2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 8;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
99 static const struct sched_class fair_sched_class;
101 /**************************************************************
102 * CFS operations on generic schedulable entities:
105 #ifdef CONFIG_FAIR_GROUP_SCHED
107 /* cpu runqueue to which this cfs_rq is attached */
108 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
113 /* An entity is a task if it doesn't "own" a runqueue */
114 #define entity_is_task(se) (!se->my_q)
116 static inline struct task_struct *task_of(struct sched_entity *se)
118 #ifdef CONFIG_SCHED_DEBUG
119 WARN_ON_ONCE(!entity_is_task(se));
121 return container_of(se, struct task_struct, se);
124 /* Walk up scheduling entities hierarchy */
125 #define for_each_sched_entity(se) \
126 for (; se; se = se->parent)
128 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
133 /* runqueue on which this entity is (to be) queued */
134 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
139 /* runqueue "owned" by this group */
140 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
145 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
146 * another cpu ('this_cpu')
148 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
150 return cfs_rq->tg->cfs_rq[this_cpu];
153 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
155 if (!cfs_rq->on_list) {
157 * Ensure we either appear before our parent (if already
158 * enqueued) or force our parent to appear after us when it is
159 * enqueued. The fact that we always enqueue bottom-up
160 * reduces this to two cases.
162 if (cfs_rq->tg->parent &&
163 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
164 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
165 &rq_of(cfs_rq)->leaf_cfs_rq_list);
167 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
168 &rq_of(cfs_rq)->leaf_cfs_rq_list);
175 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
177 if (cfs_rq->on_list) {
178 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
183 /* Iterate thr' all leaf cfs_rq's on a runqueue */
184 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
185 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
187 /* Do the two (enqueued) entities belong to the same group ? */
189 is_same_group(struct sched_entity *se, struct sched_entity *pse)
191 if (se->cfs_rq == pse->cfs_rq)
197 static inline struct sched_entity *parent_entity(struct sched_entity *se)
202 /* return depth at which a sched entity is present in the hierarchy */
203 static inline int depth_se(struct sched_entity *se)
207 for_each_sched_entity(se)
214 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
216 int se_depth, pse_depth;
219 * preemption test can be made between sibling entities who are in the
220 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
221 * both tasks until we find their ancestors who are siblings of common
225 /* First walk up until both entities are at same depth */
226 se_depth = depth_se(*se);
227 pse_depth = depth_se(*pse);
229 while (se_depth > pse_depth) {
231 *se = parent_entity(*se);
234 while (pse_depth > se_depth) {
236 *pse = parent_entity(*pse);
239 while (!is_same_group(*se, *pse)) {
240 *se = parent_entity(*se);
241 *pse = parent_entity(*pse);
245 #else /* !CONFIG_FAIR_GROUP_SCHED */
247 static inline struct task_struct *task_of(struct sched_entity *se)
249 return container_of(se, struct task_struct, se);
252 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 return container_of(cfs_rq, struct rq, cfs);
257 #define entity_is_task(se) 1
259 #define for_each_sched_entity(se) \
260 for (; se; se = NULL)
262 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
264 return &task_rq(p)->cfs;
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
269 struct task_struct *p = task_of(se);
270 struct rq *rq = task_rq(p);
275 /* runqueue "owned" by this group */
276 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
283 return &cpu_rq(this_cpu)->cfs;
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
295 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
298 is_same_group(struct sched_entity *se, struct sched_entity *pse)
303 static inline struct sched_entity *parent_entity(struct sched_entity *se)
309 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
313 #endif /* CONFIG_FAIR_GROUP_SCHED */
316 /**************************************************************
317 * Scheduling class tree data structure manipulation methods:
320 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
322 s64 delta = (s64)(vruntime - min_vruntime);
324 min_vruntime = vruntime;
329 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
331 s64 delta = (s64)(vruntime - min_vruntime);
333 min_vruntime = vruntime;
338 static inline int entity_before(struct sched_entity *a,
339 struct sched_entity *b)
341 return (s64)(a->vruntime - b->vruntime) < 0;
344 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
346 return se->vruntime - cfs_rq->min_vruntime;
349 static void update_min_vruntime(struct cfs_rq *cfs_rq)
351 u64 vruntime = cfs_rq->min_vruntime;
354 vruntime = cfs_rq->curr->vruntime;
356 if (cfs_rq->rb_leftmost) {
357 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
362 vruntime = se->vruntime;
364 vruntime = min_vruntime(vruntime, se->vruntime);
367 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
371 * Enqueue an entity into the rb-tree:
373 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
375 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
376 struct rb_node *parent = NULL;
377 struct sched_entity *entry;
378 s64 key = entity_key(cfs_rq, se);
382 * Find the right place in the rbtree:
386 entry = rb_entry(parent, struct sched_entity, run_node);
388 * We dont care about collisions. Nodes with
389 * the same key stay together.
391 if (key < entity_key(cfs_rq, entry)) {
392 link = &parent->rb_left;
394 link = &parent->rb_right;
400 * Maintain a cache of leftmost tree entries (it is frequently
404 cfs_rq->rb_leftmost = &se->run_node;
406 rb_link_node(&se->run_node, parent, link);
407 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
410 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
412 if (cfs_rq->rb_leftmost == &se->run_node) {
413 struct rb_node *next_node;
415 next_node = rb_next(&se->run_node);
416 cfs_rq->rb_leftmost = next_node;
419 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
422 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
424 struct rb_node *left = cfs_rq->rb_leftmost;
429 return rb_entry(left, struct sched_entity, run_node);
432 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
434 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
439 return rb_entry(last, struct sched_entity, run_node);
442 /**************************************************************
443 * Scheduling class statistics methods:
446 #ifdef CONFIG_SCHED_DEBUG
447 int sched_proc_update_handler(struct ctl_table *table, int write,
448 void __user *buffer, size_t *lenp,
451 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
452 int factor = get_update_sysctl_factor();
457 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
458 sysctl_sched_min_granularity);
460 #define WRT_SYSCTL(name) \
461 (normalized_sysctl_##name = sysctl_##name / (factor))
462 WRT_SYSCTL(sched_min_granularity);
463 WRT_SYSCTL(sched_latency);
464 WRT_SYSCTL(sched_wakeup_granularity);
474 static inline unsigned long
475 calc_delta_fair(unsigned long delta, struct sched_entity *se)
477 if (unlikely(se->load.weight != NICE_0_LOAD))
478 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
484 * The idea is to set a period in which each task runs once.
486 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
487 * this period because otherwise the slices get too small.
489 * p = (nr <= nl) ? l : l*nr/nl
491 static u64 __sched_period(unsigned long nr_running)
493 u64 period = sysctl_sched_latency;
494 unsigned long nr_latency = sched_nr_latency;
496 if (unlikely(nr_running > nr_latency)) {
497 period = sysctl_sched_min_granularity;
498 period *= nr_running;
505 * We calculate the wall-time slice from the period by taking a part
506 * proportional to the weight.
510 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
512 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
514 for_each_sched_entity(se) {
515 struct load_weight *load;
516 struct load_weight lw;
518 cfs_rq = cfs_rq_of(se);
519 load = &cfs_rq->load;
521 if (unlikely(!se->on_rq)) {
524 update_load_add(&lw, se->load.weight);
527 slice = calc_delta_mine(slice, se->load.weight, load);
533 * We calculate the vruntime slice of a to be inserted task
537 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
539 return calc_delta_fair(sched_slice(cfs_rq, se), se);
542 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
543 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta);
546 * Update the current task's runtime statistics. Skip current tasks that
547 * are not in our scheduling class.
550 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
551 unsigned long delta_exec)
553 unsigned long delta_exec_weighted;
555 schedstat_set(curr->statistics.exec_max,
556 max((u64)delta_exec, curr->statistics.exec_max));
558 curr->sum_exec_runtime += delta_exec;
559 schedstat_add(cfs_rq, exec_clock, delta_exec);
560 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
562 curr->vruntime += delta_exec_weighted;
563 update_min_vruntime(cfs_rq);
565 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
566 cfs_rq->load_unacc_exec_time += delta_exec;
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;
738 cfs_rq->load_stamp = now;
739 cfs_rq->load_unacc_exec_time = 0;
740 cfs_rq->load_period += delta;
742 cfs_rq->load_last = now;
743 cfs_rq->load_avg += delta * load;
746 /* consider updating load contribution on each fold or truncate */
747 if (global_update || cfs_rq->load_period > period
748 || !cfs_rq->load_period)
749 update_cfs_rq_load_contribution(cfs_rq, global_update);
751 while (cfs_rq->load_period > period) {
753 * Inline assembly required to prevent the compiler
754 * optimising this loop into a divmod call.
755 * See __iter_div_u64_rem() for another example of this.
757 asm("" : "+rm" (cfs_rq->load_period));
758 cfs_rq->load_period /= 2;
759 cfs_rq->load_avg /= 2;
762 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
763 list_del_leaf_cfs_rq(cfs_rq);
766 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 + weight_delta;
773 load_weight = atomic_read(&tg->load_weight);
774 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, 0);
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,
807 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
810 # endif /* CONFIG_SMP */
811 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
812 unsigned long weight)
815 /* commit outstanding execution time */
816 if (cfs_rq->curr == se)
818 account_entity_dequeue(cfs_rq, se);
821 update_load_set(&se->load, weight);
824 account_entity_enqueue(cfs_rq, se);
827 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
829 struct task_group *tg;
830 struct sched_entity *se;
834 se = tg->se[cpu_of(rq_of(cfs_rq))];
838 if (likely(se->load.weight == tg->shares))
841 shares = calc_cfs_shares(cfs_rq, tg, weight_delta);
843 reweight_entity(cfs_rq_of(se), se, shares);
845 #else /* CONFIG_FAIR_GROUP_SCHED */
846 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
850 static inline void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
854 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
857 #endif /* CONFIG_FAIR_GROUP_SCHED */
859 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
861 #ifdef CONFIG_SCHEDSTATS
862 struct task_struct *tsk = NULL;
864 if (entity_is_task(se))
867 if (se->statistics.sleep_start) {
868 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
873 if (unlikely(delta > se->statistics.sleep_max))
874 se->statistics.sleep_max = delta;
876 se->statistics.sleep_start = 0;
877 se->statistics.sum_sleep_runtime += delta;
880 account_scheduler_latency(tsk, delta >> 10, 1);
881 trace_sched_stat_sleep(tsk, delta);
884 if (se->statistics.block_start) {
885 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
890 if (unlikely(delta > se->statistics.block_max))
891 se->statistics.block_max = delta;
893 se->statistics.block_start = 0;
894 se->statistics.sum_sleep_runtime += delta;
897 if (tsk->in_iowait) {
898 se->statistics.iowait_sum += delta;
899 se->statistics.iowait_count++;
900 trace_sched_stat_iowait(tsk, delta);
904 * Blocking time is in units of nanosecs, so shift by
905 * 20 to get a milliseconds-range estimation of the
906 * amount of time that the task spent sleeping:
908 if (unlikely(prof_on == SLEEP_PROFILING)) {
909 profile_hits(SLEEP_PROFILING,
910 (void *)get_wchan(tsk),
913 account_scheduler_latency(tsk, delta >> 10, 0);
919 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
921 #ifdef CONFIG_SCHED_DEBUG
922 s64 d = se->vruntime - cfs_rq->min_vruntime;
927 if (d > 3*sysctl_sched_latency)
928 schedstat_inc(cfs_rq, nr_spread_over);
933 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
935 u64 vruntime = cfs_rq->min_vruntime;
938 * The 'current' period is already promised to the current tasks,
939 * however the extra weight of the new task will slow them down a
940 * little, place the new task so that it fits in the slot that
941 * stays open at the end.
