2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 8;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
99 static const struct sched_class fair_sched_class;
101 /**************************************************************
102 * CFS operations on generic schedulable entities:
105 #ifdef CONFIG_FAIR_GROUP_SCHED
107 /* cpu runqueue to which this cfs_rq is attached */
108 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
113 /* An entity is a task if it doesn't "own" a runqueue */
114 #define entity_is_task(se) (!se->my_q)
116 static inline struct task_struct *task_of(struct sched_entity *se)
118 #ifdef CONFIG_SCHED_DEBUG
119 WARN_ON_ONCE(!entity_is_task(se));
121 return container_of(se, struct task_struct, se);
124 /* Walk up scheduling entities hierarchy */
125 #define for_each_sched_entity(se) \
126 for (; se; se = se->parent)
128 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
133 /* runqueue on which this entity is (to be) queued */
134 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
139 /* runqueue "owned" by this group */
140 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
145 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
146 * another cpu ('this_cpu')
148 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
150 return cfs_rq->tg->cfs_rq[this_cpu];
153 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
155 if (!cfs_rq->on_list) {
157 * Ensure we either appear before our parent (if already
158 * enqueued) or force our parent to appear after us when it is
159 * enqueued. The fact that we always enqueue bottom-up
160 * reduces this to two cases.
162 if (cfs_rq->tg->parent &&
163 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
164 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
165 &rq_of(cfs_rq)->leaf_cfs_rq_list);
167 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
168 &rq_of(cfs_rq)->leaf_cfs_rq_list);
175 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
177 if (cfs_rq->on_list) {
178 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
183 /* Iterate thr' all leaf cfs_rq's on a runqueue */
184 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
185 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
187 /* Do the two (enqueued) entities belong to the same group ? */
189 is_same_group(struct sched_entity *se, struct sched_entity *pse)
191 if (se->cfs_rq == pse->cfs_rq)
197 static inline struct sched_entity *parent_entity(struct sched_entity *se)
202 /* return depth at which a sched entity is present in the hierarchy */
203 static inline int depth_se(struct sched_entity *se)
207 for_each_sched_entity(se)
214 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
216 int se_depth, pse_depth;
219 * preemption test can be made between sibling entities who are in the
220 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
221 * both tasks until we find their ancestors who are siblings of common
225 /* First walk up until both entities are at same depth */
226 se_depth = depth_se(*se);
227 pse_depth = depth_se(*pse);
229 while (se_depth > pse_depth) {
231 *se = parent_entity(*se);
234 while (pse_depth > se_depth) {
236 *pse = parent_entity(*pse);
239 while (!is_same_group(*se, *pse)) {
240 *se = parent_entity(*se);
241 *pse = parent_entity(*pse);
245 #else /* !CONFIG_FAIR_GROUP_SCHED */
247 static inline struct task_struct *task_of(struct sched_entity *se)
249 return container_of(se, struct task_struct, se);
252 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 return container_of(cfs_rq, struct rq, cfs);
257 #define entity_is_task(se) 1
259 #define for_each_sched_entity(se) \
260 for (; se; se = NULL)
262 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
264 return &task_rq(p)->cfs;
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
269 struct task_struct *p = task_of(se);
270 struct rq *rq = task_rq(p);
275 /* runqueue "owned" by this group */
276 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
283 return &cpu_rq(this_cpu)->cfs;
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
295 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
298 is_same_group(struct sched_entity *se, struct sched_entity *pse)
303 static inline struct sched_entity *parent_entity(struct sched_entity *se)
309 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
313 #endif /* CONFIG_FAIR_GROUP_SCHED */
316 /**************************************************************
317 * Scheduling class tree data structure manipulation methods:
320 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
322 s64 delta = (s64)(vruntime - min_vruntime);
324 min_vruntime = vruntime;
329 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
331 s64 delta = (s64)(vruntime - min_vruntime);
333 min_vruntime = vruntime;
338 static inline int entity_before(struct sched_entity *a,
339 struct sched_entity *b)
341 return (s64)(a->vruntime - b->vruntime) < 0;
344 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
346 return se->vruntime - cfs_rq->min_vruntime;
349 static void update_min_vruntime(struct cfs_rq *cfs_rq)
351 u64 vruntime = cfs_rq->min_vruntime;
354 vruntime = cfs_rq->curr->vruntime;
356 if (cfs_rq->rb_leftmost) {
357 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
362 vruntime = se->vruntime;
364 vruntime = min_vruntime(vruntime, se->vruntime);
367 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
371 * Enqueue an entity into the rb-tree:
373 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
375 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
376 struct rb_node *parent = NULL;
377 struct sched_entity *entry;
378 s64 key = entity_key(cfs_rq, se);
382 * Find the right place in the rbtree:
386 entry = rb_entry(parent, struct sched_entity, run_node);
388 * We dont care about collisions. Nodes with
389 * the same key stay together.
391 if (key < entity_key(cfs_rq, entry)) {
392 link = &parent->rb_left;
394 link = &parent->rb_right;
400 * Maintain a cache of leftmost tree entries (it is frequently
404 cfs_rq->rb_leftmost = &se->run_node;
406 rb_link_node(&se->run_node, parent, link);
407 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
410 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
412 if (cfs_rq->rb_leftmost == &se->run_node) {
413 struct rb_node *next_node;
415 next_node = rb_next(&se->run_node);
416 cfs_rq->rb_leftmost = next_node;
419 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
422 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
424 struct rb_node *left = cfs_rq->rb_leftmost;
429 return rb_entry(left, struct sched_entity, run_node);
432 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
434 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
439 return rb_entry(last, struct sched_entity, run_node);
442 /**************************************************************
443 * Scheduling class statistics methods:
446 #ifdef CONFIG_SCHED_DEBUG
447 int sched_proc_update_handler(struct ctl_table *table, int write,
448 void __user *buffer, size_t *lenp,
451 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
452 int factor = get_update_sysctl_factor();
457 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
458 sysctl_sched_min_granularity);
460 #define WRT_SYSCTL(name) \
461 (normalized_sysctl_##name = sysctl_##name / (factor))
462 WRT_SYSCTL(sched_min_granularity);
463 WRT_SYSCTL(sched_latency);
464 WRT_SYSCTL(sched_wakeup_granularity);
474 static inline unsigned long
475 calc_delta_fair(unsigned long delta, struct sched_entity *se)
477 if (unlikely(se->load.weight != NICE_0_LOAD))
478 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
484 * The idea is to set a period in which each task runs once.
486 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
487 * this period because otherwise the slices get too small.
489 * p = (nr <= nl) ? l : l*nr/nl
491 static u64 __sched_period(unsigned long nr_running)
493 u64 period = sysctl_sched_latency;
494 unsigned long nr_latency = sched_nr_latency;
496 if (unlikely(nr_running > nr_latency)) {
497 period = sysctl_sched_min_granularity;
498 period *= nr_running;
505 * We calculate the wall-time slice from the period by taking a part
506 * proportional to the weight.
510 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
512 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
514 for_each_sched_entity(se) {
515 struct load_weight *load;
516 struct load_weight lw;
518 cfs_rq = cfs_rq_of(se);
519 load = &cfs_rq->load;
521 if (unlikely(!se->on_rq)) {
524 update_load_add(&lw, se->load.weight);
527 slice = calc_delta_mine(slice, se->load.weight, load);
533 * We calculate the vruntime slice of a to be inserted task
537 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
539 return calc_delta_fair(sched_slice(cfs_rq, se), se);
542 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
543 static void update_cfs_shares(struct cfs_rq *cfs_rq);
546 * Update the current task's runtime statistics. Skip current tasks that
547 * are not in our scheduling class.
550 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
551 unsigned long delta_exec)
553 unsigned long delta_exec_weighted;
555 schedstat_set(curr->statistics.exec_max,
556 max((u64)delta_exec, curr->statistics.exec_max));
558 curr->sum_exec_runtime += delta_exec;
559 schedstat_add(cfs_rq, exec_clock, delta_exec);
560 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
562 curr->vruntime += delta_exec_weighted;
563 update_min_vruntime(cfs_rq);
565 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
566 cfs_rq->load_unacc_exec_time += delta_exec;
570 static void update_curr(struct cfs_rq *cfs_rq)
572 struct sched_entity *curr = cfs_rq->curr;
573 u64 now = rq_of(cfs_rq)->clock_task;
574 unsigned long delta_exec;
580 * Get the amount of time the current task was running
581 * since the last time we changed load (this cannot
582 * overflow on 32 bits):
584 delta_exec = (unsigned long)(now - curr->exec_start);
588 __update_curr(cfs_rq, curr, delta_exec);
589 curr->exec_start = now;
591 if (entity_is_task(curr)) {
592 struct task_struct *curtask = task_of(curr);
594 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
595 cpuacct_charge(curtask, delta_exec);
596 account_group_exec_runtime(curtask, delta_exec);
601 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
603 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
607 * Task is being enqueued - update stats:
609 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
612 * Are we enqueueing a waiting task? (for current tasks
613 * a dequeue/enqueue event is a NOP)
615 if (se != cfs_rq->curr)
616 update_stats_wait_start(cfs_rq, se);
620 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
622 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
623 rq_of(cfs_rq)->clock - se->statistics.wait_start));
624 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
625 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
626 rq_of(cfs_rq)->clock - se->statistics.wait_start);
627 #ifdef CONFIG_SCHEDSTATS
628 if (entity_is_task(se)) {
629 trace_sched_stat_wait(task_of(se),
630 rq_of(cfs_rq)->clock - se->statistics.wait_start);
633 schedstat_set(se->statistics.wait_start, 0);
637 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 * Mark the end of the wait period if dequeueing a
643 if (se != cfs_rq->curr)
644 update_stats_wait_end(cfs_rq, se);
648 * We are picking a new current task - update its stats:
651 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
654 * We are starting a new run period:
656 se->exec_start = rq_of(cfs_rq)->clock_task;
659 /**************************************************
660 * Scheduling class queueing methods:
663 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
665 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
667 cfs_rq->task_weight += weight;
671 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
677 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
679 update_load_add(&cfs_rq->load, se->load.weight);
680 if (!parent_entity(se))
681 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
682 if (entity_is_task(se)) {
683 add_cfs_task_weight(cfs_rq, se->load.weight);
684 list_add(&se->group_node, &cfs_rq->tasks);
686 cfs_rq->nr_running++;
690 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
692 update_load_sub(&cfs_rq->load, se->load.weight);
693 if (!parent_entity(se))
694 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
695 if (entity_is_task(se)) {
696 add_cfs_task_weight(cfs_rq, -se->load.weight);
697 list_del_init(&se->group_node);
699 cfs_rq->nr_running--;
702 #ifdef CONFIG_FAIR_GROUP_SCHED
704 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
707 struct task_group *tg = cfs_rq->tg;
710 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
711 load_avg -= cfs_rq->load_contribution;
713 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
714 atomic_add(load_avg, &tg->load_weight);
715 cfs_rq->load_contribution += load_avg;
719 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
721 u64 period = sysctl_sched_shares_window;
723 unsigned long load = cfs_rq->load.weight;
725 if (cfs_rq->tg == &root_task_group)
728 now = rq_of(cfs_rq)->clock_task;
729 delta = now - cfs_rq->load_stamp;
731 /* truncate load history at 4 idle periods */
732 if (cfs_rq->load_stamp > cfs_rq->load_last &&
733 now - cfs_rq->load_last > 4 * period) {
734 cfs_rq->load_period = 0;
735 cfs_rq->load_avg = 0;
739 cfs_rq->load_stamp = now;
740 cfs_rq->load_unacc_exec_time = 0;
741 cfs_rq->load_period += delta;
743 cfs_rq->load_last = now;
744 cfs_rq->load_avg += delta * load;
747 /* consider updating load contribution on each fold or truncate */
748 if (global_update || cfs_rq->load_period > period
749 || !cfs_rq->load_period)
750 update_cfs_rq_load_contribution(cfs_rq, global_update);
752 while (cfs_rq->load_period > period) {
754 * Inline assembly required to prevent the compiler
755 * optimising this loop into a divmod call.
