler that we'd really like pse to run next. */
2029 yield_task_fair(rq);
2035 /**************************************************
2036 * Fair scheduling class load-balancing methods:
2040 * pull_task - move a task from a remote runqueue to the local runqueue.
2041 * Both runqueues must be locked.
2043 static void pull_task(struct rq *src_rq, struct task_struct *p,
2044 struct rq *this_rq, int this_cpu)
2046 deactivate_task(src_rq, p, 0);
2047 set_task_cpu(p, this_cpu);
2048 activate_task(this_rq, p, 0);
2049 check_preempt_curr(this_rq, p, 0);
2053 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2056 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2057 struct sched_domain *sd, enum cpu_idle_type idle,
2060 int tsk_cache_hot = 0;
2062 * We do not migrate tasks that are:
2063 * 1) running (obviously), or
2064 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2065 * 3) are cache-hot on their current CPU.
2067 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2068 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2073 if (task_running(rq, p)) {
2074 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2079 * Aggressive migration if:
2080 * 1) task is cache cold, or
2081 * 2) too many balance attempts have failed.
2084 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2085 if (!tsk_cache_hot ||
2086 sd->nr_balance_failed > sd->cache_nice_tries) {
2087 #ifdef CONFIG_SCHEDSTATS
2088 if (tsk_cache_hot) {
2089 schedstat_inc(sd, lb_hot_gained[idle]);
2090 schedstat_inc(p, se.statistics.nr_forced_migrations);
2096 if (tsk_cache_hot) {
2097 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2104 * move_one_task tries to move exactly one task from busiest to this_rq, as
2105 * part of active balancing operations within "domain".
2106 * Returns 1 if successful and 0 otherwise.
2108 * Called with both runqueues locked.
2111 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2112 struct sched_domain *sd, enum cpu_idle_type idle)
2114 struct task_struct *p, *n;
2115 struct cfs_rq *cfs_rq;
2118 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2119 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2121 if (!can_migrate_task(p, busiest, this_cpu,
2125 pull_task(busiest, p, this_rq, this_cpu);
2127 * Right now, this is only the second place pull_task()
2128 * is called, so we can safely collect pull_task()
2129 * stats here rather than inside pull_task().
2131 schedstat_inc(sd, lb_gained[idle]);
2139 static unsigned long
2140 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2141 unsigned long max_load_move, struct sched_domain *sd,
2142 enum cpu_idle_type idle, int *all_pinned,
2143 struct cfs_rq *busiest_cfs_rq)
2145 int loops = 0, pulled = 0;
2146 long rem_load_move = max_load_move;
2147 struct task_struct *p, *n;
2149 if (max_load_move == 0)
2152 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2153 if (loops++ > sysctl_sched_nr_migrate)
2156 if ((p->se.load.weight >> 1) > rem_load_move ||
2157 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2161 pull_task(busiest, p, this_rq, this_cpu);
2163 rem_load_move -= p->se.load.weight;
2165 #ifdef CONFIG_PREEMPT
2167 * NEWIDLE balancing is a source of latency, so preemptible
2168 * kernels will stop after the first task is pulled to minimize
2169 * the critical section.
2171 if (idle == CPU_NEWLY_IDLE)
2176 * We only want to steal up to the prescribed amount of
2179 if (rem_load_move <= 0)
2184 * Right now, this is one of only two places pull_task() is called,
2185 * so we can safely collect pull_task() stats here rather than
2186 * inside pull_task().
2188 schedstat_add(sd, lb_gained[idle], pulled);
2190 return max_load_move - rem_load_move;
2193 #ifdef CONFIG_FAIR_GROUP_SCHED
2195 * update tg->load_weight by folding this cpu's load_avg
2197 static int update_shares_cpu(struct task_group *tg, int cpu)
2199 struct cfs_rq *cfs_rq;
2200 unsigned long flags;
2207 cfs_rq = tg->cfs_rq[cpu];
2209 raw_spin_lock_irqsave(&rq->lock, flags);
2211 update_rq_clock(rq);
2212 update_cfs_load(cfs_rq, 1);
2215 * We need to update shares after updating tg->load_weight in
2216 * order to adjust the weight of groups with long running tasks.
2218 update_cfs_shares(cfs_rq);
2220 raw_spin_unlock_irqrestore(&rq->lock, flags);
2225 static void update_shares(int cpu)
2227 struct cfs_rq *cfs_rq;
2228 struct rq *rq = cpu_rq(cpu);
2232 * Iterates the task_group tree in a bottom up fashion, see
2233 * list_add_leaf_cfs_rq() for details.
2235 for_each_leaf_cfs_rq(rq, cfs_rq)
2236 update_shares_cpu(cfs_rq->tg, cpu);
2241 * Compute the cpu's hierarchical load factor for each task group.
2242 * This needs to be done in a top-down fashion because the load of a child
2243 * group is a fraction of its parents load.
2245 static int tg_load_down(struct task_group *tg, void *data)
2248 long cpu = (long)data;
2251 load = cpu_rq(cpu)->load.weight;
2253 load = tg->parent->cfs_rq[cpu]->h_load;
2254 load *= tg->se[cpu]->load.weight;
2255 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2258 tg->cfs_rq[cpu]->h_load = load;
2263 static void update_h_load(long cpu)
2265 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2268 static unsigned long
2269 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2270 unsigned long max_load_move,
2271 struct sched_domain *sd, enum cpu_idle_type idle,
2274 long rem_load_move = max_load_move;
2275 struct cfs_rq *busiest_cfs_rq;
2278 update_h_load(cpu_of(busiest));
2280 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
2281 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2282 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2283 u64 rem_load, moved_load;
2288 if (!busiest_cfs_rq->task_weight)
2291 rem_load = (u64)rem_load_move * busiest_weight;
2292 rem_load = div_u64(rem_load, busiest_h_load + 1);
2294 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2295 rem_load, sd, idle, all_pinned,
2301 moved_load *= busiest_h_load;
2302 moved_load = div_u64(moved_load, busiest_weight + 1);
2304 rem_load_move -= moved_load;
2305 if (rem_load_move < 0)
2310 return max_load_move - rem_load_move;
2313 static inline void update_shares(int cpu)
2317 static unsigned long
2318 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2319 unsigned long max_load_move,
2320 struct sched_domain *sd, enum cpu_idle_type idle,
2323 return balance_tasks(this_rq, this_cpu, busiest,
2324 max_load_move, sd, idle, all_pinned,
2330 * move_tasks tries to move up to max_load_move weighted load from busiest to
2331 * this_rq, as part of a balancing operation within domain "sd".
2332 * Returns 1 if successful and 0 otherwise.
2334 * Called with both runqueues locked.
2336 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2337 unsigned long max_load_move,
2338 struct sched_domain *sd, enum cpu_idle_type idle,
2341 unsigned long total_load_moved = 0, load_moved;
2344 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2345 max_load_move - total_load_moved,
2346 sd, idle, all_pinned);
2348 total_load_moved += load_moved;
2350 #ifdef CONFIG_PREEMPT
2352 * NEWIDLE balancing is a source of latency, so preemptible
2353 * kernels will stop after the first task is pulled to minimize
2354 * the critical section.