943 if (initial && sched_feat(START_DEBIT))
944 vruntime += sched_vslice(cfs_rq, se);
946 /* sleeps up to a single latency don't count. */
948 unsigned long thresh = sysctl_sched_latency;
951 * Halve their sleep time's effect, to allow
952 * for a gentler effect of sleepers:
954 if (sched_feat(GENTLE_FAIR_SLEEPERS))
960 /* ensure we never gain time by being placed backwards. */
961 vruntime = max_vruntime(se->vruntime, vruntime);
963 se->vruntime = vruntime;
967 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
970 * Update the normalized vruntime before updating min_vruntime
971 * through callig update_curr().
973 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
974 se->vruntime += cfs_rq->min_vruntime;
977 * Update run-time statistics of the 'current'.
980 update_cfs_load(cfs_rq, 0);
981 update_cfs_shares(cfs_rq, se->load.weight);
982 account_entity_enqueue(cfs_rq, se);
984 if (flags & ENQUEUE_WAKEUP) {
985 place_entity(cfs_rq, se, 0);
986 enqueue_sleeper(cfs_rq, se);
989 update_stats_enqueue(cfs_rq, se);
990 check_spread(cfs_rq, se);
991 if (se != cfs_rq->curr)
992 __enqueue_entity(cfs_rq, se);
995 if (cfs_rq->nr_running == 1)
996 list_add_leaf_cfs_rq(cfs_rq);
999 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1001 if (!se || cfs_rq->last == se)
1002 cfs_rq->last = NULL;
1004 if (!se || cfs_rq->next == se)
1005 cfs_rq->next = NULL;
1008 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1010 for_each_sched_entity(se)
1011 __clear_buddies(cfs_rq_of(se), se);
1015 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1020 * Update run-time statistics of the 'current'.
1022 update_curr(cfs_rq);
1024 update_stats_dequeue(cfs_rq, se);
1025 if (flags & DEQUEUE_SLEEP) {
1026 #ifdef CONFIG_SCHEDSTATS
1027 if (entity_is_task(se)) {
1028 struct task_struct *tsk = task_of(se);
1030 if (tsk->state & TASK_INTERRUPTIBLE)
1031 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1032 if (tsk->state & TASK_UNINTERRUPTIBLE)
1033 se->statistics.block_start = rq_of(cfs_rq)->clock;
1038 clear_buddies(cfs_rq, se);
1040 if (se != cfs_rq->curr)
1041 __dequeue_entity(cfs_rq, se);
1043 update_cfs_load(cfs_rq, 0);
1044 account_entity_dequeue(cfs_rq, se);
1046 min_vruntime = cfs_rq->min_vruntime;
1047 update_min_vruntime(cfs_rq);
1048 update_cfs_shares(cfs_rq, 0);
1051 * Normalize the entity after updating the min_vruntime because the
1052 * update can refer to the ->curr item and we need to reflect this
1053 * movement in our normalized position.
1055 if (!(flags & DEQUEUE_SLEEP))
1056 se->vruntime -= min_vruntime;
1060 * Preempt the current task with a newly woken task if needed:
1063 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1065 unsigned long ideal_runtime, delta_exec;
1067 ideal_runtime = sched_slice(cfs_rq, curr);
1068 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1069 if (delta_exec > ideal_runtime) {
1070 resched_task(rq_of(cfs_rq)->curr);
1072 * The current task ran long enough, ensure it doesn't get
1073 * re-elected due to buddy favours.
1075 clear_buddies(cfs_rq, curr);
1080 * Ensure that a task that missed wakeup preemption by a
1081 * narrow margin doesn't have to wait for a full slice.
1082 * This also mitigates buddy induced latencies under load.
1084 if (!sched_feat(WAKEUP_PREEMPT))
1087 if (delta_exec < sysctl_sched_min_granularity)
1090 if (cfs_rq->nr_running > 1) {
1091 struct sched_entity *se = __pick_next_entity(cfs_rq);
1092 s64 delta = curr->vruntime - se->vruntime;
1097 if (delta > ideal_runtime)
1098 resched_task(rq_of(cfs_rq)->curr);
1103 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1105 /* 'current' is not kept within the tree. */
1108 * Any task has to be enqueued before it get to execute on
1109 * a CPU. So account for the time it spent waiting on the
1112 update_stats_wait_end(cfs_rq, se);
1113 __dequeue_entity(cfs_rq, se);
1116 update_stats_curr_start(cfs_rq, se);
1118 #ifdef CONFIG_SCHEDSTATS
1120 * Track our maximum slice length, if the CPU's load is at
1121 * least twice that of our own weight (i.e. dont track it
1122 * when there are only lesser-weight tasks around):
1124 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1125 se->statistics.slice_max = max(se->statistics.slice_max,
1126 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1129 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1133 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1135 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1137 struct sched_entity *se = __pick_next_entity(cfs_rq);
1138 struct sched_entity *left = se;
1140 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1144 * Prefer last buddy, try to return the CPU to a preempted task.
1146 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1149 clear_buddies(cfs_rq, se);
1154 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1157 * If still on the runqueue then deactivate_task()
1158 * was not called and update_curr() has to be done:
1161 update_curr(cfs_rq);
1163 check_spread(cfs_rq, prev);
1165 update_stats_wait_start(cfs_rq, prev);
1166 /* Put 'current' back into the tree. */
1167 __enqueue_entity(cfs_rq, prev);
1169 cfs_rq->curr = NULL;
1173 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1176 * Update run-time statistics of the 'current'.
1178 update_curr(cfs_rq);
1181 * Update share accounting for long-running entities.
1183 update_entity_shares_tick(cfs_rq);
1185 #ifdef CONFIG_SCHED_HRTICK
1187 * queued ticks are scheduled to match the slice, so don't bother
1188 * validating it and just reschedule.
1191 resched_task(rq_of(cfs_rq)->curr);
1195 * don't let the period tick interfere with the hrtick preemption
1197 if (!sched_feat(DOUBLE_TICK) &&
1198 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1202 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1203 check_preempt_tick(cfs_rq, curr);
1206 /**************************************************
1207 * CFS operations on tasks:
1210 #ifdef CONFIG_SCHED_HRTICK
1211 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1213 struct sched_entity *se = &p->se;
1214 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1216 WARN_ON(task_rq(p) != rq);
1218 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1219 u64 slice = sched_slice(cfs_rq, se);
1220 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1221 s64 delta = slice - ran;
1230 * Don't schedule slices shorter than 10000ns, that just
1231 * doesn't make sense. Rely on vruntime for fairness.
1234 delta = max_t(s64, 10000LL, delta);
1236 hrtick_start(rq, delta);
1241 * called from enqueue/dequeue and updates the hrtick when the
1242 * current task is from our class and nr_running is low enough
1245 static void hrtick_update(struct rq *rq)
1247 struct task_struct *curr = rq->curr;
1249 if (curr->sched_class != &fair_sched_class)
1252 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1253 hrtick_start_fair(rq, curr);
1255 #else /* !CONFIG_SCHED_HRTICK */
1257 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1261 static inline void hrtick_update(struct rq *rq)
1267 * The enqueue_task method is called before nr_running is
1268 * increased. Here we update the fair scheduling stats and
1269 * then put the task into the rbtree:
1272 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1274 struct cfs_rq *cfs_rq;
1275 struct sched_entity *se = &p->se;
1277 for_each_sched_entity(se) {
1280 cfs_rq = cfs_rq_of(se);
1281 enqueue_entity(cfs_rq, se, flags);
1282 flags = ENQUEUE_WAKEUP;
1285 for_each_sched_entity(se) {
1286 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1288 update_cfs_load(cfs_rq, 0);
1289 update_cfs_shares(cfs_rq, 0);
1296 * The dequeue_task method is called before nr_running is
1297 * decreased. We remove the task from the rbtree and
1298 * update the fair scheduling stats:
1300 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1302 struct cfs_rq *cfs_rq;
1303 struct sched_entity *se = &p->se;
1305 for_each_sched_entity(se) {
1306 cfs_rq = cfs_rq_of(se);
1307 dequeue_entity(cfs_rq, se, flags);
1309 /* Don't dequeue parent if it has other entities besides us */
1310 if (cfs_rq->load.weight)
1312 flags |= DEQUEUE_SLEEP;
1315 for_each_sched_entity(se) {
1316 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1318 update_cfs_load(cfs_rq, 0);
1319 update_cfs_shares(cfs_rq, 0);
1326 * sched_yield() support is very simple - we dequeue and enqueue.
1328 * If compat_yield is turned on then we requeue to the end of the tree.
1330 static void yield_task_fair(struct rq *rq)
1332 struct task_struct *curr = rq->curr;
1333 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1334 struct sched_entity *rightmost, *se = &curr->se;
1337 * Are we the only task in the tree?
1339 if (unlikely(cfs_rq->nr_running == 1))
1342 clear_buddies(cfs_rq, se);
1344 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1345 update_rq_clock(rq);
1347 * Update run-time statistics of the 'current'.
1349 update_curr(cfs_rq);
1354 * Find the rightmost entry in the rbtree:
1356 rightmost = __pick_last_entity(cfs_rq);
1358 * Already in the rightmost position?
1360 if (unlikely(!rightmost || entity_before(rightmost, se)))
1364 * Minimally necessary key value to be last in the tree:
1365 * Upon rescheduling, sched_class::put_prev_task() will place
1366 * 'current' within the tree based on its new key value.
1368 se->vruntime = rightmost->vruntime + 1;
1373 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1375 struct sched_entity *se = &p->se;
1376 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1378 se->vruntime -= cfs_rq->min_vruntime;
1381 #ifdef CONFIG_FAIR_GROUP_SCHED
1383 * effective_load() calculates the load change as seen from the root_task_group
1385 * Adding load to a group doesn't make a group heavier, but can cause movement
1386 * of group shares between cpus. Assuming the shares were perfectly aligned one
1387 * can calculate the shift in shares.
