2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
25 #include <linux/cpumask.h>
28 * Targeted preemption latency for CPU-bound tasks:
29 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
31 * NOTE: this latency value is not the same as the concept of
32 * 'timeslice length' - timeslices in CFS are of variable length
33 * and have no persistent notion like in traditional, time-slice
34 * based scheduling concepts.
36 * (to see the precise effective timeslice length of your workload,
37 * run vmstat and monitor the context-switches (cs) field)
39 unsigned int sysctl_sched_latency = 6000000ULL;
40 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
43 * The initial- and re-scaling of tunables is configurable
44 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
47 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
48 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
49 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 enum sched_tunable_scaling sysctl_sched_tunable_scaling
52 = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
56 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
58 unsigned int sysctl_sched_min_granularity = 750000ULL;
59 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
62 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
64 static unsigned int sched_nr_latency = 8;
67 * After fork, child runs first. If set to 0 (default) then
68 * parent will (try to) run first.
70 unsigned int sysctl_sched_child_runs_first __read_mostly;
73 * SCHED_OTHER wake-up granularity.
74 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
81 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
83 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
86 * The exponential sliding window over which load is averaged for shares
90 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
92 #ifdef CONFIG_CFS_BANDWIDTH
94 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
95 * each time a cfs_rq requests quota.
97 * Note: in the case that the slice exceeds the runtime remaining (either due
98 * to consumption or the quota being specified to be smaller than the slice)
99 * we will always only issue the remaining available time.
101 * default: 5 msec, units: microseconds
103 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
106 static const struct sched_class fair_sched_class;
108 /**************************************************************
109 * CFS operations on generic schedulable entities:
112 #ifdef CONFIG_FAIR_GROUP_SCHED
114 /* cpu runqueue to which this cfs_rq is attached */
115 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
120 /* An entity is a task if it doesn't "own" a runqueue */
121 #define entity_is_task(se) (!se->my_q)
123 static inline struct task_struct *task_of(struct sched_entity *se)
125 #ifdef CONFIG_SCHED_DEBUG
126 WARN_ON_ONCE(!entity_is_task(se));
128 return container_of(se, struct task_struct, se);
131 /* Walk up scheduling entities hierarchy */
132 #define for_each_sched_entity(se) \
133 for (; se; se = se->parent)
135 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
140 /* runqueue on which this entity is (to be) queued */
141 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
146 /* runqueue "owned" by this group */
147 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
152 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
154 if (!cfs_rq->on_list) {
156 * Ensure we either appear before our parent (if already
157 * enqueued) or force our parent to appear after us when it is
158 * enqueued. The fact that we always enqueue bottom-up
159 * reduces this to two cases.
161 if (cfs_rq->tg->parent &&
162 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
163 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
164 &rq_of(cfs_rq)->leaf_cfs_rq_list);
166 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
167 &rq_of(cfs_rq)->leaf_cfs_rq_list);
174 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
176 if (cfs_rq->on_list) {
177 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
182 /* Iterate thr' all leaf cfs_rq's on a runqueue */
183 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
184 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
186 /* Do the two (enqueued) entities belong to the same group ? */
188 is_same_group(struct sched_entity *se, struct sched_entity *pse)
190 if (se->cfs_rq == pse->cfs_rq)
196 static inline struct sched_entity *parent_entity(struct sched_entity *se)
201 /* return depth at which a sched entity is present in the hierarchy */
202 static inline int depth_se(struct sched_entity *se)
206 for_each_sched_entity(se)
213 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
215 int se_depth, pse_depth;
218 * preemption test can be made between sibling entities who are in the
219 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
220 * both tasks until we find their ancestors who are siblings of common
224 /* First walk up until both entities are at same depth */
225 se_depth = depth_se(*se);
226 pse_depth = depth_se(*pse);
228 while (se_depth > pse_depth) {
230 *se = parent_entity(*se);
233 while (pse_depth > se_depth) {
235 *pse = parent_entity(*pse);
238 while (!is_same_group(*se, *pse)) {
239 *se = parent_entity(*se);
240 *pse = parent_entity(*pse);
244 #else /* !CONFIG_FAIR_GROUP_SCHED */
246 static inline struct task_struct *task_of(struct sched_entity *se)
248 return container_of(se, struct task_struct, se);
251 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
253 return container_of(cfs_rq, struct rq, cfs);
256 #define entity_is_task(se) 1
258 #define for_each_sched_entity(se) \
259 for (; se; se = NULL)
261 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
263 return &task_rq(p)->cfs;
266 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268 struct task_struct *p = task_of(se);
269 struct rq *rq = task_rq(p);
274 /* runqueue "owned" by this group */
275 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
280 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
284 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
288 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
289 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
292 is_same_group(struct sched_entity *se, struct sched_entity *pse)
297 static inline struct sched_entity *parent_entity(struct sched_entity *se)
303 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
310 unsigned long delta_exec);
312 /**************************************************************
313 * Scheduling class tree data structure manipulation methods:
316 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
318 s64 delta = (s64)(vruntime - min_vruntime);
320 min_vruntime = vruntime;
325 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
327 s64 delta = (s64)(vruntime - min_vruntime);
329 min_vruntime = vruntime;
334 static inline int entity_before(struct sched_entity *a,
335 struct sched_entity *b)
337 return (s64)(a->vruntime - b->vruntime) < 0;
340 static void update_min_vruntime(struct cfs_rq *cfs_rq)
342 u64 vruntime = cfs_rq->min_vruntime;
345 vruntime = cfs_rq->curr->vruntime;
347 if (cfs_rq->rb_leftmost) {
348 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
353 vruntime = se->vruntime;
355 vruntime = min_vruntime(vruntime, se->vruntime);
358 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
361 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
366 * Enqueue an entity into the rb-tree:
368 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
370 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
371 struct rb_node *parent = NULL;
372 struct sched_entity *entry;
376 * Find the right place in the rbtree:
380 entry = rb_entry(parent, struct sched_entity, run_node);
382 * We dont care about collisions. Nodes with
383 * the same key stay together.
385 if (entity_before(se, entry)) {
386 link = &parent->rb_left;
388 link = &parent->rb_right;
394 * Maintain a cache of leftmost tree entries (it is frequently
398 cfs_rq->rb_leftmost = &se->run_node;
400 rb_link_node(&se->run_node, parent, link);
401 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
404 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
406 if (cfs_rq->rb_leftmost == &se->run_node) {
407 struct rb_node *next_node;
409 next_node = rb_next(&se->run_node);
410 cfs_rq->rb_leftmost = next_node;
413 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
416 static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
418 struct rb_node *left = cfs_rq->rb_leftmost;
423 return rb_entry(left, struct sched_entity, run_node);
426 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
428 struct rb_node *next = rb_next(&se->run_node);
433 return rb_entry(next, struct sched_entity, run_node);
436 #ifdef CONFIG_SCHED_DEBUG
437 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
439 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
444 return rb_entry(last, struct sched_entity, run_node);
447 /**************************************************************
448 * Scheduling class statistics methods:
451 int sched_proc_update_handler(struct ctl_table *table, int write,
452 void __user *buffer, size_t *lenp,
455 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
456 int factor = get_update_sysctl_factor();
461 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
462 sysctl_sched_min_granularity);
464 #define WRT_SYSCTL(name) \
465 (normalized_sysctl_##name = sysctl_##name / (factor))
466 WRT_SYSCTL(sched_min_granularity);
467 WRT_SYSCTL(sched_latency);
468 WRT_SYSCTL(sched_wakeup_granularity);
478 static inline unsigned long
479 calc_delta_fair(unsigned long delta, struct sched_entity *se)
481 if (unlikely(se->load.weight != NICE_0_LOAD))
482 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
488 * The idea is to set a period in which each task runs once.
490 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
491 * this period because otherwise the slices get too small.
493 * p = (nr <= nl) ? l : l*nr/nl
495 static u64 __sched_period(unsigned long nr_running)
497 u64 period = sysctl_sched_latency;
498 unsigned long nr_latency = sched_nr_latency;
500 if (unlikely(nr_running > nr_latency)) {
501 period = sysctl_sched_min_granularity;
502 period *= nr_running;
509 * We calculate the wall-time slice from the period by taking a part
510 * proportional to the weight.
514 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
516 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
518 for_each_sched_entity(se) {
519 struct load_weight *load;
520 struct load_weight lw;
522 cfs_rq = cfs_rq_of(se);
523 load = &cfs_rq->load;
525 if (unlikely(!se->on_rq)) {
528 update_load_add(&lw, se->load.weight);
531 slice = calc_delta_mine(slice, se->load.weight, load);
537 * We calculate the vruntime slice of a to be inserted task
541 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 return calc_delta_fair(sched_slice(cfs_rq, se), se);
546 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
547 static void update_cfs_shares(struct cfs_rq *cfs_rq);
550 * Update the current task's runtime statistics. Skip current tasks that
551 * are not in our scheduling class.
554 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
555 unsigned long delta_exec)
557 unsigned long delta_exec_weighted;
559 schedstat_set(curr->statistics.exec_max,
560 max((u64)delta_exec, curr->statistics.exec_max));
562 curr->sum_exec_runtime += delta_exec;
563 schedstat_add(cfs_rq, exec_clock, delta_exec);
564 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
566 curr->vruntime += delta_exec_weighted;
567 update_min_vruntime(cfs_rq);
569 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
570 cfs_rq->load_unacc_exec_time += delta_exec;
574 static void update_curr(struct cfs_rq *cfs_rq)
576 struct sched_entity *curr = cfs_rq->curr;
577 u64 now = rq_of(cfs_rq)->clock_task;
578 unsigned long delta_exec;
584 * Get the amount of time the current task was running
585 * since the last time we changed load (this cannot
586 * overflow on 32 bits):
588 delta_exec = (unsigned long)(now - curr->exec_start);
592 __update_curr(cfs_rq, curr, delta_exec);
593 curr->exec_start = now;
595 if (entity_is_task(curr)) {
596 struct task_struct *curtask = task_of(curr);
598 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
599 cpuacct_charge(curtask, delta_exec);
600 account_group_exec_runtime(curtask, delta_exec);
603 account_cfs_rq_runtime(cfs_rq, delta_exec);
607 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
609 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
613 * Task is being enqueued - update stats:
615 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
618 * Are we enqueueing a waiting task? (for current tasks
619 * a dequeue/enqueue event is a NOP)
621 if (se != cfs_rq->curr)
622 update_stats_wait_start(cfs_rq, se);
626 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
628 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
629 rq_of(cfs_rq)->clock - se->statistics.wait_start));
630 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
631 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
632 rq_of(cfs_rq)->clock - se->statistics.wait_start);
633 #ifdef CONFIG_SCHEDSTATS
634 if (entity_is_task(se)) {
635 trace_sched_stat_wait(task_of(se),
636 rq_of(cfs_rq)->clock - se->statistics.wait_start);
639 schedstat_set(se->statistics.wait_start, 0);
643 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
646 * Mark the end of the wait period if dequeueing a
649 if (se != cfs_rq->curr)
650 update_stats_wait_end(cfs_rq, se);
654 * We are picking a new current task - update its stats:
657 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
660 * We are starting a new run period:
662 se->exec_start = rq_of(cfs_rq)->clock_task;
665 /**************************************************
666 * Scheduling class queueing methods:
669 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
671 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
673 cfs_rq->task_weight += weight;
677 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
683 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
685 update_load_add(&cfs_rq->load, se->load.weight);
686 if (!parent_entity(se))
687 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
688 if (entity_is_task(se)) {
689 add_cfs_task_weight(cfs_rq, se->load.weight);
690 list_add(&se->group_node, &cfs_rq->tasks);
692 cfs_rq->nr_running++;
696 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
698 update_load_sub(&cfs_rq->load, se->load.weight);
699 if (!parent_entity(se))
700 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
701 if (entity_is_task(se)) {
702 add_cfs_task_weight(cfs_rq, -se->load.weight);
703 list_del_init(&se->group_node);
705 cfs_rq->nr_running--;
708 #ifdef CONFIG_FAIR_GROUP_SCHED
709 /* we need this in update_cfs_load and load-balance functions below */
710 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
712 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
715 struct task_group *tg = cfs_rq->tg;
718 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
719 load_avg -= cfs_rq->load_contribution;
721 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
722 atomic_add(load_avg, &tg->load_weight);
723 cfs_rq->load_contribution += load_avg;
727 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
729 u64 period = sysctl_sched_shares_window;
731 unsigned long load = cfs_rq->load.weight;
733 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
736 now = rq_of(cfs_rq)->clock_task;
737 delta = now - cfs_rq->load_stamp;
739 /* truncate load history at 4 idle periods */
740 if (cfs_rq->load_stamp > cfs_rq->load_last &&
741 now - cfs_rq->load_last > 4 * period) {
742 cfs_rq->load_period = 0;
743 cfs_rq->load_avg = 0;
747 cfs_rq->load_stamp = now;
748 cfs_rq->load_unacc_exec_time = 0;
749 cfs_rq->load_period += delta;
751 cfs_rq->load_last = now;
752 cfs_rq->load_avg += delta * load;
755 /* consider updating load contribution on each fold or truncate */
756 if (global_update || cfs_rq->load_period > period
757 || !cfs_rq->load_period)
758 update_cfs_rq_load_contribution(cfs_rq, global_update);
760 while (cfs_rq->load_period > period) {
762 * Inline assembly required to prevent the compiler
763 * optimising this loop into a divmod call.
764 * See __iter_div_u64_rem() for another example of this.