756 * See __iter_div_u64_rem() for another example of this.
758 asm("" : "+rm" (cfs_rq->load_period));
759 cfs_rq->load_period /= 2;
760 cfs_rq->load_avg /= 2;
763 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
764 list_del_leaf_cfs_rq(cfs_rq);
767 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
769 long load_weight, load, shares;
771 load = cfs_rq->load.weight;
773 load_weight = atomic_read(&tg->load_weight);
775 load_weight -= cfs_rq->load_contribution;
777 shares = (tg->shares * load);
779 shares /= load_weight;
781 if (shares < MIN_SHARES)
783 if (shares > tg->shares)
789 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
791 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
792 update_cfs_load(cfs_rq, 0);
793 update_cfs_shares(cfs_rq);
796 # else /* CONFIG_SMP */
797 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
801 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
806 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
809 # endif /* CONFIG_SMP */
810 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
811 unsigned long weight)
814 /* commit outstanding execution time */
815 if (cfs_rq->curr == se)
817 account_entity_dequeue(cfs_rq, se);
820 update_load_set(&se->load, weight);
823 account_entity_enqueue(cfs_rq, se);
826 static void update_cfs_shares(struct cfs_rq *cfs_rq)
828 struct task_group *tg;
829 struct sched_entity *se;
833 se = tg->se[cpu_of(rq_of(cfs_rq))];
837 if (likely(se->load.weight == tg->shares))
840 shares = calc_cfs_shares(cfs_rq, tg);
842 reweight_entity(cfs_rq_of(se), se, shares);
844 #else /* CONFIG_FAIR_GROUP_SCHED */
845 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
849 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
853 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
856 #endif /* CONFIG_FAIR_GROUP_SCHED */
858 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
860 #ifdef CONFIG_SCHEDSTATS
861 struct task_struct *tsk = NULL;
863 if (entity_is_task(se))
866 if (se->statistics.sleep_start) {
867 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
872 if (unlikely(delta > se->statistics.sleep_max))
873 se->statistics.sleep_max = delta;
875 se->statistics.sleep_start = 0;
876 se->statistics.sum_sleep_runtime += delta;
879 account_scheduler_latency(tsk, delta >> 10, 1);
880 trace_sched_stat_sleep(tsk, delta);
883 if (se->statistics.block_start) {
884 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
889 if (unlikely(delta > se->statistics.block_max))
890 se->statistics.block_max = delta;
892 se->statistics.block_start = 0;
893 se->statistics.sum_sleep_runtime += delta;
896 if (tsk->in_iowait) {
897 se->statistics.iowait_sum += delta;
898 se->statistics.iowait_count++;
899 trace_sched_stat_iowait(tsk, delta);
903 * Blocking time is in units of nanosecs, so shift by
904 * 20 to get a milliseconds-range estimation of the
905 * amount of time that the task spent sleeping:
907 if (unlikely(prof_on == SLEEP_PROFILING)) {
908 profile_hits(SLEEP_PROFILING,
909 (void *)get_wchan(tsk),
912 account_scheduler_latency(tsk, delta >> 10, 0);
918 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
920 #ifdef CONFIG_SCHED_DEBUG
921 s64 d = se->vruntime - cfs_rq->min_vruntime;
926 if (d > 3*sysctl_sched_latency)
927 schedstat_inc(cfs_rq, nr_spread_over);
932 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
934 u64 vruntime = cfs_rq->min_vruntime;
937 * The 'current' period is already promised to the current tasks,
938 * however the extra weight of the new task will slow them down a
939 * little, place the new task so that it fits in the slot that
940 * stays open at the end.
942 if (initial && sched_feat(START_DEBIT))
943 vruntime += sched_vslice(cfs_rq, se);
945 /* sleeps up to a single latency don't count. */
947 unsigned long thresh = sysctl_sched_latency;
950 * Halve their sleep time's effect, to allow
951 * for a gentler effect of sleepers:
953 if (sched_feat(GENTLE_FAIR_SLEEPERS))
959 /* ensure we never gain time by being placed backwards. */
960 vruntime = max_vruntime(se->vruntime, vruntime);
962 se->vruntime = vruntime;
966 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
969 * Update the normalized vruntime before updating min_vruntime
970 * through callig update_curr().
972 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
973 se->vruntime += cfs_rq->min_vruntime;
976 * Update run-time statistics of the 'current'.
979 update_cfs_load(cfs_rq, 0);
980 account_entity_enqueue(cfs_rq, se);
981 update_cfs_shares(cfs_rq);
983 if (flags & ENQUEUE_WAKEUP) {
984 place_entity(cfs_rq, se, 0);
985 enqueue_sleeper(cfs_rq, se);
988 update_stats_enqueue(cfs_rq, se);
989 check_spread(cfs_rq, se);
990 if (se != cfs_rq->curr)
991 __enqueue_entity(cfs_rq, se);
994 if (cfs_rq->nr_running == 1)
995 list_add_leaf_cfs_rq(cfs_rq);
998 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1000 if (!se || cfs_rq->last == se)
1001 cfs_rq->last = NULL;
1003 if (!se || cfs_rq->next == se)
1004 cfs_rq->next = NULL;
1007 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1009 for_each_sched_entity(se)
1010 __clear_buddies(cfs_rq_of(se), se);
1014 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1017 * Update run-time statistics of the 'current'.
1019 update_curr(cfs_rq);
1021 update_stats_dequeue(cfs_rq, se);
1022 if (flags & DEQUEUE_SLEEP) {
1023 #ifdef CONFIG_SCHEDSTATS
1024 if (entity_is_task(se)) {
1025 struct task_struct *tsk = task_of(se);
1027 if (tsk->state & TASK_INTERRUPTIBLE)
1028 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1029 if (tsk->state & TASK_UNINTERRUPTIBLE)
1030 se->statistics.block_start = rq_of(cfs_rq)->clock;
1035 clear_buddies(cfs_rq, se);
1037 if (se != cfs_rq->curr)
1038 __dequeue_entity(cfs_rq, se);
1040 update_cfs_load(cfs_rq, 0);
1041 account_entity_dequeue(cfs_rq, se);
1042 update_min_vruntime(cfs_rq);
1043 update_cfs_shares(cfs_rq);
1046 * Normalize the entity after updating the min_vruntime because the
1047 * update can refer to the ->curr item and we need to reflect this
1048 * movement in our normalized position.
1050 if (!(flags & DEQUEUE_SLEEP))
1051 se->vruntime -= cfs_rq->min_vruntime;
1055 * Preempt the current task with a newly woken task if needed:
1058 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1060 unsigned long ideal_runtime, delta_exec;
1062 ideal_runtime = sched_slice(cfs_rq, curr);
1063 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1064 if (delta_exec > ideal_runtime) {
1065 resched_task(rq_of(cfs_rq)->curr);
1067 * The current task ran long enough, ensure it doesn't get
1068 * re-elected due to buddy favours.
1070 clear_buddies(cfs_rq, curr);
1075 * Ensure that a task that missed wakeup preemption by a
1076 * narrow margin doesn't have to wait for a full slice.
1077 * This also mitigates buddy induced latencies under load.
1079 if (!sched_feat(WAKEUP_PREEMPT))
1082 if (delta_exec < sysctl_sched_min_granularity)
1085 if (cfs_rq->nr_running > 1) {
1086 struct sched_entity *se = __pick_next_entity(cfs_rq);
1087 s64 delta = curr->vruntime - se->vruntime;
1092 if (delta > ideal_runtime)
1093 resched_task(rq_of(cfs_rq)->curr);
1098 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1100 /* 'current' is not kept within the tree. */
1103 * Any task has to be enqueued before it get to execute on
1104 * a CPU. So account for the time it spent waiting on the
1107 update_stats_wait_end(cfs_rq, se);
1108 __dequeue_entity(cfs_rq, se);
1111 update_stats_curr_start(cfs_rq, se);
1113 #ifdef CONFIG_SCHEDSTATS
1115 * Track our maximum slice length, if the CPU's load is at
1116 * least twice that of our own weight (i.e. dont track it
1117 * when there are only lesser-weight tasks around):
1119 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1120 se->statistics.slice_max = max(se->statistics.slice_max,
1121 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1124 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1128 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1130 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1132 struct sched_entity *se = __pick_next_entity(cfs_rq);
1133 struct sched_entity *left = se;
1135 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1139 * Prefer last buddy, try to return the CPU to a preempted task.
1141 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1144 clear_buddies(cfs_rq, se);
1149 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1152 * If still on the runqueue then deactivate_task()
1153 * was not called and update_curr() has to be done:
1156 update_curr(cfs_rq);
1158 check_spread(cfs_rq, prev);
1160 update_stats_wait_start(cfs_rq, prev);
1161 /* Put 'current' back into the tree. */
1162 __enqueue_entity(cfs_rq, prev);
1164 cfs_rq->curr = NULL;
1168 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1171 * Update run-time statistics of the 'current'.
1173 update_curr(cfs_rq);
1176 * Update share accounting for long-running entities.
1178 update_entity_shares_tick(cfs_rq);
1180 #ifdef CONFIG_SCHED_HRTICK
1182 * queued ticks are scheduled to match the slice, so don't bother
1183 * validating it and just reschedule.
1186 resched_task(rq_of(cfs_rq)->curr);
1190 * don't let the period tick interfere with the hrtick preemption
1192 if (!sched_feat(DOUBLE_TICK) &&
1193 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1197 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1198 check_preempt_tick(cfs_rq, curr);
1201 /**************************************************
1202 * CFS operations on tasks:
1205 #ifdef CONFIG_SCHED_HRTICK
1206 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1208 struct sched_entity *se = &p->se;
1209 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1211 WARN_ON(task_rq(p) != rq);
1213 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1214 u64 slice = sched_slice(cfs_rq, se);
1215 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1216 s64 delta = slice - ran;
1225 * Don't schedule slices shorter than 10000ns, that just
1226 * doesn't make sense. Rely on vruntime for fairness.