2356 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2359 if (raw_spin_is_contended(&this_rq->lock) ||
2360 raw_spin_is_contended(&busiest->lock))
2363 } while (load_moved && max_load_move > total_load_moved);
2365 return total_load_moved > 0;
2368 /********** Helpers for find_busiest_group ************************/
2370 * sd_lb_stats - Structure to store the statistics of a sched_domain
2371 * during load balancing.
2373 struct sd_lb_stats {
2374 struct sched_group *busiest; /* Busiest group in this sd */
2375 struct sched_group *this; /* Local group in this sd */
2376 unsigned long total_load; /* Total load of all groups in sd */
2377 unsigned long total_pwr; /* Total power of all groups in sd */
2378 unsigned long avg_load; /* Average load across all groups in sd */
2380 /** Statistics of this group */
2381 unsigned long this_load;
2382 unsigned long this_load_per_task;
2383 unsigned long this_nr_running;
2384 unsigned long this_has_capacity;
2385 unsigned int this_idle_cpus;
2387 /* Statistics of the busiest group */
2388 unsigned int busiest_idle_cpus;
2389 unsigned long max_load;
2390 unsigned long busiest_load_per_task;
2391 unsigned long busiest_nr_running;
2392 unsigned long busiest_group_capacity;
2393 unsigned long busiest_has_capacity;
2394 unsigned int busiest_group_weight;
2396 int group_imb; /* Is there imbalance in this sd */
2397 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2398 int power_savings_balance; /* Is powersave balance needed for this sd */
2399 struct sched_group *group_min; /* Least loaded group in sd */
2400 struct sched_group *group_leader; /* Group which relieves group_min */
2401 unsigned long min_load_per_task; /* load_per_task in group_min */
2402 unsigned long leader_nr_running; /* Nr running of group_leader */
2403 unsigned long min_nr_running; /* Nr running of group_min */
2408 * sg_lb_stats - stats of a sched_group required for load_balancing
2410 struct sg_lb_stats {
2411 unsigned long avg_load; /*Avg load across the CPUs of the group */
2412 unsigned long group_load; /* Total load over the CPUs of the group */
2413 unsigned long sum_nr_running; /* Nr tasks running in the group */
2414 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2415 unsigned long group_capacity;
2416 unsigned long idle_cpus;
2417 unsigned long group_weight;
2418 int group_imb; /* Is there an imbalance in the group ? */
2419 int group_has_capacity; /* Is there extra capacity in the group? */
2423 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2424 * @group: The group whose first cpu is to be returned.
2426 static inline unsigned int group_first_cpu(struct sched_group *group)
2428 return cpumask_first(sched_group_cpus(group));
2432 * get_sd_load_idx - Obtain the load index for a given sched domain.
2433 * @sd: The sched_domain whose load_idx is to be obtained.
2434 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2436 static inline int get_sd_load_idx(struct sched_domain *sd,
2437 enum cpu_idle_type idle)
2443 load_idx = sd->busy_idx;
2446 case CPU_NEWLY_IDLE:
2447 load_idx = sd->newidle_idx;
2450 load_idx = sd->idle_idx;
2458 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2460 * init_sd_power_savings_stats - Initialize power savings statistics for
2461 * the given sched_domain, during load balancing.
2463 * @sd: Sched domain whose power-savings statistics are to be initialized.
2464 * @sds: Variable containing the statistics for sd.
2465 * @idle: Idle status of the CPU at which we're performing load-balancing.
2467 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2468 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2471 * Busy processors will not participate in power savings
2474 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2475 sds->power_savings_balance = 0;
2477 sds->power_savings_balance = 1;
2478 sds->min_nr_running = ULONG_MAX;
2479 sds->leader_nr_running = 0;
2484 * update_sd_power_savings_stats - Update the power saving stats for a
2485 * sched_domain while performing load balancing.
2487 * @group: sched_group belonging to the sched_domain under consideration.
2488 * @sds: Variable containing the statistics of the sched_domain
2489 * @local_group: Does group contain the CPU for which we're performing
2491 * @sgs: Variable containing the statistics of the group.
2493 static inline void update_sd_power_savings_stats(struct sched_group *group,
2494 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2497 if (!sds->power_savings_balance)
2501 * If the local group is idle or completely loaded
2502 * no need to do power savings balance at this domain
2504 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2505 !sds->this_nr_running))
2506 sds->power_savings_balance = 0;
2509 * If a group is already running at full capacity or idle,
2510 * don't include that group in power savings calculations
2512 if (!sds->power_savings_balance ||
2513 sgs->sum_nr_running >= sgs->group_capacity ||
2514 !sgs->sum_nr_running)
2518 * Calculate the group which has the least non-idle load.
2519 * This is the group from where we need to pick up the load
2522 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2523 (sgs->sum_nr_running == sds->min_nr_running &&
2524 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2525 sds->group_min = group;
2526 sds->min_nr_running = sgs->sum_nr_running;
2527 sds->min_load_per_task = sgs->sum_weighted_load /
2528 sgs->sum_nr_running;
2532 * Calculate the group which is almost near its
2533 * capacity but still has some space to pick up some load
2534 * from other group and save more power
2536 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2539 if (sgs->sum_nr_running > sds->leader_nr_running ||
2540 (sgs->sum_nr_running == sds->leader_nr_running &&
2541 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2542 sds->group_leader = group;
2543 sds->leader_nr_running = sgs->sum_nr_running;
2548 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2549 * @sds: Variable containing the statistics of the sched_domain
2550 * under consideration.
2551 * @this_cpu: Cpu at which we're currently performing load-balancing.
2552 * @imbalance: Variable to store the imbalance.
2555 * Check if we have potential to perform some power-savings balance.
2556 * If yes, set the busiest group to be the least loaded group in the
2557 * sched_domain, so that it's CPUs can be put to idle.
2559 * Returns 1 if there is potential to perform power-savings balance.