766 asm("" : "+rm" (cfs_rq->load_period));
767 cfs_rq->load_period /= 2;
768 cfs_rq->load_avg /= 2;
771 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
772 list_del_leaf_cfs_rq(cfs_rq);
775 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
777 long load_weight, load, shares;
779 load = cfs_rq->load.weight;
781 load_weight = atomic_read(&tg->load_weight);
783 load_weight -= cfs_rq->load_contribution;
785 shares = (tg->shares * load);
787 shares /= load_weight;
789 if (shares < MIN_SHARES)
791 if (shares > tg->shares)
797 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
799 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
800 update_cfs_load(cfs_rq, 0);
801 update_cfs_shares(cfs_rq);
804 # else /* CONFIG_SMP */
805 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
809 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
814 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
817 # endif /* CONFIG_SMP */
818 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
819 unsigned long weight)
822 /* commit outstanding execution time */
823 if (cfs_rq->curr == se)
825 account_entity_dequeue(cfs_rq, se);
828 update_load_set(&se->load, weight);
831 account_entity_enqueue(cfs_rq, se);
834 static void update_cfs_shares(struct cfs_rq *cfs_rq)
836 struct task_group *tg;
837 struct sched_entity *se;
841 se = tg->se[cpu_of(rq_of(cfs_rq))];
842 if (!se || throttled_hierarchy(cfs_rq))
845 if (likely(se->load.weight == tg->shares))
848 shares = calc_cfs_shares(cfs_rq, tg);
850 reweight_entity(cfs_rq_of(se), se, shares);
852 #else /* CONFIG_FAIR_GROUP_SCHED */
853 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
857 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
861 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
864 #endif /* CONFIG_FAIR_GROUP_SCHED */
866 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
868 #ifdef CONFIG_SCHEDSTATS
869 struct task_struct *tsk = NULL;
871 if (entity_is_task(se))
874 if (se->statistics.sleep_start) {
875 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
880 if (unlikely(delta > se->statistics.sleep_max))
881 se->statistics.sleep_max = delta;
883 se->statistics.sleep_start = 0;
884 se->statistics.sum_sleep_runtime += delta;
887 account_scheduler_latency(tsk, delta >> 10, 1);
888 trace_sched_stat_sleep(tsk, delta);
891 if (se->statistics.block_start) {
892 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
897 if (unlikely(delta > se->statistics.block_max))
898 se->statistics.block_max = delta;
900 se->statistics.block_start = 0;
901 se->statistics.sum_sleep_runtime += delta;
904 if (tsk->in_iowait) {
905 se->statistics.iowait_sum += delta;
906 se->statistics.iowait_count++;
907 trace_sched_stat_iowait(tsk, delta);
911 * Blocking time is in units of nanosecs, so shift by
912 * 20 to get a milliseconds-range estimation of the
913 * amount of time that the task spent sleeping:
915 if (unlikely(prof_on == SLEEP_PROFILING)) {
916 profile_hits(SLEEP_PROFILING,
917 (void *)get_wchan(tsk),
920 account_scheduler_latency(tsk, delta >> 10, 0);
926 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
928 #ifdef CONFIG_SCHED_DEBUG
929 s64 d = se->vruntime - cfs_rq->min_vruntime;
934 if (d > 3*sysctl_sched_latency)
935 schedstat_inc(cfs_rq, nr_spread_over);
940 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
942 u64 vruntime = cfs_rq->min_vruntime;
945 * The 'current' period is already promised to the current tasks,
946 * however the extra weight of the new task will slow them down a
947 * little, place the new task so that it fits in the slot that
948 * stays open at the end.
950 if (initial && sched_feat(START_DEBIT))
951 vruntime += sched_vslice(cfs_rq, se);
953 /* sleeps up to a single latency don't count. */
955 unsigned long thresh = sysctl_sched_latency;
958 * Halve their sleep time's effect, to allow
959 * for a gentler effect of sleepers:
961 if (sched_feat(GENTLE_FAIR_SLEEPERS))
967 /* ensure we never gain time by being placed backwards. */
968 vruntime = max_vruntime(se->vruntime, vruntime);
970 se->vruntime = vruntime;
973 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
976 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
979 * Update the normalized vruntime before updating min_vruntime
980 * through callig update_curr().
982 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
983 se->vruntime += cfs_rq->min_vruntime;
986 * Update run-time statistics of the 'current'.
989 update_cfs_load(cfs_rq, 0);
990 account_entity_enqueue(cfs_rq, se);
991 update_cfs_shares(cfs_rq);
993 if (flags & ENQUEUE_WAKEUP) {
994 place_entity(cfs_rq, se, 0);
995 enqueue_sleeper(cfs_rq, se);
998 update_stats_enqueue(cfs_rq, se);
999 check_spread(cfs_rq, se);
1000 if (se != cfs_rq->curr)
1001 __enqueue_entity(cfs_rq, se);
1004 if (cfs_rq->nr_running == 1) {
1005 list_add_leaf_cfs_rq(cfs_rq);
1006 check_enqueue_throttle(cfs_rq);
1010 static void __clear_buddies_last(struct sched_entity *se)
1012 for_each_sched_entity(se) {
1013 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1014 if (cfs_rq->last == se)
1015 cfs_rq->last = NULL;
1021 static void __clear_buddies_next(struct sched_entity *se)
1023 for_each_sched_entity(se) {
1024 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1025 if (cfs_rq->next == se)
1026 cfs_rq->next = NULL;
1032 static void __clear_buddies_skip(struct sched_entity *se)
1034 for_each_sched_entity(se) {
1035 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1036 if (cfs_rq->skip == se)
1037 cfs_rq->skip = NULL;
1043 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1045 if (cfs_rq->last == se)
1046 __clear_buddies_last(se);
1048 if (cfs_rq->next == se)
1049 __clear_buddies_next(se);
1051 if (cfs_rq->skip == se)
1052 __clear_buddies_skip(se);
1055 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1058 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1061 * Update run-time statistics of the 'current'.
1063 update_curr(cfs_rq);
1065 update_stats_dequeue(cfs_rq, se);
1066 if (flags & DEQUEUE_SLEEP) {
1067 #ifdef CONFIG_SCHEDSTATS
1068 if (entity_is_task(se)) {
1069 struct task_struct *tsk = task_of(se);
1071 if (tsk->state & TASK_INTERRUPTIBLE)
1072 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1073 if (tsk->state & TASK_UNINTERRUPTIBLE)
1074 se->statistics.block_start = rq_of(cfs_rq)->clock;
1079 clear_buddies(cfs_rq, se);
1081 if (se != cfs_rq->curr)
1082 __dequeue_entity(cfs_rq, se);
1084 update_cfs_load(cfs_rq, 0);
1085 account_entity_dequeue(cfs_rq, se);
1088 * Normalize the entity after updating the min_vruntime because the
1089 * update can refer to the ->curr item and we need to reflect this
1090 * movement in our normalized position.
1092 if (!(flags & DEQUEUE_SLEEP))
1093 se->vruntime -= cfs_rq->min_vruntime;
1095 /* return excess runtime on last dequeue */
1096 return_cfs_rq_runtime(cfs_rq);
1098 update_min_vruntime(cfs_rq);
1099 update_cfs_shares(cfs_rq);
1103 * Preempt the current task with a newly woken task if needed:
1106 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1108 unsigned long ideal_runtime, delta_exec;
1109 struct sched_entity *se;
1112 ideal_runtime = sched_slice(cfs_rq, curr);
1113 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1114 if (delta_exec > ideal_runtime) {
1115 resched_task(rq_of(cfs_rq)->curr);
1117 * The current task ran long enough, ensure it doesn't get
1118 * re-elected due to buddy favours.
1120 clear_buddies(cfs_rq, curr);
1125 * Ensure that a task that missed wakeup preemption by a
1126 * narrow margin doesn't have to wait for a full slice.
1127 * This also mitigates buddy induced latencies under load.
1129 if (delta_exec < sysctl_sched_min_granularity)
1132 se = __pick_first_entity(cfs_rq);
1133 delta = curr->vruntime - se->vruntime;
1138 if (delta > ideal_runtime)
1139 resched_task(rq_of(cfs_rq)->curr);
1143 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1145 /* 'current' is not kept within the tree. */
1148 * Any task has to be enqueued before it get to execute on
1149 * a CPU. So account for the time it spent waiting on the
1152 update_stats_wait_end(cfs_rq, se);
1153 __dequeue_entity(cfs_rq, se);
1156 update_stats_curr_start(cfs_rq, se);
1158 #ifdef CONFIG_SCHEDSTATS
1160 * Track our maximum slice length, if the CPU's load is at
1161 * least twice that of our own weight (i.e. dont track it
1162 * when there are only lesser-weight tasks around):
1164 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1165 se->statistics.slice_max = max(se->statistics.slice_max,
1166 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1169 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1173 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1176 * Pick the next process, keeping these things in mind, in this order:
1177 * 1) keep things fair between processes/task groups
1178 * 2) pick the "next" process, since someone really wants that to run
1179 * 3) pick the "last" process, for cache locality
1180 * 4) do not run the "skip" process, if something else is available
1182 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1184 struct sched_entity *se = __pick_first_entity(cfs_rq);
1185 struct sched_entity *left = se;
1188 * Avoid running the skip buddy, if running something else can
1189 * be done without getting too unfair.
1191 if (cfs_rq->skip == se) {
1192 struct sched_entity *second = __pick_next_entity(se);
1193 if (second && wakeup_preempt_entity(second, left) < 1)
1198 * Prefer last buddy, try to return the CPU to a preempted task.
1200 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1204 * Someone really wants this to run. If it's not unfair, run it.
1206 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1209 clear_buddies(cfs_rq, se);
1214 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1216 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1219 * If still on the runqueue then deactivate_task()
1220 * was not called and update_curr() has to be done:
1223 update_curr(cfs_rq);
1225 /* throttle cfs_rqs exceeding runtime */
1226 check_cfs_rq_runtime(cfs_rq);
1228 check_spread(cfs_rq, prev);
1230 update_stats_wait_start(cfs_rq, prev);
1231 /* Put 'current' back into the tree. */
1232 __enqueue_entity(cfs_rq, prev);
1234 cfs_rq->curr = NULL;
1238 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1241 * Update run-time statistics of the 'current'.
1243 update_curr(cfs_rq);
1246 * Update share accounting for long-running entities.
1248 update_entity_shares_tick(cfs_rq);
1250 #ifdef CONFIG_SCHED_HRTICK
1252 * queued ticks are scheduled to match the slice, so don't bother
1253 * validating it and just reschedule.
1256 resched_task(rq_of(cfs_rq)->curr);
1260 * don't let the period tick interfere with the hrtick preemption
1262 if (!sched_feat(DOUBLE_TICK) &&
1263 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1267 if (cfs_rq->nr_running > 1)
1268 check_preempt_tick(cfs_rq, curr);
1272 /**************************************************
1273 * CFS bandwidth control machinery
1276 #ifdef CONFIG_CFS_BANDWIDTH
1278 * default period for cfs group bandwidth.
1279 * default: 0.1s, units: nanoseconds
1281 static inline u64 default_cfs_period(void)
1283 return 100000000ULL;
1286 static inline u64 sched_cfs_bandwidth_slice(void)
1288 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1292 * Replenish runtime according to assigned quota and update expiration time.
1293 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1294 * additional synchronization around rq->lock.
1296 * requires cfs_b->lock
1298 static void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1302 if (cfs_b->quota == RUNTIME_INF)
1305 now = sched_clock_cpu(smp_processor_id());
1306 cfs_b->runtime = cfs_b->quota;
1307 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1310 /* returns 0 on failure to allocate runtime */
1311 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1313 struct task_group *tg = cfs_rq->tg;
1314 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1315 u64 amount = 0, min_amount, expires;
1317 /* note: this is a positive sum as runtime_remaining <= 0 */
1318 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1320 raw_spin_lock(&cfs_b->lock);
1321 if (cfs_b->quota == RUNTIME_INF)
1322 amount = min_amount;
1325 * If the bandwidth pool has become inactive, then at least one
1326 * period must have elapsed since the last consumption.
1327 * Refresh the global state and ensure bandwidth timer becomes
1330 if (!cfs_b->timer_active) {
1331 __refill_cfs_bandwidth_runtime(cfs_b);
1332 __start_cfs_bandwidth(cfs_b);
1335 if (cfs_b->runtime > 0) {
1336 amount = min(cfs_b->runtime, min_amount);
1337 cfs_b->runtime -= amount;
1341 expires = cfs_b->runtime_expires;
1342 raw_spin_unlock(&cfs_b->lock);
1344 cfs_rq->runtime_remaining += amount;
1346 * we may have advanced our local expiration to account for allowed
1347 * spread between our sched_clock and the one on which runtime was
1350 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1351 cfs_rq->runtime_expires = expires;
1353 return cfs_rq->runtime_remaining > 0;
1357 * Note: This depends on the synchronization provided by sched_clock and the
1358 * fact that rq->clock snapshots this value.
1360 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1362 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1363 struct rq *rq = rq_of(cfs_rq);
1365 /* if the deadline is ahead of our clock, nothing to do */
1366 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1369 if (cfs_rq->runtime_remaining < 0)
1373 * If the local deadline has passed we have to consider the
1374 * possibility that our sched_clock is 'fast' and the global deadline
1375 * has not truly expired.
1377 * Fortunately we can check determine whether this the case by checking
1378 * whether the global deadline has advanced.