1229 delta = max_t(s64, 10000LL, delta);
1231 hrtick_start(rq, delta);
1236 * called from enqueue/dequeue and updates the hrtick when the
1237 * current task is from our class and nr_running is low enough
1240 static void hrtick_update(struct rq *rq)
1242 struct task_struct *curr = rq->curr;
1244 if (curr->sched_class != &fair_sched_class)
1247 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1248 hrtick_start_fair(rq, curr);
1250 #else /* !CONFIG_SCHED_HRTICK */
1252 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1256 static inline void hrtick_update(struct rq *rq)
1262 * The enqueue_task method is called before nr_running is
1263 * increased. Here we update the fair scheduling stats and
1264 * then put the task into the rbtree:
1267 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1269 struct cfs_rq *cfs_rq;
1270 struct sched_entity *se = &p->se;
1272 for_each_sched_entity(se) {
1275 cfs_rq = cfs_rq_of(se);
1276 enqueue_entity(cfs_rq, se, flags);
1277 flags = ENQUEUE_WAKEUP;
1280 for_each_sched_entity(se) {
1281 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1283 update_cfs_load(cfs_rq, 0);
1284 update_cfs_shares(cfs_rq);
1291 * The dequeue_task method is called before nr_running is
1292 * decreased. We remove the task from the rbtree and
1293 * update the fair scheduling stats:
1295 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1297 struct cfs_rq *cfs_rq;
1298 struct sched_entity *se = &p->se;
1300 for_each_sched_entity(se) {
1301 cfs_rq = cfs_rq_of(se);
1302 dequeue_entity(cfs_rq, se, flags);
1304 /* Don't dequeue parent if it has other entities besides us */
1305 if (cfs_rq->load.weight)
1307 flags |= DEQUEUE_SLEEP;
1310 for_each_sched_entity(se) {
1311 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1313 update_cfs_load(cfs_rq, 0);
1314 update_cfs_shares(cfs_rq);
1321 * sched_yield() support is very simple - we dequeue and enqueue.
1323 * If compat_yield is turned on then we requeue to the end of the tree.
1325 static void yield_task_fair(struct rq *rq)
1327 struct task_struct *curr = rq->curr;
1328 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1329 struct sched_entity *rightmost, *se = &curr->se;
1332 * Are we the only task in the tree?
1334 if (unlikely(rq->nr_running == 1))
1337 clear_buddies(cfs_rq, se);
1339 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1340 update_rq_clock(rq);
1342 * Update run-time statistics of the 'current'.
1344 update_curr(cfs_rq);
1349 * Find the rightmost entry in the rbtree:
1351 rightmost = __pick_last_entity(cfs_rq);
1353 * Already in the rightmost position?
1355 if (unlikely(!rightmost || entity_before(rightmost, se)))
1359 * Minimally necessary key value to be last in the tree:
1360 * Upon rescheduling, sched_class::put_prev_task() will place
1361 * 'current' within the tree based on its new key value.
1363 se->vruntime = rightmost->vruntime + 1;
1368 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1370 struct sched_entity *se = &p->se;
1371 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1373 se->vruntime -= cfs_rq->min_vruntime;
1376 #ifdef CONFIG_FAIR_GROUP_SCHED
1378 * effective_load() calculates the load change as seen from the root_task_group
1380 * Adding load to a group doesn't make a group heavier, but can cause movement
1381 * of group shares between cpus. Assuming the shares were perfectly aligned one
1382 * can calculate the shift in shares.
1384 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1386 struct sched_entity *se = tg->se[cpu];
1391 for_each_sched_entity(se) {
1395 w = se->my_q->load.weight;
1397 /* use this cpu's instantaneous contribution */
1398 lw = atomic_read(&tg->load_weight);
1399 lw -= se->my_q->load_contribution;
1404 if (lw > 0 && wl < lw)
1405 wl = (wl * tg->shares) / lw;
1409 /* zero point is MIN_SHARES */
1410 if (wl < MIN_SHARES)
1412 wl -= se->load.weight;
1421 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1422 unsigned long wl, unsigned long wg)
1429 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1431 s64 this_load, load;
1432 int idx, this_cpu, prev_cpu;
1433 unsigned long tl_per_task;
1434 struct task_group *tg;
1435 unsigned long weight;
1439 this_cpu = smp_processor_id();
1440 prev_cpu = task_cpu(p);
1441 load = source_load(prev_cpu, idx);
1442 this_load = target_load(this_cpu, idx);
1445 * If sync wakeup then subtract the (maximum possible)
1446 * effect of the currently running task from the load
1447 * of the current CPU:
1451 tg = task_group(current);
1452 weight = current->se.load.weight;
1454 this_load += effective_load(tg, this_cpu, -weight, -weight);
1455 load += effective_load(tg, prev_cpu, 0, -weight);
1459 weight = p->se.load.weight;
1462 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1463 * due to the sync cause above having dropped this_load to 0, we'll
1464 * always have an imbalance, but there's really nothing you can do
1465 * about that, so that's good too.
1467 * Otherwise check if either cpus are near enough in load to allow this
1468 * task to be woken on this_cpu.
1470 if (this_load > 0) {
1471 s64 this_eff_load, prev_eff_load;
1473 this_eff_load = 100;
1474 this_eff_load *= power_of(prev_cpu);
1475 this_eff_load *= this_load +
1476 effective_load(tg, this_cpu, weight, weight);
1478 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1479 prev_eff_load *= power_of(this_cpu);
1480 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1482 balanced = this_eff_load <= prev_eff_load;
1488 * If the currently running task will sleep within
1489 * a reasonable amount of time then attract this newly
1492 if (sync && balanced)
1495 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1496 tl_per_task = cpu_avg_load_per_task(this_cpu);
1499 (this_load <= load &&
1500 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1502 * This domain has SD_WAKE_AFFINE and
1503 * p is cache cold in this domain, and
1504 * there is no bad imbalance.
1506 schedstat_inc(sd, ttwu_move_affine);
1507 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1515 * find_idlest_group finds and returns the least busy CPU group within the
1518 static struct sched_group *
1519 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1520 int this_cpu, int load_idx)
1522 struct sched_group *idlest = NULL, *group = sd->groups;
1523 unsigned long min_load = ULONG_MAX, this_load = 0;
1524 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1527 unsigned long load, avg_load;
1531 /* Skip over this group if it has no CPUs allowed */
1532 if (!cpumask_intersects(sched_group_cpus(group),
1536 local_group = cpumask_test_cpu(this_cpu,
1537 sched_group_cpus(group));
1539 /* Tally up the load of all CPUs in the group */
1542 for_each_cpu(i, sched_group_cpus(group)) {
1543 /* Bias balancing toward cpus of our domain */
1545 load = source_load(i, load_idx);
1547 load = target_load(i, load_idx);
1552 /* Adjust by relative CPU power of the group */
1553 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1556 this_load = avg_load;
1557 } else if (avg_load < min_load) {
1558 min_load = avg_load;
1561 } while (group = group->next, group != sd->groups);
1563 if (!idlest || 100*this_load < imbalance*min_load)
1569 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1572 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1574 unsigned long load, min_load = ULONG_MAX;
1578 /* Traverse only the allowed CPUs */
1579 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1580 load = weighted_cpuload(i);
1582 if (load < min_load || (load == min_load && i == this_cpu)) {
1592 * Try and locate an idle CPU in the sched_domain.
1594 static int select_idle_sibling(struct task_struct *p, int target)
1596 int cpu = smp_processor_id();
1597 int prev_cpu = task_cpu(p);
1598 struct sched_domain *sd;
1602 * If the task is going to be woken-up on this cpu and if it is
1603 * already idle, then it is the right target.
1605 if (target == cpu && idle_cpu(cpu))
1609 * If the task is going to be woken-up on the cpu where it previously
1610 * ran and if it is currently idle, then it the right target.
1612 if (target == prev_cpu && idle_cpu(prev_cpu))
1616 * Otherwise, iterate the domains and find an elegible idle cpu.
1618 for_each_domain(target, sd) {
1619 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1622 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1630 * Lets stop looking for an idle sibling when we reached
1631 * the domain that spans the current cpu and prev_cpu.
1633 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1634 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1642 * sched_balance_self: balance the current task (running on cpu) in domains
1643 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1646 * Balance, ie. select the least loaded group.
1648 * Returns the target CPU number, or the same CPU if no balancing is needed.
1650 * preempt must be disabled.
1653 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1655 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1656 int cpu = smp_processor_id();
1657 int prev_cpu = task_cpu(p);
1659 int want_affine = 0;
1661 int sync = wake_flags & WF_SYNC;
1663 if (sd_flag & SD_BALANCE_WAKE) {
1664 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1669 for_each_domain(cpu, tmp) {
1670 if (!(tmp->flags & SD_LOAD_BALANCE))
1674 * If power savings logic is enabled for a domain, see if we
1675 * are not overloaded, if so, don't balance wider.
1677 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1678 unsigned long power = 0;
1679 unsigned long nr_running = 0;
1680 unsigned long capacity;
1683 for_each_cpu(i, sched_domain_span(tmp)) {
1684 power += power_of(i);
1685 nr_running += cpu_rq(i)->cfs.nr_running;
1688 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1690 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1693 if (nr_running < capacity)
1698 * If both cpu and prev_cpu are part of this domain,
1699 * cpu is a valid SD_WAKE_AFFINE target.
1701 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1702 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1707 if (!want_sd && !want_affine)
1710 if (!(tmp->flags & sd_flag))
1718 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1719 return select_idle_sibling(p, cpu);
1721 return select_idle_sibling(p, prev_cpu);
1725 int load_idx = sd->forkexec_idx;
1726 struct sched_group *group;
1729 if (!(sd->flags & sd_flag)) {
1734 if (sd_flag & SD_BALANCE_WAKE)
1735 load_idx = sd->wake_idx;
1737 group = find_idlest_group(sd, p, cpu, load_idx);
1743 new_cpu = find_idlest_cpu(group, p, cpu);
1744 if (new_cpu == -1 || new_cpu == cpu) {
1745 /* Now try balancing at a lower domain level of cpu */
1750 /* Now try balancing at a lower domain level of new_cpu */
1752 weight = sd->span_weight;
1754 for_each_domain(cpu, tmp) {
1755 if (weight <= tmp->span_weight)
1757 if (tmp->flags & sd_flag)
1760 /* while loop will break here if sd == NULL */
1765 #endif /* CONFIG_SMP */
1767 static unsigned long
1768 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1770 unsigned long gran = sysctl_sched_wakeup_granularity;
1773 * Since its curr running now, convert the gran from real-time
1774 * to virtual-time in his units.
1776 * By using 'se' instead of 'curr' we penalize light tasks, so
1777 * they get preempted easier. That is, if 'se' < 'curr' then
1778 * the resulting gran will be larger, therefore penalizing the
1779 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1780 * be smaller, again penalizing the lighter task.
1782 * This is especially important for buddies when the leftmost
1783 * task is higher priority than the buddy.
1785 if (unlikely(se->load.weight != NICE_0_LOAD))
1786 gran = calc_delta_fair(gran, se);
1792 * Should 'se' preempt 'curr'.
1806 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1808 s64 gran, vdiff = curr->vruntime - se->vruntime;
1813 gran = wakeup_gran(curr, se);
1820 static void set_last_buddy(struct sched_entity *se)
1822 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1823 for_each_sched_entity(se)
1824 cfs_rq_of(se)->last = se;
1828 static void set_next_buddy(struct sched_entity *se)
1830 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1831 for_each_sched_entity(se)
1832 cfs_rq_of(se)->next = se;
1837 * Preempt the current task with a newly woken task if needed:
1839 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1841 struct task_struct *curr = rq->curr;
1842 struct sched_entity *se = &curr->se, *pse = &p->se;
1843 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1844 int scale = cfs_rq->nr_running >= sched_nr_latency;
1846 if (unlikely(se == pse))
1849 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1850 set_next_buddy(pse);
1853 * We can come here with TIF_NEED_RESCHED already set from new task
1856 if (test_tsk_need_resched(curr))
1860 * Batch and idle tasks do not preempt (their preemption is driven by
1863 if (unlikely(p->policy != SCHED_NORMAL))
1866 /* Idle tasks are by definition preempted by everybody. */
1867 if (unlikely(curr->policy == SCHED_IDLE))
1870 if (!sched_feat(WAKEUP_PREEMPT))
1873 update_curr(cfs_rq);
1874 find_matching_se(&se, &pse);
1876 if (wakeup_preempt_entity(se, pse) == 1)
1884 * Only set the backward buddy when the current task is still
1885 * on the rq. This can happen when a wakeup gets interleaved
1886 * with schedule on the ->pre_schedule() or idle_balance()
1887 * point, either of which can * drop the rq lock.
