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 inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
780 * Use this CPU's actual weight instead of the last load_contribution
781 * to gain a more accurate current total weight. See
782 * update_cfs_rq_load_contribution().
784 tg_weight = atomic_read(&tg->load_weight);
785 tg_weight -= cfs_rq->load_contribution;
786 tg_weight += cfs_rq->load.weight;
791 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
793 long tg_weight, load, shares;
795 tg_weight = calc_tg_weight(tg, cfs_rq);
796 load = cfs_rq->load.weight;
798 shares = (tg->shares * load);
802 if (shares < MIN_SHARES)
804 if (shares > tg->shares)
810 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
812 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
813 update_cfs_load(cfs_rq, 0);
814 update_cfs_shares(cfs_rq);
817 # else /* CONFIG_SMP */
818 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
822 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
827 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
830 # endif /* CONFIG_SMP */
831 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
832 unsigned long weight)
835 /* commit outstanding execution time */
836 if (cfs_rq->curr == se)
838 account_entity_dequeue(cfs_rq, se);
841 update_load_set(&se->load, weight);
844 account_entity_enqueue(cfs_rq, se);
847 static void update_cfs_shares(struct cfs_rq *cfs_rq)
849 struct task_group *tg;
850 struct sched_entity *se;
854 se = tg->se[cpu_of(rq_of(cfs_rq))];
855 if (!se || throttled_hierarchy(cfs_rq))
858 if (likely(se->load.weight == tg->shares))
861 shares = calc_cfs_shares(cfs_rq, tg);
863 reweight_entity(cfs_rq_of(se), se, shares);
865 #else /* CONFIG_FAIR_GROUP_SCHED */
866 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
870 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
874 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
877 #endif /* CONFIG_FAIR_GROUP_SCHED */
879 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
881 #ifdef CONFIG_SCHEDSTATS
882 struct task_struct *tsk = NULL;
884 if (entity_is_task(se))
887 if (se->statistics.sleep_start) {
888 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
893 if (unlikely(delta > se->statistics.sleep_max))
894 se->statistics.sleep_max = delta;
896 se->statistics.sleep_start = 0;
897 se->statistics.sum_sleep_runtime += delta;
900 account_scheduler_latency(tsk, delta >> 10, 1);
901 trace_sched_stat_sleep(tsk, delta);
904 if (se->statistics.block_start) {
905 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
910 if (unlikely(delta > se->statistics.block_max))
911 se->statistics.block_max = delta;
913 se->statistics.block_start = 0;
914 se->statistics.sum_sleep_runtime += delta;
917 if (tsk->in_iowait) {
918 se->statistics.iowait_sum += delta;
919 se->statistics.iowait_count++;
920 trace_sched_stat_iowait(tsk, delta);
924 * Blocking time is in units of nanosecs, so shift by
925 * 20 to get a milliseconds-range estimation of the
926 * amount of time that the task spent sleeping:
928 if (unlikely(prof_on == SLEEP_PROFILING)) {
929 profile_hits(SLEEP_PROFILING,
930 (void *)get_wchan(tsk),
933 account_scheduler_latency(tsk, delta >> 10, 0);
939 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
941 #ifdef CONFIG_SCHED_DEBUG
942 s64 d = se->vruntime - cfs_rq->min_vruntime;
947 if (d > 3*sysctl_sched_latency)
948 schedstat_inc(cfs_rq, nr_spread_over);
953 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
955 u64 vruntime = cfs_rq->min_vruntime;
958 * The 'current' period is already promised to the current tasks,
959 * however the extra weight of the new task will slow them down a
960 * little, place the new task so that it fits in the slot that
961 * stays open at the end.
963 if (initial && sched_feat(START_DEBIT))
964 vruntime += sched_vslice(cfs_rq, se);
966 /* sleeps up to a single latency don't count. */
968 unsigned long thresh = sysctl_sched_latency;
971 * Halve their sleep time's effect, to allow
972 * for a gentler effect of sleepers:
974 if (sched_feat(GENTLE_FAIR_SLEEPERS))
980 /* ensure we never gain time by being placed backwards. */
981 vruntime = max_vruntime(se->vruntime, vruntime);
983 se->vruntime = vruntime;
986 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
989 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
992 * Update the normalized vruntime before updating min_vruntime
993 * through callig update_curr().
995 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
996 se->vruntime += cfs_rq->min_vruntime;
999 * Update run-time statistics of the 'current'.
1001 update_curr(cfs_rq);
1002 update_cfs_load(cfs_rq, 0);
1003 account_entity_enqueue(cfs_rq, se);
1004 update_cfs_shares(cfs_rq);
1006 if (flags & ENQUEUE_WAKEUP) {
1007 place_entity(cfs_rq, se, 0);
1008 enqueue_sleeper(cfs_rq, se);
1011 update_stats_enqueue(cfs_rq, se);
1012 check_spread(cfs_rq, se);
1013 if (se != cfs_rq->curr)
1014 __enqueue_entity(cfs_rq, se);
1017 if (cfs_rq->nr_running == 1) {
1018 list_add_leaf_cfs_rq(cfs_rq);
1019 check_enqueue_throttle(cfs_rq);
1023 static void __clear_buddies_last(struct sched_entity *se)
1025 for_each_sched_entity(se) {
1026 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1027 if (cfs_rq->last == se)
1028 cfs_rq->last = NULL;
1034 static void __clear_buddies_next(struct sched_entity *se)
1036 for_each_sched_entity(se) {
1037 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1038 if (cfs_rq->next == se)
1039 cfs_rq->next = NULL;
1045 static void __clear_buddies_skip(struct sched_entity *se)
1047 for_each_sched_entity(se) {
1048 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1049 if (cfs_rq->skip == se)
1050 cfs_rq->skip = NULL;
1056 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1058 if (cfs_rq->last == se)
1059 __clear_buddies_last(se);
1061 if (cfs_rq->next == se)
1062 __clear_buddies_next(se);
1064 if (cfs_rq->skip == se)
1065 __clear_buddies_skip(se);
1068 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1071 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1074 * Update run-time statistics of the 'current'.
1076 update_curr(cfs_rq);
1078 update_stats_dequeue(cfs_rq, se);
1079 if (flags & DEQUEUE_SLEEP) {
1080 #ifdef CONFIG_SCHEDSTATS
1081 if (entity_is_task(se)) {
1082 struct task_struct *tsk = task_of(se);
1084 if (tsk->state & TASK_INTERRUPTIBLE)
1085 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1086 if (tsk->state & TASK_UNINTERRUPTIBLE)
1087 se->statistics.block_start = rq_of(cfs_rq)->clock;
1092 clear_buddies(cfs_rq, se);
1094 if (se != cfs_rq->curr)
1095 __dequeue_entity(cfs_rq, se);
1097 update_cfs_load(cfs_rq, 0);
1098 account_entity_dequeue(cfs_rq, se);
1101 * Normalize the entity after updating the min_vruntime because the
1102 * update can refer to the ->curr item and we need to reflect this
1103 * movement in our normalized position.
1105 if (!(flags & DEQUEUE_SLEEP))
1106 se->vruntime -= cfs_rq->min_vruntime;
1108 /* return excess runtime on last dequeue */
1109 return_cfs_rq_runtime(cfs_rq);
1111 update_min_vruntime(cfs_rq);
1112 update_cfs_shares(cfs_rq);
1116 * Preempt the current task with a newly woken task if needed:
1119 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1121 unsigned long ideal_runtime, delta_exec;
1122 struct sched_entity *se;
1125 ideal_runtime = sched_slice(cfs_rq, curr);
1126 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1127 if (delta_exec > ideal_runtime) {
1128 resched_task(rq_of(cfs_rq)->curr);
1130 * The current task ran long enough, ensure it doesn't get
1131 * re-elected due to buddy favours.
1133 clear_buddies(cfs_rq, curr);
1138 * Ensure that a task that missed wakeup preemption by a
1139 * narrow margin doesn't have to wait for a full slice.
1140 * This also mitigates buddy induced latencies under load.
1142 if (delta_exec < sysctl_sched_min_granularity)
1145 se = __pick_first_entity(cfs_rq);
1146 delta = curr->vruntime - se->vruntime;
1151 if (delta > ideal_runtime)
1152 resched_task(rq_of(cfs_rq)->curr);
1156 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1158 /* 'current' is not kept within the tree. */
1161 * Any task has to be enqueued before it get to execute on
1162 * a CPU. So account for the time it spent waiting on the
1165 update_stats_wait_end(cfs_rq, se);
1166 __dequeue_entity(cfs_rq, se);
1169 update_stats_curr_start(cfs_rq, se);
1171 #ifdef CONFIG_SCHEDSTATS
1173 * Track our maximum slice length, if the CPU's load is at
1174 * least twice that of our own weight (i.e. dont track it
1175 * when there are only lesser-weight tasks around):
1177 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1178 se->statistics.slice_max = max(se->statistics.slice_max,
1179 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1182 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1186 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1189 * Pick the next process, keeping these things in mind, in this order:
1190 * 1) keep things fair between processes/task groups
1191 * 2) pick the "next" process, since someone really wants that to run
1192 * 3) pick the "last" process, for cache locality
1193 * 4) do not run the "skip" process, if something else is available
1195 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1197 struct sched_entity *se = __pick_first_entity(cfs_rq);
1198 struct sched_entity *left = se;
1201 * Avoid running the skip buddy, if running something else can
1202 * be done without getting too unfair.