2562 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2563 int this_cpu, unsigned long *imbalance)
2565 if (!sds->power_savings_balance)
2568 if (sds->this != sds->group_leader ||
2569 sds->group_leader == sds->group_min)
2572 *imbalance = sds->min_load_per_task;
2573 sds->busiest = sds->group_min;
2578 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2579 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2580 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2585 static inline void update_sd_power_savings_stats(struct sched_group *group,
2586 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2591 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2592 int this_cpu, unsigned long *imbalance)
2596 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2599 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2601 return SCHED_POWER_SCALE;
2604 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2606 return default_scale_freq_power(sd, cpu);
2609 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2611 unsigned long weight = sd->span_weight;
2612 unsigned long smt_gain = sd->smt_gain;
2619 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2621 return default_scale_smt_power(sd, cpu);
2624 unsigned long scale_rt_power(int cpu)
2626 struct rq *rq = cpu_rq(cpu);
2627 u64 total, available;
2629 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2631 if (unlikely(total < rq->rt_avg)) {
2632 /* Ensures that power won't end up being negative */
2635 available = total - rq->rt_avg;
2638 if (unlikely((s64)total < SCHED_POWER_SCALE))
2639 total = SCHED_POWER_SCALE;
2641 total >>= SCHED_POWER_SHIFT;
2643 return div_u64(available, total);
2646 static void update_cpu_power(struct sched_domain *sd, int cpu)
2648 unsigned long weight = sd->span_weight;
2649 unsigned long power = SCHED_POWER_SCALE;
2650 struct sched_group *sdg = sd->groups;
2652 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2653 if (sched_feat(ARCH_POWER))
2654 power *= arch_scale_smt_power(sd, cpu);
2656 power *= default_scale_smt_power(sd, cpu);
2658 power >>= SCHED_POWER_SHIFT;
2661 sdg->sgp->power_orig = power;
2663 if (sched_feat(ARCH_POWER))
2664 power *= arch_scale_freq_power(sd, cpu);
2666 power *= default_scale_freq_power(sd, cpu);
2668 power >>= SCHED_POWER_SHIFT;
2670 power *= scale_rt_power(cpu);
2671 power >>= SCHED_POWER_SHIFT;
2676 cpu_rq(cpu)->cpu_power = power;
2677 sdg->sgp->power = power;
2680 static void update_group_power(struct sched_domain *sd, int cpu)
2682 struct sched_domain *child = sd->child;
2683 struct sched_group *group, *sdg = sd->groups;
2684 unsigned long power;
2687 update_cpu_power(sd, cpu);
2693 group = child->groups;
2695 power += group->sgp->power;
2696 group = group->next;
2697 } while (group != child->groups);
2699 sdg->sgp->power = power;
2703 * Try and fix up capacity for tiny siblings, this is needed when
2704 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2705 * which on its own isn't powerful enough.
2707 * See update_sd_pick_busiest() and check_asym_packing().
2710 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2713 * Only siblings can have significantly less than SCHED_POWER_SCALE
2715 if (!(sd->flags & SD_SHARE_CPUPOWER))
2719 * If ~90% of the cpu_power is still there, we're good.
2721 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
2728 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2729 * @sd: The sched_domain whose statistics are to be updated.
2730 * @group: sched_group whose statistics are to be updated.
2731 * @this_cpu: Cpu for which load balance is currently performed.
2732 * @idle: Idle status of this_cpu
2733 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2734 * @local_group: Does group contain this_cpu.
2735 * @cpus: Set of cpus considered for load balancing.
2736 * @balance: Should we balance.
2737 * @sgs: variable to hold the statistics for this group.
2739 static inline void update_sg_lb_stats(struct sched_domain *sd,
2740 struct sched_group *group, int this_cpu,
2741 enum cpu_idle_type idle, int load_idx,
2742 int local_group, const struct cpumask *cpus,
2743 int *balance, struct sg_lb_stats *sgs)
2745 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2747 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2748 unsigned long avg_load_per_task = 0;
2751 balance_cpu = group_first_cpu(group);
2753 /* Tally up the load of all CPUs in the group */
2755 min_cpu_load = ~0UL;
2758 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2759 struct rq *rq = cpu_rq(i);
2761 /* Bias balancing toward cpus of our domain */
2763 if (idle_cpu(i) && !first_idle_cpu) {
2768 load = target_load(i, load_idx);
2770 load = source_load(i, load_idx);
2771 if (load > max_cpu_load) {
2772 max_cpu_load = load;
2773 max_nr_running = rq->nr_running;
2775 if (min_cpu_load > load)
2776 min_cpu_load = load;
2779 sgs->group_load += load;
2780 sgs->sum_nr_running += rq->nr_running;
2781 sgs->sum_weighted_load += weighted_cpuload(i);
2787 * First idle cpu or the first cpu(busiest) in this sched group
2788 * is eligible for doing load balancing at this and above
2789 * domains. In the newly idle case, we will allow all the cpu's
2790 * to do the newly idle load balance.
2792 if (idle != CPU_NEWLY_IDLE && local_group) {
2793 if (balance_cpu != this_cpu) {
2797 update_group_power(sd, this_cpu);
2800 /* Adjust by relative CPU power of the group */
2801 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
2804 * Consider the group unbalanced when the imbalance is larger
2805 * than the average weight of a task.
2807 * APZ: with cgroup the avg task weight can vary wildly and
2808 * might not be a suitable number - should we keep a
2809 * normalized nr_running number somewhere that negates
2812 if (sgs->sum_nr_running)
2813 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2815 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
2818 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
2820 if (!sgs->group_capacity)
2821 sgs->group_capacity = fix_small_capacity(sd, group);
2822 sgs->group_weight = group->group_weight;
2824 if (sgs->group_capacity > sgs->sum_nr_running)
2825 sgs->group_has_capacity = 1;
2829 * update_sd_pick_busiest - return 1 on busiest group
2830 * @sd: sched_domain whose statistics are to be checked
2831 * @sds: sched_domain statistics
2832 * @sg: sched_group candidate to be checked for being the busiest
2833 * @sgs: sched_group statistics
2834 * @this_cpu: the current cpu
2836 * Determine if @sg is a busier group than the previously selected
2839 static bool update_sd_pick_busiest(struct sched_domain *sd,
2840 struct sd_lb_stats *sds,
2841 struct sched_group *sg,
2842 struct sg_lb_stats *sgs,
2845 if (sgs->avg_load <= sds->max_load)
2848 if (sgs->sum_nr_running > sgs->group_capacity)
2855 * ASYM_PACKING needs to move all the work to the lowest
2856 * numbered CPUs in the group, therefore mark all groups
2857 * higher than ourself as busy.
2859 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2860 this_cpu < group_first_cpu(sg)) {
2864 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2872 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2873 * @sd: sched_domain whose statistics are to be updated.
2874 * @this_cpu: Cpu for which load balance is currently performed.
2875 * @idle: Idle status of this_cpu
2876 * @cpus: Set of cpus considered for load balancing.
2877 * @balance: Should we balance.
2878 * @sds: variable to hold the statistics for this sched_domain.
2880 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2881 enum cpu_idle_type idle, const struct cpumask *cpus,
2882 int *balance, struct sd_lb_stats *sds)
2884 struct sched_domain *child = sd->child;
2885 struct sched_group *sg = sd->groups;
2886 struct sg_lb_stats sgs;
2887 int load_idx, prefer_sibling = 0;
2889 if (child && child->flags & SD_PREFER_SIBLING)
2892 init_sd_power_savings_stats(sd, sds, idle);
2893 load_idx = get_sd_load_idx(sd, idle);
2898 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2899 memset(&sgs, 0, sizeof(sgs));
2900 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
2901 local_group, cpus, balance, &sgs);
2903 if (local_group && !(*balance))
2906 sds->total_load += sgs.group_load;
2907 sds->total_pwr += sg->sgp->power;
2910 * In case the child domain prefers tasks go to siblings
2911 * first, lower the sg capacity to one so that we'll try
2912 * and move all the excess tasks away. We lower the capacity
2913 * of a group only if the local group has the capacity to fit
2914 * these excess tasks, i.e. nr_running < group_capacity. The
2915 * extra check prevents the case where you always pull from the
2916 * heaviest group when it is already under-utilized (possible
2917 * with a large weight task outweighs the tasks on the system).