1889 * Also, during early boot the idle thread is in the fair class,
1890 * for obvious reasons its a bad idea to schedule back to it.
1892 if (unlikely(!se->on_rq || curr == rq->idle))
1895 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1899 static struct task_struct *pick_next_task_fair(struct rq *rq)
1901 struct task_struct *p;
1902 struct cfs_rq *cfs_rq = &rq->cfs;
1903 struct sched_entity *se;
1905 if (!cfs_rq->nr_running)
1909 se = pick_next_entity(cfs_rq);
1910 set_next_entity(cfs_rq, se);
1911 cfs_rq = group_cfs_rq(se);
1915 hrtick_start_fair(rq, p);
1921 * Account for a descheduled task:
1923 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1925 struct sched_entity *se = &prev->se;
1926 struct cfs_rq *cfs_rq;
1928 for_each_sched_entity(se) {
1929 cfs_rq = cfs_rq_of(se);
1930 put_prev_entity(cfs_rq, se);
1935 /**************************************************
1936 * Fair scheduling class load-balancing methods:
1940 * pull_task - move a task from a remote runqueue to the local runqueue.
1941 * Both runqueues must be locked.
1943 static void pull_task(struct rq *src_rq, struct task_struct *p,
1944 struct rq *this_rq, int this_cpu)
1946 deactivate_task(src_rq, p, 0);
1947 set_task_cpu(p, this_cpu);
1948 activate_task(this_rq, p, 0);
1949 check_preempt_curr(this_rq, p, 0);
1953 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1956 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1957 struct sched_domain *sd, enum cpu_idle_type idle,
1960 int tsk_cache_hot = 0;
1962 * We do not migrate tasks that are:
1963 * 1) running (obviously), or
1964 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1965 * 3) are cache-hot on their current CPU.
1967 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1968 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1973 if (task_running(rq, p)) {
1974 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1979 * Aggressive migration if:
1980 * 1) task is cache cold, or
1981 * 2) too many balance attempts have failed.
1984 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
1985 if (!tsk_cache_hot ||
1986 sd->nr_balance_failed > sd->cache_nice_tries) {
1987 #ifdef CONFIG_SCHEDSTATS
1988 if (tsk_cache_hot) {
1989 schedstat_inc(sd, lb_hot_gained[idle]);
1990 schedstat_inc(p, se.statistics.nr_forced_migrations);
1996 if (tsk_cache_hot) {
1997 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2004 * move_one_task tries to move exactly one task from busiest to this_rq, as
2005 * part of active balancing operations within "domain".
2006 * Returns 1 if successful and 0 otherwise.
2008 * Called with both runqueues locked.
2011 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2012 struct sched_domain *sd, enum cpu_idle_type idle)
2014 struct task_struct *p, *n;
2015 struct cfs_rq *cfs_rq;
2018 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2019 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2021 if (!can_migrate_task(p, busiest, this_cpu,
2025 pull_task(busiest, p, this_rq, this_cpu);
2027 * Right now, this is only the second place pull_task()
2028 * is called, so we can safely collect pull_task()
2029 * stats here rather than inside pull_task().
2031 schedstat_inc(sd, lb_gained[idle]);
2039 static unsigned long
2040 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2041 unsigned long max_load_move, struct sched_domain *sd,
2042 enum cpu_idle_type idle, int *all_pinned,
2043 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
2045 int loops = 0, pulled = 0, pinned = 0;
2046 long rem_load_move = max_load_move;
2047 struct task_struct *p, *n;
2049 if (max_load_move == 0)
2054 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2055 if (loops++ > sysctl_sched_nr_migrate)
2058 if ((p->se.load.weight >> 1) > rem_load_move ||
2059 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
2062 pull_task(busiest, p, this_rq, this_cpu);
2064 rem_load_move -= p->se.load.weight;
2066 #ifdef CONFIG_PREEMPT
2068 * NEWIDLE balancing is a source of latency, so preemptible
2069 * kernels will stop after the first task is pulled to minimize
2070 * the critical section.
2072 if (idle == CPU_NEWLY_IDLE)
2077 * We only want to steal up to the prescribed amount of
2080 if (rem_load_move <= 0)
2083 if (p->prio < *this_best_prio)
2084 *this_best_prio = p->prio;
2088 * Right now, this is one of only two places pull_task() is called,
2089 * so we can safely collect pull_task() stats here rather than
2090 * inside pull_task().
2092 schedstat_add(sd, lb_gained[idle], pulled);
2095 *all_pinned = pinned;
2097 return max_load_move - rem_load_move;
2100 #ifdef CONFIG_FAIR_GROUP_SCHED
2102 * update tg->load_weight by folding this cpu's load_avg
2104 static int update_shares_cpu(struct task_group *tg, int cpu)
2106 struct cfs_rq *cfs_rq;
2107 unsigned long flags;
2114 cfs_rq = tg->cfs_rq[cpu];
2116 raw_spin_lock_irqsave(&rq->lock, flags);
2118 update_rq_clock(rq);
2119 update_cfs_load(cfs_rq, 1);
2122 * We need to update shares after updating tg->load_weight in
2123 * order to adjust the weight of groups with long running tasks.
2125 update_cfs_shares(cfs_rq);
2127 raw_spin_unlock_irqrestore(&rq->lock, flags);
2132 static void update_shares(int cpu)
2134 struct cfs_rq *cfs_rq;
2135 struct rq *rq = cpu_rq(cpu);
2138 for_each_leaf_cfs_rq(rq, cfs_rq)
2139 update_shares_cpu(cfs_rq->tg, cpu);
2143 static unsigned long
2144 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2145 unsigned long max_load_move,
2146 struct sched_domain *sd, enum cpu_idle_type idle,
2147 int *all_pinned, int *this_best_prio)
2149 long rem_load_move = max_load_move;
2150 int busiest_cpu = cpu_of(busiest);
2151 struct task_group *tg;
2154 update_h_load(busiest_cpu);
2156 list_for_each_entry_rcu(tg, &task_groups, list) {
2157 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2158 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2159 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2160 u64 rem_load, moved_load;
2165 if (!busiest_cfs_rq->task_weight)
2168 rem_load = (u64)rem_load_move * busiest_weight;
2169 rem_load = div_u64(rem_load, busiest_h_load + 1);
2171 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2172 rem_load, sd, idle, all_pinned, this_best_prio,
2178 moved_load *= busiest_h_load;
2179 moved_load = div_u64(moved_load, busiest_weight + 1);
2181 rem_load_move -= moved_load;
2182 if (rem_load_move < 0)
2187 return max_load_move - rem_load_move;
2190 static inline void update_shares(int cpu)
2194 static unsigned long
2195 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2196 unsigned long max_load_move,
2197 struct sched_domain *sd, enum cpu_idle_type idle,
2198 int *all_pinned, int *this_best_prio)
2200 return balance_tasks(this_rq, this_cpu, busiest,
2201 max_load_move, sd, idle, all_pinned,
2202 this_best_prio, &busiest->cfs);
2207 * move_tasks tries to move up to max_load_move weighted load from busiest to
2208 * this_rq, as part of a balancing operation within domain "sd".
2209 * Returns 1 if successful and 0 otherwise.
2211 * Called with both runqueues locked.
2213 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2214 unsigned long max_load_move,
2215 struct sched_domain *sd, enum cpu_idle_type idle,
2218 unsigned long total_load_moved = 0, load_moved;
2219 int this_best_prio = this_rq->curr->prio;
2222 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2223 max_load_move - total_load_moved,
2224 sd, idle, all_pinned, &this_best_prio);
2226 total_load_moved += load_moved;
2228 #ifdef CONFIG_PREEMPT
2230 * NEWIDLE balancing is a source of latency, so preemptible
2231 * kernels will stop after the first task is pulled to minimize
2232 * the critical section.
2234 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2237 if (raw_spin_is_contended(&this_rq->lock) ||
2238 raw_spin_is_contended(&busiest->lock))
2241 } while (load_moved && max_load_move > total_load_moved);
2243 return total_load_moved > 0;
2246 /********** Helpers for find_busiest_group ************************/
2248 * sd_lb_stats - Structure to store the statistics of a sched_domain
2249 * during load balancing.
2251 struct sd_lb_stats {
2252 struct sched_group *busiest; /* Busiest group in this sd */
2253 struct sched_group *this; /* Local group in this sd */
2254 unsigned long total_load; /* Total load of all groups in sd */
2255 unsigned long total_pwr; /* Total power of all groups in sd */
2256 unsigned long avg_load; /* Average load across all groups in sd */
2258 /** Statistics of this group */
2259 unsigned long this_load;
2260 unsigned long this_load_per_task;
2261 unsigned long this_nr_running;
2262 unsigned long this_has_capacity;
2263 unsigned int this_idle_cpus;
2265 /* Statistics of the busiest group */
2266 unsigned int busiest_idle_cpus;
2267 unsigned long max_load;
2268 unsigned long busiest_load_per_task;
2269 unsigned long busiest_nr_running;
2270 unsigned long busiest_group_capacity;
2271 unsigned long busiest_has_capacity;
2272 unsigned int busiest_group_weight;
2274 int group_imb; /* Is there imbalance in this sd */
2275 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2276 int power_savings_balance; /* Is powersave balance needed for this sd */
2277 struct sched_group *group_min; /* Least loaded group in sd */
2278 struct sched_group *group_leader; /* Group which relieves group_min */
2279 unsigned long min_load_per_task; /* load_per_task in group_min */
2280 unsigned long leader_nr_running; /* Nr running of group_leader */
2281 unsigned long min_nr_running; /* Nr running of group_min */
2286 * sg_lb_stats - stats of a sched_group required for load_balancing
2288 struct sg_lb_stats {
2289 unsigned long avg_load; /*Avg load across the CPUs of the group */
2290 unsigned long group_load; /* Total load over the CPUs of the group */
2291 unsigned long sum_nr_running; /* Nr tasks running in the group */
2292 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2293 unsigned long group_capacity;
2294 unsigned long idle_cpus;
2295 unsigned long group_weight;
2296 int group_imb; /* Is there an imbalance in the group ? */
2297 int group_has_capacity; /* Is there extra capacity in the group? */
2301 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2302 * @group: The group whose first cpu is to be returned.
2304 static inline unsigned int group_first_cpu(struct sched_group *group)
2306 return cpumask_first(sched_group_cpus(group));
2310 * get_sd_load_idx - Obtain the load index for a given sched domain.
2311 * @sd: The sched_domain whose load_idx is to be obtained.
2312 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2314 static inline int get_sd_load_idx(struct sched_domain *sd,
2315 enum cpu_idle_type idle)
2321 load_idx = sd->busy_idx;
2324 case CPU_NEWLY_IDLE:
2325 load_idx = sd->newidle_idx;
2328 load_idx = sd->idle_idx;
2336 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2338 * init_sd_power_savings_stats - Initialize power savings statistics for
2339 * the given sched_domain, during load balancing.
2341 * @sd: Sched domain whose power-savings statistics are to be initialized.
2342 * @sds: Variable containing the statistics for sd.