1381 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1382 /* extend local deadline, drift is bounded above by 2 ticks */
1383 cfs_rq->runtime_expires += TICK_NSEC;
1385 /* global deadline is ahead, expiration has passed */
1386 cfs_rq->runtime_remaining = 0;
1390 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1391 unsigned long delta_exec)
1393 /* dock delta_exec before expiring quota (as it could span periods) */
1394 cfs_rq->runtime_remaining -= delta_exec;
1395 expire_cfs_rq_runtime(cfs_rq);
1397 if (likely(cfs_rq->runtime_remaining > 0))
1401 * if we're unable to extend our runtime we resched so that the active
1402 * hierarchy can be throttled
1404 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1405 resched_task(rq_of(cfs_rq)->curr);
1408 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1409 unsigned long delta_exec)
1411 if (!cfs_rq->runtime_enabled)
1414 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1417 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1419 return cfs_rq->throttled;
1422 /* check whether cfs_rq, or any parent, is throttled */
1423 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1425 return cfs_rq->throttle_count;
1429 * Ensure that neither of the group entities corresponding to src_cpu or
1430 * dest_cpu are members of a throttled hierarchy when performing group
1431 * load-balance operations.
1433 static inline int throttled_lb_pair(struct task_group *tg,
1434 int src_cpu, int dest_cpu)
1436 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1438 src_cfs_rq = tg->cfs_rq[src_cpu];
1439 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1441 return throttled_hierarchy(src_cfs_rq) ||
1442 throttled_hierarchy(dest_cfs_rq);
1445 /* updated child weight may affect parent so we have to do this bottom up */
1446 static int tg_unthrottle_up(struct task_group *tg, void *data)
1448 struct rq *rq = data;
1449 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1451 cfs_rq->throttle_count--;
1453 if (!cfs_rq->throttle_count) {
1454 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1456 /* leaving throttled state, advance shares averaging windows */
1457 cfs_rq->load_stamp += delta;
1458 cfs_rq->load_last += delta;
1460 /* update entity weight now that we are on_rq again */
1461 update_cfs_shares(cfs_rq);
1468 static int tg_throttle_down(struct task_group *tg, void *data)
1470 struct rq *rq = data;
1471 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1473 /* group is entering throttled state, record last load */
1474 if (!cfs_rq->throttle_count)
1475 update_cfs_load(cfs_rq, 0);
1476 cfs_rq->throttle_count++;
1481 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1483 struct rq *rq = rq_of(cfs_rq);
1484 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1485 struct sched_entity *se;
1486 long task_delta, dequeue = 1;
1488 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1490 /* account load preceding throttle */
1492 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1495 task_delta = cfs_rq->h_nr_running;
1496 for_each_sched_entity(se) {
1497 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1498 /* throttled entity or throttle-on-deactivate */
1503 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1504 qcfs_rq->h_nr_running -= task_delta;
1506 if (qcfs_rq->load.weight)
1511 rq->nr_running -= task_delta;
1513 cfs_rq->throttled = 1;
1514 cfs_rq->throttled_timestamp = rq->clock;
1515 raw_spin_lock(&cfs_b->lock);
1516 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1517 raw_spin_unlock(&cfs_b->lock);
1520 static void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1522 struct rq *rq = rq_of(cfs_rq);
1523 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1524 struct sched_entity *se;
1528 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1530 cfs_rq->throttled = 0;
1531 raw_spin_lock(&cfs_b->lock);
1532 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1533 list_del_rcu(&cfs_rq->throttled_list);
1534 raw_spin_unlock(&cfs_b->lock);
1535 cfs_rq->throttled_timestamp = 0;
1537 update_rq_clock(rq);
1538 /* update hierarchical throttle state */
1539 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1541 if (!cfs_rq->load.weight)
1544 task_delta = cfs_rq->h_nr_running;
1545 for_each_sched_entity(se) {
1549 cfs_rq = cfs_rq_of(se);
1551 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1552 cfs_rq->h_nr_running += task_delta;
1554 if (cfs_rq_throttled(cfs_rq))
1559 rq->nr_running += task_delta;
1561 /* determine whether we need to wake up potentially idle cpu */
1562 if (rq->curr == rq->idle && rq->cfs.nr_running)
1563 resched_task(rq->curr);
1566 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1567 u64 remaining, u64 expires)
1569 struct cfs_rq *cfs_rq;
1570 u64 runtime = remaining;
1573 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1575 struct rq *rq = rq_of(cfs_rq);
1577 raw_spin_lock(&rq->lock);
1578 if (!cfs_rq_throttled(cfs_rq))
1581 runtime = -cfs_rq->runtime_remaining + 1;
1582 if (runtime > remaining)
1583 runtime = remaining;
1584 remaining -= runtime;
1586 cfs_rq->runtime_remaining += runtime;
1587 cfs_rq->runtime_expires = expires;
1589 /* we check whether we're throttled above */
1590 if (cfs_rq->runtime_remaining > 0)
1591 unthrottle_cfs_rq(cfs_rq);
1594 raw_spin_unlock(&rq->lock);
1605 * Responsible for refilling a task_group's bandwidth and unthrottling its
1606 * cfs_rqs as appropriate. If there has been no activity within the last
1607 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1608 * used to track this state.
1610 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1612 u64 runtime, runtime_expires;
1613 int idle = 1, throttled;
1615 raw_spin_lock(&cfs_b->lock);
1616 /* no need to continue the timer with no bandwidth constraint */
1617 if (cfs_b->quota == RUNTIME_INF)
1620 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1621 /* idle depends on !throttled (for the case of a large deficit) */
1622 idle = cfs_b->idle && !throttled;
1623 cfs_b->nr_periods += overrun;
1625 /* if we're going inactive then everything else can be deferred */
1629 __refill_cfs_bandwidth_runtime(cfs_b);
1632 /* mark as potentially idle for the upcoming period */
1637 /* account preceding periods in which throttling occurred */
1638 cfs_b->nr_throttled += overrun;
1641 * There are throttled entities so we must first use the new bandwidth
1642 * to unthrottle them before making it generally available. This
1643 * ensures that all existing debts will be paid before a new cfs_rq is
1646 runtime = cfs_b->runtime;
1647 runtime_expires = cfs_b->runtime_expires;
1651 * This check is repeated as we are holding onto the new bandwidth
1652 * while we unthrottle. This can potentially race with an unthrottled
1653 * group trying to acquire new bandwidth from the global pool.
1655 while (throttled && runtime > 0) {
1656 raw_spin_unlock(&cfs_b->lock);
1657 /* we can't nest cfs_b->lock while distributing bandwidth */
1658 runtime = distribute_cfs_runtime(cfs_b, runtime,
1660 raw_spin_lock(&cfs_b->lock);
1662 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1665 /* return (any) remaining runtime */
1666 cfs_b->runtime = runtime;
1668 * While we are ensured activity in the period following an
1669 * unthrottle, this also covers the case in which the new bandwidth is
1670 * insufficient to cover the existing bandwidth deficit. (Forcing the
1671 * timer to remain active while there are any throttled entities.)
1676 cfs_b->timer_active = 0;
1677 raw_spin_unlock(&cfs_b->lock);
1682 /* a cfs_rq won't donate quota below this amount */
1683 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1684 /* minimum remaining period time to redistribute slack quota */
1685 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1686 /* how long we wait to gather additional slack before distributing */
1687 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1689 /* are we near the end of the current quota period? */
1690 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
1692 struct hrtimer *refresh_timer = &cfs_b->period_timer;
1695 /* if the call-back is running a quota refresh is already occurring */
1696 if (hrtimer_callback_running(refresh_timer))
1699 /* is a quota refresh about to occur? */
1700 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
1701 if (remaining < min_expire)
1707 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
1709 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
1711 /* if there's a quota refresh soon don't bother with slack */
1712 if (runtime_refresh_within(cfs_b, min_left))
1715 start_bandwidth_timer(&cfs_b->slack_timer,
1716 ns_to_ktime(cfs_bandwidth_slack_period));
1719 /* we know any runtime found here is valid as update_curr() precedes return */
1720 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1722 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1723 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
1725 if (slack_runtime <= 0)
1728 raw_spin_lock(&cfs_b->lock);
1729 if (cfs_b->quota != RUNTIME_INF &&
1730 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
1731 cfs_b->runtime += slack_runtime;
1733 /* we are under rq->lock, defer unthrottling using a timer */
1734 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
1735 !list_empty(&cfs_b->throttled_cfs_rq))
1736 start_cfs_slack_bandwidth(cfs_b);
1738 raw_spin_unlock(&cfs_b->lock);
1740 /* even if it's not valid for return we don't want to try again */
1741 cfs_rq->runtime_remaining -= slack_runtime;
1744 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1746 if (!cfs_rq->runtime_enabled || !cfs_rq->nr_running)
1749 __return_cfs_rq_runtime(cfs_rq);
1753 * This is done with a timer (instead of inline with bandwidth return) since
1754 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
1756 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
1758 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
1761 /* confirm we're still not at a refresh boundary */
1762 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
1765 raw_spin_lock(&cfs_b->lock);
1766 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
1767 runtime = cfs_b->runtime;
1770 expires = cfs_b->runtime_expires;
1771 raw_spin_unlock(&cfs_b->lock);
1776 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
1778 raw_spin_lock(&cfs_b->lock);
1779 if (expires == cfs_b->runtime_expires)
1780 cfs_b->runtime = runtime;
1781 raw_spin_unlock(&cfs_b->lock);
1785 * When a group wakes up we want to make sure that its quota is not already
1786 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1787 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1789 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
1791 /* an active group must be handled by the update_curr()->put() path */
1792 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
1795 /* ensure the group is not already throttled */
1796 if (cfs_rq_throttled(cfs_rq))
1799 /* update runtime allocation */
1800 account_cfs_rq_runtime(cfs_rq, 0);
1801 if (cfs_rq->runtime_remaining <= 0)
1802 throttle_cfs_rq(cfs_rq);
1805 /* conditionally throttle active cfs_rq's from put_prev_entity() */
1806 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1808 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
1812 * it's possible for a throttled entity to be forced into a running
1813 * state (e.g. set_curr_task), in this case we're finished.
1815 if (cfs_rq_throttled(cfs_rq))
1818 throttle_cfs_rq(cfs_rq);
1821 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1822 unsigned long delta_exec) {}
1823 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
1824 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
1825 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
1827 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1832 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1837 static inline int throttled_lb_pair(struct task_group *tg,
1838 int src_cpu, int dest_cpu)
1844 /**************************************************
1845 * CFS operations on tasks:
1848 #ifdef CONFIG_SCHED_HRTICK
1849 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1851 struct sched_entity *se = &p->se;
1852 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1854 WARN_ON(task_rq(p) != rq);
1856 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1857 u64 slice = sched_slice(cfs_rq, se);
1858 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1859 s64 delta = slice - ran;
1868 * Don't schedule slices shorter than 10000ns, that just
1869 * doesn't make sense. Rely on vruntime for fairness.
1872 delta = max_t(s64, 10000LL, delta);
1874 hrtick_start(rq, delta);
1879 * called from enqueue/dequeue and updates the hrtick when the
1880 * current task is from our class and nr_running is low enough
1883 static void hrtick_update(struct rq *rq)
1885 struct task_struct *curr = rq->curr;
1887 if (curr->sched_class != &fair_sched_class)
1890 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1891 hrtick_start_fair(rq, curr);
1893 #else /* !CONFIG_SCHED_HRTICK */
1895 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1899 static inline void hrtick_update(struct rq *rq)
1905 * The enqueue_task method is called before nr_running is
1906 * increased. Here we update the fair scheduling stats and
1907 * then put the task into the rbtree:
1910 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1912 struct cfs_rq *cfs_rq;
1913 struct sched_entity *se = &p->se;
1915 for_each_sched_entity(se) {
1918 cfs_rq = cfs_rq_of(se);
1919 enqueue_entity(cfs_rq, se, flags);
1922 * end evaluation on encountering a throttled cfs_rq
1924 * note: in the case of encountering a throttled cfs_rq we will
1925 * post the final h_nr_running increment below.
1927 if (cfs_rq_throttled(cfs_rq))
1929 cfs_rq->h_nr_running++;
1931 flags = ENQUEUE_WAKEUP;
1934 for_each_sched_entity(se) {
1935 cfs_rq = cfs_rq_of(se);
1936 cfs_rq->h_nr_running++;
1938 if (cfs_rq_throttled(cfs_rq))
1941 update_cfs_load(cfs_rq, 0);
1942 update_cfs_shares(cfs_rq);
1950 static void set_next_buddy(struct sched_entity *se);
1953 * The dequeue_task method is called before nr_running is
1954 * decreased. We remove the task from the rbtree and
1955 * update the fair scheduling stats:
1957 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1959 struct cfs_rq *cfs_rq;
1960 struct sched_entity *se = &p->se;
1961 int task_sleep = flags & DEQUEUE_SLEEP;
1963 for_each_sched_entity(se) {
1964 cfs_rq = cfs_rq_of(se);
1965 dequeue_entity(cfs_rq, se, flags);
1968 * end evaluation on encountering a throttled cfs_rq
1970 * note: in the case of encountering a throttled cfs_rq we will
1971 * post the final h_nr_running decrement below.
1973 if (cfs_rq_throttled(cfs_rq))
1975 cfs_rq->h_nr_running--;
1977 /* Don't dequeue parent if it has other entities besides us */
1978 if (cfs_rq->load.weight) {
1980 * Bias pick_next to pick a task from this cfs_rq, as
1981 * p is sleeping when it is within its sched_slice.
1983 if (task_sleep && parent_entity(se))
1984 set_next_buddy(parent_entity(se));
1986 /* avoid re-evaluating load for this entity */
1987 se = parent_entity(se);
1990 flags |= DEQUEUE_SLEEP;
1993 for_each_sched_entity(se) {
1994 cfs_rq = cfs_rq_of(se);
1995 cfs_rq->h_nr_running--;
1997 if (cfs_rq_throttled(cfs_rq))
2000 update_cfs_load(cfs_rq, 0);
2001 update_cfs_shares(cfs_rq);
2011 static void task_waking_fair(struct task_struct *p)
2013 struct sched_entity *se = &p->se;
2014 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2017 #ifndef CONFIG_64BIT
2018 u64 min_vruntime_copy;
2021 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2023 min_vruntime = cfs_rq->min_vruntime;
2024 } while (min_vruntime != min_vruntime_copy);
2026 min_vruntime = cfs_rq->min_vruntime;
2029 se->vruntime -= min_vruntime;
2032 #ifdef CONFIG_FAIR_GROUP_SCHED
2034 * effective_load() calculates the load change as seen from the root_task_group
2036 * Adding load to a group doesn't make a group heavier, but can cause movement
2037 * of group shares between cpus. Assuming the shares were perfectly aligned one
2038 * can calculate the shift in shares.