2919 if (prefer_sibling && !local_group && sds->this_has_capacity)
2920 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2923 sds->this_load = sgs.avg_load;
2925 sds->this_nr_running = sgs.sum_nr_running;
2926 sds->this_load_per_task = sgs.sum_weighted_load;
2927 sds->this_has_capacity = sgs.group_has_capacity;
2928 sds->this_idle_cpus = sgs.idle_cpus;
2929 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2930 sds->max_load = sgs.avg_load;
2932 sds->busiest_nr_running = sgs.sum_nr_running;
2933 sds->busiest_idle_cpus = sgs.idle_cpus;
2934 sds->busiest_group_capacity = sgs.group_capacity;
2935 sds->busiest_load_per_task = sgs.sum_weighted_load;
2936 sds->busiest_has_capacity = sgs.group_has_capacity;
2937 sds->busiest_group_weight = sgs.group_weight;
2938 sds->group_imb = sgs.group_imb;
2941 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2943 } while (sg != sd->groups);
2946 int __weak arch_sd_sibling_asym_packing(void)
2948 return 0*SD_ASYM_PACKING;
2952 * check_asym_packing - Check to see if the group is packed into the
2955 * This is primarily intended to used at the sibling level. Some
2956 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2957 * case of POWER7, it can move to lower SMT modes only when higher
2958 * threads are idle. When in lower SMT modes, the threads will
2959 * perform better since they share less core resources. Hence when we
2960 * have idle threads, we want them to be the higher ones.
2962 * This packing function is run on idle threads. It checks to see if
2963 * the busiest CPU in this domain (core in the P7 case) has a higher
2964 * CPU number than the packing function is being run on. Here we are
2965 * assuming lower CPU number will be equivalent to lower a SMT thread
2968 * Returns 1 when packing is required and a task should be moved to
2969 * this CPU. The amount of the imbalance is returned in *imbalance.
2971 * @sd: The sched_domain whose packing is to be checked.
2972 * @sds: Statistics of the sched_domain which is to be packed
2973 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2974 * @imbalance: returns amount of imbalanced due to packing.
2976 static int check_asym_packing(struct sched_domain *sd,
2977 struct sd_lb_stats *sds,
2978 int this_cpu, unsigned long *imbalance)
2982 if (!(sd->flags & SD_ASYM_PACKING))
2988 busiest_cpu = group_first_cpu(sds->busiest);
2989 if (this_cpu > busiest_cpu)
2992 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
2998 * fix_small_imbalance - Calculate the minor imbalance that exists
2999 * amongst the groups of a sched_domain, during
3001 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3002 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3003 * @imbalance: Variable to store the imbalance.
3005 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3006 int this_cpu, unsigned long *imbalance)
3008 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3009 unsigned int imbn = 2;
3010 unsigned long scaled_busy_load_per_task;
3012 if (sds->this_nr_running) {
3013 sds->this_load_per_task /= sds->this_nr_running;
3014 if (sds->busiest_load_per_task >
3015 sds->this_load_per_task)
3018 sds->this_load_per_task =
3019 cpu_avg_load_per_task(this_cpu);
3021 scaled_busy_load_per_task = sds->busiest_load_per_task
3022 * SCHED_POWER_SCALE;
3023 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3025 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3026 (scaled_busy_load_per_task * imbn)) {
3027 *imbalance = sds->busiest_load_per_task;
3032 * OK, we don't have enough imbalance to justify moving tasks,
3033 * however we may be able to increase total CPU power used by
3037 pwr_now += sds->busiest->sgp->power *
3038 min(sds->busiest_load_per_task, sds->max_load);
3039 pwr_now += sds->this->sgp->power *
3040 min(sds->this_load_per_task, sds->this_load);
3041 pwr_now /= SCHED_POWER_SCALE;
3043 /* Amount of load we'd subtract */
3044 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3045 sds->busiest->sgp->power;
3046 if (sds->max_load > tmp)
3047 pwr_move += sds->busiest->sgp->power *
3048 min(sds->busiest_load_per_task, sds->max_load - tmp);
3050 /* Amount of load we'd add */
3051 if (sds->max_load * sds->busiest->sgp->power <
3052 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3053 tmp = (sds->max_load * sds->busiest->sgp->power) /
3054 sds->this->sgp->power;
3056 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3057 sds->this->sgp->power;
3058 pwr_move += sds->this->sgp->power *
3059 min(sds->this_load_per_task, sds->this_load + tmp);
3060 pwr_move /= SCHED_POWER_SCALE;
3062 /* Move if we gain throughput */
3063 if (pwr_move > pwr_now)
3064 *imbalance = sds->busiest_load_per_task;
3068 * calculate_imbalance - Calculate the amount of imbalance present within the
3069 * groups of a given sched_domain during load balance.
3070 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3071 * @this_cpu: Cpu for which currently load balance is being performed.
3072 * @imbalance: The variable to store the imbalance.
3074 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3075 unsigned long *imbalance)
3077 unsigned long max_pull, load_above_capacity = ~0UL;
3079 sds->busiest_load_per_task /= sds->busiest_nr_running;
3080 if (sds->group_imb) {
3081 sds->busiest_load_per_task =
3082 min(sds->busiest_load_per_task, sds->avg_load);
3086 * In the presence of smp nice balancing, certain scenarios can have
3087 * max load less than avg load(as we skip the groups at or below
3088 * its cpu_power, while calculating max_load..)
3090 if (sds->max_load < sds->avg_load) {
3092 return fix_small_imbalance(sds, this_cpu, imbalance);
3095 if (!sds->group_imb) {
3097 * Don't want to pull so many tasks that a group would go idle.
3099 load_above_capacity = (sds->busiest_nr_running -
3100 sds->busiest_group_capacity);
3102 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3104 load_above_capacity /= sds->busiest->sgp->power;
3108 * We're trying to get all the cpus to the average_load, so we don't
3109 * want to push ourselves above the average load, nor do we wish to
3110 * reduce the max loaded cpu below the average load. At the same time,
3111 * we also don't want to reduce the group load below the group capacity
3112 * (so that we can implement power-savings policies etc). Thus we look
3113 * for the minimum possible imbalance.
3114 * Be careful of negative numbers as they'll appear as very large values
3115 * with unsigned longs.
3117 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3119 /* How much load to actually move to equalise the imbalance */
3120 *imbalance = min(max_pull * sds->busiest->sgp->power,
3121 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3122 / SCHED_POWER_SCALE;
3125 * if *imbalance is less than the average load per runnable task
3126 * there is no guarantee that any tasks will be moved so we'll have
3127 * a think about bumping its value to force at least one task to be
3130 if (*imbalance < sds->busiest_load_per_task)
3131 return fix_small_imbalance(sds, this_cpu, imbalance);
3135 /******* find_busiest_group() helpers end here *********************/
3138 * find_busiest_group - Returns the busiest group within the sched_domain
3139 * if there is an imbalance. If there isn't an imbalance, and
3140 * the user has opted for power-savings, it returns a group whose
3141 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3142 * such a group exists.