1204 if (cfs_rq->skip == se) {
1205 struct sched_entity *second = __pick_next_entity(se);
1206 if (second && wakeup_preempt_entity(second, left) < 1)
1211 * Prefer last buddy, try to return the CPU to a preempted task.
1213 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1217 * Someone really wants this to run. If it's not unfair, run it.
1219 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1222 clear_buddies(cfs_rq, se);
1227 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1229 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1232 * If still on the runqueue then deactivate_task()
1233 * was not called and update_curr() has to be done:
1236 update_curr(cfs_rq);
1238 /* throttle cfs_rqs exceeding runtime */
1239 check_cfs_rq_runtime(cfs_rq);
1241 check_spread(cfs_rq, prev);
1243 update_stats_wait_start(cfs_rq, prev);
1244 /* Put 'current' back into the tree. */
1245 __enqueue_entity(cfs_rq, prev);
1247 cfs_rq->curr = NULL;
1251 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1254 * Update run-time statistics of the 'current'.
1256 update_curr(cfs_rq);
1259 * Update share accounting for long-running entities.
1261 update_entity_shares_tick(cfs_rq);
1263 #ifdef CONFIG_SCHED_HRTICK
1265 * queued ticks are scheduled to match the slice, so don't bother
1266 * validating it and just reschedule.
1269 resched_task(rq_of(cfs_rq)->curr);
1273 * don't let the period tick interfere with the hrtick preemption
1275 if (!sched_feat(DOUBLE_TICK) &&
1276 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1280 if (cfs_rq->nr_running > 1)
1281 check_preempt_tick(cfs_rq, curr);
1285 /**************************************************
1286 * CFS bandwidth control machinery
1289 #ifdef CONFIG_CFS_BANDWIDTH
1291 * default period for cfs group bandwidth.
1292 * default: 0.1s, units: nanoseconds
1294 static inline u64 default_cfs_period(void)
1296 return 100000000ULL;
1299 static inline u64 sched_cfs_bandwidth_slice(void)
1301 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1305 * Replenish runtime according to assigned quota and update expiration time.
1306 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1307 * additional synchronization around rq->lock.
1309 * requires cfs_b->lock
1311 static void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1315 if (cfs_b->quota == RUNTIME_INF)
1318 now = sched_clock_cpu(smp_processor_id());
1319 cfs_b->runtime = cfs_b->quota;
1320 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1323 /* returns 0 on failure to allocate runtime */
1324 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1326 struct task_group *tg = cfs_rq->tg;
1327 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1328 u64 amount = 0, min_amount, expires;
1330 /* note: this is a positive sum as runtime_remaining <= 0 */
1331 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1333 raw_spin_lock(&cfs_b->lock);
1334 if (cfs_b->quota == RUNTIME_INF)
1335 amount = min_amount;
1338 * If the bandwidth pool has become inactive, then at least one
1339 * period must have elapsed since the last consumption.
1340 * Refresh the global state and ensure bandwidth timer becomes
1343 if (!cfs_b->timer_active) {
1344 __refill_cfs_bandwidth_runtime(cfs_b);
1345 __start_cfs_bandwidth(cfs_b);
1348 if (cfs_b->runtime > 0) {
1349 amount = min(cfs_b->runtime, min_amount);
1350 cfs_b->runtime -= amount;
1354 expires = cfs_b->runtime_expires;
1355 raw_spin_unlock(&cfs_b->lock);
1357 cfs_rq->runtime_remaining += amount;
1359 * we may have advanced our local expiration to account for allowed
1360 * spread between our sched_clock and the one on which runtime was
1363 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1364 cfs_rq->runtime_expires = expires;
1366 return cfs_rq->runtime_remaining > 0;
1370 * Note: This depends on the synchronization provided by sched_clock and the
1371 * fact that rq->clock snapshots this value.
1373 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1375 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1376 struct rq *rq = rq_of(cfs_rq);
1378 /* if the deadline is ahead of our clock, nothing to do */
1379 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1382 if (cfs_rq->runtime_remaining < 0)
1386 * If the local deadline has passed we have to consider the
1387 * possibility that our sched_clock is 'fast' and the global deadline
1388 * has not truly expired.
1390 * Fortunately we can check determine whether this the case by checking
1391 * whether the global deadline has advanced.
1394 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1395 /* extend local deadline, drift is bounded above by 2 ticks */
1396 cfs_rq->runtime_expires += TICK_NSEC;
1398 /* global deadline is ahead, expiration has passed */
1399 cfs_rq->runtime_remaining = 0;
1403 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1404 unsigned long delta_exec)
1406 /* dock delta_exec before expiring quota (as it could span periods) */
1407 cfs_rq->runtime_remaining -= delta_exec;
1408 expire_cfs_rq_runtime(cfs_rq);
1410 if (likely(cfs_rq->runtime_remaining > 0))
1414 * if we're unable to extend our runtime we resched so that the active
1415 * hierarchy can be throttled
1417 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1418 resched_task(rq_of(cfs_rq)->curr);
1421 static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1422 unsigned long delta_exec)
1424 if (!cfs_rq->runtime_enabled)
1427 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1430 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1432 return cfs_rq->throttled;
1435 /* check whether cfs_rq, or any parent, is throttled */
1436 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1438 return cfs_rq->throttle_count;
1442 * Ensure that neither of the group entities corresponding to src_cpu or
1443 * dest_cpu are members of a throttled hierarchy when performing group
1444 * load-balance operations.
1446 static inline int throttled_lb_pair(struct task_group *tg,
1447 int src_cpu, int dest_cpu)
1449 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1451 src_cfs_rq = tg->cfs_rq[src_cpu];
1452 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1454 return throttled_hierarchy(src_cfs_rq) ||
1455 throttled_hierarchy(dest_cfs_rq);
1458 /* updated child weight may affect parent so we have to do this bottom up */
1459 static int tg_unthrottle_up(struct task_group *tg, void *data)
1461 struct rq *rq = data;
1462 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1464 cfs_rq->throttle_count--;
1466 if (!cfs_rq->throttle_count) {
1467 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1469 /* leaving throttled state, advance shares averaging windows */
1470 cfs_rq->load_stamp += delta;
1471 cfs_rq->load_last += delta;
1473 /* update entity weight now that we are on_rq again */
1474 update_cfs_shares(cfs_rq);
1481 static int tg_throttle_down(struct task_group *tg, void *data)
1483 struct rq *rq = data;
1484 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1486 /* group is entering throttled state, record last load */
1487 if (!cfs_rq->throttle_count)
1488 update_cfs_load(cfs_rq, 0);
1489 cfs_rq->throttle_count++;
1494 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1496 struct rq *rq = rq_of(cfs_rq);
1497 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1498 struct sched_entity *se;
1499 long task_delta, dequeue = 1;
1501 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1503 /* account load preceding throttle */
1505 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1508 task_delta = cfs_rq->h_nr_running;
1509 for_each_sched_entity(se) {
1510 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1511 /* throttled entity or throttle-on-deactivate */
1516 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1517 qcfs_rq->h_nr_running -= task_delta;
1519 if (qcfs_rq->load.weight)
1524 rq->nr_running -= task_delta;
1526 cfs_rq->throttled = 1;
1527 cfs_rq->throttled_timestamp = rq->clock;
1528 raw_spin_lock(&cfs_b->lock);
1529 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1530 raw_spin_unlock(&cfs_b->lock);
1533 static void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1535 struct rq *rq = rq_of(cfs_rq);
1536 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1537 struct sched_entity *se;
1541 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1543 cfs_rq->throttled = 0;
1544 raw_spin_lock(&cfs_b->lock);
1545 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1546 list_del_rcu(&cfs_rq->throttled_list);
1547 raw_spin_unlock(&cfs_b->lock);
1548 cfs_rq->throttled_timestamp = 0;
1550 update_rq_clock(rq);
1551 /* update hierarchical throttle state */
1552 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1554 if (!cfs_rq->load.weight)
1557 task_delta = cfs_rq->h_nr_running;
1558 for_each_sched_entity(se) {
1562 cfs_rq = cfs_rq_of(se);
1564 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1565 cfs_rq->h_nr_running += task_delta;
1567 if (cfs_rq_throttled(cfs_rq))
1572 rq->nr_running += task_delta;
1574 /* determine whether we need to wake up potentially idle cpu */
1575 if (rq->curr == rq->idle && rq->cfs.nr_running)
1576 resched_task(rq->curr);
1579 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1580 u64 remaining, u64 expires)
1582 struct cfs_rq *cfs_rq;
1583 u64 runtime = remaining;
1586 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1588 struct rq *rq = rq_of(cfs_rq);
1590 raw_spin_lock(&rq->lock);
1591 if (!cfs_rq_throttled(cfs_rq))
1594 runtime = -cfs_rq->runtime_remaining + 1;
1595 if (runtime > remaining)
1596 runtime = remaining;
1597 remaining -= runtime;
1599 cfs_rq->runtime_remaining += runtime;
1600 cfs_rq->runtime_expires = expires;
1602 /* we check whether we're throttled above */
1603 if (cfs_rq->runtime_remaining > 0)
1604 unthrottle_cfs_rq(cfs_rq);
1607 raw_spin_unlock(&rq->lock);
1618 * Responsible for refilling a task_group's bandwidth and unthrottling its
1619 * cfs_rqs as appropriate. If there has been no activity within the last
1620 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1621 * used to track this state.