1389 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1391 struct sched_entity *se = tg->se[cpu];
1396 for_each_sched_entity(se) {
1400 w = se->my_q->load.weight;
1402 /* use this cpu's instantaneous contribution */
1403 lw = atomic_read(&tg->load_weight);
1404 lw -= se->my_q->load_contribution;
1409 if (lw > 0 && wl < lw)
1410 wl = (wl * tg->shares) / lw;
1414 /* zero point is MIN_SHARES */
1415 if (wl < MIN_SHARES)
1417 wl -= se->load.weight;
1426 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1427 unsigned long wl, unsigned long wg)
1434 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1436 s64 this_load, load;
1437 int idx, this_cpu, prev_cpu;
1438 unsigned long tl_per_task;
1439 struct task_group *tg;
1440 unsigned long weight;
1444 this_cpu = smp_processor_id();
1445 prev_cpu = task_cpu(p);
1446 load = source_load(prev_cpu, idx);
1447 this_load = target_load(this_cpu, idx);
1450 * If sync wakeup then subtract the (maximum possible)
1451 * effect of the currently running task from the load
1452 * of the current CPU:
1456 tg = task_group(current);
1457 weight = current->se.load.weight;
1459 this_load += effective_load(tg, this_cpu, -weight, -weight);
1460 load += effective_load(tg, prev_cpu, 0, -weight);
1464 weight = p->se.load.weight;
1467 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1468 * due to the sync cause above having dropped this_load to 0, we'll
1469 * always have an imbalance, but there's really nothing you can do
1470 * about that, so that's good too.
1472 * Otherwise check if either cpus are near enough in load to allow this
1473 * task to be woken on this_cpu.
1475 if (this_load > 0) {
1476 s64 this_eff_load, prev_eff_load;
1478 this_eff_load = 100;
1479 this_eff_load *= power_of(prev_cpu);
1480 this_eff_load *= this_load +
1481 effective_load(tg, this_cpu, weight, weight);
1483 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1484 prev_eff_load *= power_of(this_cpu);
1485 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1487 balanced = this_eff_load <= prev_eff_load;
1493 * If the currently running task will sleep within
1494 * a reasonable amount of time then attract this newly
1497 if (sync && balanced)
1500 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1501 tl_per_task = cpu_avg_load_per_task(this_cpu);
1504 (this_load <= load &&
1505 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1507 * This domain has SD_WAKE_AFFINE and
1508 * p is cache cold in this domain, and
1509 * there is no bad imbalance.
1511 schedstat_inc(sd, ttwu_move_affine);
1512 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1520 * find_idlest_group finds and returns the least busy CPU group within the
1523 static struct sched_group *
1524 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1525 int this_cpu, int load_idx)
1527 struct sched_group *idlest = NULL, *group = sd->groups;
1528 unsigned long min_load = ULONG_MAX, this_load = 0;
1529 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1532 unsigned long load, avg_load;
1536 /* Skip over this group if it has no CPUs allowed */
1537 if (!cpumask_intersects(sched_group_cpus(group),
1541 local_group = cpumask_test_cpu(this_cpu,
1542 sched_group_cpus(group));
1544 /* Tally up the load of all CPUs in the group */
1547 for_each_cpu(i, sched_group_cpus(group)) {
1548 /* Bias balancing toward cpus of our domain */
1550 load = source_load(i, load_idx);
1552 load = target_load(i, load_idx);
1557 /* Adjust by relative CPU power of the group */
1558 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1561 this_load = avg_load;
1562 } else if (avg_load < min_load) {
1563 min_load = avg_load;
1566 } while (group = group->next, group != sd->groups);
1568 if (!idlest || 100*this_load < imbalance*min_load)
1574 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1577 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1579 unsigned long load, min_load = ULONG_MAX;
1583 /* Traverse only the allowed CPUs */
1584 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1585 load = weighted_cpuload(i);
1587 if (load < min_load || (load == min_load && i == this_cpu)) {
1597 * Try and locate an idle CPU in the sched_domain.
1599 static int select_idle_sibling(struct task_struct *p, int target)
1601 int cpu = smp_processor_id();
1602 int prev_cpu = task_cpu(p);
1603 struct sched_domain *sd;
1607 * If the task is going to be woken-up on this cpu and if it is
1608 * already idle, then it is the right target.
1610 if (target == cpu && idle_cpu(cpu))
1614 * If the task is going to be woken-up on the cpu where it previously
1615 * ran and if it is currently idle, then it the right target.
1617 if (target == prev_cpu && idle_cpu(prev_cpu))
1621 * Otherwise, iterate the domains and find an elegible idle cpu.
1623 for_each_domain(target, sd) {
1624 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1627 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1635 * Lets stop looking for an idle sibling when we reached
1636 * the domain that spans the current cpu and prev_cpu.
1638 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1639 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1647 * sched_balance_self: balance the current task (running on cpu) in domains
1648 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1651 * Balance, ie. select the least loaded group.
1653 * Returns the target CPU number, or the same CPU if no balancing is needed.
1655 * preempt must be disabled.
1658 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1660 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1661 int cpu = smp_processor_id();
1662 int prev_cpu = task_cpu(p);
1664 int want_affine = 0;
1666 int sync = wake_flags & WF_SYNC;
1668 if (sd_flag & SD_BALANCE_WAKE) {
1669 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1674 for_each_domain(cpu, tmp) {
1675 if (!(tmp->flags & SD_LOAD_BALANCE))
1679 * If power savings logic is enabled for a domain, see if we
1680 * are not overloaded, if so, don't balance wider.
1682 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1683 unsigned long power = 0;
1684 unsigned long nr_running = 0;
1685 unsigned long capacity;
1688 for_each_cpu(i, sched_domain_span(tmp)) {
1689 power += power_of(i);
1690 nr_running += cpu_rq(i)->cfs.nr_running;
1693 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1695 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1698 if (nr_running < capacity)
1703 * If both cpu and prev_cpu are part of this domain,
1704 * cpu is a valid SD_WAKE_AFFINE target.
1706 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1707 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1712 if (!want_sd && !want_affine)
1715 if (!(tmp->flags & sd_flag))
1723 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1724 return select_idle_sibling(p, cpu);
1726 return select_idle_sibling(p, prev_cpu);
1730 int load_idx = sd->forkexec_idx;
1731 struct sched_group *group;
1734 if (!(sd->flags & sd_flag)) {
1739 if (sd_flag & SD_BALANCE_WAKE)
1740 load_idx = sd->wake_idx;
1742 group = find_idlest_group(sd, p, cpu, load_idx);
1748 new_cpu = find_idlest_cpu(group, p, cpu);
1749 if (new_cpu == -1 || new_cpu == cpu) {
1750 /* Now try balancing at a lower domain level of cpu */
1755 /* Now try balancing at a lower domain level of new_cpu */
1757 weight = sd->span_weight;
1759 for_each_domain(cpu, tmp) {
1760 if (weight <= tmp->span_weight)
1762 if (tmp->flags & sd_flag)
1765 /* while loop will break here if sd == NULL */
1770 #endif /* CONFIG_SMP */
1772 static unsigned long
1773 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1775 unsigned long gran = sysctl_sched_wakeup_granularity;
1778 * Since its curr running now, convert the gran from real-time
1779 * to virtual-time in his units.
1781 * By using 'se' instead of 'curr' we penalize light tasks, so
1782 * they get preempted easier. That is, if 'se' < 'curr' then
1783 * the resulting gran will be larger, therefore penalizing the
1784 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1785 * be smaller, again penalizing the lighter task.
1787 * This is especially important for buddies when the leftmost
1788 * task is higher priority than the buddy.
1790 if (unlikely(se->load.weight != NICE_0_LOAD))
1791 gran = calc_delta_fair(gran, se);
1797 * Should 'se' preempt 'curr'.
1811 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1813 s64 gran, vdiff = curr->vruntime - se->vruntime;
1818 gran = wakeup_gran(curr, se);
1825 static void set_last_buddy(struct sched_entity *se)
1827 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1828 for_each_sched_entity(se)
1829 cfs_rq_of(se)->last = se;
1833 static void set_next_buddy(struct sched_entity *se)
1835 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1836 for_each_sched_entity(se)
1837 cfs_rq_of(se)->next = se;
1842 * Preempt the current task with a newly woken task if needed:
1844 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1846 struct task_struct *curr = rq->curr;
1847 struct sched_entity *se = &curr->se, *pse = &p->se;
1848 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1849 int scale = cfs_rq->nr_running >= sched_nr_latency;
1851 if (unlikely(se == pse))
1854 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1855 set_next_buddy(pse);
1858 * We can come here with TIF_NEED_RESCHED already set from new task
1861 if (test_tsk_need_resched(curr))
1865 * Batch and idle tasks do not preempt (their preemption is driven by
1868 if (unlikely(p->policy != SCHED_NORMAL))
1871 /* Idle tasks are by definition preempted by everybody. */
1872 if (unlikely(curr->policy == SCHED_IDLE))
1875 if (!sched_feat(WAKEUP_PREEMPT))
1878 update_curr(cfs_rq);
1879 find_matching_se(&se, &pse);
1881 if (wakeup_preempt_entity(se, pse) == 1)
1889 * Only set the backward buddy when the current task is still
1890 * on the rq. This can happen when a wakeup gets interleaved
1891 * with schedule on the ->pre_schedule() or idle_balance()
1892 * point, either of which can * drop the rq lock.
1894 * Also, during early boot the idle thread is in the fair class,
1895 * for obvious reasons its a bad idea to schedule back to it.
1897 if (unlikely(!se->on_rq || curr == rq->idle))
1900 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1904 static struct task_struct *pick_next_task_fair(struct rq *rq)
1906 struct task_struct *p;
1907 struct cfs_rq *cfs_rq = &rq->cfs;
1908 struct sched_entity *se;
1910 if (!cfs_rq->nr_running)
1914 se = pick_next_entity(cfs_rq);
1915 set_next_entity(cfs_rq, se);
1916 cfs_rq = group_cfs_rq(se);
1920 hrtick_start_fair(rq, p);
1926 * Account for a descheduled task:
1928 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1930 struct sched_entity *se = &prev->se;
1931 struct cfs_rq *cfs_rq;
1933 for_each_sched_entity(se) {
1934 cfs_rq = cfs_rq_of(se);
1935 put_prev_entity(cfs_rq, se);
1940 /**************************************************
1941 * Fair scheduling class load-balancing methods:
1945 * pull_task - move a task from a remote runqueue to the local runqueue.
1946 * Both runqueues must be locked.
1948 static void pull_task(struct rq *src_rq, struct task_struct *p,
1949 struct rq *this_rq, int this_cpu)
1951 deactivate_task(src_rq, p, 0);
1952 set_task_cpu(p, this_cpu);
1953 activate_task(this_rq, p, 0);
1954 check_preempt_curr(this_rq, p, 0);
1958 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1961 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1962 struct sched_domain *sd, enum cpu_idle_type idle,
1965 int tsk_cache_hot = 0;
1967 * We do not migrate tasks that are:
1968 * 1) running (obviously), or
1969 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1970 * 3) are cache-hot on their current CPU.