3144 * Also calculates the amount of weighted load which should be moved
3145 * to restore balance.
3147 * @sd: The sched_domain whose busiest group is to be returned.
3148 * @this_cpu: The cpu for which load balancing is currently being performed.
3149 * @imbalance: Variable which stores amount of weighted load which should
3150 * be moved to restore balance/put a group to idle.
3151 * @idle: The idle status of this_cpu.
3152 * @cpus: The set of CPUs under consideration for load-balancing.
3153 * @balance: Pointer to a variable indicating if this_cpu
3154 * is the appropriate cpu to perform load balancing at this_level.
3156 * Returns: - the busiest group if imbalance exists.
3157 * - If no imbalance and user has opted for power-savings balance,
3158 * return the least loaded group whose CPUs can be
3159 * put to idle by rebalancing its tasks onto our group.
3161 static struct sched_group *
3162 find_busiest_group(struct sched_domain *sd, int this_cpu,
3163 unsigned long *imbalance, enum cpu_idle_type idle,
3164 const struct cpumask *cpus, int *balance)
3166 struct sd_lb_stats sds;
3168 memset(&sds, 0, sizeof(sds));
3171 * Compute the various statistics relavent for load balancing at
3174 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3177 * this_cpu is not the appropriate cpu to perform load balancing at
3183 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3184 check_asym_packing(sd, &sds, this_cpu, imbalance))
3187 /* There is no busy sibling group to pull tasks from */
3188 if (!sds.busiest || sds.busiest_nr_running == 0)
3191 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3194 * If the busiest group is imbalanced the below checks don't
3195 * work because they assumes all things are equal, which typically
3196 * isn't true due to cpus_allowed constraints and the like.
3201 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3202 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3203 !sds.busiest_has_capacity)
3207 * If the local group is more busy than the selected busiest group
3208 * don't try and pull any tasks.
3210 if (sds.this_load >= sds.max_load)
3214 * Don't pull any tasks if this group is already above the domain
3217 if (sds.this_load >= sds.avg_load)
3220 if (idle == CPU_IDLE) {
3222 * This cpu is idle. If the busiest group load doesn't
3223 * have more tasks than the number of available cpu's and
3224 * there is no imbalance between this and busiest group
3225 * wrt to idle cpu's, it is balanced.
3227 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3228 sds.busiest_nr_running <= sds.busiest_group_weight)
3232 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3233 * imbalance_pct to be conservative.
3235 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3240 /* Looks like there is an imbalance. Compute it */
3241 calculate_imbalance(&sds, this_cpu, imbalance);
3246 * There is no obvious imbalance. But check if we can do some balancing
3249 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3257 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3260 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3261 enum cpu_idle_type idle, unsigned long imbalance,
3262 const struct cpumask *cpus)
3264 struct rq *busiest = NULL, *rq;
3265 unsigned long max_load = 0;
3268 for_each_cpu(i, sched_group_cpus(group)) {
3269 unsigned long power = power_of(i);
3270 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3275 capacity = fix_small_capacity(sd, group);
3277 if (!cpumask_test_cpu(i, cpus))
3281 wl = weighted_cpuload(i);
3284 * When comparing with imbalance, use weighted_cpuload()
3285 * which is not scaled with the cpu power.
3287 if (capacity && rq->nr_running == 1 && wl > imbalance)
3291 * For the load comparisons with the other cpu's, consider
3292 * the weighted_cpuload() scaled with the cpu power, so that
3293 * the load can be moved away from the cpu that is potentially
3294 * running at a lower capacity.
3296 wl = (wl * SCHED_POWER_SCALE) / power;
3298 if (wl > max_load) {
3308 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3309 * so long as it is large enough.
3311 #define MAX_PINNED_INTERVAL 512
3313 /* Working cpumask for load_balance and load_balance_newidle. */
3314 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3316 static int need_active_balance(struct sched_domain *sd, int idle,
3317 int busiest_cpu, int this_cpu)
3319 if (idle == CPU_NEWLY_IDLE) {
3322 * ASYM_PACKING needs to force migrate tasks from busy but
3323 * higher numbered CPUs in order to pack all tasks in the
3324 * lowest numbered CPUs.
3326 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3330 * The only task running in a non-idle cpu can be moved to this
3331 * cpu in an attempt to completely freeup the other CPU
3334 * The package power saving logic comes from
3335 * find_busiest_group(). If there are no imbalance, then
3336 * f_b_g() will return NULL. However when sched_mc={1,2} then
3337 * f_b_g() will select a group from which a running task may be
3338 * pulled to this cpu in order to make the other package idle.
3339 * If there is no opportunity to make a package idle and if
3340 * there are no imbalance, then f_b_g() will return NULL and no
3341 * action will be taken in load_balance_newidle().
3343 * Under normal task pull operation due to imbalance, there
3344 * will be more than one task in the source run queue and
3345 * move_tasks() will succeed. ld_moved will be true and this
3346 * active balance code will not be triggered.
3348 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3352 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3355 static int active_load_balance_cpu_stop(void *data);
3358 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3359 * tasks if there is an imbalance.
3361 static int load_balance(int this_cpu, struct rq *this_rq,
3362 struct sched_domain *sd, enum cpu_idle_type idle,
3365 int ld_moved, all_pinned = 0, active_balance = 0;
3366 struct sched_group *group;
3367 unsigned long imbalance;
3369 unsigned long flags;
3370 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3372 cpumask_copy(cpus, cpu_active_mask);
3374 schedstat_inc(sd, lb_count[idle]);
3377 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
3384 schedstat_inc(sd, lb_nobusyg[idle]);
3388 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3390 schedstat_inc(sd, lb_nobusyq[idle]);
3394 BUG_ON(busiest == this_rq);
3396 schedstat_add(sd, lb_imbalance[idle], imbalance);
3399 if (busiest->nr_running > 1) {
3401 * Attempt to move tasks. If find_busiest_group has found
3402 * an imbalance but busiest->nr_running <= 1, the group is
3403 * still unbalanced. ld_moved simply stays zero, so it is
3404 * correctly treated as an imbalance.
3407 local_irq_save(flags);
3408 double_rq_lock(this_rq, busiest);
3409 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3410 imbalance, sd, idle, &all_pinned);
3411 double_rq_unlock(this_rq, busiest);
3412 local_irq_restore(flags);
3415 * some other cpu did the load balance for us.
3417 if (ld_moved && this_cpu != smp_processor_id())
3418 resched_cpu(this_cpu);
3420 /* All tasks on this runqueue were pinned by CPU affinity */
3421 if (unlikely(all_pinned)) {
3422 cpumask_clear_cpu(cpu_of(busiest), cpus);
3423 if (!cpumask_empty(cpus))
3430 schedstat_inc(sd, lb_failed[idle]);
3432 * Increment the failure counter only on periodic balance.