2040 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2042 struct sched_entity *se = tg->se[cpu];
2047 for_each_sched_entity(se) {
2051 w = se->my_q->load.weight;
2053 /* use this cpu's instantaneous contribution */
2054 lw = atomic_read(&tg->load_weight);
2055 lw -= se->my_q->load_contribution;
2060 if (lw > 0 && wl < lw)
2061 wl = (wl * tg->shares) / lw;
2065 /* zero point is MIN_SHARES */
2066 if (wl < MIN_SHARES)
2068 wl -= se->load.weight;
2076 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2077 unsigned long wl, unsigned long wg)
2084 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2086 s64 this_load, load;
2087 int idx, this_cpu, prev_cpu;
2088 unsigned long tl_per_task;
2089 struct task_group *tg;
2090 unsigned long weight;
2094 this_cpu = smp_processor_id();
2095 prev_cpu = task_cpu(p);
2096 load = source_load(prev_cpu, idx);
2097 this_load = target_load(this_cpu, idx);
2100 * If sync wakeup then subtract the (maximum possible)
2101 * effect of the currently running task from the load
2102 * of the current CPU:
2105 tg = task_group(current);
2106 weight = current->se.load.weight;
2108 this_load += effective_load(tg, this_cpu, -weight, -weight);
2109 load += effective_load(tg, prev_cpu, 0, -weight);
2113 weight = p->se.load.weight;
2116 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2117 * due to the sync cause above having dropped this_load to 0, we'll
2118 * always have an imbalance, but there's really nothing you can do
2119 * about that, so that's good too.
2121 * Otherwise check if either cpus are near enough in load to allow this
2122 * task to be woken on this_cpu.
2124 if (this_load > 0) {
2125 s64 this_eff_load, prev_eff_load;
2127 this_eff_load = 100;
2128 this_eff_load *= power_of(prev_cpu);
2129 this_eff_load *= this_load +
2130 effective_load(tg, this_cpu, weight, weight);
2132 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2133 prev_eff_load *= power_of(this_cpu);
2134 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2136 balanced = this_eff_load <= prev_eff_load;
2141 * If the currently running task will sleep within
2142 * a reasonable amount of time then attract this newly
2145 if (sync && balanced)
2148 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2149 tl_per_task = cpu_avg_load_per_task(this_cpu);
2152 (this_load <= load &&
2153 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2155 * This domain has SD_WAKE_AFFINE and
2156 * p is cache cold in this domain, and
2157 * there is no bad imbalance.
2159 schedstat_inc(sd, ttwu_move_affine);
2160 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2168 * find_idlest_group finds and returns the least busy CPU group within the
2171 static struct sched_group *
2172 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2173 int this_cpu, int load_idx)
2175 struct sched_group *idlest = NULL, *group = sd->groups;
2176 unsigned long min_load = ULONG_MAX, this_load = 0;
2177 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2180 unsigned long load, avg_load;
2184 /* Skip over this group if it has no CPUs allowed */
2185 if (!cpumask_intersects(sched_group_cpus(group),
2186 tsk_cpus_allowed(p)))
2189 local_group = cpumask_test_cpu(this_cpu,
2190 sched_group_cpus(group));
2192 /* Tally up the load of all CPUs in the group */
2195 for_each_cpu(i, sched_group_cpus(group)) {
2196 /* Bias balancing toward cpus of our domain */
2198 load = source_load(i, load_idx);
2200 load = target_load(i, load_idx);
2205 /* Adjust by relative CPU power of the group */
2206 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2209 this_load = avg_load;
2210 } else if (avg_load < min_load) {
2211 min_load = avg_load;
2214 } while (group = group->next, group != sd->groups);
2216 if (!idlest || 100*this_load < imbalance*min_load)
2222 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2225 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2227 unsigned long load, min_load = ULONG_MAX;
2231 /* Traverse only the allowed CPUs */
2232 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2233 load = weighted_cpuload(i);
2235 if (load < min_load || (load == min_load && i == this_cpu)) {
2245 * Try and locate an idle CPU in the sched_domain.
2247 static int select_idle_sibling(struct task_struct *p, int target)
2249 int cpu = smp_processor_id();
2250 int prev_cpu = task_cpu(p);
2251 struct sched_domain *sd;
2255 * If the task is going to be woken-up on this cpu and if it is
2256 * already idle, then it is the right target.
2258 if (target == cpu && idle_cpu(cpu))
2262 * If the task is going to be woken-up on the cpu where it previously
2263 * ran and if it is currently idle, then it the right target.
2265 if (target == prev_cpu && idle_cpu(prev_cpu))
2269 * Otherwise, iterate the domains and find an elegible idle cpu.
2272 for_each_domain(target, sd) {
2273 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
2276 for_each_cpu_and(i, sched_domain_span(sd), tsk_cpus_allowed(p)) {
2284 * Lets stop looking for an idle sibling when we reached
2285 * the domain that spans the current cpu and prev_cpu.
2287 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
2288 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
2297 * sched_balance_self: balance the current task (running on cpu) in domains
2298 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2301 * Balance, ie. select the least loaded group.
2303 * Returns the target CPU number, or the same CPU if no balancing is needed.
2305 * preempt must be disabled.
2308 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2310 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2311 int cpu = smp_processor_id();
2312 int prev_cpu = task_cpu(p);
2314 int want_affine = 0;
2316 int sync = wake_flags & WF_SYNC;
2318 if (sd_flag & SD_BALANCE_WAKE) {
2319 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2325 for_each_domain(cpu, tmp) {
2326 if (!(tmp->flags & SD_LOAD_BALANCE))
2330 * If power savings logic is enabled for a domain, see if we
2331 * are not overloaded, if so, don't balance wider.
2333 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2334 unsigned long power = 0;
2335 unsigned long nr_running = 0;
2336 unsigned long capacity;
2339 for_each_cpu(i, sched_domain_span(tmp)) {
2340 power += power_of(i);
2341 nr_running += cpu_rq(i)->cfs.nr_running;
2344 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2346 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2349 if (nr_running < capacity)
2354 * If both cpu and prev_cpu are part of this domain,
2355 * cpu is a valid SD_WAKE_AFFINE target.
2357 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2358 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2363 if (!want_sd && !want_affine)
2366 if (!(tmp->flags & sd_flag))
2374 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2377 new_cpu = select_idle_sibling(p, prev_cpu);
2382 int load_idx = sd->forkexec_idx;
2383 struct sched_group *group;
2386 if (!(sd->flags & sd_flag)) {
2391 if (sd_flag & SD_BALANCE_WAKE)
2392 load_idx = sd->wake_idx;
2394 group = find_idlest_group(sd, p, cpu, load_idx);
2400 new_cpu = find_idlest_cpu(group, p, cpu);
2401 if (new_cpu == -1 || new_cpu == cpu) {
2402 /* Now try balancing at a lower domain level of cpu */
2407 /* Now try balancing at a lower domain level of new_cpu */
2409 weight = sd->span_weight;
2411 for_each_domain(cpu, tmp) {
2412 if (weight <= tmp->span_weight)
2414 if (tmp->flags & sd_flag)
2417 /* while loop will break here if sd == NULL */
2424 #endif /* CONFIG_SMP */
2426 static unsigned long
2427 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2429 unsigned long gran = sysctl_sched_wakeup_granularity;
2432 * Since its curr running now, convert the gran from real-time
2433 * to virtual-time in his units.
2435 * By using 'se' instead of 'curr' we penalize light tasks, so
2436 * they get preempted easier. That is, if 'se' < 'curr' then
2437 * the resulting gran will be larger, therefore penalizing the
2438 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2439 * be smaller, again penalizing the lighter task.
2441 * This is especially important for buddies when the leftmost
2442 * task is higher priority than the buddy.
2444 return calc_delta_fair(gran, se);
2448 * Should 'se' preempt 'curr'.
2462 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2464 s64 gran, vdiff = curr->vruntime - se->vruntime;
2469 gran = wakeup_gran(curr, se);
2476 static void set_last_buddy(struct sched_entity *se)
2478 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2481 for_each_sched_entity(se)
2482 cfs_rq_of(se)->last = se;
2485 static void set_next_buddy(struct sched_entity *se)
2487 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2490 for_each_sched_entity(se)
2491 cfs_rq_of(se)->next = se;
2494 static void set_skip_buddy(struct sched_entity *se)
2496 for_each_sched_entity(se)
2497 cfs_rq_of(se)->skip = se;
2501 * Preempt the current task with a newly woken task if needed:
2503 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2505 struct task_struct *curr = rq->curr;
2506 struct sched_entity *se = &curr->se, *pse = &p->se;
2507 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2508 int scale = cfs_rq->nr_running >= sched_nr_latency;
2509 int next_buddy_marked = 0;
2511 if (unlikely(se == pse))
2515 * This is possible from callers such as pull_task(), in which we
2516 * unconditionally check_prempt_curr() after an enqueue (which may have
2517 * lead to a throttle). This both saves work and prevents false
2518 * next-buddy nomination below.
2520 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
2523 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2524 set_next_buddy(pse);
2525 next_buddy_marked = 1;
2529 * We can come here with TIF_NEED_RESCHED already set from new task
2532 * Note: this also catches the edge-case of curr being in a throttled
2533 * group (e.g. via set_curr_task), since update_curr() (in the
2534 * enqueue of curr) will have resulted in resched being set. This
2535 * prevents us from potentially nominating it as a false LAST_BUDDY
2538 if (test_tsk_need_resched(curr))
2541 /* Idle tasks are by definition preempted by non-idle tasks. */
2542 if (unlikely(curr->policy == SCHED_IDLE) &&
2543 likely(p->policy != SCHED_IDLE))
2547 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2548 * is driven by the tick):
2550 if (unlikely(p->policy != SCHED_NORMAL))
2553 find_matching_se(&se, &pse);
2554 update_curr(cfs_rq_of(se));
2556 if (wakeup_preempt_entity(se, pse) == 1) {
2558 * Bias pick_next to pick the sched entity that is
2559 * triggering this preemption.
2561 if (!next_buddy_marked)
2562 set_next_buddy(pse);
2571 * Only set the backward buddy when the current task is still
2572 * on the rq. This can happen when a wakeup gets interleaved
2573 * with schedule on the ->pre_schedule() or idle_balance()
2574 * point, either of which can * drop the rq lock.
2576 * Also, during early boot the idle thread is in the fair class,
2577 * for obvious reasons its a bad idea to schedule back to it.
2579 if (unlikely(!se->on_rq || curr == rq->idle))
2582 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2586 static struct task_struct *pick_next_task_fair(struct rq *rq)
2588 struct task_struct *p;
2589 struct cfs_rq *cfs_rq = &rq->cfs;
2590 struct sched_entity *se;
2592 if (!cfs_rq->nr_running)
2596 se = pick_next_entity(cfs_rq);
2597 set_next_entity(cfs_rq, se);
2598 cfs_rq = group_cfs_rq(se);
2602 hrtick_start_fair(rq, p);
2608 * Account for a descheduled task:
2610 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
2612 struct sched_entity *se = &prev->se;
2613 struct cfs_rq *cfs_rq;
2615 for_each_sched_entity(se) {
2616 cfs_rq = cfs_rq_of(se);
2617 put_prev_entity(cfs_rq, se);
2622 * sched_yield() is very simple
2624 * The magic of dealing with the ->skip buddy is in pick_next_entity.
2626 static void yield_task_fair(struct rq *rq)
2628 struct task_struct *curr = rq->curr;
2629 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2630 struct sched_entity *se = &curr->se;
2633 * Are we the only task in the tree?
2635 if (unlikely(rq->nr_running == 1))
2638 clear_buddies(cfs_rq, se);
2640 if (curr->policy != SCHED_BATCH) {
2641 update_rq_clock(rq);
2643 * Update run-time statistics of the 'current'.
2645 update_curr(cfs_rq);
2651 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2653 struct sched_entity *se = &p->se;
2655 /* throttled hierarchies are not runnable */
2656 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
2659 /* Tell the scheduler that we'd really like pse to run next. */
2662 yield_task_fair(rq);
2668 /**************************************************
2669 * Fair scheduling class load-balancing methods:
2673 * pull_task - move a task from a remote runqueue to the local runqueue.
2674 * Both runqueues must be locked.
2676 static void pull_task(struct rq *src_rq, struct task_struct *p,
2677 struct rq *this_rq, int this_cpu)
2679 deactivate_task(src_rq, p, 0);
2680 set_task_cpu(p, this_cpu);
2681 activate_task(this_rq, p, 0);
2682 check_preempt_curr(this_rq, p, 0);
2686 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2689 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2690 struct sched_domain *sd, enum cpu_idle_type idle,
2693 int tsk_cache_hot = 0;
2695 * We do not migrate tasks that are:
2696 * 1) running (obviously), or
2697 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2698 * 3) are cache-hot on their current CPU.
2700 if (!cpumask_test_cpu(this_cpu, tsk_cpus_allowed(p))) {
2701 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2706 if (task_running(rq, p)) {
2707 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2712 * Aggressive migration if:
2713 * 1) task is cache cold, or
2714 * 2) too many balance attempts have failed.