1972 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1973 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1978 if (task_running(rq, p)) {
1979 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1984 * Aggressive migration if:
1985 * 1) task is cache cold, or
1986 * 2) too many balance attempts have failed.
1989 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
1990 if (!tsk_cache_hot ||
1991 sd->nr_balance_failed > sd->cache_nice_tries) {
1992 #ifdef CONFIG_SCHEDSTATS
1993 if (tsk_cache_hot) {
1994 schedstat_inc(sd, lb_hot_gained[idle]);
1995 schedstat_inc(p, se.statistics.nr_forced_migrations);
2001 if (tsk_cache_hot) {
2002 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2009 * move_one_task tries to move exactly one task from busiest to this_rq, as
2010 * part of active balancing operations within "domain".
2011 * Returns 1 if successful and 0 otherwise.
2013 * Called with both runqueues locked.
2016 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2017 struct sched_domain *sd, enum cpu_idle_type idle)
2019 struct task_struct *p, *n;
2020 struct cfs_rq *cfs_rq;
2023 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2024 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2026 if (!can_migrate_task(p, busiest, this_cpu,
2030 pull_task(busiest, p, this_rq, this_cpu);
2032 * Right now, this is only the second place pull_task()
2033 * is called, so we can safely collect pull_task()
2034 * stats here rather than inside pull_task().
2036 schedstat_inc(sd, lb_gained[idle]);
2044 static unsigned long
2045 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2046 unsigned long max_load_move, struct sched_domain *sd,
2047 enum cpu_idle_type idle, int *all_pinned,
2048 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
2050 int loops = 0, pulled = 0, pinned = 0;
2051 long rem_load_move = max_load_move;
2052 struct task_struct *p, *n;
2054 if (max_load_move == 0)
2059 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2060 if (loops++ > sysctl_sched_nr_migrate)
2063 if ((p->se.load.weight >> 1) > rem_load_move ||
2064 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
2067 pull_task(busiest, p, this_rq, this_cpu);
2069 rem_load_move -= p->se.load.weight;
2071 #ifdef CONFIG_PREEMPT
2073 * NEWIDLE balancing is a source of latency, so preemptible
2074 * kernels will stop after the first task is pulled to minimize
2075 * the critical section.
2077 if (idle == CPU_NEWLY_IDLE)
2082 * We only want to steal up to the prescribed amount of
2085 if (rem_load_move <= 0)
2088 if (p->prio < *this_best_prio)
2089 *this_best_prio = p->prio;
2093 * Right now, this is one of only two places pull_task() is called,
2094 * so we can safely collect pull_task() stats here rather than
2095 * inside pull_task().
2097 schedstat_add(sd, lb_gained[idle], pulled);
2100 *all_pinned = pinned;
2102 return max_load_move - rem_load_move;
2105 #ifdef CONFIG_FAIR_GROUP_SCHED
2107 * update tg->load_weight by folding this cpu's load_avg
2109 static int update_shares_cpu(struct task_group *tg, int cpu)
2111 struct cfs_rq *cfs_rq;
2112 unsigned long flags;
2119 cfs_rq = tg->cfs_rq[cpu];
2121 raw_spin_lock_irqsave(&rq->lock, flags);
2123 update_rq_clock(rq);
2124 update_cfs_load(cfs_rq, 1);
2127 * We need to update shares after updating tg->load_weight in
2128 * order to adjust the weight of groups with long running tasks.
2130 update_cfs_shares(cfs_rq, 0);
2132 raw_spin_unlock_irqrestore(&rq->lock, flags);
2137 static void update_shares(int cpu)
2139 struct cfs_rq *cfs_rq;
2140 struct rq *rq = cpu_rq(cpu);
2143 for_each_leaf_cfs_rq(rq, cfs_rq)
2144 update_shares_cpu(cfs_rq->tg, cpu);
2148 static unsigned long
2149 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2150 unsigned long max_load_move,
2151 struct sched_domain *sd, enum cpu_idle_type idle,
2152 int *all_pinned, int *this_best_prio)
2154 long rem_load_move = max_load_move;
2155 int busiest_cpu = cpu_of(busiest);
2156 struct task_group *tg;
2159 update_h_load(busiest_cpu);
2161 list_for_each_entry_rcu(tg, &task_groups, list) {
2162 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2163 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2164 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2165 u64 rem_load, moved_load;
2170 if (!busiest_cfs_rq->task_weight)
2173 rem_load = (u64)rem_load_move * busiest_weight;
2174 rem_load = div_u64(rem_load, busiest_h_load + 1);
2176 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2177 rem_load, sd, idle, all_pinned, this_best_prio,
2183 moved_load *= busiest_h_load;
2184 moved_load = div_u64(moved_load, busiest_weight + 1);
2186 rem_load_move -= moved_load;
2187 if (rem_load_move < 0)
2192 return max_load_move - rem_load_move;
2195 static inline void update_shares(int cpu)
2199 static unsigned long
2200 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2201 unsigned long max_load_move,
2202 struct sched_domain *sd, enum cpu_idle_type idle,
2203 int *all_pinned, int *this_best_prio)
2205 return balance_tasks(this_rq, this_cpu, busiest,
2206 max_load_move, sd, idle, all_pinned,
2207 this_best_prio, &busiest->cfs);
2212 * move_tasks tries to move up to max_load_move weighted load from busiest to
2213 * this_rq, as part of a balancing operation within domain "sd".
2214 * Returns 1 if successful and 0 otherwise.
2216 * Called with both runqueues locked.
2218 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2219 unsigned long max_load_move,
2220 struct sched_domain *sd, enum cpu_idle_type idle,
2223 unsigned long total_load_moved = 0, load_moved;
2224 int this_best_prio = this_rq->curr->prio;
2227 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2228 max_load_move - total_load_moved,
2229 sd, idle, all_pinned, &this_best_prio);
2231 total_load_moved += load_moved;
2233 #ifdef CONFIG_PREEMPT
2235 * NEWIDLE balancing is a source of latency, so preemptible
2236 * kernels will stop after the first task is pulled to minimize
2237 * the critical section.
2239 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2242 if (raw_spin_is_contended(&this_rq->lock) ||
2243 raw_spin_is_contended(&busiest->lock))
2246 } while (load_moved && max_load_move > total_load_moved);
2248 return total_load_moved > 0;
2251 /********** Helpers for find_busiest_group ************************/
2253 * sd_lb_stats - Structure to store the statistics of a sched_domain
2254 * during load balancing.
2256 struct sd_lb_stats {
2257 struct sched_group *busiest; /* Busiest group in this sd */
2258 struct sched_group *this; /* Local group in this sd */
2259 unsigned long total_load; /* Total load of all groups in sd */
2260 unsigned long total_pwr; /* Total power of all groups in sd */
2261 unsigned long avg_load; /* Average load across all groups in sd */
2263 /** Statistics of this group */
2264 unsigned long this_load;
2265 unsigned long this_load_per_task;
2266 unsigned long this_nr_running;
2267 unsigned long this_has_capacity;
2268 unsigned int this_idle_cpus;
2270 /* Statistics of the busiest group */
2271 unsigned int busiest_idle_cpus;
2272 unsigned long max_load;
2273 unsigned long busiest_load_per_task;
2274 unsigned long busiest_nr_running;
2275 unsigned long busiest_group_capacity;
2276 unsigned long busiest_has_capacity;
2277 unsigned int busiest_group_weight;
2279 int group_imb; /* Is there imbalance in this sd */
2280 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2281 int power_savings_balance; /* Is powersave balance needed for this sd */
2282 struct sched_group *group_min; /* Least loaded group in sd */
2283 struct sched_group *group_leader; /* Group which relieves group_min */
2284 unsigned long min_load_per_task; /* load_per_task in group_min */
2285 unsigned long leader_nr_running; /* Nr running of group_leader */
2286 unsigned long min_nr_running; /* Nr running of group_min */
2291 * sg_lb_stats - stats of a sched_group required for load_balancing
2293 struct sg_lb_stats {
2294 unsigned long avg_load; /*Avg load across the CPUs of the group */
2295 unsigned long group_load; /* Total load over the CPUs of the group */
2296 unsigned long sum_nr_running; /* Nr tasks running in the group */
2297 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2298 unsigned long group_capacity;
2299 unsigned long idle_cpus;
2300 unsigned long group_weight;
2301 int group_imb; /* Is there an imbalance in the group ? */
2302 int group_has_capacity; /* Is there extra capacity in the group? */
2306 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2307 * @group: The group whose first cpu is to be returned.
2309 static inline unsigned int group_first_cpu(struct sched_group *group)
2311 return cpumask_first(sched_group_cpus(group));
2315 * get_sd_load_idx - Obtain the load index for a given sched domain.
2316 * @sd: The sched_domain whose load_idx is to be obtained.
2317 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2319 static inline int get_sd_load_idx(struct sched_domain *sd,
2320 enum cpu_idle_type idle)
2326 load_idx = sd->busy_idx;
2329 case CPU_NEWLY_IDLE:
2330 load_idx = sd->newidle_idx;
2333 load_idx = sd->idle_idx;
2341 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2343 * init_sd_power_savings_stats - Initialize power savings statistics for
2344 * the given sched_domain, during load balancing.
2346 * @sd: Sched domain whose power-savings statistics are to be initialized.
2347 * @sds: Variable containing the statistics for sd.
2348 * @idle: Idle status of the CPU at which we're performing load-balancing.
2350 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2351 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2354 * Busy processors will not participate in power savings
2357 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2358 sds->power_savings_balance = 0;
2360 sds->power_savings_balance = 1;
2361 sds->min_nr_running = ULONG_MAX;
2362 sds->leader_nr_running = 0;
2367 * update_sd_power_savings_stats - Update the power saving stats for a
2368 * sched_domain while performing load balancing.
2370 * @group: sched_group belonging to the sched_domain under consideration.
2371 * @sds: Variable containing the statistics of the sched_domain
2372 * @local_group: Does group contain the CPU for which we're performing
2374 * @sgs: Variable containing the statistics of the group.
2376 static inline void update_sd_power_savings_stats(struct sched_group *group,
2377 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2380 if (!sds->power_savings_balance)
2384 * If the local group is idle or completely loaded
2385 * no need to do power savings balance at this domain
2387 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2388 !sds->this_nr_running))
2389 sds->power_savings_balance = 0;
2392 * If a group is already running at full capacity or idle,
2393 * don't include that group in power savings calculations
2395 if (!sds->power_savings_balance ||
2396 sgs->sum_nr_running >= sgs->group_capacity ||
2397 !sgs->sum_nr_running)
2401 * Calculate the group which has the least non-idle load.