3433 * We do not want newidle balance, which can be very
3434 * frequent, pollute the failure counter causing
3435 * excessive cache_hot migrations and active balances.
3437 if (idle != CPU_NEWLY_IDLE)
3438 sd->nr_balance_failed++;
3440 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
3441 raw_spin_lock_irqsave(&busiest->lock, flags);
3443 /* don't kick the active_load_balance_cpu_stop,
3444 * if the curr task on busiest cpu can't be
3447 if (!cpumask_test_cpu(this_cpu,
3448 &busiest->curr->cpus_allowed)) {
3449 raw_spin_unlock_irqrestore(&busiest->lock,
3452 goto out_one_pinned;
3456 * ->active_balance synchronizes accesses to
3457 * ->active_balance_work. Once set, it's cleared
3458 * only after active load balance is finished.
3460 if (!busiest->active_balance) {
3461 busiest->active_balance = 1;
3462 busiest->push_cpu = this_cpu;
3465 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3468 stop_one_cpu_nowait(cpu_of(busiest),
3469 active_load_balance_cpu_stop, busiest,
3470 &busiest->active_balance_work);
3473 * We've kicked active balancing, reset the failure
3476 sd->nr_balance_failed = sd->cache_nice_tries+1;
3479 sd->nr_balance_failed = 0;
3481 if (likely(!active_balance)) {
3482 /* We were unbalanced, so reset the balancing interval */
3483 sd->balance_interval = sd->min_interval;
3486 * If we've begun active balancing, start to back off. This
3487 * case may not be covered by the all_pinned logic if there
3488 * is only 1 task on the busy runqueue (because we don't call
3491 if (sd->balance_interval < sd->max_interval)
3492 sd->balance_interval *= 2;
3498 schedstat_inc(sd, lb_balanced[idle]);
3500 sd->nr_balance_failed = 0;
3503 /* tune up the balancing interval */
3504 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3505 (sd->balance_interval < sd->max_interval))
3506 sd->balance_interval *= 2;
3514 * idle_balance is called by schedule() if this_cpu is about to become
3515 * idle. Attempts to pull tasks from other CPUs.
3517 static void idle_balance(int this_cpu, struct rq *this_rq)
3519 struct sched_domain *sd;
3520 int pulled_task = 0;
3521 unsigned long next_balance = jiffies + HZ;
3523 this_rq->idle_stamp = this_rq->clock;
3525 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3529 * Drop the rq->lock, but keep IRQ/preempt disabled.
3531 raw_spin_unlock(&this_rq->lock);
3533 update_shares(this_cpu);
3535 for_each_domain(this_cpu, sd) {
3536 unsigned long interval;
3539 if (!(sd->flags & SD_LOAD_BALANCE))
3542 if (sd->flags & SD_BALANCE_NEWIDLE) {
3543 /* If we've pulled tasks over stop searching: */
3544 pulled_task = load_balance(this_cpu, this_rq,
3545 sd, CPU_NEWLY_IDLE, &balance);
3548 interval = msecs_to_jiffies(sd->balance_interval);
3549 if (time_after(next_balance, sd->last_balance + interval))
3550 next_balance = sd->last_balance + interval;
3552 this_rq->idle_stamp = 0;
3558 raw_spin_lock(&this_rq->lock);
3560 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3562 * We are going idle. next_balance may be set based on
3563 * a busy processor. So reset next_balance.
3565 this_rq->next_balance = next_balance;
3570 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3571 * running tasks off the busiest CPU onto idle CPUs. It requires at
3572 * least 1 task to be running on each physical CPU where possible, and
3573 * avoids physical / logical imbalances.
3575 static int active_load_balance_cpu_stop(void *data)
3577 struct rq *busiest_rq = data;
3578 int busiest_cpu = cpu_of(busiest_rq);
3579 int target_cpu = busiest_rq->push_cpu;
3580 struct rq *target_rq = cpu_rq(target_cpu);
3581 struct sched_domain *sd;
3583 raw_spin_lock_irq(&busiest_rq->lock);
3585 /* make sure the requested cpu hasn't gone down in the meantime */
3586 if (unlikely(busiest_cpu != smp_processor_id() ||
3587 !busiest_rq->active_balance))
3590 /* Is there any task to move? */
3591 if (busiest_rq->nr_running <= 1)
3595 * This condition is "impossible", if it occurs
3596 * we need to fix it. Originally reported by
3597 * Bjorn Helgaas on a 128-cpu setup.
3599 BUG_ON(busiest_rq == target_rq);
3601 /* move a task from busiest_rq to target_rq */
3602 double_lock_balance(busiest_rq, target_rq);
3604 /* Search for an sd spanning us and the target CPU. */
3606 for_each_domain(target_cpu, sd) {
3607 if ((sd->flags & SD_LOAD_BALANCE) &&
3608 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3613 schedstat_inc(sd, alb_count);
3615 if (move_one_task(target_rq, target_cpu, busiest_rq,
3617 schedstat_inc(sd, alb_pushed);
3619 schedstat_inc(sd, alb_failed);
3622 double_unlock_balance(busiest_rq, target_rq);
3624 busiest_rq->active_balance = 0;
3625 raw_spin_unlock_irq(&busiest_rq->lock);
3631 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3633 static void trigger_sched_softirq(void *data)
3635 raise_softirq_irqoff(SCHED_SOFTIRQ);
3638 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3640 csd->func = trigger_sched_softirq;
3647 * idle load balancing details
3648 * - One of the idle CPUs nominates itself as idle load_balancer, while
3650 * - This idle load balancer CPU will also go into tickless mode when
3651 * it is idle, just like all other idle CPUs
3652 * - When one of the busy CPUs notice that there may be an idle rebalancing
3653 * needed, they will kick the idle load balancer, which then does idle
3654 * load balancing for all the idle CPUs.
3657 atomic_t load_balancer;
3658 atomic_t first_pick_cpu;
3659 atomic_t second_pick_cpu;
3660 cpumask_var_t idle_cpus_mask;
3661 cpumask_var_t grp_idle_mask;
3662 unsigned long next_balance; /* in jiffy units */
3663 } nohz ____cacheline_aligned;
3665 int get_nohz_load_balancer(void)
3667 return atomic_read(&nohz.load_balancer);
3670 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3672 * lowest_flag_domain - Return lowest sched_domain containing flag.
3673 * @cpu: The cpu whose lowest level of sched domain is to
3675 * @flag: The flag to check for the lowest sched_domain
3676 * for the given cpu.
3678 * Returns the lowest sched_domain of a cpu which contains the given flag.
3680 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3682 struct sched_domain *sd;
3684 for_each_domain(cpu, sd)
3685 if (sd->flags & flag)
3692 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3693 * @cpu: The cpu whose domains we're iterating over.
3694 * @sd: variable holding the value of the power_savings_sd
3696 * @flag: The flag to filter the sched_domains to be iterated.
3698 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3699 * set, starting from the lowest sched_domain to the highest.
3701 #define for_each_flag_domain(cpu, sd, flag) \
3702 for (sd = lowest_flag_domain(cpu, flag); \
3703 (sd && (sd->flags & flag)); sd = sd->parent)
3706 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3707 * @ilb_group: group to be checked for semi-idleness
3709 * Returns: 1 if the group is semi-idle. 0 otherwise.