1623 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1625 u64 runtime, runtime_expires;
1626 int idle = 1, throttled;
1628 raw_spin_lock(&cfs_b->lock);
1629 /* no need to continue the timer with no bandwidth constraint */
1630 if (cfs_b->quota == RUNTIME_INF)
1633 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1634 /* idle depends on !throttled (for the case of a large deficit) */
1635 idle = cfs_b->idle && !throttled;
1636 cfs_b->nr_periods += overrun;
1638 /* if we're going inactive then everything else can be deferred */
1642 __refill_cfs_bandwidth_runtime(cfs_b);
1645 /* mark as potentially idle for the upcoming period */
1650 /* account preceding periods in which throttling occurred */
1651 cfs_b->nr_throttled += overrun;
1654 * There are throttled entities so we must first use the new bandwidth
1655 * to unthrottle them before making it generally available. This
1656 * ensures that all existing debts will be paid before a new cfs_rq is
1659 runtime = cfs_b->runtime;
1660 runtime_expires = cfs_b->runtime_expires;
1664 * This check is repeated as we are holding onto the new bandwidth
1665 * while we unthrottle. This can potentially race with an unthrottled
1666 * group trying to acquire new bandwidth from the global pool.
1668 while (throttled && runtime > 0) {
1669 raw_spin_unlock(&cfs_b->lock);
1670 /* we can't nest cfs_b->lock while distributing bandwidth */
1671 runtime = distribute_cfs_runtime(cfs_b, runtime,
1673 raw_spin_lock(&cfs_b->lock);
1675 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1678 /* return (any) remaining runtime */
1679 cfs_b->runtime = runtime;
1681 * While we are ensured activity in the period following an
1682 * unthrottle, this also covers the case in which the new bandwidth is
1683 * insufficient to cover the existing bandwidth deficit. (Forcing the
1684 * timer to remain active while there are any throttled entities.)
1689 cfs_b->timer_active = 0;
1690 raw_spin_unlock(&cfs_b->lock);
1695 /* a cfs_rq won't donate quota below this amount */
1696 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1697 /* minimum remaining period time to redistribute slack quota */
1698 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1699 /* how long we wait to gather additional slack before distributing */
1700 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1702 /* are we near the end of the current quota period? */
1703 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
1705 struct hrtimer *refresh_timer = &cfs_b->period_timer;
1708 /* if the call-back is running a quota refresh is already occurring */
1709 if (hrtimer_callback_running(refresh_timer))
1712 /* is a quota refresh about to occur? */
1713 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
1714 if (remaining < min_expire)
1720 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
1722 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
1724 /* if there's a quota refresh soon don't bother with slack */
1725 if (runtime_refresh_within(cfs_b, min_left))
1728 start_bandwidth_timer(&cfs_b->slack_timer,
1729 ns_to_ktime(cfs_bandwidth_slack_period));
1732 /* we know any runtime found here is valid as update_curr() precedes return */
1733 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1735 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1736 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
1738 if (slack_runtime <= 0)
1741 raw_spin_lock(&cfs_b->lock);
1742 if (cfs_b->quota != RUNTIME_INF &&
1743 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
1744 cfs_b->runtime += slack_runtime;
1746 /* we are under rq->lock, defer unthrottling using a timer */
1747 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
1748 !list_empty(&cfs_b->throttled_cfs_rq))
1749 start_cfs_slack_bandwidth(cfs_b);
1751 raw_spin_unlock(&cfs_b->lock);
1753 /* even if it's not valid for return we don't want to try again */
1754 cfs_rq->runtime_remaining -= slack_runtime;
1757 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1759 if (!cfs_rq->runtime_enabled || !cfs_rq->nr_running)
1762 __return_cfs_rq_runtime(cfs_rq);
1766 * This is done with a timer (instead of inline with bandwidth return) since
1767 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
1769 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
1771 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
1774 /* confirm we're still not at a refresh boundary */
1775 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
1778 raw_spin_lock(&cfs_b->lock);
1779 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
1780 runtime = cfs_b->runtime;
1783 expires = cfs_b->runtime_expires;
1784 raw_spin_unlock(&cfs_b->lock);
1789 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
1791 raw_spin_lock(&cfs_b->lock);
1792 if (expires == cfs_b->runtime_expires)
1793 cfs_b->runtime = runtime;
1794 raw_spin_unlock(&cfs_b->lock);
1798 * When a group wakes up we want to make sure that its quota is not already
1799 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1800 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1802 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
1804 /* an active group must be handled by the update_curr()->put() path */
1805 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
1808 /* ensure the group is not already throttled */
1809 if (cfs_rq_throttled(cfs_rq))
1812 /* update runtime allocation */
1813 account_cfs_rq_runtime(cfs_rq, 0);
1814 if (cfs_rq->runtime_remaining <= 0)
1815 throttle_cfs_rq(cfs_rq);
1818 /* conditionally throttle active cfs_rq's from put_prev_entity() */
1819 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1821 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
1825 * it's possible for a throttled entity to be forced into a running
1826 * state (e.g. set_curr_task), in this case we're finished.
1828 if (cfs_rq_throttled(cfs_rq))
1831 throttle_cfs_rq(cfs_rq);
1834 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1835 unsigned long delta_exec) {}
1836 static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
1837 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
1838 static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
1840 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1845 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1850 static inline int throttled_lb_pair(struct task_group *tg,
1851 int src_cpu, int dest_cpu)
1857 /**************************************************
1858 * CFS operations on tasks:
1861 #ifdef CONFIG_SCHED_HRTICK
1862 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1864 struct sched_entity *se = &p->se;
1865 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1867 WARN_ON(task_rq(p) != rq);
1869 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1870 u64 slice = sched_slice(cfs_rq, se);
1871 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1872 s64 delta = slice - ran;
1881 * Don't schedule slices shorter than 10000ns, that just
1882 * doesn't make sense. Rely on vruntime for fairness.
1885 delta = max_t(s64, 10000LL, delta);
1887 hrtick_start(rq, delta);
1892 * called from enqueue/dequeue and updates the hrtick when the
1893 * current task is from our class and nr_running is low enough
1896 static void hrtick_update(struct rq *rq)
1898 struct task_struct *curr = rq->curr;
1900 if (curr->sched_class != &fair_sched_class)
1903 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1904 hrtick_start_fair(rq, curr);
1906 #else /* !CONFIG_SCHED_HRTICK */
1908 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1912 static inline void hrtick_update(struct rq *rq)
1918 * The enqueue_task method is called before nr_running is
1919 * increased. Here we update the fair scheduling stats and
1920 * then put the task into the rbtree:
1923 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1925 struct cfs_rq *cfs_rq;
1926 struct sched_entity *se = &p->se;
1928 for_each_sched_entity(se) {
1931 cfs_rq = cfs_rq_of(se);
1932 enqueue_entity(cfs_rq, se, flags);
1935 * end evaluation on encountering a throttled cfs_rq
1937 * note: in the case of encountering a throttled cfs_rq we will
1938 * post the final h_nr_running increment below.
1940 if (cfs_rq_throttled(cfs_rq))
1942 cfs_rq->h_nr_running++;
1944 flags = ENQUEUE_WAKEUP;
1947 for_each_sched_entity(se) {
1948 cfs_rq = cfs_rq_of(se);
1949 cfs_rq->h_nr_running++;
1951 if (cfs_rq_throttled(cfs_rq))
1954 update_cfs_load(cfs_rq, 0);
1955 update_cfs_shares(cfs_rq);
1963 static void set_next_buddy(struct sched_entity *se);
1966 * The dequeue_task method is called before nr_running is
1967 * decreased. We remove the task from the rbtree and
1968 * update the fair scheduling stats:
1970 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1972 struct cfs_rq *cfs_rq;
1973 struct sched_entity *se = &p->se;
1974 int task_sleep = flags & DEQUEUE_SLEEP;
1976 for_each_sched_entity(se) {
1977 cfs_rq = cfs_rq_of(se);
1978 dequeue_entity(cfs_rq, se, flags);
1981 * end evaluation on encountering a throttled cfs_rq
1983 * note: in the case of encountering a throttled cfs_rq we will
1984 * post the final h_nr_running decrement below.