2717 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2718 if (!tsk_cache_hot ||
2719 sd->nr_balance_failed > sd->cache_nice_tries) {
2720 #ifdef CONFIG_SCHEDSTATS
2721 if (tsk_cache_hot) {
2722 schedstat_inc(sd, lb_hot_gained[idle]);
2723 schedstat_inc(p, se.statistics.nr_forced_migrations);
2729 if (tsk_cache_hot) {
2730 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2737 * move_one_task tries to move exactly one task from busiest to this_rq, as
2738 * part of active balancing operations within "domain".
2739 * Returns 1 if successful and 0 otherwise.
2741 * Called with both runqueues locked.
2744 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2745 struct sched_domain *sd, enum cpu_idle_type idle)
2747 struct task_struct *p, *n;
2748 struct cfs_rq *cfs_rq;
2751 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2752 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2753 if (throttled_lb_pair(task_group(p),
2754 busiest->cpu, this_cpu))
2757 if (!can_migrate_task(p, busiest, this_cpu,
2761 pull_task(busiest, p, this_rq, this_cpu);
2763 * Right now, this is only the second place pull_task()
2764 * is called, so we can safely collect pull_task()
2765 * stats here rather than inside pull_task().
2767 schedstat_inc(sd, lb_gained[idle]);
2775 static unsigned long
2776 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2777 unsigned long max_load_move, struct sched_domain *sd,
2778 enum cpu_idle_type idle, int *all_pinned,
2779 struct cfs_rq *busiest_cfs_rq)
2781 int loops = 0, pulled = 0;
2782 long rem_load_move = max_load_move;
2783 struct task_struct *p, *n;
2785 if (max_load_move == 0)
2788 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2789 if (loops++ > sysctl_sched_nr_migrate)
2792 if ((p->se.load.weight >> 1) > rem_load_move ||
2793 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2797 pull_task(busiest, p, this_rq, this_cpu);
2799 rem_load_move -= p->se.load.weight;
2801 #ifdef CONFIG_PREEMPT
2803 * NEWIDLE balancing is a source of latency, so preemptible
2804 * kernels will stop after the first task is pulled to minimize
2805 * the critical section.
2807 if (idle == CPU_NEWLY_IDLE)
2812 * We only want to steal up to the prescribed amount of
2815 if (rem_load_move <= 0)
2820 * Right now, this is one of only two places pull_task() is called,
2821 * so we can safely collect pull_task() stats here rather than
2822 * inside pull_task().
2824 schedstat_add(sd, lb_gained[idle], pulled);
2826 return max_load_move - rem_load_move;
2829 #ifdef CONFIG_FAIR_GROUP_SCHED
2831 * update tg->load_weight by folding this cpu's load_avg
2833 static int update_shares_cpu(struct task_group *tg, int cpu)
2835 struct cfs_rq *cfs_rq;
2836 unsigned long flags;
2843 cfs_rq = tg->cfs_rq[cpu];
2845 raw_spin_lock_irqsave(&rq->lock, flags);
2847 update_rq_clock(rq);
2848 update_cfs_load(cfs_rq, 1);
2851 * We need to update shares after updating tg->load_weight in
2852 * order to adjust the weight of groups with long running tasks.
2854 update_cfs_shares(cfs_rq);
2856 raw_spin_unlock_irqrestore(&rq->lock, flags);
2861 static void update_shares(int cpu)
2863 struct cfs_rq *cfs_rq;
2864 struct rq *rq = cpu_rq(cpu);
2868 * Iterates the task_group tree in a bottom up fashion, see
2869 * list_add_leaf_cfs_rq() for details.
2871 for_each_leaf_cfs_rq(rq, cfs_rq) {
2872 /* throttled entities do not contribute to load */
2873 if (throttled_hierarchy(cfs_rq))
2876 update_shares_cpu(cfs_rq->tg, cpu);
2882 * Compute the cpu's hierarchical load factor for each task group.
2883 * This needs to be done in a top-down fashion because the load of a child
2884 * group is a fraction of its parents load.
2886 static int tg_load_down(struct task_group *tg, void *data)
2889 long cpu = (long)data;
2892 load = cpu_rq(cpu)->load.weight;
2894 load = tg->parent->cfs_rq[cpu]->h_load;
2895 load *= tg->se[cpu]->load.weight;
2896 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2899 tg->cfs_rq[cpu]->h_load = load;
2904 static void update_h_load(long cpu)
2906 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2909 static unsigned long
2910 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2911 unsigned long max_load_move,
2912 struct sched_domain *sd, enum cpu_idle_type idle,
2915 long rem_load_move = max_load_move;
2916 struct cfs_rq *busiest_cfs_rq;
2919 update_h_load(cpu_of(busiest));
2921 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
2922 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2923 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2924 u64 rem_load, moved_load;
2927 * empty group or part of a throttled hierarchy
2929 if (!busiest_cfs_rq->task_weight ||
2930 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu))
2933 rem_load = (u64)rem_load_move * busiest_weight;
2934 rem_load = div_u64(rem_load, busiest_h_load + 1);
2936 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2937 rem_load, sd, idle, all_pinned,
2943 moved_load *= busiest_h_load;
2944 moved_load = div_u64(moved_load, busiest_weight + 1);
2946 rem_load_move -= moved_load;
2947 if (rem_load_move < 0)
2952 return max_load_move - rem_load_move;
2955 static inline void update_shares(int cpu)
2959 static unsigned long
2960 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2961 unsigned long max_load_move,
2962 struct sched_domain *sd, enum cpu_idle_type idle,
2965 return balance_tasks(this_rq, this_cpu, busiest,
2966 max_load_move, sd, idle, all_pinned,
2972 * move_tasks tries to move up to max_load_move weighted load from busiest to
2973 * this_rq, as part of a balancing operation within domain "sd".
2974 * Returns 1 if successful and 0 otherwise.
2976 * Called with both runqueues locked.
2978 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2979 unsigned long max_load_move,
2980 struct sched_domain *sd, enum cpu_idle_type idle,
2983 unsigned long total_load_moved = 0, load_moved;
2986 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2987 max_load_move - total_load_moved,
2988 sd, idle, all_pinned);
2990 total_load_moved += load_moved;
2992 #ifdef CONFIG_PREEMPT
2994 * NEWIDLE balancing is a source of latency, so preemptible
2995 * kernels will stop after the first task is pulled to minimize
2996 * the critical section.
2998 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3001 if (raw_spin_is_contended(&this_rq->lock) ||
3002 raw_spin_is_contended(&busiest->lock))
3005 } while (load_moved && max_load_move > total_load_moved);
3007 return total_load_moved > 0;
3010 /********** Helpers for find_busiest_group ************************/
3012 * sd_lb_stats - Structure to store the statistics of a sched_domain
3013 * during load balancing.
3015 struct sd_lb_stats {
3016 struct sched_group *busiest; /* Busiest group in this sd */
3017 struct sched_group *this; /* Local group in this sd */
3018 unsigned long total_load; /* Total load of all groups in sd */
3019 unsigned long total_pwr; /* Total power of all groups in sd */
3020 unsigned long avg_load; /* Average load across all groups in sd */
3022 /** Statistics of this group */
3023 unsigned long this_load;
3024 unsigned long this_load_per_task;
3025 unsigned long this_nr_running;
3026 unsigned long this_has_capacity;
3027 unsigned int this_idle_cpus;
3029 /* Statistics of the busiest group */
3030 unsigned int busiest_idle_cpus;
3031 unsigned long max_load;
3032 unsigned long busiest_load_per_task;
3033 unsigned long busiest_nr_running;
3034 unsigned long busiest_group_capacity;
3035 unsigned long busiest_has_capacity;
3036 unsigned int busiest_group_weight;
3038 int group_imb; /* Is there imbalance in this sd */
3039 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3040 int power_savings_balance; /* Is powersave balance needed for this sd */
3041 struct sched_group *group_min; /* Least loaded group in sd */
3042 struct sched_group *group_leader; /* Group which relieves group_min */
3043 unsigned long min_load_per_task; /* load_per_task in group_min */
3044 unsigned long leader_nr_running; /* Nr running of group_leader */
3045 unsigned long min_nr_running; /* Nr running of group_min */
3050 * sg_lb_stats - stats of a sched_group required for load_balancing
3052 struct sg_lb_stats {
3053 unsigned long avg_load; /*Avg load across the CPUs of the group */
3054 unsigned long group_load; /* Total load over the CPUs of the group */
3055 unsigned long sum_nr_running; /* Nr tasks running in the group */
3056 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3057 unsigned long group_capacity;
3058 unsigned long idle_cpus;
3059 unsigned long group_weight;
3060 int group_imb; /* Is there an imbalance in the group ? */
3061 int group_has_capacity; /* Is there extra capacity in the group? */
3065 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3066 * @group: The group whose first cpu is to be returned.
3068 static inline unsigned int group_first_cpu(struct sched_group *group)
3070 return cpumask_first(sched_group_cpus(group));
3074 * get_sd_load_idx - Obtain the load index for a given sched domain.
3075 * @sd: The sched_domain whose load_idx is to be obtained.
3076 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3078 static inline int get_sd_load_idx(struct sched_domain *sd,
3079 enum cpu_idle_type idle)
3085 load_idx = sd->busy_idx;
3088 case CPU_NEWLY_IDLE:
3089 load_idx = sd->newidle_idx;
3092 load_idx = sd->idle_idx;
3100 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3102 * init_sd_power_savings_stats - Initialize power savings statistics for
3103 * the given sched_domain, during load balancing.
3105 * @sd: Sched domain whose power-savings statistics are to be initialized.
3106 * @sds: Variable containing the statistics for sd.
3107 * @idle: Idle status of the CPU at which we're performing load-balancing.
3109 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3110 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3113 * Busy processors will not participate in power savings
3116 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3117 sds->power_savings_balance = 0;
3119 sds->power_savings_balance = 1;
3120 sds->min_nr_running = ULONG_MAX;
3121 sds->leader_nr_running = 0;
3126 * update_sd_power_savings_stats - Update the power saving stats for a
3127 * sched_domain while performing load balancing.
3129 * @group: sched_group belonging to the sched_domain under consideration.
3130 * @sds: Variable containing the statistics of the sched_domain
3131 * @local_group: Does group contain the CPU for which we're performing
3133 * @sgs: Variable containing the statistics of the group.
3135 static inline void update_sd_power_savings_stats(struct sched_group *group,
3136 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3139 if (!sds->power_savings_balance)
3143 * If the local group is idle or completely loaded
3144 * no need to do power savings balance at this domain
3146 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3147 !sds->this_nr_running))
3148 sds->power_savings_balance = 0;
3151 * If a group is already running at full capacity or idle,
3152 * don't include that group in power savings calculations
3154 if (!sds->power_savings_balance ||
3155 sgs->sum_nr_running >= sgs->group_capacity ||
3156 !sgs->sum_nr_running)
3160 * Calculate the group which has the least non-idle load.
3161 * This is the group from where we need to pick up the load
3164 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3165 (sgs->sum_nr_running == sds->min_nr_running &&
3166 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3167 sds->group_min = group;
3168 sds->min_nr_running = sgs->sum_nr_running;
3169 sds->min_load_per_task = sgs->sum_weighted_load /
3170 sgs->sum_nr_running;
3174 * Calculate the group which is almost near its
3175 * capacity but still has some space to pick up some load
3176 * from other group and save more power
3178 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3181 if (sgs->sum_nr_running > sds->leader_nr_running ||
3182 (sgs->sum_nr_running == sds->leader_nr_running &&
3183 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3184 sds->group_leader = group;
3185 sds->leader_nr_running = sgs->sum_nr_running;
3190 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3191 * @sds: Variable containing the statistics of the sched_domain
3192 * under consideration.
3193 * @this_cpu: Cpu at which we're currently performing load-balancing.
3194 * @imbalance: Variable to store the imbalance.
3197 * Check if we have potential to perform some power-savings balance.
3198 * If yes, set the busiest group to be the least loaded group in the
3199 * sched_domain, so that it's CPUs can be put to idle.
3201 * Returns 1 if there is potential to perform power-savings balance.
3204 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3205 int this_cpu, unsigned long *imbalance)
3207 if (!sds->power_savings_balance)
3210 if (sds->this != sds->group_leader ||
3211 sds->group_leader == sds->group_min)
3214 *imbalance = sds->min_load_per_task;
3215 sds->busiest = sds->group_min;
3220 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3221 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3222 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3227 static inline void update_sd_power_savings_stats(struct sched_group *group,
3228 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3233 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3234 int this_cpu, unsigned long *imbalance)
3238 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3241 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3243 return SCHED_POWER_SCALE;
3246 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3248 return default_scale_freq_power(sd, cpu);
3251 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3253 unsigned long weight = sd->span_weight;
3254 unsigned long smt_gain = sd->smt_gain;
3261 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3263 return default_scale_smt_power(sd, cpu);
3266 unsigned long scale_rt_power(int cpu)
3268 struct rq *rq = cpu_rq(cpu);
3269 u64 total, available;
3271 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3273 if (unlikely(total < rq->rt_avg)) {
3274 /* Ensures that power won't end up being negative */
3277 available = total - rq->rt_avg;
3280 if (unlikely((s64)total < SCHED_POWER_SCALE))
3281 total = SCHED_POWER_SCALE;
3283 total >>= SCHED_POWER_SHIFT;
3285 return div_u64(available, total);
3288 static void update_cpu_power(struct sched_domain *sd, int cpu)
3290 unsigned long weight = sd->span_weight;
3291 unsigned long power = SCHED_POWER_SCALE;
3292 struct sched_group *sdg = sd->groups;
3294 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3295 if (sched_feat(ARCH_POWER))
3296 power *= arch_scale_smt_power(sd, cpu);
3298 power *= default_scale_smt_power(sd, cpu);
3300 power >>= SCHED_POWER_SHIFT;
3303 sdg->sgp->power_orig = power;
3305 if (sched_feat(ARCH_POWER))
3306 power *= arch_scale_freq_power(sd, cpu);
3308 power *= default_scale_freq_power(sd, cpu);
3310 power >>= SCHED_POWER_SHIFT;
3312 power *= scale_rt_power(cpu);
3313 power >>= SCHED_POWER_SHIFT;
3318 cpu_rq(cpu)->cpu_power = power;
3319 sdg->sgp->power = power;
3322 static void update_group_power(struct sched_domain *sd, int cpu)
3324 struct sched_domain *child = sd->child;
3325 struct sched_group *group, *sdg = sd->groups;
3326 unsigned long power;
3329 update_cpu_power(sd, cpu);
3335 group = child->groups;
3337 power += group->sgp->power;
3338 group = group->next;
3339 } while (group != child->groups);
3341 sdg->sgp->power = power;
3345 * Try and fix up capacity for tiny siblings, this is needed when
3346 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3347 * which on its own isn't powerful enough.