2402 * This is the group from where we need to pick up the load
2405 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2406 (sgs->sum_nr_running == sds->min_nr_running &&
2407 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2408 sds->group_min = group;
2409 sds->min_nr_running = sgs->sum_nr_running;
2410 sds->min_load_per_task = sgs->sum_weighted_load /
2411 sgs->sum_nr_running;
2415 * Calculate the group which is almost near its
2416 * capacity but still has some space to pick up some load
2417 * from other group and save more power
2419 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2422 if (sgs->sum_nr_running > sds->leader_nr_running ||
2423 (sgs->sum_nr_running == sds->leader_nr_running &&
2424 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2425 sds->group_leader = group;
2426 sds->leader_nr_running = sgs->sum_nr_running;
2431 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2432 * @sds: Variable containing the statistics of the sched_domain
2433 * under consideration.
2434 * @this_cpu: Cpu at which we're currently performing load-balancing.
2435 * @imbalance: Variable to store the imbalance.
2438 * Check if we have potential to perform some power-savings balance.
2439 * If yes, set the busiest group to be the least loaded group in the
2440 * sched_domain, so that it's CPUs can be put to idle.
2442 * Returns 1 if there is potential to perform power-savings balance.
2445 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2446 int this_cpu, unsigned long *imbalance)
2448 if (!sds->power_savings_balance)
2451 if (sds->this != sds->group_leader ||
2452 sds->group_leader == sds->group_min)
2455 *imbalance = sds->min_load_per_task;
2456 sds->busiest = sds->group_min;
2461 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2462 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2463 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2468 static inline void update_sd_power_savings_stats(struct sched_group *group,
2469 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2474 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2475 int this_cpu, unsigned long *imbalance)
2479 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2482 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2484 return SCHED_LOAD_SCALE;
2487 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2489 return default_scale_freq_power(sd, cpu);
2492 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2494 unsigned long weight = sd->span_weight;
2495 unsigned long smt_gain = sd->smt_gain;
2502 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2504 return default_scale_smt_power(sd, cpu);
2507 unsigned long scale_rt_power(int cpu)
2509 struct rq *rq = cpu_rq(cpu);
2510 u64 total, available;
2512 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2514 if (unlikely(total < rq->rt_avg)) {
2515 /* Ensures that power won't end up being negative */
2518 available = total - rq->rt_avg;
2521 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2522 total = SCHED_LOAD_SCALE;
2524 total >>= SCHED_LOAD_SHIFT;
2526 return div_u64(available, total);
2529 static void update_cpu_power(struct sched_domain *sd, int cpu)
2531 unsigned long weight = sd->span_weight;
2532 unsigned long power = SCHED_LOAD_SCALE;
2533 struct sched_group *sdg = sd->groups;
2535 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2536 if (sched_feat(ARCH_POWER))
2537 power *= arch_scale_smt_power(sd, cpu);
2539 power *= default_scale_smt_power(sd, cpu);
2541 power >>= SCHED_LOAD_SHIFT;
2544 sdg->cpu_power_orig = power;
2546 if (sched_feat(ARCH_POWER))
2547 power *= arch_scale_freq_power(sd, cpu);
2549 power *= default_scale_freq_power(sd, cpu);
2551 power >>= SCHED_LOAD_SHIFT;
2553 power *= scale_rt_power(cpu);
2554 power >>= SCHED_LOAD_SHIFT;
2559 cpu_rq(cpu)->cpu_power = power;
2560 sdg->cpu_power = power;
2563 static void update_group_power(struct sched_domain *sd, int cpu)
2565 struct sched_domain *child = sd->child;
2566 struct sched_group *group, *sdg = sd->groups;
2567 unsigned long power;
2570 update_cpu_power(sd, cpu);
2576 group = child->groups;
2578 power += group->cpu_power;
2579 group = group->next;
2580 } while (group != child->groups);
2582 sdg->cpu_power = power;
2586 * Try and fix up capacity for tiny siblings, this is needed when
2587 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2588 * which on its own isn't powerful enough.
2590 * See update_sd_pick_busiest() and check_asym_packing().
2593 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2596 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2598 if (sd->level != SD_LV_SIBLING)
2602 * If ~90% of the cpu_power is still there, we're good.
2604 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2611 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2612 * @sd: The sched_domain whose statistics are to be updated.
2613 * @group: sched_group whose statistics are to be updated.
2614 * @this_cpu: Cpu for which load balance is currently performed.
2615 * @idle: Idle status of this_cpu
2616 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2617 * @sd_idle: Idle status of the sched_domain containing group.
2618 * @local_group: Does group contain this_cpu.
2619 * @cpus: Set of cpus considered for load balancing.
2620 * @balance: Should we balance.
2621 * @sgs: variable to hold the statistics for this group.
2623 static inline void update_sg_lb_stats(struct sched_domain *sd,
2624 struct sched_group *group, int this_cpu,
2625 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2626 int local_group, const struct cpumask *cpus,
2627 int *balance, struct sg_lb_stats *sgs)
2629 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2631 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2632 unsigned long avg_load_per_task = 0;
2635 balance_cpu = group_first_cpu(group);
2637 /* Tally up the load of all CPUs in the group */
2639 min_cpu_load = ~0UL;
2642 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2643 struct rq *rq = cpu_rq(i);
2645 if (*sd_idle && rq->nr_running)
2648 /* Bias balancing toward cpus of our domain */
2650 if (idle_cpu(i) && !first_idle_cpu) {
2655 load = target_load(i, load_idx);
2657 load = source_load(i, load_idx);
2658 if (load > max_cpu_load) {
2659 max_cpu_load = load;
2660 max_nr_running = rq->nr_running;
2662 if (min_cpu_load > load)
2663 min_cpu_load = load;
2666 sgs->group_load += load;
2667 sgs->sum_nr_running += rq->nr_running;
2668 sgs->sum_weighted_load += weighted_cpuload(i);
2674 * First idle cpu or the first cpu(busiest) in this sched group
2675 * is eligible for doing load balancing at this and above
2676 * domains. In the newly idle case, we will allow all the cpu's
2677 * to do the newly idle load balance.
2679 if (idle != CPU_NEWLY_IDLE && local_group) {
2680 if (balance_cpu != this_cpu) {
2684 update_group_power(sd, this_cpu);
2687 /* Adjust by relative CPU power of the group */
2688 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2691 * Consider the group unbalanced when the imbalance is larger
2692 * than the average weight of two tasks.
2694 * APZ: with cgroup the avg task weight can vary wildly and
2695 * might not be a suitable number - should we keep a
2696 * normalized nr_running number somewhere that negates
2699 if (sgs->sum_nr_running)
2700 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2702 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2705 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2706 if (!sgs->group_capacity)
2707 sgs->group_capacity = fix_small_capacity(sd, group);
2708 sgs->group_weight = group->group_weight;
2710 if (sgs->group_capacity > sgs->sum_nr_running)
2711 sgs->group_has_capacity = 1;
2715 * update_sd_pick_busiest - return 1 on busiest group
2716 * @sd: sched_domain whose statistics are to be checked
2717 * @sds: sched_domain statistics
2718 * @sg: sched_group candidate to be checked for being the busiest
2719 * @sgs: sched_group statistics
2720 * @this_cpu: the current cpu
2722 * Determine if @sg is a busier group than the previously selected
2725 static bool update_sd_pick_busiest(struct sched_domain *sd,
2726 struct sd_lb_stats *sds,
2727 struct sched_group *sg,
2728 struct sg_lb_stats *sgs,
2731 if (sgs->avg_load <= sds->max_load)
2734 if (sgs->sum_nr_running > sgs->group_capacity)
2741 * ASYM_PACKING needs to move all the work to the lowest
2742 * numbered CPUs in the group, therefore mark all groups
2743 * higher than ourself as busy.
2745 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2746 this_cpu < group_first_cpu(sg)) {
2750 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2758 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2759 * @sd: sched_domain whose statistics are to be updated.
2760 * @this_cpu: Cpu for which load balance is currently performed.
2761 * @idle: Idle status of this_cpu
2762 * @sd_idle: Idle status of the sched_domain containing sg.
2763 * @cpus: Set of cpus considered for load balancing.
2764 * @balance: Should we balance.
2765 * @sds: variable to hold the statistics for this sched_domain.
2767 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2768 enum cpu_idle_type idle, int *sd_idle,
2769 const struct cpumask *cpus, int *balance,
2770 struct sd_lb_stats *sds)
2772 struct sched_domain *child = sd->child;
2773 struct sched_group *sg = sd->groups;
2774 struct sg_lb_stats sgs;
2775 int load_idx, prefer_sibling = 0;
2777 if (child && child->flags & SD_PREFER_SIBLING)
2780 init_sd_power_savings_stats(sd, sds, idle);
2781 load_idx = get_sd_load_idx(sd, idle);
2786 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2787 memset(&sgs, 0, sizeof(sgs));
2788 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2789 local_group, cpus, balance, &sgs);
2791 if (local_group && !(*balance))
2794 sds->total_load += sgs.group_load;
2795 sds->total_pwr += sg->cpu_power;
2798 * In case the child domain prefers tasks go to siblings
2799 * first, lower the sg capacity to one so that we'll try
2800 * and move all the excess tasks away. We lower the capacity
2801 * of a group only if the local group has the capacity to fit
2802 * these excess tasks, i.e. nr_running < group_capacity. The
2803 * extra check prevents the case where you always pull from the
2804 * heaviest group when it is already under-utilized (possible
2805 * with a large weight task outweighs the tasks on the system).
2807 if (prefer_sibling && !local_group && sds->this_has_capacity)
2808 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2811 sds->this_load = sgs.avg_load;
2813 sds->this_nr_running = sgs.sum_nr_running;
2814 sds->this_load_per_task = sgs.sum_weighted_load;
2815 sds->this_has_capacity = sgs.group_has_capacity;
2816 sds->this_idle_cpus = sgs.idle_cpus;
2817 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2818 sds->max_load = sgs.avg_load;
2820 sds->busiest_nr_running = sgs.sum_nr_running;
2821 sds->busiest_idle_cpus = sgs.idle_cpus;
2822 sds->busiest_group_capacity = sgs.group_capacity;
2823 sds->busiest_load_per_task = sgs.sum_weighted_load;
2824 sds->busiest_has_capacity = sgs.group_has_capacity;
2825 sds->busiest_group_weight = sgs.group_weight;
2826 sds->group_imb = sgs.group_imb;
2829 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2831 } while (sg != sd->groups);
2834 int __weak arch_sd_sibling_asym_packing(void)
2836 return 0*SD_ASYM_PACKING;
2840 * check_asym_packing - Check to see if the group is packed into the
2843 * This is primarily intended to used at the sibling level. Some
2844 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2845 * case of POWER7, it can move to lower SMT modes only when higher
2846 * threads are idle. When in lower SMT modes, the threads will
2847 * perform better since they share less core resources. Hence when we
2848 * have idle threads, we want them to be the higher ones.
2850 * This packing function is run on idle threads. It checks to see if
2851 * the busiest CPU in this domain (core in the P7 case) has a higher
2852 * CPU number than the packing function is being run on. Here we are
2853 * assuming lower CPU number will be equivalent to lower a SMT thread
2856 * Returns 1 when packing is required and a task should be moved to
2857 * this CPU. The amount of the imbalance is returned in *imbalance.