1986 if (cfs_rq_throttled(cfs_rq))
1988 cfs_rq->h_nr_running--;
1990 /* Don't dequeue parent if it has other entities besides us */
1991 if (cfs_rq->load.weight) {
1993 * Bias pick_next to pick a task from this cfs_rq, as
1994 * p is sleeping when it is within its sched_slice.
1996 if (task_sleep && parent_entity(se))
1997 set_next_buddy(parent_entity(se));
1999 /* avoid re-evaluating load for this entity */
2000 se = parent_entity(se);
2003 flags |= DEQUEUE_SLEEP;
2006 for_each_sched_entity(se) {
2007 cfs_rq = cfs_rq_of(se);
2008 cfs_rq->h_nr_running--;
2010 if (cfs_rq_throttled(cfs_rq))
2013 update_cfs_load(cfs_rq, 0);
2014 update_cfs_shares(cfs_rq);
2024 static void task_waking_fair(struct task_struct *p)
2026 struct sched_entity *se = &p->se;
2027 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2030 #ifndef CONFIG_64BIT
2031 u64 min_vruntime_copy;
2034 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2036 min_vruntime = cfs_rq->min_vruntime;
2037 } while (min_vruntime != min_vruntime_copy);
2039 min_vruntime = cfs_rq->min_vruntime;
2042 se->vruntime -= min_vruntime;
2045 #ifdef CONFIG_FAIR_GROUP_SCHED
2047 * effective_load() calculates the load change as seen from the root_task_group
2049 * Adding load to a group doesn't make a group heavier, but can cause movement
2050 * of group shares between cpus. Assuming the shares were perfectly aligned one
2051 * can calculate the shift in shares.
2053 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2054 * on this @cpu and results in a total addition (subtraction) of @wg to the
2055 * total group weight.
2057 * Given a runqueue weight distribution (rw_i) we can compute a shares
2058 * distribution (s_i) using:
2060 * s_i = rw_i / \Sum rw_j (1)
2062 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2063 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2064 * shares distribution (s_i):
2066 * rw_i = { 2, 4, 1, 0 }
2067 * s_i = { 2/7, 4/7, 1/7, 0 }
2069 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2070 * task used to run on and the CPU the waker is running on), we need to
2071 * compute the effect of waking a task on either CPU and, in case of a sync
2072 * wakeup, compute the effect of the current task going to sleep.
2074 * So for a change of @wl to the local @cpu with an overall group weight change
2075 * of @wl we can compute the new shares distribution (s'_i) using:
2077 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2079 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2080 * differences in waking a task to CPU 0. The additional task changes the
2081 * weight and shares distributions like:
2083 * rw'_i = { 3, 4, 1, 0 }
2084 * s'_i = { 3/8, 4/8, 1/8, 0 }
2086 * We can then compute the difference in effective weight by using:
2088 * dw_i = S * (s'_i - s_i) (3)
2090 * Where 'S' is the group weight as seen by its parent.
2092 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2093 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2094 * 4/7) times the weight of the group.
2096 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2098 struct sched_entity *se = tg->se[cpu];
2100 if (!tg->parent) /* the trivial, non-cgroup case */
2103 for_each_sched_entity(se) {
2109 * W = @wg + \Sum rw_j
2111 W = wg + calc_tg_weight(tg, se->my_q);
2116 w = se->my_q->load.weight + wl;
2119 * wl = S * s'_i; see (2)
2122 wl = (w * tg->shares) / W;
2127 * Per the above, wl is the new se->load.weight value; since
2128 * those are clipped to [MIN_SHARES, ...) do so now. See
2129 * calc_cfs_shares().
2131 if (wl < MIN_SHARES)
2135 * wl = dw_i = S * (s'_i - s_i); see (3)
2137 wl -= se->load.weight;
2140 * Recursively apply this logic to all parent groups to compute
2141 * the final effective load change on the root group. Since
2142 * only the @tg group gets extra weight, all parent groups can
2143 * only redistribute existing shares. @wl is the shift in shares
2144 * resulting from this level per the above.
2153 static inline unsigned long effective_load(struct task_group *tg, int cpu,
2154 unsigned long wl, unsigned long wg)
2161 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2163 s64 this_load, load;
2164 int idx, this_cpu, prev_cpu;
2165 unsigned long tl_per_task;
2166 struct task_group *tg;
2167 unsigned long weight;
2171 this_cpu = smp_processor_id();
2172 prev_cpu = task_cpu(p);
2173 load = source_load(prev_cpu, idx);
2174 this_load = target_load(this_cpu, idx);
2177 * If sync wakeup then subtract the (maximum possible)
2178 * effect of the currently running task from the load
2179 * of the current CPU:
2182 tg = task_group(current);
2183 weight = current->se.load.weight;
2185 this_load += effective_load(tg, this_cpu, -weight, -weight);
2186 load += effective_load(tg, prev_cpu, 0, -weight);
2190 weight = p->se.load.weight;
2193 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2194 * due to the sync cause above having dropped this_load to 0, we'll
2195 * always have an imbalance, but there's really nothing you can do
2196 * about that, so that's good too.
2198 * Otherwise check if either cpus are near enough in load to allow this
2199 * task to be woken on this_cpu.
2201 if (this_load > 0) {
2202 s64 this_eff_load, prev_eff_load;
2204 this_eff_load = 100;
2205 this_eff_load *= power_of(prev_cpu);
2206 this_eff_load *= this_load +
2207 effective_load(tg, this_cpu, weight, weight);
2209 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2210 prev_eff_load *= power_of(this_cpu);
2211 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2213 balanced = this_eff_load <= prev_eff_load;
2218 * If the currently running task will sleep within
2219 * a reasonable amount of time then attract this newly
2222 if (sync && balanced)
2225 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2226 tl_per_task = cpu_avg_load_per_task(this_cpu);
2229 (this_load <= load &&
2230 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2232 * This domain has SD_WAKE_AFFINE and
2233 * p is cache cold in this domain, and
2234 * there is no bad imbalance.
2236 schedstat_inc(sd, ttwu_move_affine);
2237 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2245 * find_idlest_group finds and returns the least busy CPU group within the
2248 static struct sched_group *
2249 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2250 int this_cpu, int load_idx)
2252 struct sched_group *idlest = NULL, *group = sd->groups;
2253 unsigned long min_load = ULONG_MAX, this_load = 0;
2254 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2257 unsigned long load, avg_load;
2261 /* Skip over this group if it has no CPUs allowed */
2262 if (!cpumask_intersects(sched_group_cpus(group),
2263 tsk_cpus_allowed(p)))
2266 local_group = cpumask_test_cpu(this_cpu,
2267 sched_group_cpus(group));
2269 /* Tally up the load of all CPUs in the group */
2272 for_each_cpu(i, sched_group_cpus(group)) {
2273 /* Bias balancing toward cpus of our domain */
2275 load = source_load(i, load_idx);
2277 load = target_load(i, load_idx);
2282 /* Adjust by relative CPU power of the group */
2283 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2286 this_load = avg_load;
2287 } else if (avg_load < min_load) {
2288 min_load = avg_load;
2291 } while (group = group->next, group != sd->groups);
2293 if (!idlest || 100*this_load < imbalance*min_load)
2299 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2302 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2304 unsigned long load, min_load = ULONG_MAX;
2308 /* Traverse only the allowed CPUs */
2309 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2310 load = weighted_cpuload(i);
2312 if (load < min_load || (load == min_load && i == this_cpu)) {
2322 * Try and locate an idle CPU in the sched_domain.
2324 static int select_idle_sibling(struct task_struct *p, int target)
2326 int cpu = smp_processor_id();
2327 int prev_cpu = task_cpu(p);
2328 struct sched_domain *sd;
2332 * If the task is going to be woken-up on this cpu and if it is
2333 * already idle, then it is the right target.
2335 if (target == cpu && idle_cpu(cpu))
2339 * If the task is going to be woken-up on the cpu where it previously
2340 * ran and if it is currently idle, then it the right target.
2342 if (target == prev_cpu && idle_cpu(prev_cpu))
2346 * Otherwise, iterate the domains and find an elegible idle cpu.
2349 for_each_domain(target, sd) {
2350 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
2353 for_each_cpu_and(i, sched_domain_span(sd), tsk_cpus_allowed(p)) {
2361 * Lets stop looking for an idle sibling when we reached
2362 * the domain that spans the current cpu and prev_cpu.
2364 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
2365 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
2374 * sched_balance_self: balance the current task (running on cpu) in domains
2375 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2378 * Balance, ie. select the least loaded group.
2380 * Returns the target CPU number, or the same CPU if no balancing is needed.
2382 * preempt must be disabled.