3349 * See update_sd_pick_busiest() and check_asym_packing().
3352 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3355 * Only siblings can have significantly less than SCHED_POWER_SCALE
3357 if (!(sd->flags & SD_SHARE_CPUPOWER))
3361 * If ~90% of the cpu_power is still there, we're good.
3363 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3370 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3371 * @sd: The sched_domain whose statistics are to be updated.
3372 * @group: sched_group whose statistics are to be updated.
3373 * @this_cpu: Cpu for which load balance is currently performed.
3374 * @idle: Idle status of this_cpu
3375 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3376 * @local_group: Does group contain this_cpu.
3377 * @cpus: Set of cpus considered for load balancing.
3378 * @balance: Should we balance.
3379 * @sgs: variable to hold the statistics for this group.
3381 static inline void update_sg_lb_stats(struct sched_domain *sd,
3382 struct sched_group *group, int this_cpu,
3383 enum cpu_idle_type idle, int load_idx,
3384 int local_group, const struct cpumask *cpus,
3385 int *balance, struct sg_lb_stats *sgs)
3387 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3389 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3390 unsigned long avg_load_per_task = 0;
3393 balance_cpu = group_first_cpu(group);
3395 /* Tally up the load of all CPUs in the group */
3397 min_cpu_load = ~0UL;
3400 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3401 struct rq *rq = cpu_rq(i);
3403 /* Bias balancing toward cpus of our domain */
3405 if (idle_cpu(i) && !first_idle_cpu) {
3410 load = target_load(i, load_idx);
3412 load = source_load(i, load_idx);
3413 if (load > max_cpu_load) {
3414 max_cpu_load = load;
3415 max_nr_running = rq->nr_running;
3417 if (min_cpu_load > load)
3418 min_cpu_load = load;
3421 sgs->group_load += load;
3422 sgs->sum_nr_running += rq->nr_running;
3423 sgs->sum_weighted_load += weighted_cpuload(i);
3429 * First idle cpu or the first cpu(busiest) in this sched group
3430 * is eligible for doing load balancing at this and above
3431 * domains. In the newly idle case, we will allow all the cpu's
3432 * to do the newly idle load balance.
3434 if (idle != CPU_NEWLY_IDLE && local_group) {
3435 if (balance_cpu != this_cpu) {
3439 update_group_power(sd, this_cpu);
3442 /* Adjust by relative CPU power of the group */
3443 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3446 * Consider the group unbalanced when the imbalance is larger
3447 * than the average weight of a task.
3449 * APZ: with cgroup the avg task weight can vary wildly and
3450 * might not be a suitable number - should we keep a
3451 * normalized nr_running number somewhere that negates
3454 if (sgs->sum_nr_running)
3455 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3457 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3460 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3462 if (!sgs->group_capacity)
3463 sgs->group_capacity = fix_small_capacity(sd, group);
3464 sgs->group_weight = group->group_weight;
3466 if (sgs->group_capacity > sgs->sum_nr_running)
3467 sgs->group_has_capacity = 1;
3471 * update_sd_pick_busiest - return 1 on busiest group
3472 * @sd: sched_domain whose statistics are to be checked
3473 * @sds: sched_domain statistics
3474 * @sg: sched_group candidate to be checked for being the busiest
3475 * @sgs: sched_group statistics
3476 * @this_cpu: the current cpu
3478 * Determine if @sg is a busier group than the previously selected
3481 static bool update_sd_pick_busiest(struct sched_domain *sd,
3482 struct sd_lb_stats *sds,
3483 struct sched_group *sg,
3484 struct sg_lb_stats *sgs,
3487 if (sgs->avg_load <= sds->max_load)
3490 if (sgs->sum_nr_running > sgs->group_capacity)
3497 * ASYM_PACKING needs to move all the work to the lowest
3498 * numbered CPUs in the group, therefore mark all groups
3499 * higher than ourself as busy.
3501 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3502 this_cpu < group_first_cpu(sg)) {
3506 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3514 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3515 * @sd: sched_domain whose statistics are to be updated.
3516 * @this_cpu: Cpu for which load balance is currently performed.
3517 * @idle: Idle status of this_cpu
3518 * @cpus: Set of cpus considered for load balancing.
3519 * @balance: Should we balance.
3520 * @sds: variable to hold the statistics for this sched_domain.
3522 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3523 enum cpu_idle_type idle, const struct cpumask *cpus,
3524 int *balance, struct sd_lb_stats *sds)
3526 struct sched_domain *child = sd->child;
3527 struct sched_group *sg = sd->groups;
3528 struct sg_lb_stats sgs;
3529 int load_idx, prefer_sibling = 0;
3531 if (child && child->flags & SD_PREFER_SIBLING)
3534 init_sd_power_savings_stats(sd, sds, idle);
3535 load_idx = get_sd_load_idx(sd, idle);
3540 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3541 memset(&sgs, 0, sizeof(sgs));
3542 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3543 local_group, cpus, balance, &sgs);
3545 if (local_group && !(*balance))
3548 sds->total_load += sgs.group_load;
3549 sds->total_pwr += sg->sgp->power;
3552 * In case the child domain prefers tasks go to siblings
3553 * first, lower the sg capacity to one so that we'll try
3554 * and move all the excess tasks away. We lower the capacity
3555 * of a group only if the local group has the capacity to fit
3556 * these excess tasks, i.e. nr_running < group_capacity. The
3557 * extra check prevents the case where you always pull from the
3558 * heaviest group when it is already under-utilized (possible
3559 * with a large weight task outweighs the tasks on the system).
3561 if (prefer_sibling && !local_group && sds->this_has_capacity)
3562 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3565 sds->this_load = sgs.avg_load;
3567 sds->this_nr_running = sgs.sum_nr_running;
3568 sds->this_load_per_task = sgs.sum_weighted_load;
3569 sds->this_has_capacity = sgs.group_has_capacity;
3570 sds->this_idle_cpus = sgs.idle_cpus;
3571 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
3572 sds->max_load = sgs.avg_load;
3574 sds->busiest_nr_running = sgs.sum_nr_running;
3575 sds->busiest_idle_cpus = sgs.idle_cpus;
3576 sds->busiest_group_capacity = sgs.group_capacity;
3577 sds->busiest_load_per_task = sgs.sum_weighted_load;
3578 sds->busiest_has_capacity = sgs.group_has_capacity;
3579 sds->busiest_group_weight = sgs.group_weight;
3580 sds->group_imb = sgs.group_imb;
3583 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
3585 } while (sg != sd->groups);
3588 int __weak arch_sd_sibling_asym_packing(void)
3590 return 0*SD_ASYM_PACKING;
3594 * check_asym_packing - Check to see if the group is packed into the
3597 * This is primarily intended to used at the sibling level. Some
3598 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3599 * case of POWER7, it can move to lower SMT modes only when higher
3600 * threads are idle. When in lower SMT modes, the threads will
3601 * perform better since they share less core resources. Hence when we
3602 * have idle threads, we want them to be the higher ones.
3604 * This packing function is run on idle threads. It checks to see if
3605 * the busiest CPU in this domain (core in the P7 case) has a higher
3606 * CPU number than the packing function is being run on. Here we are
3607 * assuming lower CPU number will be equivalent to lower a SMT thread
3610 * Returns 1 when packing is required and a task should be moved to
3611 * this CPU. The amount of the imbalance is returned in *imbalance.
3613 * @sd: The sched_domain whose packing is to be checked.
3614 * @sds: Statistics of the sched_domain which is to be packed
3615 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3616 * @imbalance: returns amount of imbalanced due to packing.
3618 static int check_asym_packing(struct sched_domain *sd,
3619 struct sd_lb_stats *sds,
3620 int this_cpu, unsigned long *imbalance)
3624 if (!(sd->flags & SD_ASYM_PACKING))
3630 busiest_cpu = group_first_cpu(sds->busiest);
3631 if (this_cpu > busiest_cpu)
3634 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
3640 * fix_small_imbalance - Calculate the minor imbalance that exists
3641 * amongst the groups of a sched_domain, during
3643 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3644 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3645 * @imbalance: Variable to store the imbalance.
3647 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3648 int this_cpu, unsigned long *imbalance)
3650 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3651 unsigned int imbn = 2;
3652 unsigned long scaled_busy_load_per_task;
3654 if (sds->this_nr_running) {
3655 sds->this_load_per_task /= sds->this_nr_running;
3656 if (sds->busiest_load_per_task >
3657 sds->this_load_per_task)
3660 sds->this_load_per_task =
3661 cpu_avg_load_per_task(this_cpu);
3663 scaled_busy_load_per_task = sds->busiest_load_per_task
3664 * SCHED_POWER_SCALE;
3665 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3667 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3668 (scaled_busy_load_per_task * imbn)) {
3669 *imbalance = sds->busiest_load_per_task;
3674 * OK, we don't have enough imbalance to justify moving tasks,
3675 * however we may be able to increase total CPU power used by
3679 pwr_now += sds->busiest->sgp->power *
3680 min(sds->busiest_load_per_task, sds->max_load);
3681 pwr_now += sds->this->sgp->power *
3682 min(sds->this_load_per_task, sds->this_load);
3683 pwr_now /= SCHED_POWER_SCALE;
3685 /* Amount of load we'd subtract */
3686 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3687 sds->busiest->sgp->power;
3688 if (sds->max_load > tmp)
3689 pwr_move += sds->busiest->sgp->power *
3690 min(sds->busiest_load_per_task, sds->max_load - tmp);
3692 /* Amount of load we'd add */
3693 if (sds->max_load * sds->busiest->sgp->power <
3694 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3695 tmp = (sds->max_load * sds->busiest->sgp->power) /
3696 sds->this->sgp->power;
3698 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3699 sds->this->sgp->power;
3700 pwr_move += sds->this->sgp->power *
3701 min(sds->this_load_per_task, sds->this_load + tmp);
3702 pwr_move /= SCHED_POWER_SCALE;
3704 /* Move if we gain throughput */
3705 if (pwr_move > pwr_now)
3706 *imbalance = sds->busiest_load_per_task;
3710 * calculate_imbalance - Calculate the amount of imbalance present within the
3711 * groups of a given sched_domain during load balance.
3712 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3713 * @this_cpu: Cpu for which currently load balance is being performed.
3714 * @imbalance: The variable to store the imbalance.
3716 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3717 unsigned long *imbalance)
3719 unsigned long max_pull, load_above_capacity = ~0UL;
3721 sds->busiest_load_per_task /= sds->busiest_nr_running;
3722 if (sds->group_imb) {
3723 sds->busiest_load_per_task =
3724 min(sds->busiest_load_per_task, sds->avg_load);
3728 * In the presence of smp nice balancing, certain scenarios can have
3729 * max load less than avg load(as we skip the groups at or below
3730 * its cpu_power, while calculating max_load..)
3732 if (sds->max_load < sds->avg_load) {
3734 return fix_small_imbalance(sds, this_cpu, imbalance);
3737 if (!sds->group_imb) {
3739 * Don't want to pull so many tasks that a group would go idle.
3741 load_above_capacity = (sds->busiest_nr_running -
3742 sds->busiest_group_capacity);
3744 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3746 load_above_capacity /= sds->busiest->sgp->power;
3750 * We're trying to get all the cpus to the average_load, so we don't
3751 * want to push ourselves above the average load, nor do we wish to
3752 * reduce the max loaded cpu below the average load. At the same time,
3753 * we also don't want to reduce the group load below the group capacity
3754 * (so that we can implement power-savings policies etc). Thus we look
3755 * for the minimum possible imbalance.
3756 * Be careful of negative numbers as they'll appear as very large values
3757 * with unsigned longs.
3759 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3761 /* How much load to actually move to equalise the imbalance */
3762 *imbalance = min(max_pull * sds->busiest->sgp->power,
3763 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3764 / SCHED_POWER_SCALE;
3767 * if *imbalance is less than the average load per runnable task
3768 * there is no guarantee that any tasks will be moved so we'll have
3769 * a think about bumping its value to force at least one task to be
3772 if (*imbalance < sds->busiest_load_per_task)
3773 return fix_small_imbalance(sds, this_cpu, imbalance);
3777 /******* find_busiest_group() helpers end here *********************/
3780 * find_busiest_group - Returns the busiest group within the sched_domain
3781 * if there is an imbalance. If there isn't an imbalance, and
3782 * the user has opted for power-savings, it returns a group whose
3783 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3784 * such a group exists.
3786 * Also calculates the amount of weighted load which should be moved
3787 * to restore balance.
3789 * @sd: The sched_domain whose busiest group is to be returned.
3790 * @this_cpu: The cpu for which load balancing is currently being performed.
3791 * @imbalance: Variable which stores amount of weighted load which should
3792 * be moved to restore balance/put a group to idle.