3711 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3712 * and atleast one non-idle CPU. This helper function checks if the given
3713 * sched_group is semi-idle or not.
3715 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3717 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3718 sched_group_cpus(ilb_group));
3721 * A sched_group is semi-idle when it has atleast one busy cpu
3722 * and atleast one idle cpu.
3724 if (cpumask_empty(nohz.grp_idle_mask))
3727 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3733 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3734 * @cpu: The cpu which is nominating a new idle_load_balancer.
3736 * Returns: Returns the id of the idle load balancer if it exists,
3737 * Else, returns >= nr_cpu_ids.
3739 * This algorithm picks the idle load balancer such that it belongs to a
3740 * semi-idle powersavings sched_domain. The idea is to try and avoid
3741 * completely idle packages/cores just for the purpose of idle load balancing
3742 * when there are other idle cpu's which are better suited for that job.
3744 static int find_new_ilb(int cpu)
3746 struct sched_domain *sd;
3747 struct sched_group *ilb_group;
3748 int ilb = nr_cpu_ids;
3751 * Have idle load balancer selection from semi-idle packages only
3752 * when power-aware load balancing is enabled
3754 if (!(sched_smt_power_savings || sched_mc_power_savings))
3758 * Optimize for the case when we have no idle CPUs or only one
3759 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3761 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3765 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3766 ilb_group = sd->groups;
3769 if (is_semi_idle_group(ilb_group)) {
3770 ilb = cpumask_first(nohz.grp_idle_mask);
3774 ilb_group = ilb_group->next;
3776 } while (ilb_group != sd->groups);
3784 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3785 static inline int find_new_ilb(int call_cpu)
3792 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3793 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3794 * CPU (if there is one).
3796 static void nohz_balancer_kick(int cpu)
3800 nohz.next_balance++;
3802 ilb_cpu = get_nohz_load_balancer();
3804 if (ilb_cpu >= nr_cpu_ids) {
3805 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3806 if (ilb_cpu >= nr_cpu_ids)
3810 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3811 struct call_single_data *cp;
3813 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3814 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3815 __smp_call_function_single(ilb_cpu, cp, 0);
3821 * This routine will try to nominate the ilb (idle load balancing)
3822 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3823 * load balancing on behalf of all those cpus.
3825 * When the ilb owner becomes busy, we will not have new ilb owner until some
3826 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3827 * idle load balancing by kicking one of the idle CPUs.
3829 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3830 * ilb owner CPU in future (when there is a need for idle load balancing on
3831 * behalf of all idle CPUs).
3833 void select_nohz_load_balancer(int stop_tick)
3835 int cpu = smp_processor_id();
3838 if (!cpu_active(cpu)) {
3839 if (atomic_read(&nohz.load_balancer) != cpu)
3843 * If we are going offline and still the leader,
3846 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3853 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3855 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3856 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3857 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3858 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3860 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3863 /* make me the ilb owner */
3864 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3869 * Check to see if there is a more power-efficient
3872 new_ilb = find_new_ilb(cpu);
3873 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3874 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3875 resched_cpu(new_ilb);
3881 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3884 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3886 if (atomic_read(&nohz.load_balancer) == cpu)
3887 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3895 static DEFINE_SPINLOCK(balancing);
3897 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3900 * Scale the max load_balance interval with the number of CPUs in the system.
3901 * This trades load-balance latency on larger machines for less cross talk.
3903 static void update_max_interval(void)
3905 max_load_balance_interval = HZ*num_online_cpus()/10;
3909 * It checks each scheduling domain to see if it is due to be balanced,
3910 * and initiates a balancing operation if so.
3912 * Balancing parameters are set up in arch_init_sched_domains.
3914 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3917 struct rq *rq = cpu_rq(cpu);
3918 unsigned long interval;
3919 struct sched_domain *sd;
3920 /* Earliest time when we have to do rebalance again */
3921 unsigned long next_balance = jiffies + 60*HZ;
3922 int update_next_balance = 0;
3928 for_each_domain(cpu, sd) {
3929 if (!(sd->flags & SD_LOAD_BALANCE))
3932 interval = sd->balance_interval;
3933 if (idle != CPU_IDLE)
3934 interval *= sd->busy_factor;
3936 /* scale ms to jiffies */
3937 interval = msecs_to_jiffies(interval);
3938 interval = clamp(interval, 1UL, max_load_balance_interval);
3940 need_serialize = sd->flags & SD_SERIALIZE;
3942 if (need_serialize) {
3943 if (!spin_trylock(&balancing))
3947 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3948 if (load_balance(cpu, rq, sd, idle, &balance)) {
3950 * We've pulled tasks over so either we're no
3953 idle = CPU_NOT_IDLE;
3955 sd->last_balance = jiffies;
3958 spin_unlock(&balancing);
3960 if (time_after(next_balance, sd->last_balance + interval)) {
3961 next_balance = sd->last_balance + interval;
3962 update_next_balance = 1;
3966 * Stop the load balance at this level. There is another
3967 * CPU in our sched group which is doing load balancing more
3976 * next_balance will be updated only when there is a need.
3977 * When the cpu is attached to null domain for ex, it will not be
3980 if (likely(update_next_balance))
3981 rq->next_balance = next_balance;
3986 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3987 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3989 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3991 struct rq *this_rq = cpu_rq(this_cpu);
3995 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3998 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3999 if (balance_cpu == this_cpu)
4003 * If this cpu gets work to do, stop the load balancing
4004 * work being done for other cpus. Next load
4005 * balancing owner will pick it up.
4007 if (need_resched()) {
4008 this_rq->nohz_balance_kick = 0;
4012 raw_spin_lock_irq(&this_rq->lock);
4013 update_rq_clock(this_rq);
4014 update_cpu_load(this_rq);
4015 raw_spin_unlock_irq(&this_rq->lock);
4017 rebalance_domains(balance_cpu, CPU_IDLE);
4019 rq = cpu_rq(balance_cpu);
4020 if (time_after(this_rq->next_balance, rq->next_balance))
4021 this_rq->next_balance = rq->next_balance;
4023 nohz.next_balance = this_rq->next_balance;
4024 this_rq->nohz_balance_kick = 0;
4028 * Current heuristic for kicking the idle load balancer
4029 * - first_pick_cpu is the one of the busy CPUs. It will kick
4030 * idle load balancer when it has more than one process active. This
4031 * eliminates the need for idle load balancing altogether when we have
4032 * only one running process in the system (common case).
4033 * - If there are more than one busy CPU, idle load balancer may have
4034 * to run for active_load_balance to happen (i.e., two busy CPUs are
4035 * SMT or core siblings and can run better if they move to different
4036 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4037 * which will kick idle load balancer as soon as it has any load.