2385 select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2387 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2388 int cpu = smp_processor_id();
2389 int prev_cpu = task_cpu(p);
2391 int want_affine = 0;
2393 int sync = wake_flags & WF_SYNC;
2395 if (sd_flag & SD_BALANCE_WAKE) {
2396 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2402 for_each_domain(cpu, tmp) {
2403 if (!(tmp->flags & SD_LOAD_BALANCE))
2407 * If power savings logic is enabled for a domain, see if we
2408 * are not overloaded, if so, don't balance wider.
2410 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2411 unsigned long power = 0;
2412 unsigned long nr_running = 0;
2413 unsigned long capacity;
2416 for_each_cpu(i, sched_domain_span(tmp)) {
2417 power += power_of(i);
2418 nr_running += cpu_rq(i)->cfs.nr_running;
2421 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2423 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2426 if (nr_running < capacity)
2431 * If both cpu and prev_cpu are part of this domain,
2432 * cpu is a valid SD_WAKE_AFFINE target.
2434 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2435 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2440 if (!want_sd && !want_affine)
2443 if (!(tmp->flags & sd_flag))
2451 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2454 new_cpu = select_idle_sibling(p, prev_cpu);
2459 int load_idx = sd->forkexec_idx;
2460 struct sched_group *group;
2463 if (!(sd->flags & sd_flag)) {
2468 if (sd_flag & SD_BALANCE_WAKE)
2469 load_idx = sd->wake_idx;
2471 group = find_idlest_group(sd, p, cpu, load_idx);
2477 new_cpu = find_idlest_cpu(group, p, cpu);
2478 if (new_cpu == -1 || new_cpu == cpu) {
2479 /* Now try balancing at a lower domain level of cpu */
2484 /* Now try balancing at a lower domain level of new_cpu */
2486 weight = sd->span_weight;
2488 for_each_domain(cpu, tmp) {
2489 if (weight <= tmp->span_weight)
2491 if (tmp->flags & sd_flag)
2494 /* while loop will break here if sd == NULL */
2501 #endif /* CONFIG_SMP */
2503 static unsigned long
2504 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2506 unsigned long gran = sysctl_sched_wakeup_granularity;
2509 * Since its curr running now, convert the gran from real-time
2510 * to virtual-time in his units.
2512 * By using 'se' instead of 'curr' we penalize light tasks, so
2513 * they get preempted easier. That is, if 'se' < 'curr' then
2514 * the resulting gran will be larger, therefore penalizing the
2515 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2516 * be smaller, again penalizing the lighter task.
2518 * This is especially important for buddies when the leftmost
2519 * task is higher priority than the buddy.
2521 return calc_delta_fair(gran, se);
2525 * Should 'se' preempt 'curr'.
2539 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2541 s64 gran, vdiff = curr->vruntime - se->vruntime;
2546 gran = wakeup_gran(curr, se);
2553 static void set_last_buddy(struct sched_entity *se)
2555 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2558 for_each_sched_entity(se)
2559 cfs_rq_of(se)->last = se;
2562 static void set_next_buddy(struct sched_entity *se)
2564 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2567 for_each_sched_entity(se)
2568 cfs_rq_of(se)->next = se;
2571 static void set_skip_buddy(struct sched_entity *se)
2573 for_each_sched_entity(se)
2574 cfs_rq_of(se)->skip = se;
2578 * Preempt the current task with a newly woken task if needed:
2580 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2582 struct task_struct *curr = rq->curr;
2583 struct sched_entity *se = &curr->se, *pse = &p->se;
2584 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2585 int scale = cfs_rq->nr_running >= sched_nr_latency;
2586 int next_buddy_marked = 0;
2588 if (unlikely(se == pse))
2592 * This is possible from callers such as pull_task(), in which we
2593 * unconditionally check_prempt_curr() after an enqueue (which may have
2594 * lead to a throttle). This both saves work and prevents false
2595 * next-buddy nomination below.
2597 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
2600 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2601 set_next_buddy(pse);
2602 next_buddy_marked = 1;
2606 * We can come here with TIF_NEED_RESCHED already set from new task
2609 * Note: this also catches the edge-case of curr being in a throttled
2610 * group (e.g. via set_curr_task), since update_curr() (in the
2611 * enqueue of curr) will have resulted in resched being set. This
2612 * prevents us from potentially nominating it as a false LAST_BUDDY
2615 if (test_tsk_need_resched(curr))
2618 /* Idle tasks are by definition preempted by non-idle tasks. */
2619 if (unlikely(curr->policy == SCHED_IDLE) &&
2620 likely(p->policy != SCHED_IDLE))
2624 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2625 * is driven by the tick):
2627 if (unlikely(p->policy != SCHED_NORMAL))
2630 find_matching_se(&se, &pse);
2631 update_curr(cfs_rq_of(se));
2633 if (wakeup_preempt_entity(se, pse) == 1) {
2635 * Bias pick_next to pick the sched entity that is
2636 * triggering this preemption.
2638 if (!next_buddy_marked)
2639 set_next_buddy(pse);
2648 * Only set the backward buddy when the current task is still
2649 * on the rq. This can happen when a wakeup gets interleaved
2650 * with schedule on the ->pre_schedule() or idle_balance()
2651 * point, either of which can * drop the rq lock.
2653 * Also, during early boot the idle thread is in the fair class,
2654 * for obvious reasons its a bad idea to schedule back to it.
2656 if (unlikely(!se->on_rq || curr == rq->idle))
2659 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2663 static struct task_struct *pick_next_task_fair(struct rq *rq)
2665 struct task_struct *p;
2666 struct cfs_rq *cfs_rq = &rq->cfs;
2667 struct sched_entity *se;
2669 if (!cfs_rq->nr_running)
2673 se = pick_next_entity(cfs_rq);
2674 set_next_entity(cfs_rq, se);
2675 cfs_rq = group_cfs_rq(se);
2679 hrtick_start_fair(rq, p);
2685 * Account for a descheduled task:
2687 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
2689 struct sched_entity *se = &prev->se;
2690 struct cfs_rq *cfs_rq;
2692 for_each_sched_entity(se) {
2693 cfs_rq = cfs_rq_of(se);
2694 put_prev_entity(cfs_rq, se);
2699 * sched_yield() is very simple
2701 * The magic of dealing with the ->skip buddy is in pick_next_entity.
2703 static void yield_task_fair(struct rq *rq)
2705 struct task_struct *curr = rq->curr;
2706 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2707 struct sched_entity *se = &curr->se;
2710 * Are we the only task in the tree?
2712 if (unlikely(rq->nr_running == 1))
2715 clear_buddies(cfs_rq, se);
2717 if (curr->policy != SCHED_BATCH) {
2718 update_rq_clock(rq);
2720 * Update run-time statistics of the 'current'.
2722 update_curr(cfs_rq);
2728 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
2730 struct sched_entity *se = &p->se;
2732 /* throttled hierarchies are not runnable */
2733 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
2736 /* Tell the scheduler that we'd really like pse to run next. */
2739 yield_task_fair(rq);
2745 /**************************************************
2746 * Fair scheduling class load-balancing methods:
2750 * pull_task - move a task from a remote runqueue to the local runqueue.
2751 * Both runqueues must be locked.
2753 static void pull_task(struct rq *src_rq, struct task_struct *p,
2754 struct rq *this_rq, int this_cpu)
2756 deactivate_task(src_rq, p, 0);
2757 set_task_cpu(p, this_cpu);
2758 activate_task(this_rq, p, 0);
2759 check_preempt_curr(this_rq, p, 0);
2763 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2766 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2767 struct sched_domain *sd, enum cpu_idle_type idle,
2770 int tsk_cache_hot = 0;
2772 * We do not migrate tasks that are:
2773 * 1) running (obviously), or
2774 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2775 * 3) are cache-hot on their current CPU.
2777 if (!cpumask_test_cpu(this_cpu, tsk_cpus_allowed(p))) {
2778 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
2783 if (task_running(rq, p)) {
2784 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
2789 * Aggressive migration if:
2790 * 1) task is cache cold, or
2791 * 2) too many balance attempts have failed.
2794 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
2795 if (!tsk_cache_hot ||
2796 sd->nr_balance_failed > sd->cache_nice_tries) {
2797 #ifdef CONFIG_SCHEDSTATS
2798 if (tsk_cache_hot) {
2799 schedstat_inc(sd, lb_hot_gained[idle]);
2800 schedstat_inc(p, se.statistics.nr_forced_migrations);
2806 if (tsk_cache_hot) {
2807 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
2814 * move_one_task tries to move exactly one task from busiest to this_rq, as
2815 * part of active balancing operations within "domain".
2816 * Returns 1 if successful and 0 otherwise.
2818 * Called with both runqueues locked.
2821 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2822 struct sched_domain *sd, enum cpu_idle_type idle)
2824 struct task_struct *p, *n;
2825 struct cfs_rq *cfs_rq;
2828 for_each_leaf_cfs_rq(busiest, cfs_rq) {
2829 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
2830 if (throttled_lb_pair(task_group(p),
2831 busiest->cpu, this_cpu))
2834 if (!can_migrate_task(p, busiest, this_cpu,
2838 pull_task(busiest, p, this_rq, this_cpu);
2840 * Right now, this is only the second place pull_task()
2841 * is called, so we can safely collect pull_task()
2842 * stats here rather than inside pull_task().