3793 * @idle: The idle status of this_cpu.
3794 * @cpus: The set of CPUs under consideration for load-balancing.
3795 * @balance: Pointer to a variable indicating if this_cpu
3796 * is the appropriate cpu to perform load balancing at this_level.
3798 * Returns: - the busiest group if imbalance exists.
3799 * - If no imbalance and user has opted for power-savings balance,
3800 * return the least loaded group whose CPUs can be
3801 * put to idle by rebalancing its tasks onto our group.
3803 static struct sched_group *
3804 find_busiest_group(struct sched_domain *sd, int this_cpu,
3805 unsigned long *imbalance, enum cpu_idle_type idle,
3806 const struct cpumask *cpus, int *balance)
3808 struct sd_lb_stats sds;
3810 memset(&sds, 0, sizeof(sds));
3813 * Compute the various statistics relavent for load balancing at
3816 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3819 * this_cpu is not the appropriate cpu to perform load balancing at
3825 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3826 check_asym_packing(sd, &sds, this_cpu, imbalance))
3829 /* There is no busy sibling group to pull tasks from */
3830 if (!sds.busiest || sds.busiest_nr_running == 0)
3833 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3836 * If the busiest group is imbalanced the below checks don't
3837 * work because they assumes all things are equal, which typically
3838 * isn't true due to cpus_allowed constraints and the like.
3843 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3844 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3845 !sds.busiest_has_capacity)
3849 * If the local group is more busy than the selected busiest group
3850 * don't try and pull any tasks.
3852 if (sds.this_load >= sds.max_load)
3856 * Don't pull any tasks if this group is already above the domain
3859 if (sds.this_load >= sds.avg_load)
3862 if (idle == CPU_IDLE) {
3864 * This cpu is idle. If the busiest group load doesn't
3865 * have more tasks than the number of available cpu's and
3866 * there is no imbalance between this and busiest group
3867 * wrt to idle cpu's, it is balanced.
3869 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3870 sds.busiest_nr_running <= sds.busiest_group_weight)
3874 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3875 * imbalance_pct to be conservative.
3877 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3882 /* Looks like there is an imbalance. Compute it */
3883 calculate_imbalance(&sds, this_cpu, imbalance);
3888 * There is no obvious imbalance. But check if we can do some balancing
3891 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3899 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3902 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3903 enum cpu_idle_type idle, unsigned long imbalance,
3904 const struct cpumask *cpus)
3906 struct rq *busiest = NULL, *rq;
3907 unsigned long max_load = 0;
3910 for_each_cpu(i, sched_group_cpus(group)) {
3911 unsigned long power = power_of(i);
3912 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3917 capacity = fix_small_capacity(sd, group);
3919 if (!cpumask_test_cpu(i, cpus))
3923 wl = weighted_cpuload(i);
3926 * When comparing with imbalance, use weighted_cpuload()
3927 * which is not scaled with the cpu power.
3929 if (capacity && rq->nr_running == 1 && wl > imbalance)
3933 * For the load comparisons with the other cpu's, consider
3934 * the weighted_cpuload() scaled with the cpu power, so that
3935 * the load can be moved away from the cpu that is potentially
3936 * running at a lower capacity.
3938 wl = (wl * SCHED_POWER_SCALE) / power;
3940 if (wl > max_load) {
3950 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3951 * so long as it is large enough.
3953 #define MAX_PINNED_INTERVAL 512
3955 /* Working cpumask for load_balance and load_balance_newidle. */
3956 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3958 static int need_active_balance(struct sched_domain *sd, int idle,
3959 int busiest_cpu, int this_cpu)
3961 if (idle == CPU_NEWLY_IDLE) {
3964 * ASYM_PACKING needs to force migrate tasks from busy but
3965 * higher numbered CPUs in order to pack all tasks in the
3966 * lowest numbered CPUs.
3968 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3972 * The only task running in a non-idle cpu can be moved to this
3973 * cpu in an attempt to completely freeup the other CPU
3976 * The package power saving logic comes from
3977 * find_busiest_group(). If there are no imbalance, then
3978 * f_b_g() will return NULL. However when sched_mc={1,2} then
3979 * f_b_g() will select a group from which a running task may be
3980 * pulled to this cpu in order to make the other package idle.
3981 * If there is no opportunity to make a package idle and if
3982 * there are no imbalance, then f_b_g() will return NULL and no
3983 * action will be taken in load_balance_newidle().
3985 * Under normal task pull operation due to imbalance, there
3986 * will be more than one task in the source run queue and
3987 * move_tasks() will succeed. ld_moved will be true and this
3988 * active balance code will not be triggered.
3990 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3994 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3997 static int active_load_balance_cpu_stop(void *data);
4000 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4001 * tasks if there is an imbalance.
4003 static int load_balance(int this_cpu, struct rq *this_rq,
4004 struct sched_domain *sd, enum cpu_idle_type idle,
4007 int ld_moved, all_pinned = 0, active_balance = 0;
4008 struct sched_group *group;
4009 unsigned long imbalance;
4011 unsigned long flags;
4012 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4014 cpumask_copy(cpus, cpu_active_mask);
4016 schedstat_inc(sd, lb_count[idle]);
4019 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
4026 schedstat_inc(sd, lb_nobusyg[idle]);
4030 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
4032 schedstat_inc(sd, lb_nobusyq[idle]);
4036 BUG_ON(busiest == this_rq);
4038 schedstat_add(sd, lb_imbalance[idle], imbalance);
4041 if (busiest->nr_running > 1) {
4043 * Attempt to move tasks. If find_busiest_group has found
4044 * an imbalance but busiest->nr_running <= 1, the group is
4045 * still unbalanced. ld_moved simply stays zero, so it is
4046 * correctly treated as an imbalance.
4049 local_irq_save(flags);
4050 double_rq_lock(this_rq, busiest);
4051 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4052 imbalance, sd, idle, &all_pinned);
4053 double_rq_unlock(this_rq, busiest);
4054 local_irq_restore(flags);
4057 * some other cpu did the load balance for us.
4059 if (ld_moved && this_cpu != smp_processor_id())
4060 resched_cpu(this_cpu);
4062 /* All tasks on this runqueue were pinned by CPU affinity */
4063 if (unlikely(all_pinned)) {
4064 cpumask_clear_cpu(cpu_of(busiest), cpus);
4065 if (!cpumask_empty(cpus))
4072 schedstat_inc(sd, lb_failed[idle]);
4074 * Increment the failure counter only on periodic balance.
4075 * We do not want newidle balance, which can be very
4076 * frequent, pollute the failure counter causing
4077 * excessive cache_hot migrations and active balances.
4079 if (idle != CPU_NEWLY_IDLE)
4080 sd->nr_balance_failed++;
4082 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
4083 raw_spin_lock_irqsave(&busiest->lock, flags);
4085 /* don't kick the active_load_balance_cpu_stop,
4086 * if the curr task on busiest cpu can't be
4089 if (!cpumask_test_cpu(this_cpu,
4090 tsk_cpus_allowed(busiest->curr))) {
4091 raw_spin_unlock_irqrestore(&busiest->lock,
4094 goto out_one_pinned;
4098 * ->active_balance synchronizes accesses to
4099 * ->active_balance_work. Once set, it's cleared
4100 * only after active load balance is finished.
4102 if (!busiest->active_balance) {
4103 busiest->active_balance = 1;
4104 busiest->push_cpu = this_cpu;
4107 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4110 stop_one_cpu_nowait(cpu_of(busiest),
4111 active_load_balance_cpu_stop, busiest,
4112 &busiest->active_balance_work);
4115 * We've kicked active balancing, reset the failure
4118 sd->nr_balance_failed = sd->cache_nice_tries+1;
4121 sd->nr_balance_failed = 0;
4123 if (likely(!active_balance)) {
4124 /* We were unbalanced, so reset the balancing interval */
4125 sd->balance_interval = sd->min_interval;
4128 * If we've begun active balancing, start to back off. This
4129 * case may not be covered by the all_pinned logic if there
4130 * is only 1 task on the busy runqueue (because we don't call
4133 if (sd->balance_interval < sd->max_interval)
4134 sd->balance_interval *= 2;
4140 schedstat_inc(sd, lb_balanced[idle]);
4142 sd->nr_balance_failed = 0;
4145 /* tune up the balancing interval */
4146 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4147 (sd->balance_interval < sd->max_interval))
4148 sd->balance_interval *= 2;
4156 * idle_balance is called by schedule() if this_cpu is about to become
4157 * idle. Attempts to pull tasks from other CPUs.
4159 static void idle_balance(int this_cpu, struct rq *this_rq)
4161 struct sched_domain *sd;
4162 int pulled_task = 0;
4163 unsigned long next_balance = jiffies + HZ;
4165 this_rq->idle_stamp = this_rq->clock;
4167 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4171 * Drop the rq->lock, but keep IRQ/preempt disabled.
4173 raw_spin_unlock(&this_rq->lock);
4175 update_shares(this_cpu);
4177 for_each_domain(this_cpu, sd) {
4178 unsigned long interval;
4181 if (!(sd->flags & SD_LOAD_BALANCE))
4184 if (sd->flags & SD_BALANCE_NEWIDLE) {
4185 /* If we've pulled tasks over stop searching: */
4186 pulled_task = load_balance(this_cpu, this_rq,
4187 sd, CPU_NEWLY_IDLE, &balance);
4190 interval = msecs_to_jiffies(sd->balance_interval);
4191 if (time_after(next_balance, sd->last_balance + interval))
4192 next_balance = sd->last_balance + interval;
4194 this_rq->idle_stamp = 0;
4200 raw_spin_lock(&this_rq->lock);
4202 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4204 * We are going idle. next_balance may be set based on
4205 * a busy processor. So reset next_balance.
4207 this_rq->next_balance = next_balance;
4212 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4213 * running tasks off the busiest CPU onto idle CPUs. It requires at
4214 * least 1 task to be running on each physical CPU where possible, and
4215 * avoids physical / logical imbalances.
4217 static int active_load_balance_cpu_stop(void *data)
4219 struct rq *busiest_rq = data;
4220 int busiest_cpu = cpu_of(busiest_rq);
4221 int target_cpu = busiest_rq->push_cpu;
4222 struct rq *target_rq = cpu_rq(target_cpu);
4223 struct sched_domain *sd;
4225 raw_spin_lock_irq(&busiest_rq->lock);
4227 /* make sure the requested cpu hasn't gone down in the meantime */
4228 if (unlikely(busiest_cpu != smp_processor_id() ||
4229 !busiest_rq->active_balance))
4232 /* Is there any task to move? */
4233 if (busiest_rq->nr_running <= 1)
4237 * This condition is "impossible", if it occurs
4238 * we need to fix it. Originally reported by
4239 * Bjorn Helgaas on a 128-cpu setup.
4241 BUG_ON(busiest_rq == target_rq);
4243 /* move a task from busiest_rq to target_rq */
4244 double_lock_balance(busiest_rq, target_rq);
4246 /* Search for an sd spanning us and the target CPU. */
4248 for_each_domain(target_cpu, sd) {
4249 if ((sd->flags & SD_LOAD_BALANCE) &&
4250 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4255 schedstat_inc(sd, alb_count);
4257 if (move_one_task(target_rq, target_cpu, busiest_rq,
4259 schedstat_inc(sd, alb_pushed);
4261 schedstat_inc(sd, alb_failed);
4264 double_unlock_balance(busiest_rq, target_rq);
4266 busiest_rq->active_balance = 0;
4267 raw_spin_unlock_irq(&busiest_rq->lock);
4273 * idle load balancing details
4274 * - One of the idle CPUs nominates itself as idle load_balancer, while
4276 * - This idle load balancer CPU will also go into tickless mode when
4277 * it is idle, just like all other idle CPUs
4278 * - When one of the busy CPUs notice that there may be an idle rebalancing
4279 * needed, they will kick the idle load balancer, which then does idle
4280 * load balancing for all the idle CPUs.
4283 atomic_t load_balancer;
4284 atomic_t first_pick_cpu;
4285 atomic_t second_pick_cpu;
4286 cpumask_var_t idle_cpus_mask;
4287 cpumask_var_t grp_idle_mask;
4288 unsigned long next_balance; /* in jiffy units */
4289 } nohz ____cacheline_aligned;
4291 int get_nohz_load_balancer(void)
4293 return atomic_read(&nohz.load_balancer);
4296 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4298 * lowest_flag_domain - Return lowest sched_domain containing flag.
4299 * @cpu: The cpu whose lowest level of sched domain is to
4301 * @flag: The flag to check for the lowest sched_domain
4302 * for the given cpu.
4304 * Returns the lowest sched_domain of a cpu which contains the given flag.
4306 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4308 struct sched_domain *sd;
4310 for_each_domain(cpu, sd)
4311 if (sd->flags & flag)
4318 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4319 * @cpu: The cpu whose domains we're iterating over.
4320 * @sd: variable holding the value of the power_savings_sd
4322 * @flag: The flag to filter the sched_domains to be iterated.
4324 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4325 * set, starting from the lowest sched_domain to the highest.
4327 #define for_each_flag_domain(cpu, sd, flag) \
4328 for (sd = lowest_flag_domain(cpu, flag); \
4329 (sd && (sd->flags & flag)); sd = sd->parent)
4332 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4333 * @ilb_group: group to be checked for semi-idleness
4335 * Returns: 1 if the group is semi-idle. 0 otherwise.
4337 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4338 * and atleast one non-idle CPU. This helper function checks if the given
4339 * sched_group is semi-idle or not.
4341 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4343 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
4344 sched_group_cpus(ilb_group));
4347 * A sched_group is semi-idle when it has atleast one busy cpu
4348 * and atleast one idle cpu.