2343 * @idle: Idle status of the CPU at which we're performing load-balancing.
2345 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2346 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2349 * Busy processors will not participate in power savings
2352 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2353 sds->power_savings_balance = 0;
2355 sds->power_savings_balance = 1;
2356 sds->min_nr_running = ULONG_MAX;
2357 sds->leader_nr_running = 0;
2362 * update_sd_power_savings_stats - Update the power saving stats for a
2363 * sched_domain while performing load balancing.
2365 * @group: sched_group belonging to the sched_domain under consideration.
2366 * @sds: Variable containing the statistics of the sched_domain
2367 * @local_group: Does group contain the CPU for which we're performing
2369 * @sgs: Variable containing the statistics of the group.
2371 static inline void update_sd_power_savings_stats(struct sched_group *group,
2372 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2375 if (!sds->power_savings_balance)
2379 * If the local group is idle or completely loaded
2380 * no need to do power savings balance at this domain
2382 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2383 !sds->this_nr_running))
2384 sds->power_savings_balance = 0;
2387 * If a group is already running at full capacity or idle,
2388 * don't include that group in power savings calculations
2390 if (!sds->power_savings_balance ||
2391 sgs->sum_nr_running >= sgs->group_capacity ||
2392 !sgs->sum_nr_running)
2396 * Calculate the group which has the least non-idle load.
2397 * This is the group from where we need to pick up the load
2400 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2401 (sgs->sum_nr_running == sds->min_nr_running &&
2402 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2403 sds->group_min = group;
2404 sds->min_nr_running = sgs->sum_nr_running;
2405 sds->min_load_per_task = sgs->sum_weighted_load /
2406 sgs->sum_nr_running;
2410 * Calculate the group which is almost near its
2411 * capacity but still has some space to pick up some load
2412 * from other group and save more power
2414 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2417 if (sgs->sum_nr_running > sds->leader_nr_running ||
2418 (sgs->sum_nr_running == sds->leader_nr_running &&
2419 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2420 sds->group_leader = group;
2421 sds->leader_nr_running = sgs->sum_nr_running;
2426 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2427 * @sds: Variable containing the statistics of the sched_domain
2428 * under consideration.
2429 * @this_cpu: Cpu at which we're currently performing load-balancing.
2430 * @imbalance: Variable to store the imbalance.
2433 * Check if we have potential to perform some power-savings balance.
2434 * If yes, set the busiest group to be the least loaded group in the
2435 * sched_domain, so that it's CPUs can be put to idle.
2437 * Returns 1 if there is potential to perform power-savings balance.
2440 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2441 int this_cpu, unsigned long *imbalance)
2443 if (!sds->power_savings_balance)
2446 if (sds->this != sds->group_leader ||
2447 sds->group_leader == sds->group_min)
2450 *imbalance = sds->min_load_per_task;
2451 sds->busiest = sds->group_min;
2456 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2457 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2458 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2463 static inline void update_sd_power_savings_stats(struct sched_group *group,
2464 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2469 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2470 int this_cpu, unsigned long *imbalance)
2474 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2477 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2479 return SCHED_LOAD_SCALE;
2482 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2484 return default_scale_freq_power(sd, cpu);
2487 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2489 unsigned long weight = sd->span_weight;
2490 unsigned long smt_gain = sd->smt_gain;
2497 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2499 return default_scale_smt_power(sd, cpu);
2502 unsigned long scale_rt_power(int cpu)
2504 struct rq *rq = cpu_rq(cpu);
2505 u64 total, available;
2507 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2509 if (unlikely(total < rq->rt_avg)) {
2510 /* Ensures that power won't end up being negative */
2513 available = total - rq->rt_avg;
2516 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2517 total = SCHED_LOAD_SCALE;
2519 total >>= SCHED_LOAD_SHIFT;
2521 return div_u64(available, total);
2524 static void update_cpu_power(struct sched_domain *sd, int cpu)
2526 unsigned long weight = sd->span_weight;
2527 unsigned long power = SCHED_LOAD_SCALE;
2528 struct sched_group *sdg = sd->groups;
2530 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2531 if (sched_feat(ARCH_POWER))
2532 power *= arch_scale_smt_power(sd, cpu);
2534 power *= default_scale_smt_power(sd, cpu);
2536 power >>= SCHED_LOAD_SHIFT;
2539 sdg->cpu_power_orig = power;
2541 if (sched_feat(ARCH_POWER))
2542 power *= arch_scale_freq_power(sd, cpu);
2544 power *= default_scale_freq_power(sd, cpu);
2546 power >>= SCHED_LOAD_SHIFT;
2548 power *= scale_rt_power(cpu);
2549 power >>= SCHED_LOAD_SHIFT;
2554 cpu_rq(cpu)->cpu_power = power;
2555 sdg->cpu_power = power;
2558 static void update_group_power(struct sched_domain *sd, int cpu)
2560 struct sched_domain *child = sd->child;
2561 struct sched_group *group, *sdg = sd->groups;
2562 unsigned long power;
2565 update_cpu_power(sd, cpu);
2571 group = child->groups;
2573 power += group->cpu_power;
2574 group = group->next;
2575 } while (group != child->groups);
2577 sdg->cpu_power = power;
2581 * Try and fix up capacity for tiny siblings, this is needed when
2582 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2583 * which on its own isn't powerful enough.
2585 * See update_sd_pick_busiest() and check_asym_packing().
2588 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2591 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2593 if (sd->level != SD_LV_SIBLING)
2597 * If ~90% of the cpu_power is still there, we're good.
2599 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2606 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2607 * @sd: The sched_domain whose statistics are to be updated.
2608 * @group: sched_group whose statistics are to be updated.
2609 * @this_cpu: Cpu for which load balance is currently performed.
2610 * @idle: Idle status of this_cpu
2611 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2612 * @sd_idle: Idle status of the sched_domain containing group.
2613 * @local_group: Does group contain this_cpu.
2614 * @cpus: Set of cpus considered for load balancing.
2615 * @balance: Should we balance.
2616 * @sgs: variable to hold the statistics for this group.
2618 static inline void update_sg_lb_stats(struct sched_domain *sd,
2619 struct sched_group *group, int this_cpu,
2620 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2621 int local_group, const struct cpumask *cpus,
2622 int *balance, struct sg_lb_stats *sgs)
2624 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2626 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2627 unsigned long avg_load_per_task = 0;
2630 balance_cpu = group_first_cpu(group);
2632 /* Tally up the load of all CPUs in the group */
2634 min_cpu_load = ~0UL;
2637 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2638 struct rq *rq = cpu_rq(i);
2640 if (*sd_idle && rq->nr_running)
2643 /* Bias balancing toward cpus of our domain */
2645 if (idle_cpu(i) && !first_idle_cpu) {
2650 load = target_load(i, load_idx);
2652 load = source_load(i, load_idx);
2653 if (load > max_cpu_load) {
2654 max_cpu_load = load;
2655 max_nr_running = rq->nr_running;
2657 if (min_cpu_load > load)
2658 min_cpu_load = load;
2661 sgs->group_load += load;
2662 sgs->sum_nr_running += rq->nr_running;
2663 sgs->sum_weighted_load += weighted_cpuload(i);
2669 * First idle cpu or the first cpu(busiest) in this sched group
2670 * is eligible for doing load balancing at this and above
2671 * domains. In the newly idle case, we will allow all the cpu's
2672 * to do the newly idle load balance.
2674 if (idle != CPU_NEWLY_IDLE && local_group) {
2675 if (balance_cpu != this_cpu) {
2679 update_group_power(sd, this_cpu);
2682 /* Adjust by relative CPU power of the group */
2683 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2686 * Consider the group unbalanced when the imbalance is larger
2687 * than the average weight of two tasks.
2689 * APZ: with cgroup the avg task weight can vary wildly and
2690 * might not be a suitable number - should we keep a
2691 * normalized nr_running number somewhere that negates
2694 if (sgs->sum_nr_running)
2695 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2697 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2700 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2701 if (!sgs->group_capacity)
2702 sgs->group_capacity = fix_small_capacity(sd, group);
2703 sgs->group_weight = group->group_weight;
2705 if (sgs->group_capacity > sgs->sum_nr_running)
2706 sgs->group_has_capacity = 1;
2710 * update_sd_pick_busiest - return 1 on busiest group
2711 * @sd: sched_domain whose statistics are to be checked
2712 * @sds: sched_domain statistics
2713 * @sg: sched_group candidate to be checked for being the busiest
2714 * @sgs: sched_group statistics
2715 * @this_cpu: the current cpu
2717 * Determine if @sg is a busier group than the previously selected
2720 static bool update_sd_pick_busiest(struct sched_domain *sd,
2721 struct sd_lb_stats *sds,
2722 struct sched_group *sg,
2723 struct sg_lb_stats *sgs,
2726 if (sgs->avg_load <= sds->max_load)
2729 if (sgs->sum_nr_running > sgs->group_capacity)
2736 * ASYM_PACKING needs to move all the work to the lowest
2737 * numbered CPUs in the group, therefore mark all groups
2738 * higher than ourself as busy.
2740 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2741 this_cpu < group_first_cpu(sg)) {
2745 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2753 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2754 * @sd: sched_domain whose statistics are to be updated.
2755 * @this_cpu: Cpu for which load balance is currently performed.
2756 * @idle: Idle status of this_cpu
2757 * @sd_idle: Idle status of the sched_domain containing sg.
2758 * @cpus: Set of cpus considered for load balancing.
2759 * @balance: Should we balance.
2760 * @sds: variable to hold the statistics for this sched_domain.
2762 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2763 enum cpu_idle_type idle, int *sd_idle,
2764 const struct cpumask *cpus, int *balance,
2765 struct sd_lb_stats *sds)
2767 struct sched_domain *child = sd->child;
2768 struct sched_group *sg = sd->groups;
2769 struct sg_lb_stats sgs;
2770 int load_idx, prefer_sibling = 0;
2772 if (child && child->flags & SD_PREFER_SIBLING)
2775 init_sd_power_savings_stats(sd, sds, idle);
2776 load_idx = get_sd_load_idx(sd, idle);
2781 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2782 memset(&sgs, 0, sizeof(sgs));
2783 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2784 local_group, cpus, balance, &sgs);
2786 if (local_group && !(*balance))
2789 sds->total_load += sgs.group_load;
2790 sds->total_pwr += sg->cpu_power;
2793 * In case the child domain prefers tasks go to siblings
2794 * first, lower the sg capacity to one so that we'll try
2795 * and move all the excess tasks away. We lower the capacity
2796 * of a group only if the local group has the capacity to fit
2797 * these excess tasks, i.e. nr_running < group_capacity. The
2798 * extra check prevents the case where you always pull from the
2799 * heaviest group when it is already under-utilized (possible
2800 * with a large weight task outweighs the tasks on the system).