2844 schedstat_inc(sd, lb_gained[idle]);
2852 static unsigned long
2853 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2854 unsigned long max_load_move, struct sched_domain *sd,
2855 enum cpu_idle_type idle, int *all_pinned,
2856 struct cfs_rq *busiest_cfs_rq)
2858 int loops = 0, pulled = 0;
2859 long rem_load_move = max_load_move;
2860 struct task_struct *p, *n;
2862 if (max_load_move == 0)
2865 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2866 if (loops++ > sysctl_sched_nr_migrate)
2869 if ((p->se.load.weight >> 1) > rem_load_move ||
2870 !can_migrate_task(p, busiest, this_cpu, sd, idle,
2874 pull_task(busiest, p, this_rq, this_cpu);
2876 rem_load_move -= p->se.load.weight;
2878 #ifdef CONFIG_PREEMPT
2880 * NEWIDLE balancing is a source of latency, so preemptible
2881 * kernels will stop after the first task is pulled to minimize
2882 * the critical section.
2884 if (idle == CPU_NEWLY_IDLE)
2889 * We only want to steal up to the prescribed amount of
2892 if (rem_load_move <= 0)
2897 * Right now, this is one of only two places pull_task() is called,
2898 * so we can safely collect pull_task() stats here rather than
2899 * inside pull_task().
2901 schedstat_add(sd, lb_gained[idle], pulled);
2903 return max_load_move - rem_load_move;
2906 #ifdef CONFIG_FAIR_GROUP_SCHED
2908 * update tg->load_weight by folding this cpu's load_avg
2910 static int update_shares_cpu(struct task_group *tg, int cpu)
2912 struct cfs_rq *cfs_rq;
2913 unsigned long flags;
2920 cfs_rq = tg->cfs_rq[cpu];
2922 raw_spin_lock_irqsave(&rq->lock, flags);
2924 update_rq_clock(rq);
2925 update_cfs_load(cfs_rq, 1);
2928 * We need to update shares after updating tg->load_weight in
2929 * order to adjust the weight of groups with long running tasks.
2931 update_cfs_shares(cfs_rq);
2933 raw_spin_unlock_irqrestore(&rq->lock, flags);
2938 static void update_shares(int cpu)
2940 struct cfs_rq *cfs_rq;
2941 struct rq *rq = cpu_rq(cpu);
2945 * Iterates the task_group tree in a bottom up fashion, see
2946 * list_add_leaf_cfs_rq() for details.
2948 for_each_leaf_cfs_rq(rq, cfs_rq) {
2949 /* throttled entities do not contribute to load */
2950 if (throttled_hierarchy(cfs_rq))
2953 update_shares_cpu(cfs_rq->tg, cpu);
2959 * Compute the cpu's hierarchical load factor for each task group.
2960 * This needs to be done in a top-down fashion because the load of a child
2961 * group is a fraction of its parents load.
2963 static int tg_load_down(struct task_group *tg, void *data)
2966 long cpu = (long)data;
2969 load = cpu_rq(cpu)->load.weight;
2971 load = tg->parent->cfs_rq[cpu]->h_load;
2972 load *= tg->se[cpu]->load.weight;
2973 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
2976 tg->cfs_rq[cpu]->h_load = load;
2981 static void update_h_load(long cpu)
2983 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
2986 static unsigned long
2987 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2988 unsigned long max_load_move,
2989 struct sched_domain *sd, enum cpu_idle_type idle,
2992 long rem_load_move = max_load_move;
2993 struct cfs_rq *busiest_cfs_rq;
2996 update_h_load(cpu_of(busiest));
2998 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
2999 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
3000 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
3001 u64 rem_load, moved_load;
3004 * empty group or part of a throttled hierarchy
3006 if (!busiest_cfs_rq->task_weight ||
3007 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu))
3010 rem_load = (u64)rem_load_move * busiest_weight;
3011 rem_load = div_u64(rem_load, busiest_h_load + 1);
3013 moved_load = balance_tasks(this_rq, this_cpu, busiest,
3014 rem_load, sd, idle, all_pinned,
3020 moved_load *= busiest_h_load;
3021 moved_load = div_u64(moved_load, busiest_weight + 1);
3023 rem_load_move -= moved_load;
3024 if (rem_load_move < 0)
3029 return max_load_move - rem_load_move;
3032 static inline void update_shares(int cpu)
3036 static unsigned long
3037 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
3038 unsigned long max_load_move,
3039 struct sched_domain *sd, enum cpu_idle_type idle,
3042 return balance_tasks(this_rq, this_cpu, busiest,
3043 max_load_move, sd, idle, all_pinned,
3049 * move_tasks tries to move up to max_load_move weighted load from busiest to
3050 * this_rq, as part of a balancing operation within domain "sd".
3051 * Returns 1 if successful and 0 otherwise.
3053 * Called with both runqueues locked.
3055 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3056 unsigned long max_load_move,
3057 struct sched_domain *sd, enum cpu_idle_type idle,
3060 unsigned long total_load_moved = 0, load_moved;
3063 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
3064 max_load_move - total_load_moved,
3065 sd, idle, all_pinned);
3067 total_load_moved += load_moved;
3069 #ifdef CONFIG_PREEMPT
3071 * NEWIDLE balancing is a source of latency, so preemptible
3072 * kernels will stop after the first task is pulled to minimize
3073 * the critical section.
3075 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3078 if (raw_spin_is_contended(&this_rq->lock) ||
3079 raw_spin_is_contended(&busiest->lock))
3082 } while (load_moved && max_load_move > total_load_moved);
3084 return total_load_moved > 0;
3087 /********** Helpers for find_busiest_group ************************/
3089 * sd_lb_stats - Structure to store the statistics of a sched_domain
3090 * during load balancing.
3092 struct sd_lb_stats {
3093 struct sched_group *busiest; /* Busiest group in this sd */
3094 struct sched_group *this; /* Local group in this sd */
3095 unsigned long total_load; /* Total load of all groups in sd */
3096 unsigned long total_pwr; /* Total power of all groups in sd */
3097 unsigned long avg_load; /* Average load across all groups in sd */
3099 /** Statistics of this group */
3100 unsigned long this_load;
3101 unsigned long this_load_per_task;
3102 unsigned long this_nr_running;
3103 unsigned long this_has_capacity;
3104 unsigned int this_idle_cpus;
3106 /* Statistics of the busiest group */
3107 unsigned int busiest_idle_cpus;
3108 unsigned long max_load;
3109 unsigned long busiest_load_per_task;
3110 unsigned long busiest_nr_running;
3111 unsigned long busiest_group_capacity;
3112 unsigned long busiest_has_capacity;
3113 unsigned int busiest_group_weight;
3115 int group_imb; /* Is there imbalance in this sd */
3116 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3117 int power_savings_balance; /* Is powersave balance needed for this sd */
3118 struct sched_group *group_min; /* Least loaded group in sd */
3119 struct sched_group *group_leader; /* Group which relieves group_min */
3120 unsigned long min_load_per_task; /* load_per_task in group_min */
3121 unsigned long leader_nr_running; /* Nr running of group_leader */
3122 unsigned long min_nr_running; /* Nr running of group_min */
3127 * sg_lb_stats - stats of a sched_group required for load_balancing
3129 struct sg_lb_stats {
3130 unsigned long avg_load; /*Avg load across the CPUs of the group */
3131 unsigned long group_load; /* Total load over the CPUs of the group */
3132 unsigned long sum_nr_running; /* Nr tasks running in the group */
3133 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3134 unsigned long group_capacity;
3135 unsigned long idle_cpus;
3136 unsigned long group_weight;
3137 int group_imb; /* Is there an imbalance in the group ? */
3138 int group_has_capacity; /* Is there extra capacity in the group? */
3142 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3143 * @group: The group whose first cpu is to be returned.
3145 static inline unsigned int group_first_cpu(struct sched_group *group)
3147 return cpumask_first(sched_group_cpus(group));
3151 * get_sd_load_idx - Obtain the load index for a given sched domain.
3152 * @sd: The sched_domain whose load_idx is to be obtained.
3153 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3155 static inline int get_sd_load_idx(struct sched_domain *sd,
3156 enum cpu_idle_type idle)
3162 load_idx = sd->busy_idx;
3165 case CPU_NEWLY_IDLE:
3166 load_idx = sd->newidle_idx;
3169 load_idx = sd->idle_idx;
3177 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3179 * init_sd_power_savings_stats - Initialize power savings statistics for
3180 * the given sched_domain, during load balancing.
3182 * @sd: Sched domain whose power-savings statistics are to be initialized.
3183 * @sds: Variable containing the statistics for sd.
3184 * @idle: Idle status of the CPU at which we're performing load-balancing.