4039 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4041 unsigned long now = jiffies;
4043 int first_pick_cpu, second_pick_cpu;
4045 if (time_before(now, nohz.next_balance))
4048 if (rq->idle_at_tick)
4051 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4052 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4054 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4055 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4058 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4059 if (ret == nr_cpu_ids || ret == cpu) {
4060 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4061 if (rq->nr_running > 1)
4064 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4065 if (ret == nr_cpu_ids || ret == cpu) {
4073 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4077 * run_rebalance_domains is triggered when needed from the scheduler tick.
4078 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4080 static void run_rebalance_domains(struct softirq_action *h)
4082 int this_cpu = smp_processor_id();
4083 struct rq *this_rq = cpu_rq(this_cpu);
4084 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4085 CPU_IDLE : CPU_NOT_IDLE;
4087 rebalance_domains(this_cpu, idle);
4090 * If this cpu has a pending nohz_balance_kick, then do the
4091 * balancing on behalf of the other idle cpus whose ticks are
4094 nohz_idle_balance(this_cpu, idle);
4097 static inline int on_null_domain(int cpu)
4099 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4103 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4105 static inline void trigger_load_balance(struct rq *rq, int cpu)
4107 /* Don't need to rebalance while attached to NULL domain */
4108 if (time_after_eq(jiffies, rq->next_balance) &&
4109 likely(!on_null_domain(cpu)))
4110 raise_softirq(SCHED_SOFTIRQ);
4112 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4113 nohz_balancer_kick(cpu);
4117 static void rq_online_fair(struct rq *rq)
4122 static void rq_offline_fair(struct rq *rq)
4127 #else /* CONFIG_SMP */
4130 * on UP we do not need to balance between CPUs:
4132 static inline void idle_balance(int cpu, struct rq *rq)
4136 #endif /* CONFIG_SMP */
4139 * scheduler tick hitting a task of our scheduling class:
4141 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4143 struct cfs_rq *cfs_rq;
4144 struct sched_entity *se = &curr->se;
4146 for_each_sched_entity(se) {
4147 cfs_rq = cfs_rq_of(se);
4148 entity_tick(cfs_rq, se, queued);
4153 * called on fork with the child task as argument from the parent's context
4154 * - child not yet on the tasklist
4155 * - preemption disabled
4157 static void task_fork_fair(struct task_struct *p)
4159 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4160 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4161 int this_cpu = smp_processor_id();
4162 struct rq *rq = this_rq();
4163 unsigned long flags;
4165 raw_spin_lock_irqsave(&rq->lock, flags);
4167 update_rq_clock(rq);
4169 if (unlikely(task_cpu(p) != this_cpu)) {
4171 __set_task_cpu(p, this_cpu);
4175 update_curr(cfs_rq);
4178 se->vruntime = curr->vruntime;
4179 place_entity(cfs_rq, se, 1);
4181 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4183 * Upon rescheduling, sched_class::put_prev_task() will place
4184 * 'current' within the tree based on its new key value.
4186 swap(curr->vruntime, se->vruntime);
4187 resched_task(rq->curr);
4190 se->vruntime -= cfs_rq->min_vruntime;
4192 raw_spin_unlock_irqrestore(&rq->lock, flags);
4196 * Priority of the task has changed. Check to see if we preempt
4200 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4206 * Reschedule if we are currently running on this runqueue and
4207 * our priority decreased, or if we are not currently running on
4208 * this runqueue and our priority is higher than the current's
4210 if (rq->curr == p) {
4211 if (p->prio > oldprio)
4212 resched_task(rq->curr);
4214 check_preempt_curr(rq, p, 0);
4217 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4219 struct sched_entity *se = &p->se;
4220 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4223 * Ensure the task's vruntime is normalized, so that when its
4224 * switched back to the fair class the enqueue_entity(.flags=0) will
4225 * do the right thing.
4227 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4228 * have normalized the vruntime, if it was !on_rq, then only when
4229 * the task is sleeping will it still have non-normalized vruntime.
4231 if (!se->on_rq && p->state != TASK_RUNNING) {
4233 * Fix up our vruntime so that the current sleep doesn't
4234 * cause 'unlimited' sleep bonus.
4236 place_entity(cfs_rq, se, 0);
4237 se->vruntime -= cfs_rq->min_vruntime;
4242 * We switched to the sched_fair class.
4244 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4250 * We were most likely switched from sched_rt, so
4251 * kick off the schedule if running, otherwise just see
4252 * if we can still preempt the current task.
4255 resched_task(rq->curr);
4257 check_preempt_curr(rq, p, 0);
4260 /* Account for a task changing its policy or group.
4262 * This routine is mostly called to set cfs_rq->curr field when a task
4263 * migrates between groups/classes.
4265 static void set_curr_task_fair(struct rq *rq)
4267 struct sched_entity *se = &rq->curr->se;
4269 for_each_sched_entity(se)
4270 set_next_entity(cfs_rq_of(se), se);
4273 #ifdef CONFIG_FAIR_GROUP_SCHED
4274 static void task_move_group_fair(struct task_struct *p, int on_rq)
4277 * If the task was not on the rq at the time of this cgroup movement
4278 * it must have been asleep, sleeping tasks keep their ->vruntime
4279 * absolute on their old rq until wakeup (needed for the fair sleeper
4280 * bonus in place_entity()).
4282 * If it was on the rq, we've just 'preempted' it, which does convert
4283 * ->vruntime to a relative base.
4285 * Make sure both cases convert their relative position when migrating
4286 * to another cgroup's rq. This does somewhat interfere with the
4287 * fair sleeper stuff for the first placement, but who cares.
4290 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4291 set_task_rq(p, task_cpu(p));
4293 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4297 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4299 struct sched_entity *se = &task->se;
4300 unsigned int rr_interval = 0;
4303 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4306 if (rq->cfs.load.weight)
4307 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4313 * All the scheduling class methods:
4315 static const struct sched_class fair_sched_class = {
4316 .next = &idle_sched_class,
4317 .enqueue_task = enqueue_task_fair,
4318 .dequeue_task = dequeue_task_fair,
4319 .yield_task = yield_task_fair,
4320 .yield_to_task = yield_to_task_fair,
4322 .check_preempt_curr = check_preempt_wakeup,
4324 .pick_next_task = pick_next_task_fair,
4325 .put_prev_task = put_prev_task_fair,
4328 .select_task_rq = select_task_rq_fair,
4330 .rq_online = rq_online_fair,
4331 .rq_offline = rq_offline_fair,
4333 .task_waking = task_waking_fair,
4336 .set_curr_task = set_curr_task_fair,
4337 .task_tick = task_tick_fair,
4338 .task_fork = task_fork_fair,
4340 .prio_changed = prio_changed_fair,
4341 .switched_from = switched_from_fair,
4342 .switched_to = switched_to_fair,
4344 .get_rr_interval = get_rr_interval_fair,
4346 #ifdef CONFIG_FAIR_GROUP_SCHED
4347 .task_move_group = task_move_group_fair,
4351 #ifdef CONFIG_SCHED_DEBUG
4352 static void print_cfs_stats(struct seq_file *m, int cpu)
4354 struct cfs_rq *cfs_rq;
4357 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4358 print_cfs_rq(m, cpu, cfs_rq);