4350 if (cpumask_empty(nohz.grp_idle_mask))
4353 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
4359 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4360 * @cpu: The cpu which is nominating a new idle_load_balancer.
4362 * Returns: Returns the id of the idle load balancer if it exists,
4363 * Else, returns >= nr_cpu_ids.
4365 * This algorithm picks the idle load balancer such that it belongs to a
4366 * semi-idle powersavings sched_domain. The idea is to try and avoid
4367 * completely idle packages/cores just for the purpose of idle load balancing
4368 * when there are other idle cpu's which are better suited for that job.
4370 static int find_new_ilb(int cpu)
4372 struct sched_domain *sd;
4373 struct sched_group *ilb_group;
4374 int ilb = nr_cpu_ids;
4377 * Have idle load balancer selection from semi-idle packages only
4378 * when power-aware load balancing is enabled
4380 if (!(sched_smt_power_savings || sched_mc_power_savings))
4384 * Optimize for the case when we have no idle CPUs or only one
4385 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4387 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4391 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4392 ilb_group = sd->groups;
4395 if (is_semi_idle_group(ilb_group)) {
4396 ilb = cpumask_first(nohz.grp_idle_mask);
4400 ilb_group = ilb_group->next;
4402 } while (ilb_group != sd->groups);
4410 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4411 static inline int find_new_ilb(int call_cpu)
4418 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4419 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4420 * CPU (if there is one).
4422 static void nohz_balancer_kick(int cpu)
4426 nohz.next_balance++;
4428 ilb_cpu = get_nohz_load_balancer();
4430 if (ilb_cpu >= nr_cpu_ids) {
4431 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4432 if (ilb_cpu >= nr_cpu_ids)
4436 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4437 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4441 * Use smp_send_reschedule() instead of resched_cpu().
4442 * This way we generate a sched IPI on the target cpu which
4443 * is idle. And the softirq performing nohz idle load balance
4444 * will be run before returning from the IPI.
4446 smp_send_reschedule(ilb_cpu);
4452 * This routine will try to nominate the ilb (idle load balancing)
4453 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4454 * load balancing on behalf of all those cpus.
4456 * When the ilb owner becomes busy, we will not have new ilb owner until some
4457 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4458 * idle load balancing by kicking one of the idle CPUs.
4460 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4461 * ilb owner CPU in future (when there is a need for idle load balancing on
4462 * behalf of all idle CPUs).
4464 void select_nohz_load_balancer(int stop_tick)
4466 int cpu = smp_processor_id();
4469 if (!cpu_active(cpu)) {
4470 if (atomic_read(&nohz.load_balancer) != cpu)
4474 * If we are going offline and still the leader,
4477 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4484 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4486 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4487 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4488 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4489 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4491 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4494 /* make me the ilb owner */
4495 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4500 * Check to see if there is a more power-efficient
4503 new_ilb = find_new_ilb(cpu);
4504 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4505 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4506 resched_cpu(new_ilb);
4512 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4515 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4517 if (atomic_read(&nohz.load_balancer) == cpu)
4518 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4526 static DEFINE_SPINLOCK(balancing);
4528 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4531 * Scale the max load_balance interval with the number of CPUs in the system.
4532 * This trades load-balance latency on larger machines for less cross talk.
4534 static void update_max_interval(void)
4536 max_load_balance_interval = HZ*num_online_cpus()/10;
4540 * It checks each scheduling domain to see if it is due to be balanced,
4541 * and initiates a balancing operation if so.
4543 * Balancing parameters are set up in arch_init_sched_domains.
4545 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4548 struct rq *rq = cpu_rq(cpu);
4549 unsigned long interval;
4550 struct sched_domain *sd;
4551 /* Earliest time when we have to do rebalance again */
4552 unsigned long next_balance = jiffies + 60*HZ;
4553 int update_next_balance = 0;
4559 for_each_domain(cpu, sd) {
4560 if (!(sd->flags & SD_LOAD_BALANCE))
4563 interval = sd->balance_interval;
4564 if (idle != CPU_IDLE)
4565 interval *= sd->busy_factor;
4567 /* scale ms to jiffies */
4568 interval = msecs_to_jiffies(interval);
4569 interval = clamp(interval, 1UL, max_load_balance_interval);
4571 need_serialize = sd->flags & SD_SERIALIZE;
4573 if (need_serialize) {
4574 if (!spin_trylock(&balancing))
4578 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4579 if (load_balance(cpu, rq, sd, idle, &balance)) {
4581 * We've pulled tasks over so either we're no
4584 idle = CPU_NOT_IDLE;
4586 sd->last_balance = jiffies;
4589 spin_unlock(&balancing);
4591 if (time_after(next_balance, sd->last_balance + interval)) {
4592 next_balance = sd->last_balance + interval;
4593 update_next_balance = 1;
4597 * Stop the load balance at this level. There is another
4598 * CPU in our sched group which is doing load balancing more
4607 * next_balance will be updated only when there is a need.
4608 * When the cpu is attached to null domain for ex, it will not be
4611 if (likely(update_next_balance))
4612 rq->next_balance = next_balance;
4617 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4618 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4620 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4622 struct rq *this_rq = cpu_rq(this_cpu);
4626 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
4629 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4630 if (balance_cpu == this_cpu)
4634 * If this cpu gets work to do, stop the load balancing
4635 * work being done for other cpus. Next load
4636 * balancing owner will pick it up.
4638 if (need_resched()) {
4639 this_rq->nohz_balance_kick = 0;
4643 raw_spin_lock_irq(&this_rq->lock);
4644 update_rq_clock(this_rq);
4645 update_cpu_load(this_rq);
4646 raw_spin_unlock_irq(&this_rq->lock);
4648 rebalance_domains(balance_cpu, CPU_IDLE);
4650 rq = cpu_rq(balance_cpu);
4651 if (time_after(this_rq->next_balance, rq->next_balance))
4652 this_rq->next_balance = rq->next_balance;
4654 nohz.next_balance = this_rq->next_balance;
4655 this_rq->nohz_balance_kick = 0;
4659 * Current heuristic for kicking the idle load balancer
4660 * - first_pick_cpu is the one of the busy CPUs. It will kick
4661 * idle load balancer when it has more than one process active. This
4662 * eliminates the need for idle load balancing altogether when we have
4663 * only one running process in the system (common case).
4664 * - If there are more than one busy CPU, idle load balancer may have
4665 * to run for active_load_balance to happen (i.e., two busy CPUs are
4666 * SMT or core siblings and can run better if they move to different
4667 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4668 * which will kick idle load balancer as soon as it has any load.
4670 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4672 unsigned long now = jiffies;
4674 int first_pick_cpu, second_pick_cpu;
4676 if (time_before(now, nohz.next_balance))
4682 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4683 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4685 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4686 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4689 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4690 if (ret == nr_cpu_ids || ret == cpu) {
4691 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4692 if (rq->nr_running > 1)
4695 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4696 if (ret == nr_cpu_ids || ret == cpu) {
4704 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4708 * run_rebalance_domains is triggered when needed from the scheduler tick.
4709 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4711 static void run_rebalance_domains(struct softirq_action *h)
4713 int this_cpu = smp_processor_id();
4714 struct rq *this_rq = cpu_rq(this_cpu);
4715 enum cpu_idle_type idle = this_rq->idle_balance ?
4716 CPU_IDLE : CPU_NOT_IDLE;
4718 rebalance_domains(this_cpu, idle);
4721 * If this cpu has a pending nohz_balance_kick, then do the
4722 * balancing on behalf of the other idle cpus whose ticks are
4725 nohz_idle_balance(this_cpu, idle);
4728 static inline int on_null_domain(int cpu)
4730 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4734 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4736 static inline void trigger_load_balance(struct rq *rq, int cpu)
4738 /* Don't need to rebalance while attached to NULL domain */
4739 if (time_after_eq(jiffies, rq->next_balance) &&
4740 likely(!on_null_domain(cpu)))
4741 raise_softirq(SCHED_SOFTIRQ);
4743 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4744 nohz_balancer_kick(cpu);
4748 static void rq_online_fair(struct rq *rq)
4753 static void rq_offline_fair(struct rq *rq)
4758 #else /* CONFIG_SMP */
4761 * on UP we do not need to balance between CPUs:
4763 static inline void idle_balance(int cpu, struct rq *rq)
4767 #endif /* CONFIG_SMP */
4770 * scheduler tick hitting a task of our scheduling class:
4772 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4774 struct cfs_rq *cfs_rq;
4775 struct sched_entity *se = &curr->se;
4777 for_each_sched_entity(se) {
4778 cfs_rq = cfs_rq_of(se);
4779 entity_tick(cfs_rq, se, queued);
4784 * called on fork with the child task as argument from the parent's context
4785 * - child not yet on the tasklist
4786 * - preemption disabled
4788 static void task_fork_fair(struct task_struct *p)
4790 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4791 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4792 int this_cpu = smp_processor_id();
4793 struct rq *rq = this_rq();
4794 unsigned long flags;
4796 raw_spin_lock_irqsave(&rq->lock, flags);
4798 update_rq_clock(rq);
4800 if (unlikely(task_cpu(p) != this_cpu)) {
4802 __set_task_cpu(p, this_cpu);
4806 update_curr(cfs_rq);
4809 se->vruntime = curr->vruntime;
4810 place_entity(cfs_rq, se, 1);
4812 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4814 * Upon rescheduling, sched_class::put_prev_task() will place
4815 * 'current' within the tree based on its new key value.
4817 swap(curr->vruntime, se->vruntime);
4818 resched_task(rq->curr);
4821 se->vruntime -= cfs_rq->min_vruntime;
4823 raw_spin_unlock_irqrestore(&rq->lock, flags);
4827 * Priority of the task has changed. Check to see if we preempt
4831 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4837 * Reschedule if we are currently running on this runqueue and
4838 * our priority decreased, or if we are not currently running on
4839 * this runqueue and our priority is higher than the current's
4841 if (rq->curr == p) {
4842 if (p->prio > oldprio)
4843 resched_task(rq->curr);
4845 check_preempt_curr(rq, p, 0);
4848 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4850 struct sched_entity *se = &p->se;
4851 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4854 * Ensure the task's vruntime is normalized, so that when its
4855 * switched back to the fair class the enqueue_entity(.flags=0) will
4856 * do the right thing.
4858 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4859 * have normalized the vruntime, if it was !on_rq, then only when
4860 * the task is sleeping will it still have non-normalized vruntime.
4862 if (!se->on_rq && p->state != TASK_RUNNING) {
4864 * Fix up our vruntime so that the current sleep doesn't
4865 * cause 'unlimited' sleep bonus.
4867 place_entity(cfs_rq, se, 0);
4868 se->vruntime -= cfs_rq->min_vruntime;
4873 * We switched to the sched_fair class.
4875 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4881 * We were most likely switched from sched_rt, so
4882 * kick off the schedule if running, otherwise just see
4883 * if we can still preempt the current task.
4886 resched_task(rq->curr);
4888 check_preempt_curr(rq, p, 0);
4891 /* Account for a task changing its policy or group.
4893 * This routine is mostly called to set cfs_rq->curr field when a task
4894 * migrates between groups/classes.
4896 static void set_curr_task_fair(struct rq *rq)
4898 struct sched_entity *se = &rq->curr->se;
4900 for_each_sched_entity(se) {
4901 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4903 set_next_entity(cfs_rq, se);
4904 /* ensure bandwidth has been allocated on our new cfs_rq */
4905 account_cfs_rq_runtime(cfs_rq, 0);
4909 #ifdef CONFIG_FAIR_GROUP_SCHED
4910 static void task_move_group_fair(struct task_struct *p, int on_rq)
4913 * If the task was not on the rq at the time of this cgroup movement
4914 * it must have been asleep, sleeping tasks keep their ->vruntime
4915 * absolute on their old rq until wakeup (needed for the fair sleeper
4916 * bonus in place_entity()).
4918 * If it was on the rq, we've just 'preempted' it, which does convert
4919 * ->vruntime to a relative base.
4921 * Make sure both cases convert their relative position when migrating
4922 * to another cgroup's rq. This does somewhat interfere with the
4923 * fair sleeper stuff for the first placement, but who cares.
4926 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4927 set_task_rq(p, task_cpu(p));
4929 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4933 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4935 struct sched_entity *se = &task->se;
4936 unsigned int rr_interval = 0;
4939 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4942 if (rq->cfs.load.weight)
4943 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4949 * All the scheduling class methods:
4951 static const struct sched_class fair_sched_class = {
4952 .next = &idle_sched_class,
4953 .enqueue_task = enqueue_task_fair,
4954 .dequeue_task = dequeue_task_fair,
4955 .yield_task = yield_task_fair,
4956 .yield_to_task = yield_to_task_fair,
4958 .check_preempt_curr = check_preempt_wakeup,
4960 .pick_next_task = pick_next_task_fair,
4961 .put_prev_task = put_prev_task_fair,
4964 .select_task_rq = select_task_rq_fair,
4966 .rq_online = rq_online_fair,
4967 .rq_offline = rq_offline_fair,
4969 .task_waking = task_waking_fair,
4972 .set_curr_task = set_curr_task_fair,
4973 .task_tick = task_tick_fair,
4974 .task_fork = task_fork_fair,
4976 .prio_changed = prio_changed_fair,
4977 .switched_from = switched_from_fair,
4978 .switched_to = switched_to_fair,
4980 .get_rr_interval = get_rr_interval_fair,
4982 #ifdef CONFIG_FAIR_GROUP_SCHED
4983 .task_move_group = task_move_group_fair,
4987 #ifdef CONFIG_SCHED_DEBUG
4988 static void print_cfs_stats(struct seq_file *m, int cpu)
4990 struct cfs_rq *cfs_rq;
4993 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4994 print_cfs_rq(m, cpu, cfs_rq);