2859 * @sd: The sched_domain whose packing is to be checked.
2860 * @sds: Statistics of the sched_domain which is to be packed
2861 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2862 * @imbalance: returns amount of imbalanced due to packing.
2864 static int check_asym_packing(struct sched_domain *sd,
2865 struct sd_lb_stats *sds,
2866 int this_cpu, unsigned long *imbalance)
2870 if (!(sd->flags & SD_ASYM_PACKING))
2876 busiest_cpu = group_first_cpu(sds->busiest);
2877 if (this_cpu > busiest_cpu)
2880 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2886 * fix_small_imbalance - Calculate the minor imbalance that exists
2887 * amongst the groups of a sched_domain, during
2889 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2890 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2891 * @imbalance: Variable to store the imbalance.
2893 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2894 int this_cpu, unsigned long *imbalance)
2896 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2897 unsigned int imbn = 2;
2898 unsigned long scaled_busy_load_per_task;
2900 if (sds->this_nr_running) {
2901 sds->this_load_per_task /= sds->this_nr_running;
2902 if (sds->busiest_load_per_task >
2903 sds->this_load_per_task)
2906 sds->this_load_per_task =
2907 cpu_avg_load_per_task(this_cpu);
2909 scaled_busy_load_per_task = sds->busiest_load_per_task
2911 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2913 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2914 (scaled_busy_load_per_task * imbn)) {
2915 *imbalance = sds->busiest_load_per_task;
2920 * OK, we don't have enough imbalance to justify moving tasks,
2921 * however we may be able to increase total CPU power used by
2925 pwr_now += sds->busiest->cpu_power *
2926 min(sds->busiest_load_per_task, sds->max_load);
2927 pwr_now += sds->this->cpu_power *
2928 min(sds->this_load_per_task, sds->this_load);
2929 pwr_now /= SCHED_LOAD_SCALE;
2931 /* Amount of load we'd subtract */
2932 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2933 sds->busiest->cpu_power;
2934 if (sds->max_load > tmp)
2935 pwr_move += sds->busiest->cpu_power *
2936 min(sds->busiest_load_per_task, sds->max_load - tmp);
2938 /* Amount of load we'd add */
2939 if (sds->max_load * sds->busiest->cpu_power <
2940 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2941 tmp = (sds->max_load * sds->busiest->cpu_power) /
2942 sds->this->cpu_power;
2944 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2945 sds->this->cpu_power;
2946 pwr_move += sds->this->cpu_power *
2947 min(sds->this_load_per_task, sds->this_load + tmp);
2948 pwr_move /= SCHED_LOAD_SCALE;
2950 /* Move if we gain throughput */
2951 if (pwr_move > pwr_now)
2952 *imbalance = sds->busiest_load_per_task;
2956 * calculate_imbalance - Calculate the amount of imbalance present within the
2957 * groups of a given sched_domain during load balance.
2958 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2959 * @this_cpu: Cpu for which currently load balance is being performed.
2960 * @imbalance: The variable to store the imbalance.
2962 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2963 unsigned long *imbalance)
2965 unsigned long max_pull, load_above_capacity = ~0UL;
2967 sds->busiest_load_per_task /= sds->busiest_nr_running;
2968 if (sds->group_imb) {
2969 sds->busiest_load_per_task =
2970 min(sds->busiest_load_per_task, sds->avg_load);
2974 * In the presence of smp nice balancing, certain scenarios can have
2975 * max load less than avg load(as we skip the groups at or below
2976 * its cpu_power, while calculating max_load..)
2978 if (sds->max_load < sds->avg_load) {
2980 return fix_small_imbalance(sds, this_cpu, imbalance);
2983 if (!sds->group_imb) {
2985 * Don't want to pull so many tasks that a group would go idle.
2987 load_above_capacity = (sds->busiest_nr_running -
2988 sds->busiest_group_capacity);
2990 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2992 load_above_capacity /= sds->busiest->cpu_power;
2996 * We're trying to get all the cpus to the average_load, so we don't
2997 * want to push ourselves above the average load, nor do we wish to
2998 * reduce the max loaded cpu below the average load. At the same time,
2999 * we also don't want to reduce the group load below the group capacity
3000 * (so that we can implement power-savings policies etc). Thus we look
3001 * for the minimum possible imbalance.
3002 * Be careful of negative numbers as they'll appear as very large values
3003 * with unsigned longs.
3005 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3007 /* How much load to actually move to equalise the imbalance */
3008 *imbalance = min(max_pull * sds->busiest->cpu_power,
3009 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3013 * if *imbalance is less than the average load per runnable task
3014 * there is no gaurantee that any tasks will be moved so we'll have
3015 * a think about bumping its value to force at least one task to be
3018 if (*imbalance < sds->busiest_load_per_task)
3019 return fix_small_imbalance(sds, this_cpu, imbalance);
3023 /******* find_busiest_group() helpers end here *********************/
3026 * find_busiest_group - Returns the busiest group within the sched_domain
3027 * if there is an imbalance. If there isn't an imbalance, and
3028 * the user has opted for power-savings, it returns a group whose
3029 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3030 * such a group exists.
3032 * Also calculates the amount of weighted load which should be moved
3033 * to restore balance.
3035 * @sd: The sched_domain whose busiest group is to be returned.
3036 * @this_cpu: The cpu for which load balancing is currently being performed.
3037 * @imbalance: Variable which stores amount of weighted load which should
3038 * be moved to restore balance/put a group to idle.
3039 * @idle: The idle status of this_cpu.
3040 * @sd_idle: The idleness of sd
3041 * @cpus: The set of CPUs under consideration for load-balancing.
3042 * @balance: Pointer to a variable indicating if this_cpu
3043 * is the appropriate cpu to perform load balancing at this_level.
3045 * Returns: - the busiest group if imbalance exists.
3046 * - If no imbalance and user has opted for power-savings balance,
3047 * return the least loaded group whose CPUs can be
3048 * put to idle by rebalancing its tasks onto our group.
3050 static struct sched_group *
3051 find_busiest_group(struct sched_domain *sd, int this_cpu,
3052 unsigned long *imbalance, enum cpu_idle_type idle,
3053 int *sd_idle, const struct cpumask *cpus, int *balance)
3055 struct sd_lb_stats sds;
3057 memset(&sds, 0, sizeof(sds));
3060 * Compute the various statistics relavent for load balancing at
3063 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3066 /* Cases where imbalance does not exist from POV of this_cpu */
3067 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3069 * 2) There is no busy sibling group to pull from.
3070 * 3) This group is the busiest group.
3071 * 4) This group is more busy than the avg busieness at this
3073 * 5) The imbalance is within the specified limit.
3075 * Note: when doing newidle balance, if the local group has excess
3076 * capacity (i.e. nr_running < group_capacity) and the busiest group
3077 * does not have any capacity, we force a load balance to pull tasks
3078 * to the local group. In this case, we skip past checks 3, 4 and 5.
3083 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3084 check_asym_packing(sd, &sds, this_cpu, imbalance))
3087 if (!sds.busiest || sds.busiest_nr_running == 0)
3090 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3091 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3092 !sds.busiest_has_capacity)
3095 if (sds.this_load >= sds.max_load)
3098 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3100 if (sds.this_load >= sds.avg_load)
3104 * In the CPU_NEWLY_IDLE, use imbalance_pct to be conservative.
3105 * And to check for busy balance use !idle_cpu instead of
3106 * CPU_NOT_IDLE. This is because HT siblings will use CPU_NOT_IDLE
3107 * even when they are idle.
3109 if (idle == CPU_NEWLY_IDLE || !idle_cpu(this_cpu)) {
3110 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3114 * This cpu is idle. If the busiest group load doesn't
3115 * have more tasks than the number of available cpu's and
3116 * there is no imbalance between this and busiest group
3117 * wrt to idle cpu's, it is balanced.
3119 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3120 sds.busiest_nr_running <= sds.busiest_group_weight)
3125 /* Looks like there is an imbalance. Compute it */
3126 calculate_imbalance(&sds, this_cpu, imbalance);
3131 * There is no obvious imbalance. But check if we can do some balancing
3134 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3142 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3145 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3146 enum cpu_idle_type idle, unsigned long imbalance,
3147 const struct cpumask *cpus)
3149 struct rq *busiest = NULL, *rq;
3150 unsigned long max_load = 0;
3153 for_each_cpu(i, sched_group_cpus(group)) {
3154 unsigned long power = power_of(i);
3155 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3159 capacity = fix_small_capacity(sd, group);
3161 if (!cpumask_test_cpu(i, cpus))
3165 wl = weighted_cpuload(i);
3168 * When comparing with imbalance, use weighted_cpuload()
3169 * which is not scaled with the cpu power.
3171 if (capacity && rq->nr_running == 1 && wl > imbalance)
3175 * For the load comparisons with the other cpu's, consider
3176 * the weighted_cpuload() scaled with the cpu power, so that
3177 * the load can be moved away from the cpu that is potentially
3178 * running at a lower capacity.
3180 wl = (wl * SCHED_LOAD_SCALE) / power;
3182 if (wl > max_load) {
3192 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3193 * so long as it is large enough.
3195 #define MAX_PINNED_INTERVAL 512
3197 /* Working cpumask for load_balance and load_balance_newidle. */
3198 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3200 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
3201 int busiest_cpu, int this_cpu)
3203 if (idle == CPU_NEWLY_IDLE) {
3206 * ASYM_PACKING needs to force migrate tasks from busy but
3207 * higher numbered CPUs in order to pack all tasks in the
3208 * lowest numbered CPUs.
3210 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3214 * The only task running in a non-idle cpu can be moved to this
3215 * cpu in an attempt to completely freeup the other CPU
3218 * The package power saving logic comes from
3219 * find_busiest_group(). If there are no imbalance, then
3220 * f_b_g() will return NULL. However when sched_mc={1,2} then
3221 * f_b_g() will select a group from which a running task may be
3222 * pulled to this cpu in order to make the other package idle.
3223 * If there is no opportunity to make a package idle and if
3224 * there are no imbalance, then f_b_g() will return NULL and no
3225 * action will be taken in load_balance_newidle().
3227 * Under normal task pull operation due to imbalance, there
3228 * will be more than one task in the source run queue and
3229 * move_tasks() will succeed. ld_moved will be true and this
3230 * active balance code will not be triggered.
3232 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3233 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3236 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3240 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3243 static int active_load_balance_cpu_stop(void *data);
3246 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3247 * tasks if there is an imbalance.
3249 static int load_balance(int this_cpu, struct rq *this_rq,
3250 struct sched_domain *sd, enum cpu_idle_type idle,
3253 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3254 struct sched_group *group;
3255 unsigned long imbalance;
3257 unsigned long flags;
3258 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3260 cpumask_copy(cpus, cpu_active_mask);
3263 * When power savings policy is enabled for the parent domain, idle
3264 * sibling can pick up load irrespective of busy siblings. In this case,
3265 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3266 * portraying it as CPU_NOT_IDLE.