3186 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3187 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3190 * Busy processors will not participate in power savings
3193 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3194 sds->power_savings_balance = 0;
3196 sds->power_savings_balance = 1;
3197 sds->min_nr_running = ULONG_MAX;
3198 sds->leader_nr_running = 0;
3203 * update_sd_power_savings_stats - Update the power saving stats for a
3204 * sched_domain while performing load balancing.
3206 * @group: sched_group belonging to the sched_domain under consideration.
3207 * @sds: Variable containing the statistics of the sched_domain
3208 * @local_group: Does group contain the CPU for which we're performing
3210 * @sgs: Variable containing the statistics of the group.
3212 static inline void update_sd_power_savings_stats(struct sched_group *group,
3213 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3216 if (!sds->power_savings_balance)
3220 * If the local group is idle or completely loaded
3221 * no need to do power savings balance at this domain
3223 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3224 !sds->this_nr_running))
3225 sds->power_savings_balance = 0;
3228 * If a group is already running at full capacity or idle,
3229 * don't include that group in power savings calculations
3231 if (!sds->power_savings_balance ||
3232 sgs->sum_nr_running >= sgs->group_capacity ||
3233 !sgs->sum_nr_running)
3237 * Calculate the group which has the least non-idle load.
3238 * This is the group from where we need to pick up the load
3241 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3242 (sgs->sum_nr_running == sds->min_nr_running &&
3243 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3244 sds->group_min = group;
3245 sds->min_nr_running = sgs->sum_nr_running;
3246 sds->min_load_per_task = sgs->sum_weighted_load /
3247 sgs->sum_nr_running;
3251 * Calculate the group which is almost near its
3252 * capacity but still has some space to pick up some load
3253 * from other group and save more power
3255 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3258 if (sgs->sum_nr_running > sds->leader_nr_running ||
3259 (sgs->sum_nr_running == sds->leader_nr_running &&
3260 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3261 sds->group_leader = group;
3262 sds->leader_nr_running = sgs->sum_nr_running;
3267 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3268 * @sds: Variable containing the statistics of the sched_domain
3269 * under consideration.
3270 * @this_cpu: Cpu at which we're currently performing load-balancing.
3271 * @imbalance: Variable to store the imbalance.
3274 * Check if we have potential to perform some power-savings balance.
3275 * If yes, set the busiest group to be the least loaded group in the
3276 * sched_domain, so that it's CPUs can be put to idle.
3278 * Returns 1 if there is potential to perform power-savings balance.
3281 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3282 int this_cpu, unsigned long *imbalance)
3284 if (!sds->power_savings_balance)
3287 if (sds->this != sds->group_leader ||
3288 sds->group_leader == sds->group_min)
3291 *imbalance = sds->min_load_per_task;
3292 sds->busiest = sds->group_min;
3297 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3298 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3299 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3304 static inline void update_sd_power_savings_stats(struct sched_group *group,
3305 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3310 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3311 int this_cpu, unsigned long *imbalance)
3315 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3318 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3320 return SCHED_POWER_SCALE;
3323 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3325 return default_scale_freq_power(sd, cpu);
3328 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3330 unsigned long weight = sd->span_weight;
3331 unsigned long smt_gain = sd->smt_gain;
3338 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3340 return default_scale_smt_power(sd, cpu);
3343 unsigned long scale_rt_power(int cpu)
3345 struct rq *rq = cpu_rq(cpu);
3346 u64 total, available;
3348 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3350 if (unlikely(total < rq->rt_avg)) {
3351 /* Ensures that power won't end up being negative */
3354 available = total - rq->rt_avg;
3357 if (unlikely((s64)total < SCHED_POWER_SCALE))
3358 total = SCHED_POWER_SCALE;
3360 total >>= SCHED_POWER_SHIFT;
3362 return div_u64(available, total);
3365 static void update_cpu_power(struct sched_domain *sd, int cpu)
3367 unsigned long weight = sd->span_weight;
3368 unsigned long power = SCHED_POWER_SCALE;
3369 struct sched_group *sdg = sd->groups;
3371 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3372 if (sched_feat(ARCH_POWER))
3373 power *= arch_scale_smt_power(sd, cpu);
3375 power *= default_scale_smt_power(sd, cpu);
3377 power >>= SCHED_POWER_SHIFT;
3380 sdg->sgp->power_orig = power;
3382 if (sched_feat(ARCH_POWER))
3383 power *= arch_scale_freq_power(sd, cpu);
3385 power *= default_scale_freq_power(sd, cpu);
3387 power >>= SCHED_POWER_SHIFT;
3389 power *= scale_rt_power(cpu);
3390 power >>= SCHED_POWER_SHIFT;
3395 cpu_rq(cpu)->cpu_power = power;
3396 sdg->sgp->power = power;
3399 static void update_group_power(struct sched_domain *sd, int cpu)
3401 struct sched_domain *child = sd->child;
3402 struct sched_group *group, *sdg = sd->groups;
3403 unsigned long power;
3406 update_cpu_power(sd, cpu);
3412 group = child->groups;
3414 power += group->sgp->power;
3415 group = group->next;
3416 } while (group != child->groups);
3418 sdg->sgp->power = power;
3422 * Try and fix up capacity for tiny siblings, this is needed when
3423 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3424 * which on its own isn't powerful enough.
3426 * See update_sd_pick_busiest() and check_asym_packing().
3429 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3432 * Only siblings can have significantly less than SCHED_POWER_SCALE
3434 if (!(sd->flags & SD_SHARE_CPUPOWER))
3438 * If ~90% of the cpu_power is still there, we're good.
3440 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3447 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3448 * @sd: The sched_domain whose statistics are to be updated.
3449 * @group: sched_group whose statistics are to be updated.
3450 * @this_cpu: Cpu for which load balance is currently performed.
3451 * @idle: Idle status of this_cpu
3452 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3453 * @local_group: Does group contain this_cpu.
3454 * @cpus: Set of cpus considered for load balancing.
3455 * @balance: Should we balance.
3456 * @sgs: variable to hold the statistics for this group.
3458 static inline void update_sg_lb_stats(struct sched_domain *sd,
3459 struct sched_group *group, int this_cpu,
3460 enum cpu_idle_type idle, int load_idx,
3461 int local_group, const struct cpumask *cpus,
3462 int *balance, struct sg_lb_stats *sgs)
3464 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3466 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3467 unsigned long avg_load_per_task = 0;
3470 balance_cpu = group_first_cpu(group);
3472 /* Tally up the load of all CPUs in the group */
3474 min_cpu_load = ~0UL;
3477 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3478 struct rq *rq = cpu_rq(i);
3480 /* Bias balancing toward cpus of our domain */
3482 if (idle_cpu(i) && !first_idle_cpu) {
3487 load = target_load(i, load_idx);
3489 load = source_load(i, load_idx);
3490 if (load > max_cpu_load) {
3491 max_cpu_load = load;
3492 max_nr_running = rq->nr_running;
3494 if (min_cpu_load > load)
3495 min_cpu_load = load;
3498 sgs->group_load += load;
3499 sgs->sum_nr_running += rq->nr_running;
3500 sgs->sum_weighted_load += weighted_cpuload(i);
3506 * First idle cpu or the first cpu(busiest) in this sched group
3507 * is eligible for doing load balancing at this and above
3508 * domains. In the newly idle case, we will allow all the cpu's
3509 * to do the newly idle load balance.
3511 if (idle != CPU_NEWLY_IDLE && local_group) {
3512 if (balance_cpu != this_cpu) {
3516 update_group_power(sd, this_cpu);
3519 /* Adjust by relative CPU power of the group */
3520 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3523 * Consider the group unbalanced when the imbalance is larger
3524 * than the average weight of a task.
3526 * APZ: with cgroup the avg task weight can vary wildly and
3527 * might not be a suitable number - should we keep a
3528 * normalized nr_running number somewhere that negates
3531 if (sgs->sum_nr_running)
3532 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3534 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3537 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3539 if (!sgs->group_capacity)
3540 sgs->group_capacity = fix_small_capacity(sd, group);
3541 sgs->group_weight = group->group_weight;
3543 if (sgs->group_capacity > sgs->sum_nr_running)
3544 sgs->group_has_capacity = 1;
3548 * update_sd_pick_busiest - return 1 on busiest group
3549 * @sd: sched_domain whose statistics are to be checked
3550 * @sds: sched_domain statistics
3551 * @sg: sched_group candidate to be checked for being the busiest
3552 * @sgs: sched_group statistics
3553 * @this_cpu: the current cpu
3555 * Determine if @sg is a busier group than the previously selected
3558 static bool update_sd_pick_busiest(struct sched_domain *sd,
3559 struct sd_lb_stats *sds,
3560 struct sched_group *sg,
3561 struct sg_lb_stats *sgs,
3564 if (sgs->avg_load <= sds->max_load)
3567 if (sgs->sum_nr_running > sgs->group_capacity)
3574 * ASYM_PACKING needs to move all the work to the lowest
3575 * numbered CPUs in the group, therefore mark all groups
3576 * higher than ourself as busy.