2802 if (prefer_sibling && !local_group && sds->this_has_capacity)
2803 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2806 sds->this_load = sgs.avg_load;
2808 sds->this_nr_running = sgs.sum_nr_running;
2809 sds->this_load_per_task = sgs.sum_weighted_load;
2810 sds->this_has_capacity = sgs.group_has_capacity;
2811 sds->this_idle_cpus = sgs.idle_cpus;
2812 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2813 sds->max_load = sgs.avg_load;
2815 sds->busiest_nr_running = sgs.sum_nr_running;
2816 sds->busiest_idle_cpus = sgs.idle_cpus;
2817 sds->busiest_group_capacity = sgs.group_capacity;
2818 sds->busiest_load_per_task = sgs.sum_weighted_load;
2819 sds->busiest_has_capacity = sgs.group_has_capacity;
2820 sds->busiest_group_weight = sgs.group_weight;
2821 sds->group_imb = sgs.group_imb;
2824 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2826 } while (sg != sd->groups);
2829 int __weak arch_sd_sibling_asym_packing(void)
2831 return 0*SD_ASYM_PACKING;
2835 * check_asym_packing - Check to see if the group is packed into the
2838 * This is primarily intended to used at the sibling level. Some
2839 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2840 * case of POWER7, it can move to lower SMT modes only when higher
2841 * threads are idle. When in lower SMT modes, the threads will
2842 * perform better since they share less core resources. Hence when we
2843 * have idle threads, we want them to be the higher ones.
2845 * This packing function is run on idle threads. It checks to see if
2846 * the busiest CPU in this domain (core in the P7 case) has a higher
2847 * CPU number than the packing function is being run on. Here we are
2848 * assuming lower CPU number will be equivalent to lower a SMT thread
2851 * Returns 1 when packing is required and a task should be moved to
2852 * this CPU. The amount of the imbalance is returned in *imbalance.
2854 * @sd: The sched_domain whose packing is to be checked.
2855 * @sds: Statistics of the sched_domain which is to be packed
2856 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2857 * @imbalance: returns amount of imbalanced due to packing.
2859 static int check_asym_packing(struct sched_domain *sd,
2860 struct sd_lb_stats *sds,
2861 int this_cpu, unsigned long *imbalance)
2865 if (!(sd->flags & SD_ASYM_PACKING))
2871 busiest_cpu = group_first_cpu(sds->busiest);
2872 if (this_cpu > busiest_cpu)
2875 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2881 * fix_small_imbalance - Calculate the minor imbalance that exists
2882 * amongst the groups of a sched_domain, during
2884 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2885 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2886 * @imbalance: Variable to store the imbalance.
2888 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2889 int this_cpu, unsigned long *imbalance)
2891 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2892 unsigned int imbn = 2;
2893 unsigned long scaled_busy_load_per_task;
2895 if (sds->this_nr_running) {
2896 sds->this_load_per_task /= sds->this_nr_running;
2897 if (sds->busiest_load_per_task >
2898 sds->this_load_per_task)
2901 sds->this_load_per_task =
2902 cpu_avg_load_per_task(this_cpu);
2904 scaled_busy_load_per_task = sds->busiest_load_per_task
2906 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2908 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2909 (scaled_busy_load_per_task * imbn)) {
2910 *imbalance = sds->busiest_load_per_task;
2915 * OK, we don't have enough imbalance to justify moving tasks,
2916 * however we may be able to increase total CPU power used by
2920 pwr_now += sds->busiest->cpu_power *
2921 min(sds->busiest_load_per_task, sds->max_load);
2922 pwr_now += sds->this->cpu_power *
2923 min(sds->this_load_per_task, sds->this_load);
2924 pwr_now /= SCHED_LOAD_SCALE;
2926 /* Amount of load we'd subtract */
2927 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2928 sds->busiest->cpu_power;
2929 if (sds->max_load > tmp)
2930 pwr_move += sds->busiest->cpu_power *
2931 min(sds->busiest_load_per_task, sds->max_load - tmp);
2933 /* Amount of load we'd add */
2934 if (sds->max_load * sds->busiest->cpu_power <
2935 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2936 tmp = (sds->max_load * sds->busiest->cpu_power) /
2937 sds->this->cpu_power;
2939 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2940 sds->this->cpu_power;
2941 pwr_move += sds->this->cpu_power *
2942 min(sds->this_load_per_task, sds->this_load + tmp);
2943 pwr_move /= SCHED_LOAD_SCALE;
2945 /* Move if we gain throughput */
2946 if (pwr_move > pwr_now)
2947 *imbalance = sds->busiest_load_per_task;
2951 * calculate_imbalance - Calculate the amount of imbalance present within the
2952 * groups of a given sched_domain during load balance.
2953 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2954 * @this_cpu: Cpu for which currently load balance is being performed.
2955 * @imbalance: The variable to store the imbalance.
2957 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2958 unsigned long *imbalance)
2960 unsigned long max_pull, load_above_capacity = ~0UL;
2962 sds->busiest_load_per_task /= sds->busiest_nr_running;
2963 if (sds->group_imb) {
2964 sds->busiest_load_per_task =
2965 min(sds->busiest_load_per_task, sds->avg_load);
2969 * In the presence of smp nice balancing, certain scenarios can have
2970 * max load less than avg load(as we skip the groups at or below
2971 * its cpu_power, while calculating max_load..)
2973 if (sds->max_load < sds->avg_load) {
2975 return fix_small_imbalance(sds, this_cpu, imbalance);
2978 if (!sds->group_imb) {
2980 * Don't want to pull so many tasks that a group would go idle.
2982 load_above_capacity = (sds->busiest_nr_running -
2983 sds->busiest_group_capacity);
2985 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2987 load_above_capacity /= sds->busiest->cpu_power;
2991 * We're trying to get all the cpus to the average_load, so we don't
2992 * want to push ourselves above the average load, nor do we wish to
2993 * reduce the max loaded cpu below the average load. At the same time,
2994 * we also don't want to reduce the group load below the group capacity
2995 * (so that we can implement power-savings policies etc). Thus we look
2996 * for the minimum possible imbalance.
2997 * Be careful of negative numbers as they'll appear as very large values
2998 * with unsigned longs.
3000 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3002 /* How much load to actually move to equalise the imbalance */
3003 *imbalance = min(max_pull * sds->busiest->cpu_power,
3004 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3008 * if *imbalance is less than the average load per runnable task
3009 * there is no gaurantee that any tasks will be moved so we'll have
3010 * a think about bumping its value to force at least one task to be
3013 if (*imbalance < sds->busiest_load_per_task)
3014 return fix_small_imbalance(sds, this_cpu, imbalance);
3018 /******* find_busiest_group() helpers end here *********************/
3021 * find_busiest_group - Returns the busiest group within the sched_domain
3022 * if there is an imbalance. If there isn't an imbalance, and
3023 * the user has opted for power-savings, it returns a group whose
3024 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3025 * such a group exists.
3027 * Also calculates the amount of weighted load which should be moved
3028 * to restore balance.
3030 * @sd: The sched_domain whose busiest group is to be returned.
3031 * @this_cpu: The cpu for which load balancing is currently being performed.
3032 * @imbalance: Variable which stores amount of weighted load which should
3033 * be moved to restore balance/put a group to idle.
3034 * @idle: The idle status of this_cpu.
3035 * @sd_idle: The idleness of sd
3036 * @cpus: The set of CPUs under consideration for load-balancing.
3037 * @balance: Pointer to a variable indicating if this_cpu
3038 * is the appropriate cpu to perform load balancing at this_level.
3040 * Returns: - the busiest group if imbalance exists.
3041 * - If no imbalance and user has opted for power-savings balance,
3042 * return the least loaded group whose CPUs can be
3043 * put to idle by rebalancing its tasks onto our group.
3045 static struct sched_group *
3046 find_busiest_group(struct sched_domain *sd, int this_cpu,
3047 unsigned long *imbalance, enum cpu_idle_type idle,
3048 int *sd_idle, const struct cpumask *cpus, int *balance)
3050 struct sd_lb_stats sds;
3052 memset(&sds, 0, sizeof(sds));
3055 * Compute the various statistics relavent for load balancing at
3058 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3061 /* Cases where imbalance does not exist from POV of this_cpu */
3062 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3064 * 2) There is no busy sibling group to pull from.
3065 * 3) This group is the busiest group.
3066 * 4) This group is more busy than the avg busieness at this
3068 * 5) The imbalance is within the specified limit.
3070 * Note: when doing newidle balance, if the local group has excess
3071 * capacity (i.e. nr_running < group_capacity) and the busiest group
3072 * does not have any capacity, we force a load balance to pull tasks
3073 * to the local group. In this case, we skip past checks 3, 4 and 5.
3078 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3079 check_asym_packing(sd, &sds, this_cpu, imbalance))
3082 if (!sds.busiest || sds.busiest_nr_running == 0)
3085 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3086 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3087 !sds.busiest_has_capacity)
3090 if (sds.this_load >= sds.max_load)
3093 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3095 if (sds.this_load >= sds.avg_load)
3099 * In the CPU_NEWLY_IDLE, use imbalance_pct to be conservative.
3100 * And to check for busy balance use !idle_cpu instead of
3101 * CPU_NOT_IDLE. This is because HT siblings will use CPU_NOT_IDLE
3102 * even when they are idle.
3104 if (idle == CPU_NEWLY_IDLE || !idle_cpu(this_cpu)) {
3105 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3109 * This cpu is idle. If the busiest group load doesn't
3110 * have more tasks than the number of available cpu's and
3111 * there is no imbalance between this and busiest group
3112 * wrt to idle cpu's, it is balanced.
3114 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3115 sds.busiest_nr_running <= sds.busiest_group_weight)
3120 /* Looks like there is an imbalance. Compute it */
3121 calculate_imbalance(&sds, this_cpu, imbalance);
3126 * There is no obvious imbalance. But check if we can do some balancing
3129 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3137 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3140 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3141 enum cpu_idle_type idle, unsigned long imbalance,
3142 const struct cpumask *cpus)
3144 struct rq *busiest = NULL, *rq;
3145 unsigned long max_load = 0;
3148 for_each_cpu(i, sched_group_cpus(group)) {
3149 unsigned long power = power_of(i);
3150 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3154 capacity = fix_small_capacity(sd, group);
3156 if (!cpumask_test_cpu(i, cpus))
3160 wl = weighted_cpuload(i);
3163 * When comparing with imbalance, use weighted_cpuload()
3164 * which is not scaled with the cpu power.
3166 if (capacity && rq->nr_running == 1 && wl > imbalance)
3170 * For the load comparisons with the other cpu's, consider
3171 * the weighted_cpuload() scaled with the cpu power, so that
3172 * the load can be moved away from the cpu that is potentially
3173 * running at a lower capacity.
3175 wl = (wl * SCHED_LOAD_SCALE) / power;
3177 if (wl > max_load) {
3187 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3188 * so long as it is large enough.
3190 #define MAX_PINNED_INTERVAL 512
3192 /* Working cpumask for load_balance and load_balance_newidle. */
3193 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3195 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
3196 int busiest_cpu, int this_cpu)
3198 if (idle == CPU_NEWLY_IDLE) {
3201 * ASYM_PACKING needs to force migrate tasks from busy but
3202 * higher numbered CPUs in order to pack all tasks in the
3203 * lowest numbered CPUs.
3205 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3209 * The only task running in a non-idle cpu can be moved to this
3210 * cpu in an attempt to completely freeup the other CPU
3213 * The package power saving logic comes from
3214 * find_busiest_group(). If there are no imbalance, then
3215 * f_b_g() will return NULL. However when sched_mc={1,2} then
3216 * f_b_g() will select a group from which a running task may be
3217 * pulled to this cpu in order to make the other package idle.
3218 * If there is no opportunity to make a package idle and if
3219 * there are no imbalance, then f_b_g() will return NULL and no
3220 * action will be taken in load_balance_newidle().
3222 * Under normal task pull operation due to imbalance, there
3223 * will be more than one task in the source run queue and
3224 * move_tasks() will succeed. ld_moved will be true and this
3225 * active balance code will not be triggered.