3268 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3269 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3272 schedstat_inc(sd, lb_count[idle]);
3275 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3282 schedstat_inc(sd, lb_nobusyg[idle]);
3286 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3288 schedstat_inc(sd, lb_nobusyq[idle]);
3292 BUG_ON(busiest == this_rq);
3294 schedstat_add(sd, lb_imbalance[idle], imbalance);
3297 if (busiest->nr_running > 1) {
3299 * Attempt to move tasks. If find_busiest_group has found
3300 * an imbalance but busiest->nr_running <= 1, the group is
3301 * still unbalanced. ld_moved simply stays zero, so it is
3302 * correctly treated as an imbalance.
3304 local_irq_save(flags);
3305 double_rq_lock(this_rq, busiest);
3306 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3307 imbalance, sd, idle, &all_pinned);
3308 double_rq_unlock(this_rq, busiest);
3309 local_irq_restore(flags);
3312 * some other cpu did the load balance for us.
3314 if (ld_moved && this_cpu != smp_processor_id())
3315 resched_cpu(this_cpu);
3317 /* All tasks on this runqueue were pinned by CPU affinity */
3318 if (unlikely(all_pinned)) {
3319 cpumask_clear_cpu(cpu_of(busiest), cpus);
3320 if (!cpumask_empty(cpus))
3327 schedstat_inc(sd, lb_failed[idle]);
3329 * Increment the failure counter only on periodic balance.
3330 * We do not want newidle balance, which can be very
3331 * frequent, pollute the failure counter causing
3332 * excessive cache_hot migrations and active balances.
3334 if (idle != CPU_NEWLY_IDLE)
3335 sd->nr_balance_failed++;
3337 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3339 raw_spin_lock_irqsave(&busiest->lock, flags);
3341 /* don't kick the active_load_balance_cpu_stop,
3342 * if the curr task on busiest cpu can't be
3345 if (!cpumask_test_cpu(this_cpu,
3346 &busiest->curr->cpus_allowed)) {
3347 raw_spin_unlock_irqrestore(&busiest->lock,
3350 goto out_one_pinned;
3354 * ->active_balance synchronizes accesses to
3355 * ->active_balance_work. Once set, it's cleared
3356 * only after active load balance is finished.
3358 if (!busiest->active_balance) {
3359 busiest->active_balance = 1;
3360 busiest->push_cpu = this_cpu;
3363 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3366 stop_one_cpu_nowait(cpu_of(busiest),
3367 active_load_balance_cpu_stop, busiest,
3368 &busiest->active_balance_work);
3371 * We've kicked active balancing, reset the failure
3374 sd->nr_balance_failed = sd->cache_nice_tries+1;
3377 sd->nr_balance_failed = 0;
3379 if (likely(!active_balance)) {
3380 /* We were unbalanced, so reset the balancing interval */
3381 sd->balance_interval = sd->min_interval;
3384 * If we've begun active balancing, start to back off. This
3385 * case may not be covered by the all_pinned logic if there
3386 * is only 1 task on the busy runqueue (because we don't call
3389 if (sd->balance_interval < sd->max_interval)
3390 sd->balance_interval *= 2;
3393 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3394 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3400 schedstat_inc(sd, lb_balanced[idle]);
3402 sd->nr_balance_failed = 0;
3405 /* tune up the balancing interval */
3406 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3407 (sd->balance_interval < sd->max_interval))
3408 sd->balance_interval *= 2;
3410 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3411 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3420 * idle_balance is called by schedule() if this_cpu is about to become
3421 * idle. Attempts to pull tasks from other CPUs.
3423 static void idle_balance(int this_cpu, struct rq *this_rq)
3425 struct sched_domain *sd;
3426 int pulled_task = 0;
3427 unsigned long next_balance = jiffies + HZ;
3429 this_rq->idle_stamp = this_rq->clock;
3431 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3435 * Drop the rq->lock, but keep IRQ/preempt disabled.
3437 raw_spin_unlock(&this_rq->lock);
3439 update_shares(this_cpu);
3440 for_each_domain(this_cpu, sd) {
3441 unsigned long interval;
3444 if (!(sd->flags & SD_LOAD_BALANCE))
3447 if (sd->flags & SD_BALANCE_NEWIDLE) {
3448 /* If we've pulled tasks over stop searching: */
3449 pulled_task = load_balance(this_cpu, this_rq,
3450 sd, CPU_NEWLY_IDLE, &balance);
3453 interval = msecs_to_jiffies(sd->balance_interval);
3454 if (time_after(next_balance, sd->last_balance + interval))
3455 next_balance = sd->last_balance + interval;
3457 this_rq->idle_stamp = 0;
3462 raw_spin_lock(&this_rq->lock);
3464 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3466 * We are going idle. next_balance may be set based on
3467 * a busy processor. So reset next_balance.
3469 this_rq->next_balance = next_balance;
3474 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3475 * running tasks off the busiest CPU onto idle CPUs. It requires at
3476 * least 1 task to be running on each physical CPU where possible, and
3477 * avoids physical / logical imbalances.
3479 static int active_load_balance_cpu_stop(void *data)
3481 struct rq *busiest_rq = data;
3482 int busiest_cpu = cpu_of(busiest_rq);
3483 int target_cpu = busiest_rq->push_cpu;
3484 struct rq *target_rq = cpu_rq(target_cpu);
3485 struct sched_domain *sd;
3487 raw_spin_lock_irq(&busiest_rq->lock);
3489 /* make sure the requested cpu hasn't gone down in the meantime */
3490 if (unlikely(busiest_cpu != smp_processor_id() ||
3491 !busiest_rq->active_balance))
3494 /* Is there any task to move? */
3495 if (busiest_rq->nr_running <= 1)
3499 * This condition is "impossible", if it occurs
3500 * we need to fix it. Originally reported by
3501 * Bjorn Helgaas on a 128-cpu setup.
3503 BUG_ON(busiest_rq == target_rq);
3505 /* move a task from busiest_rq to target_rq */
3506 double_lock_balance(busiest_rq, target_rq);
3508 /* Search for an sd spanning us and the target CPU. */
3509 for_each_domain(target_cpu, sd) {
3510 if ((sd->flags & SD_LOAD_BALANCE) &&
3511 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3516 schedstat_inc(sd, alb_count);
3518 if (move_one_task(target_rq, target_cpu, busiest_rq,
3520 schedstat_inc(sd, alb_pushed);
3522 schedstat_inc(sd, alb_failed);
3524 double_unlock_balance(busiest_rq, target_rq);
3526 busiest_rq->active_balance = 0;
3527 raw_spin_unlock_irq(&busiest_rq->lock);
3533 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3535 static void trigger_sched_softirq(void *data)
3537 raise_softirq_irqoff(SCHED_SOFTIRQ);
3540 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3542 csd->func = trigger_sched_softirq;
3549 * idle load balancing details
3550 * - One of the idle CPUs nominates itself as idle load_balancer, while
3552 * - This idle load balancer CPU will also go into tickless mode when
3553 * it is idle, just like all other idle CPUs
3554 * - When one of the busy CPUs notice that there may be an idle rebalancing
3555 * needed, they will kick the idle load balancer, which then does idle
3556 * load balancing for all the idle CPUs.
3559 atomic_t load_balancer;
3560 atomic_t first_pick_cpu;
3561 atomic_t second_pick_cpu;
3562 cpumask_var_t idle_cpus_mask;
3563 cpumask_var_t grp_idle_mask;
3564 unsigned long next_balance; /* in jiffy units */
3565 } nohz ____cacheline_aligned;
3567 int get_nohz_load_balancer(void)
3569 return atomic_read(&nohz.load_balancer);
3572 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3574 * lowest_flag_domain - Return lowest sched_domain containing flag.
3575 * @cpu: The cpu whose lowest level of sched domain is to
3577 * @flag: The flag to check for the lowest sched_domain
3578 * for the given cpu.
3580 * Returns the lowest sched_domain of a cpu which contains the given flag.
3582 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3584 struct sched_domain *sd;
3586 for_each_domain(cpu, sd)
3587 if (sd && (sd->flags & flag))
3594 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3595 * @cpu: The cpu whose domains we're iterating over.
3596 * @sd: variable holding the value of the power_savings_sd
3598 * @flag: The flag to filter the sched_domains to be iterated.
3600 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3601 * set, starting from the lowest sched_domain to the highest.
3603 #define for_each_flag_domain(cpu, sd, flag) \
3604 for (sd = lowest_flag_domain(cpu, flag); \
3605 (sd && (sd->flags & flag)); sd = sd->parent)
3608 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3609 * @ilb_group: group to be checked for semi-idleness
3611 * Returns: 1 if the group is semi-idle. 0 otherwise.
3613 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3614 * and atleast one non-idle CPU. This helper function checks if the given
3615 * sched_group is semi-idle or not.
3617 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3619 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3620 sched_group_cpus(ilb_group));
3623 * A sched_group is semi-idle when it has atleast one busy cpu
3624 * and atleast one idle cpu.
3626 if (cpumask_empty(nohz.grp_idle_mask))
3629 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3635 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3636 * @cpu: The cpu which is nominating a new idle_load_balancer.
3638 * Returns: Returns the id of the idle load balancer if it exists,
3639 * Else, returns >= nr_cpu_ids.
3641 * This algorithm picks the idle load balancer such that it belongs to a
3642 * semi-idle powersavings sched_domain. The idea is to try and avoid
3643 * completely idle packages/cores just for the purpose of idle load balancing
3644 * when there are other idle cpu's which are better suited for that job.
3646 static int find_new_ilb(int cpu)
3648 struct sched_domain *sd;
3649 struct sched_group *ilb_group;
3652 * Have idle load balancer selection from semi-idle packages only
3653 * when power-aware load balancing is enabled
3655 if (!(sched_smt_power_savings || sched_mc_power_savings))
3659 * Optimize for the case when we have no idle CPUs or only one
3660 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3662 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3665 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3666 ilb_group = sd->groups;
3669 if (is_semi_idle_group(ilb_group))
3670 return cpumask_first(nohz.grp_idle_mask);
3672 ilb_group = ilb_group->next;
3674 } while (ilb_group != sd->groups);
3680 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3681 static inline int find_new_ilb(int call_cpu)
3688 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3689 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3690 * CPU (if there is one).
3692 static void nohz_balancer_kick(int cpu)
3696 nohz.next_balance++;
3698 ilb_cpu = get_nohz_load_balancer();
3700 if (ilb_cpu >= nr_cpu_ids) {
3701 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3702 if (ilb_cpu >= nr_cpu_ids)
3706 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3707 struct call_single_data *cp;
3709 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3710 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3711 __smp_call_function_single(ilb_cpu, cp, 0);
3717 * This routine will try to nominate the ilb (idle load balancing)
3718 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3719 * load balancing on behalf of all those cpus.