3227 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3228 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3231 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3235 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3238 static int active_load_balance_cpu_stop(void *data);
3241 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3242 * tasks if there is an imbalance.
3244 static int load_balance(int this_cpu, struct rq *this_rq,
3245 struct sched_domain *sd, enum cpu_idle_type idle,
3248 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3249 struct sched_group *group;
3250 unsigned long imbalance;
3252 unsigned long flags;
3253 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3255 cpumask_copy(cpus, cpu_active_mask);
3258 * When power savings policy is enabled for the parent domain, idle
3259 * sibling can pick up load irrespective of busy siblings. In this case,
3260 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3261 * portraying it as CPU_NOT_IDLE.
3263 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3264 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3267 schedstat_inc(sd, lb_count[idle]);
3270 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3277 schedstat_inc(sd, lb_nobusyg[idle]);
3281 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3283 schedstat_inc(sd, lb_nobusyq[idle]);
3287 BUG_ON(busiest == this_rq);
3289 schedstat_add(sd, lb_imbalance[idle], imbalance);
3292 if (busiest->nr_running > 1) {
3294 * Attempt to move tasks. If find_busiest_group has found
3295 * an imbalance but busiest->nr_running <= 1, the group is
3296 * still unbalanced. ld_moved simply stays zero, so it is
3297 * correctly treated as an imbalance.
3299 local_irq_save(flags);
3300 double_rq_lock(this_rq, busiest);
3301 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3302 imbalance, sd, idle, &all_pinned);
3303 double_rq_unlock(this_rq, busiest);
3304 local_irq_restore(flags);
3307 * some other cpu did the load balance for us.
3309 if (ld_moved && this_cpu != smp_processor_id())
3310 resched_cpu(this_cpu);
3312 /* All tasks on this runqueue were pinned by CPU affinity */
3313 if (unlikely(all_pinned)) {
3314 cpumask_clear_cpu(cpu_of(busiest), cpus);
3315 if (!cpumask_empty(cpus))
3322 schedstat_inc(sd, lb_failed[idle]);
3324 * Increment the failure counter only on periodic balance.
3325 * We do not want newidle balance, which can be very
3326 * frequent, pollute the failure counter causing
3327 * excessive cache_hot migrations and active balances.
3329 if (idle != CPU_NEWLY_IDLE)
3330 sd->nr_balance_failed++;
3332 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3334 raw_spin_lock_irqsave(&busiest->lock, flags);
3336 /* don't kick the active_load_balance_cpu_stop,
3337 * if the curr task on busiest cpu can't be
3340 if (!cpumask_test_cpu(this_cpu,
3341 &busiest->curr->cpus_allowed)) {
3342 raw_spin_unlock_irqrestore(&busiest->lock,
3345 goto out_one_pinned;
3349 * ->active_balance synchronizes accesses to
3350 * ->active_balance_work. Once set, it's cleared
3351 * only after active load balance is finished.
3353 if (!busiest->active_balance) {
3354 busiest->active_balance = 1;
3355 busiest->push_cpu = this_cpu;
3358 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3361 stop_one_cpu_nowait(cpu_of(busiest),
3362 active_load_balance_cpu_stop, busiest,
3363 &busiest->active_balance_work);
3366 * We've kicked active balancing, reset the failure
3369 sd->nr_balance_failed = sd->cache_nice_tries+1;
3372 sd->nr_balance_failed = 0;
3374 if (likely(!active_balance)) {
3375 /* We were unbalanced, so reset the balancing interval */
3376 sd->balance_interval = sd->min_interval;
3379 * If we've begun active balancing, start to back off. This
3380 * case may not be covered by the all_pinned logic if there
3381 * is only 1 task on the busy runqueue (because we don't call
3384 if (sd->balance_interval < sd->max_interval)
3385 sd->balance_interval *= 2;
3388 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3389 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3395 schedstat_inc(sd, lb_balanced[idle]);
3397 sd->nr_balance_failed = 0;
3400 /* tune up the balancing interval */
3401 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3402 (sd->balance_interval < sd->max_interval))
3403 sd->balance_interval *= 2;
3405 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3406 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3415 * idle_balance is called by schedule() if this_cpu is about to become
3416 * idle. Attempts to pull tasks from other CPUs.
3418 static void idle_balance(int this_cpu, struct rq *this_rq)
3420 struct sched_domain *sd;
3421 int pulled_task = 0;
3422 unsigned long next_balance = jiffies + HZ;
3424 this_rq->idle_stamp = this_rq->clock;
3426 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3430 * Drop the rq->lock, but keep IRQ/preempt disabled.
3432 raw_spin_unlock(&this_rq->lock);
3434 update_shares(this_cpu);
3435 for_each_domain(this_cpu, sd) {
3436 unsigned long interval;
3439 if (!(sd->flags & SD_LOAD_BALANCE))
3442 if (sd->flags & SD_BALANCE_NEWIDLE) {
3443 /* If we've pulled tasks over stop searching: */
3444 pulled_task = load_balance(this_cpu, this_rq,
3445 sd, CPU_NEWLY_IDLE, &balance);
3448 interval = msecs_to_jiffies(sd->balance_interval);
3449 if (time_after(next_balance, sd->last_balance + interval))
3450 next_balance = sd->last_balance + interval;
3452 this_rq->idle_stamp = 0;
3457 raw_spin_lock(&this_rq->lock);
3459 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3461 * We are going idle. next_balance may be set based on
3462 * a busy processor. So reset next_balance.
3464 this_rq->next_balance = next_balance;
3469 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3470 * running tasks off the busiest CPU onto idle CPUs. It requires at
3471 * least 1 task to be running on each physical CPU where possible, and
3472 * avoids physical / logical imbalances.
3474 static int active_load_balance_cpu_stop(void *data)
3476 struct rq *busiest_rq = data;
3477 int busiest_cpu = cpu_of(busiest_rq);
3478 int target_cpu = busiest_rq->push_cpu;
3479 struct rq *target_rq = cpu_rq(target_cpu);
3480 struct sched_domain *sd;
3482 raw_spin_lock_irq(&busiest_rq->lock);
3484 /* make sure the requested cpu hasn't gone down in the meantime */
3485 if (unlikely(busiest_cpu != smp_processor_id() ||
3486 !busiest_rq->active_balance))
3489 /* Is there any task to move? */
3490 if (busiest_rq->nr_running <= 1)
3494 * This condition is "impossible", if it occurs
3495 * we need to fix it. Originally reported by
3496 * Bjorn Helgaas on a 128-cpu setup.
3498 BUG_ON(busiest_rq == target_rq);
3500 /* move a task from busiest_rq to target_rq */
3501 double_lock_balance(busiest_rq, target_rq);
3503 /* Search for an sd spanning us and the target CPU. */
3504 for_each_domain(target_cpu, sd) {
3505 if ((sd->flags & SD_LOAD_BALANCE) &&
3506 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3511 schedstat_inc(sd, alb_count);
3513 if (move_one_task(target_rq, target_cpu, busiest_rq,
3515 schedstat_inc(sd, alb_pushed);
3517 schedstat_inc(sd, alb_failed);
3519 double_unlock_balance(busiest_rq, target_rq);
3521 busiest_rq->active_balance = 0;
3522 raw_spin_unlock_irq(&busiest_rq->lock);
3528 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3530 static void trigger_sched_softirq(void *data)
3532 raise_softirq_irqoff(SCHED_SOFTIRQ);
3535 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3537 csd->func = trigger_sched_softirq;
3544 * idle load balancing details
3545 * - One of the idle CPUs nominates itself as idle load_balancer, while
3547 * - This idle load balancer CPU will also go into tickless mode when
3548 * it is idle, just like all other idle CPUs
3549 * - When one of the busy CPUs notice that there may be an idle rebalancing
3550 * needed, they will kick the idle load balancer, which then does idle
3551 * load balancing for all the idle CPUs.
3554 atomic_t load_balancer;
3555 atomic_t first_pick_cpu;
3556 atomic_t second_pick_cpu;
3557 cpumask_var_t idle_cpus_mask;
3558 cpumask_var_t grp_idle_mask;
3559 unsigned long next_balance; /* in jiffy units */
3560 } nohz ____cacheline_aligned;
3562 int get_nohz_load_balancer(void)
3564 return atomic_read(&nohz.load_balancer);
3567 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3569 * lowest_flag_domain - Return lowest sched_domain containing flag.
3570 * @cpu: The cpu whose lowest level of sched domain is to
3572 * @flag: The flag to check for the lowest sched_domain
3573 * for the given cpu.
3575 * Returns the lowest sched_domain of a cpu which contains the given flag.
3577 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3579 struct sched_domain *sd;
3581 for_each_domain(cpu, sd)
3582 if (sd && (sd->flags & flag))
3589 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3590 * @cpu: The cpu whose domains we're iterating over.
3591 * @sd: variable holding the value of the power_savings_sd
3593 * @flag: The flag to filter the sched_domains to be iterated.
3595 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3596 * set, starting from the lowest sched_domain to the highest.
3598 #define for_each_flag_domain(cpu, sd, flag) \
3599 for (sd = lowest_flag_domain(cpu, flag); \
3600 (sd && (sd->flags & flag)); sd = sd->parent)
3603 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3604 * @ilb_group: group to be checked for semi-idleness
3606 * Returns: 1 if the group is semi-idle. 0 otherwise.
3608 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3609 * and atleast one non-idle CPU. This helper function checks if the given
3610 * sched_group is semi-idle or not.
3612 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3614 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3615 sched_group_cpus(ilb_group));
3618 * A sched_group is semi-idle when it has atleast one busy cpu
3619 * and atleast one idle cpu.
3621 if (cpumask_empty(nohz.grp_idle_mask))
3624 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3630 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3631 * @cpu: The cpu which is nominating a new idle_load_balancer.
3633 * Returns: Returns the id of the idle load balancer if it exists,
3634 * Else, returns >= nr_cpu_ids.
3636 * This algorithm picks the idle load balancer such that it belongs to a
3637 * semi-idle powersavings sched_domain. The idea is to try and avoid
3638 * completely idle packages/cores just for the purpose of idle load balancing
3639 * when there are other idle cpu's which are better suited for that job.
3641 static int find_new_ilb(int cpu)
3643 struct sched_domain *sd;
3644 struct sched_group *ilb_group;
3647 * Have idle load balancer selection from semi-idle packages only
3648 * when power-aware load balancing is enabled
3650 if (!(sched_smt_power_savings || sched_mc_power_savings))
3654 * Optimize for the case when we have no idle CPUs or only one
3655 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3657 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3660 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3661 ilb_group = sd->groups;
3664 if (is_semi_idle_group(ilb_group))
3665 return cpumask_first(nohz.grp_idle_mask);
3667 ilb_group = ilb_group->next;
3669 } while (ilb_group != sd->groups);
3675 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3676 static inline int find_new_ilb(int call_cpu)
3683 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3684 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3685 * CPU (if there is one).
3687 static void nohz_balancer_kick(int cpu)
3691 nohz.next_balance++;
3693 ilb_cpu = get_nohz_load_balancer();
3695 if (ilb_cpu >= nr_cpu_ids) {
3696 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3697 if (ilb_cpu >= nr_cpu_ids)
3701 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3702 struct call_single_data *cp;
3704 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3705 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3706 __smp_call_function_single(ilb_cpu, cp, 0);
3712 * This routine will try to nominate the ilb (idle load balancing)
3713 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3714 * load balancing on behalf of all those cpus.