3721 * When the ilb owner becomes busy, we will not have new ilb owner until some
3722 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3723 * idle load balancing by kicking one of the idle CPUs.
3725 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3726 * ilb owner CPU in future (when there is a need for idle load balancing on
3727 * behalf of all idle CPUs).
3729 void select_nohz_load_balancer(int stop_tick)
3731 int cpu = smp_processor_id();
3734 if (!cpu_active(cpu)) {
3735 if (atomic_read(&nohz.load_balancer) != cpu)
3739 * If we are going offline and still the leader,
3742 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3749 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3751 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3752 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3753 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3754 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3756 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3759 /* make me the ilb owner */
3760 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3765 * Check to see if there is a more power-efficient
3768 new_ilb = find_new_ilb(cpu);
3769 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3770 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3771 resched_cpu(new_ilb);
3777 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3780 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3782 if (atomic_read(&nohz.load_balancer) == cpu)
3783 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3791 static DEFINE_SPINLOCK(balancing);
3794 * It checks each scheduling domain to see if it is due to be balanced,
3795 * and initiates a balancing operation if so.
3797 * Balancing parameters are set up in arch_init_sched_domains.
3799 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3802 struct rq *rq = cpu_rq(cpu);
3803 unsigned long interval;
3804 struct sched_domain *sd;
3805 /* Earliest time when we have to do rebalance again */
3806 unsigned long next_balance = jiffies + 60*HZ;
3807 int update_next_balance = 0;
3812 for_each_domain(cpu, sd) {
3813 if (!(sd->flags & SD_LOAD_BALANCE))
3816 interval = sd->balance_interval;
3817 if (idle != CPU_IDLE)
3818 interval *= sd->busy_factor;
3820 /* scale ms to jiffies */
3821 interval = msecs_to_jiffies(interval);
3822 if (unlikely(!interval))
3824 if (interval > HZ*NR_CPUS/10)
3825 interval = HZ*NR_CPUS/10;
3827 need_serialize = sd->flags & SD_SERIALIZE;
3829 if (need_serialize) {
3830 if (!spin_trylock(&balancing))
3834 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3835 if (load_balance(cpu, rq, sd, idle, &balance)) {
3837 * We've pulled tasks over so either we're no
3838 * longer idle, or one of our SMT siblings is
3841 idle = CPU_NOT_IDLE;
3843 sd->last_balance = jiffies;
3846 spin_unlock(&balancing);
3848 if (time_after(next_balance, sd->last_balance + interval)) {
3849 next_balance = sd->last_balance + interval;
3850 update_next_balance = 1;
3854 * Stop the load balance at this level. There is another
3855 * CPU in our sched group which is doing load balancing more
3863 * next_balance will be updated only when there is a need.
3864 * When the cpu is attached to null domain for ex, it will not be
3867 if (likely(update_next_balance))
3868 rq->next_balance = next_balance;
3873 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3874 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3876 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3878 struct rq *this_rq = cpu_rq(this_cpu);
3882 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3885 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3886 if (balance_cpu == this_cpu)
3890 * If this cpu gets work to do, stop the load balancing
3891 * work being done for other cpus. Next load
3892 * balancing owner will pick it up.
3894 if (need_resched()) {
3895 this_rq->nohz_balance_kick = 0;
3899 raw_spin_lock_irq(&this_rq->lock);
3900 update_rq_clock(this_rq);
3901 update_cpu_load(this_rq);
3902 raw_spin_unlock_irq(&this_rq->lock);
3904 rebalance_domains(balance_cpu, CPU_IDLE);
3906 rq = cpu_rq(balance_cpu);
3907 if (time_after(this_rq->next_balance, rq->next_balance))
3908 this_rq->next_balance = rq->next_balance;
3910 nohz.next_balance = this_rq->next_balance;
3911 this_rq->nohz_balance_kick = 0;
3915 * Current heuristic for kicking the idle load balancer
3916 * - first_pick_cpu is the one of the busy CPUs. It will kick
3917 * idle load balancer when it has more than one process active. This
3918 * eliminates the need for idle load balancing altogether when we have
3919 * only one running process in the system (common case).
3920 * - If there are more than one busy CPU, idle load balancer may have
3921 * to run for active_load_balance to happen (i.e., two busy CPUs are
3922 * SMT or core siblings and can run better if they move to different
3923 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3924 * which will kick idle load balancer as soon as it has any load.
3926 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3928 unsigned long now = jiffies;
3930 int first_pick_cpu, second_pick_cpu;
3932 if (time_before(now, nohz.next_balance))
3935 if (rq->idle_at_tick)
3938 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3939 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3941 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3942 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3945 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3946 if (ret == nr_cpu_ids || ret == cpu) {
3947 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3948 if (rq->nr_running > 1)
3951 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3952 if (ret == nr_cpu_ids || ret == cpu) {
3960 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3964 * run_rebalance_domains is triggered when needed from the scheduler tick.
3965 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3967 static void run_rebalance_domains(struct softirq_action *h)
3969 int this_cpu = smp_processor_id();
3970 struct rq *this_rq = cpu_rq(this_cpu);
3971 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3972 CPU_IDLE : CPU_NOT_IDLE;
3974 rebalance_domains(this_cpu, idle);
3977 * If this cpu has a pending nohz_balance_kick, then do the
3978 * balancing on behalf of the other idle cpus whose ticks are
3981 nohz_idle_balance(this_cpu, idle);
3984 static inline int on_null_domain(int cpu)
3986 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3990 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3992 static inline void trigger_load_balance(struct rq *rq, int cpu)
3994 /* Don't need to rebalance while attached to NULL domain */
3995 if (time_after_eq(jiffies, rq->next_balance) &&
3996 likely(!on_null_domain(cpu)))
3997 raise_softirq(SCHED_SOFTIRQ);
3999 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4000 nohz_balancer_kick(cpu);
4004 static void rq_online_fair(struct rq *rq)
4009 static void rq_offline_fair(struct rq *rq)
4014 #else /* CONFIG_SMP */
4017 * on UP we do not need to balance between CPUs:
4019 static inline void idle_balance(int cpu, struct rq *rq)
4023 #endif /* CONFIG_SMP */
4026 * scheduler tick hitting a task of our scheduling class:
4028 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4030 struct cfs_rq *cfs_rq;
4031 struct sched_entity *se = &curr->se;
4033 for_each_sched_entity(se) {
4034 cfs_rq = cfs_rq_of(se);
4035 entity_tick(cfs_rq, se, queued);
4040 * called on fork with the child task as argument from the parent's context
4041 * - child not yet on the tasklist
4042 * - preemption disabled
4044 static void task_fork_fair(struct task_struct *p)
4046 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4047 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4048 int this_cpu = smp_processor_id();
4049 struct rq *rq = this_rq();
4050 unsigned long flags;
4052 raw_spin_lock_irqsave(&rq->lock, flags);
4054 update_rq_clock(rq);
4056 if (unlikely(task_cpu(p) != this_cpu)) {
4058 __set_task_cpu(p, this_cpu);
4062 update_curr(cfs_rq);
4065 se->vruntime = curr->vruntime;
4066 place_entity(cfs_rq, se, 1);
4068 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4070 * Upon rescheduling, sched_class::put_prev_task() will place
4071 * 'current' within the tree based on its new key value.
4073 swap(curr->vruntime, se->vruntime);
4074 resched_task(rq->curr);
4077 se->vruntime -= cfs_rq->min_vruntime;
4079 raw_spin_unlock_irqrestore(&rq->lock, flags);
4083 * Priority of the task has changed. Check to see if we preempt
4086 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
4087 int oldprio, int running)
4090 * Reschedule if we are currently running on this runqueue and
4091 * our priority decreased, or if we are not currently running on
4092 * this runqueue and our priority is higher than the current's
4095 if (p->prio > oldprio)
4096 resched_task(rq->curr);
4098 check_preempt_curr(rq, p, 0);
4102 * We switched to the sched_fair class.
4104 static void switched_to_fair(struct rq *rq, struct task_struct *p,
4108 * We were most likely switched from sched_rt, so
4109 * kick off the schedule if running, otherwise just see
4110 * if we can still preempt the current task.
4113 resched_task(rq->curr);
4115 check_preempt_curr(rq, p, 0);
4118 /* Account for a task changing its policy or group.
4120 * This routine is mostly called to set cfs_rq->curr field when a task
4121 * migrates between groups/classes.
4123 static void set_curr_task_fair(struct rq *rq)
4125 struct sched_entity *se = &rq->curr->se;
4127 for_each_sched_entity(se)
4128 set_next_entity(cfs_rq_of(se), se);
4131 #ifdef CONFIG_FAIR_GROUP_SCHED
4132 static void task_move_group_fair(struct task_struct *p, int on_rq)
4135 * If the task was not on the rq at the time of this cgroup movement
4136 * it must have been asleep, sleeping tasks keep their ->vruntime
4137 * absolute on their old rq until wakeup (needed for the fair sleeper
4138 * bonus in place_entity()).
4140 * If it was on the rq, we've just 'preempted' it, which does convert
4141 * ->vruntime to a relative base.
4143 * Make sure both cases convert their relative position when migrating
4144 * to another cgroup's rq. This does somewhat interfere with the
4145 * fair sleeper stuff for the first placement, but who cares.
4148 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4149 set_task_rq(p, task_cpu(p));
4151 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4155 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4157 struct sched_entity *se = &task->se;
4158 unsigned int rr_interval = 0;
4161 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4164 if (rq->cfs.load.weight)
4165 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4171 * All the scheduling class methods:
4173 static const struct sched_class fair_sched_class = {
4174 .next = &idle_sched_class,
4175 .enqueue_task = enqueue_task_fair,
4176 .dequeue_task = dequeue_task_fair,
4177 .yield_task = yield_task_fair,
4179 .check_preempt_curr = check_preempt_wakeup,
4181 .pick_next_task = pick_next_task_fair,
4182 .put_prev_task = put_prev_task_fair,
4185 .select_task_rq = select_task_rq_fair,
4187 .rq_online = rq_online_fair,
4188 .rq_offline = rq_offline_fair,
4190 .task_waking = task_waking_fair,
4193 .set_curr_task = set_curr_task_fair,
4194 .task_tick = task_tick_fair,
4195 .task_fork = task_fork_fair,
4197 .prio_changed = prio_changed_fair,
4198 .switched_to = switched_to_fair,
4200 .get_rr_interval = get_rr_interval_fair,
4202 #ifdef CONFIG_FAIR_GROUP_SCHED
4203 .task_move_group = task_move_group_fair,
4207 #ifdef CONFIG_SCHED_DEBUG
4208 static void print_cfs_stats(struct seq_file *m, int cpu)
4210 struct cfs_rq *cfs_rq;
4213 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4214 print_cfs_rq(m, cpu, cfs_rq);