3716 * When the ilb owner becomes busy, we will not have new ilb owner until some
3717 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3718 * idle load balancing by kicking one of the idle CPUs.
3720 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3721 * ilb owner CPU in future (when there is a need for idle load balancing on
3722 * behalf of all idle CPUs).
3724 void select_nohz_load_balancer(int stop_tick)
3726 int cpu = smp_processor_id();
3729 if (!cpu_active(cpu)) {
3730 if (atomic_read(&nohz.load_balancer) != cpu)
3734 * If we are going offline and still the leader,
3737 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3744 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3746 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3747 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3748 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3749 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3751 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3754 /* make me the ilb owner */
3755 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3760 * Check to see if there is a more power-efficient
3763 new_ilb = find_new_ilb(cpu);
3764 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3765 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3766 resched_cpu(new_ilb);
3772 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3775 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3777 if (atomic_read(&nohz.load_balancer) == cpu)
3778 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3786 static DEFINE_SPINLOCK(balancing);
3789 * It checks each scheduling domain to see if it is due to be balanced,
3790 * and initiates a balancing operation if so.
3792 * Balancing parameters are set up in arch_init_sched_domains.
3794 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3797 struct rq *rq = cpu_rq(cpu);
3798 unsigned long interval;
3799 struct sched_domain *sd;
3800 /* Earliest time when we have to do rebalance again */
3801 unsigned long next_balance = jiffies + 60*HZ;
3802 int update_next_balance = 0;
3807 for_each_domain(cpu, sd) {
3808 if (!(sd->flags & SD_LOAD_BALANCE))
3811 interval = sd->balance_interval;
3812 if (idle != CPU_IDLE)
3813 interval *= sd->busy_factor;
3815 /* scale ms to jiffies */
3816 interval = msecs_to_jiffies(interval);
3817 if (unlikely(!interval))
3819 if (interval > HZ*NR_CPUS/10)
3820 interval = HZ*NR_CPUS/10;
3822 need_serialize = sd->flags & SD_SERIALIZE;
3824 if (need_serialize) {
3825 if (!spin_trylock(&balancing))
3829 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3830 if (load_balance(cpu, rq, sd, idle, &balance)) {
3832 * We've pulled tasks over so either we're no
3833 * longer idle, or one of our SMT siblings is
3836 idle = CPU_NOT_IDLE;
3838 sd->last_balance = jiffies;
3841 spin_unlock(&balancing);
3843 if (time_after(next_balance, sd->last_balance + interval)) {
3844 next_balance = sd->last_balance + interval;
3845 update_next_balance = 1;
3849 * Stop the load balance at this level. There is another
3850 * CPU in our sched group which is doing load balancing more
3858 * next_balance will be updated only when there is a need.
3859 * When the cpu is attached to null domain for ex, it will not be
3862 if (likely(update_next_balance))
3863 rq->next_balance = next_balance;
3868 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3869 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3871 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3873 struct rq *this_rq = cpu_rq(this_cpu);
3877 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3880 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3881 if (balance_cpu == this_cpu)
3885 * If this cpu gets work to do, stop the load balancing
3886 * work being done for other cpus. Next load
3887 * balancing owner will pick it up.
3889 if (need_resched()) {
3890 this_rq->nohz_balance_kick = 0;
3894 raw_spin_lock_irq(&this_rq->lock);
3895 update_rq_clock(this_rq);
3896 update_cpu_load(this_rq);
3897 raw_spin_unlock_irq(&this_rq->lock);
3899 rebalance_domains(balance_cpu, CPU_IDLE);
3901 rq = cpu_rq(balance_cpu);
3902 if (time_after(this_rq->next_balance, rq->next_balance))
3903 this_rq->next_balance = rq->next_balance;
3905 nohz.next_balance = this_rq->next_balance;
3906 this_rq->nohz_balance_kick = 0;
3910 * Current heuristic for kicking the idle load balancer
3911 * - first_pick_cpu is the one of the busy CPUs. It will kick
3912 * idle load balancer when it has more than one process active. This
3913 * eliminates the need for idle load balancing altogether when we have
3914 * only one running process in the system (common case).
3915 * - If there are more than one busy CPU, idle load balancer may have
3916 * to run for active_load_balance to happen (i.e., two busy CPUs are
3917 * SMT or core siblings and can run better if they move to different
3918 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3919 * which will kick idle load balancer as soon as it has any load.
3921 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3923 unsigned long now = jiffies;
3925 int first_pick_cpu, second_pick_cpu;
3927 if (time_before(now, nohz.next_balance))
3930 if (rq->idle_at_tick)
3933 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3934 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3936 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3937 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3940 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3941 if (ret == nr_cpu_ids || ret == cpu) {
3942 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3943 if (rq->nr_running > 1)
3946 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3947 if (ret == nr_cpu_ids || ret == cpu) {
3955 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3959 * run_rebalance_domains is triggered when needed from the scheduler tick.
3960 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3962 static void run_rebalance_domains(struct softirq_action *h)
3964 int this_cpu = smp_processor_id();
3965 struct rq *this_rq = cpu_rq(this_cpu);
3966 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3967 CPU_IDLE : CPU_NOT_IDLE;
3969 rebalance_domains(this_cpu, idle);
3972 * If this cpu has a pending nohz_balance_kick, then do the
3973 * balancing on behalf of the other idle cpus whose ticks are
3976 nohz_idle_balance(this_cpu, idle);
3979 static inline int on_null_domain(int cpu)
3981 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3985 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3987 static inline void trigger_load_balance(struct rq *rq, int cpu)
3989 /* Don't need to rebalance while attached to NULL domain */
3990 if (time_after_eq(jiffies, rq->next_balance) &&
3991 likely(!on_null_domain(cpu)))
3992 raise_softirq(SCHED_SOFTIRQ);
3994 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
3995 nohz_balancer_kick(cpu);
3999 static void rq_online_fair(struct rq *rq)
4004 static void rq_offline_fair(struct rq *rq)
4009 #else /* CONFIG_SMP */
4012 * on UP we do not need to balance between CPUs:
4014 static inline void idle_balance(int cpu, struct rq *rq)
4018 #endif /* CONFIG_SMP */
4021 * scheduler tick hitting a task of our scheduling class:
4023 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4025 struct cfs_rq *cfs_rq;
4026 struct sched_entity *se = &curr->se;
4028 for_each_sched_entity(se) {
4029 cfs_rq = cfs_rq_of(se);
4030 entity_tick(cfs_rq, se, queued);
4035 * called on fork with the child task as argument from the parent's context
4036 * - child not yet on the tasklist
4037 * - preemption disabled
4039 static void task_fork_fair(struct task_struct *p)
4041 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4042 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4043 int this_cpu = smp_processor_id();
4044 struct rq *rq = this_rq();
4045 unsigned long flags;
4047 raw_spin_lock_irqsave(&rq->lock, flags);
4049 update_rq_clock(rq);
4051 if (unlikely(task_cpu(p) != this_cpu)) {
4053 __set_task_cpu(p, this_cpu);
4057 update_curr(cfs_rq);
4060 se->vruntime = curr->vruntime;
4061 place_entity(cfs_rq, se, 1);
4063 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4065 * Upon rescheduling, sched_class::put_prev_task() will place
4066 * 'current' within the tree based on its new key value.
4068 swap(curr->vruntime, se->vruntime);
4069 resched_task(rq->curr);
4072 se->vruntime -= cfs_rq->min_vruntime;
4074 raw_spin_unlock_irqrestore(&rq->lock, flags);
4078 * Priority of the task has changed. Check to see if we preempt
4082 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4088 * Reschedule if we are currently running on this runqueue and
4089 * our priority decreased, or if we are not currently running on
4090 * this runqueue and our priority is higher than the current's
4092 if (rq->curr == p) {
4093 if (p->prio > oldprio)
4094 resched_task(rq->curr);
4096 check_preempt_curr(rq, p, 0);
4099 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4101 struct sched_entity *se = &p->se;
4102 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4105 * Ensure the task's vruntime is normalized, so that when its
4106 * switched back to the fair class the enqueue_entity(.flags=0) will
4107 * do the right thing.
4109 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4110 * have normalized the vruntime, if it was !on_rq, then only when
4111 * the task is sleeping will it still have non-normalized vruntime.
4113 if (!se->on_rq && p->state != TASK_RUNNING) {
4115 * Fix up our vruntime so that the current sleep doesn't
4116 * cause 'unlimited' sleep bonus.
4118 place_entity(cfs_rq, se, 0);
4119 se->vruntime -= cfs_rq->min_vruntime;
4124 * We switched to the sched_fair class.
4126 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4132 * We were most likely switched from sched_rt, so
4133 * kick off the schedule if running, otherwise just see
4134 * if we can still preempt the current task.
4137 resched_task(rq->curr);
4139 check_preempt_curr(rq, p, 0);
4142 /* Account for a task changing its policy or group.
4144 * This routine is mostly called to set cfs_rq->curr field when a task
4145 * migrates between groups/classes.
4147 static void set_curr_task_fair(struct rq *rq)
4149 struct sched_entity *se = &rq->curr->se;
4151 for_each_sched_entity(se)
4152 set_next_entity(cfs_rq_of(se), se);
4155 #ifdef CONFIG_FAIR_GROUP_SCHED
4156 static void task_move_group_fair(struct task_struct *p, int on_rq)
4159 * If the task was not on the rq at the time of this cgroup movement
4160 * it must have been asleep, sleeping tasks keep their ->vruntime
4161 * absolute on their old rq until wakeup (needed for the fair sleeper
4162 * bonus in place_entity()).
4164 * If it was on the rq, we've just 'preempted' it, which does convert
4165 * ->vruntime to a relative base.
4167 * Make sure both cases convert their relative position when migrating
4168 * to another cgroup's rq. This does somewhat interfere with the
4169 * fair sleeper stuff for the first placement, but who cares.
4172 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4173 set_task_rq(p, task_cpu(p));
4175 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4179 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4181 struct sched_entity *se = &task->se;
4182 unsigned int rr_interval = 0;
4185 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4188 if (rq->cfs.load.weight)
4189 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4195 * All the scheduling class methods:
4197 static const struct sched_class fair_sched_class = {
4198 .next = &idle_sched_class,
4199 .enqueue_task = enqueue_task_fair,
4200 .dequeue_task = dequeue_task_fair,
4201 .yield_task = yield_task_fair,
4203 .check_preempt_curr = check_preempt_wakeup,
4205 .pick_next_task = pick_next_task_fair,
4206 .put_prev_task = put_prev_task_fair,
4209 .select_task_rq = select_task_rq_fair,
4211 .rq_online = rq_online_fair,
4212 .rq_offline = rq_offline_fair,
4214 .task_waking = task_waking_fair,
4217 .set_curr_task = set_curr_task_fair,
4218 .task_tick = task_tick_fair,
4219 .task_fork = task_fork_fair,
4221 .prio_changed = prio_changed_fair,
4222 .switched_from = switched_from_fair,
4223 .switched_to = switched_to_fair,
4225 .get_rr_interval = get_rr_interval_fair,
4227 #ifdef CONFIG_FAIR_GROUP_SCHED
4228 .task_move_group = task_move_group_fair,
4232 #ifdef CONFIG_SCHED_DEBUG
4233 static void print_cfs_stats(struct seq_file *m, int cpu)
4235 struct cfs_rq *cfs_rq;
4238 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4239 print_cfs_rq(m, cpu, cfs_rq);