3578 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3579 this_cpu < group_first_cpu(sg)) {
3583 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3591 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3592 * @sd: sched_domain whose statistics are to be updated.
3593 * @this_cpu: Cpu for which load balance is currently performed.
3594 * @idle: Idle status of this_cpu
3595 * @cpus: Set of cpus considered for load balancing.
3596 * @balance: Should we balance.
3597 * @sds: variable to hold the statistics for this sched_domain.
3599 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3600 enum cpu_idle_type idle, const struct cpumask *cpus,
3601 int *balance, struct sd_lb_stats *sds)
3603 struct sched_domain *child = sd->child;
3604 struct sched_group *sg = sd->groups;
3605 struct sg_lb_stats sgs;
3606 int load_idx, prefer_sibling = 0;
3608 if (child && child->flags & SD_PREFER_SIBLING)
3611 init_sd_power_savings_stats(sd, sds, idle);
3612 load_idx = get_sd_load_idx(sd, idle);
3617 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3618 memset(&sgs, 0, sizeof(sgs));
3619 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3620 local_group, cpus, balance, &sgs);
3622 if (local_group && !(*balance))
3625 sds->total_load += sgs.group_load;
3626 sds->total_pwr += sg->sgp->power;
3629 * In case the child domain prefers tasks go to siblings
3630 * first, lower the sg capacity to one so that we'll try
3631 * and move all the excess tasks away. We lower the capacity
3632 * of a group only if the local group has the capacity to fit
3633 * these excess tasks, i.e. nr_running < group_capacity. The
3634 * extra check prevents the case where you always pull from the
3635 * heaviest group when it is already under-utilized (possible
3636 * with a large weight task outweighs the tasks on the system).
3638 if (prefer_sibling && !local_group && sds->this_has_capacity)
3639 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3642 sds->this_load = sgs.avg_load;
3644 sds->this_nr_running = sgs.sum_nr_running;
3645 sds->this_load_per_task = sgs.sum_weighted_load;
3646 sds->this_has_capacity = sgs.group_has_capacity;
3647 sds->this_idle_cpus = sgs.idle_cpus;
3648 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
3649 sds->max_load = sgs.avg_load;
3651 sds->busiest_nr_running = sgs.sum_nr_running;
3652 sds->busiest_idle_cpus = sgs.idle_cpus;
3653 sds->busiest_group_capacity = sgs.group_capacity;
3654 sds->busiest_load_per_task = sgs.sum_weighted_load;
3655 sds->busiest_has_capacity = sgs.group_has_capacity;
3656 sds->busiest_group_weight = sgs.group_weight;
3657 sds->group_imb = sgs.group_imb;
3660 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
3662 } while (sg != sd->groups);
3665 int __weak arch_sd_sibling_asym_packing(void)
3667 return 0*SD_ASYM_PACKING;
3671 * check_asym_packing - Check to see if the group is packed into the
3674 * This is primarily intended to used at the sibling level. Some
3675 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3676 * case of POWER7, it can move to lower SMT modes only when higher
3677 * threads are idle. When in lower SMT modes, the threads will
3678 * perform better since they share less core resources. Hence when we
3679 * have idle threads, we want them to be the higher ones.
3681 * This packing function is run on idle threads. It checks to see if
3682 * the busiest CPU in this domain (core in the P7 case) has a higher
3683 * CPU number than the packing function is being run on. Here we are
3684 * assuming lower CPU number will be equivalent to lower a SMT thread
3687 * Returns 1 when packing is required and a task should be moved to
3688 * this CPU. The amount of the imbalance is returned in *imbalance.
3690 * @sd: The sched_domain whose packing is to be checked.
3691 * @sds: Statistics of the sched_domain which is to be packed
3692 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3693 * @imbalance: returns amount of imbalanced due to packing.
3695 static int check_asym_packing(struct sched_domain *sd,
3696 struct sd_lb_stats *sds,
3697 int this_cpu, unsigned long *imbalance)
3701 if (!(sd->flags & SD_ASYM_PACKING))
3707 busiest_cpu = group_first_cpu(sds->busiest);
3708 if (this_cpu > busiest_cpu)
3711 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
3717 * fix_small_imbalance - Calculate the minor imbalance that exists
3718 * amongst the groups of a sched_domain, during
3720 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3721 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3722 * @imbalance: Variable to store the imbalance.
3724 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3725 int this_cpu, unsigned long *imbalance)
3727 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3728 unsigned int imbn = 2;
3729 unsigned long scaled_busy_load_per_task;
3731 if (sds->this_nr_running) {
3732 sds->this_load_per_task /= sds->this_nr_running;
3733 if (sds->busiest_load_per_task >
3734 sds->this_load_per_task)
3737 sds->this_load_per_task =
3738 cpu_avg_load_per_task(this_cpu);
3740 scaled_busy_load_per_task = sds->busiest_load_per_task
3741 * SCHED_POWER_SCALE;
3742 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3744 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3745 (scaled_busy_load_per_task * imbn)) {
3746 *imbalance = sds->busiest_load_per_task;
3751 * OK, we don't have enough imbalance to justify moving tasks,
3752 * however we may be able to increase total CPU power used by
3756 pwr_now += sds->busiest->sgp->power *
3757 min(sds->busiest_load_per_task, sds->max_load);
3758 pwr_now += sds->this->sgp->power *
3759 min(sds->this_load_per_task, sds->this_load);
3760 pwr_now /= SCHED_POWER_SCALE;
3762 /* Amount of load we'd subtract */
3763 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3764 sds->busiest->sgp->power;
3765 if (sds->max_load > tmp)
3766 pwr_move += sds->busiest->sgp->power *
3767 min(sds->busiest_load_per_task, sds->max_load - tmp);
3769 /* Amount of load we'd add */
3770 if (sds->max_load * sds->busiest->sgp->power <
3771 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3772 tmp = (sds->max_load * sds->busiest->sgp->power) /
3773 sds->this->sgp->power;
3775 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3776 sds->this->sgp->power;
3777 pwr_move += sds->this->sgp->power *
3778 min(sds->this_load_per_task, sds->this_load + tmp);
3779 pwr_move /= SCHED_POWER_SCALE;
3781 /* Move if we gain throughput */
3782 if (pwr_move > pwr_now)
3783 *imbalance = sds->busiest_load_per_task;
3787 * calculate_imbalance - Calculate the amount of imbalance present within the
3788 * groups of a given sched_domain during load balance.
3789 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3790 * @this_cpu: Cpu for which currently load balance is being performed.
3791 * @imbalance: The variable to store the imbalance.
3793 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3794 unsigned long *imbalance)
3796 unsigned long max_pull, load_above_capacity = ~0UL;
3798 sds->busiest_load_per_task /= sds->busiest_nr_running;
3799 if (sds->group_imb) {
3800 sds->busiest_load_per_task =
3801 min(sds->busiest_load_per_task, sds->avg_load);
3805 * In the presence of smp nice balancing, certain scenarios can have
3806 * max load less than avg load(as we skip the groups at or below
3807 * its cpu_power, while calculating max_load..)
3809 if (sds->max_load < sds->avg_load) {
3811 return fix_small_imbalance(sds, this_cpu, imbalance);
3814 if (!sds->group_imb) {
3816 * Don't want to pull so many tasks that a group would go idle.
3818 load_above_capacity = (sds->busiest_nr_running -
3819 sds->busiest_group_capacity);
3821 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
3823 load_above_capacity /= sds->busiest->sgp->power;
3827 * We're trying to get all the cpus to the average_load, so we don't
3828 * want to push ourselves above the average load, nor do we wish to
3829 * reduce the max loaded cpu below the average load. At the same time,
3830 * we also don't want to reduce the group load below the group capacity
3831 * (so that we can implement power-savings policies etc). Thus we look
3832 * for the minimum possible imbalance.
3833 * Be careful of negative numbers as they'll appear as very large values
3834 * with unsigned longs.
3836 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
3838 /* How much load to actually move to equalise the imbalance */
3839 *imbalance = min(max_pull * sds->busiest->sgp->power,
3840 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
3841 / SCHED_POWER_SCALE;
3844 * if *imbalance is less than the average load per runnable task
3845 * there is no guarantee that any tasks will be moved so we'll have
3846 * a think about bumping its value to force at least one task to be
3849 if (*imbalance < sds->busiest_load_per_task)
3850 return fix_small_imbalance(sds, this_cpu, imbalance);
3854 /******* find_busiest_group() helpers end here *********************/
3857 * find_busiest_group - Returns the busiest group within the sched_domain
3858 * if there is an imbalance. If there isn't an imbalance, and
3859 * the user has opted for power-savings, it returns a group whose
3860 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3861 * such a group exists.
3863 * Also calculates the amount of weighted load which should be moved
3864 * to restore balance.
3866 * @sd: The sched_domain whose busiest group is to be returned.
3867 * @this_cpu: The cpu for which load balancing is currently being performed.
3868 * @imbalance: Variable which stores amount of weighted load which should
3869 * be moved to restore balance/put a group to idle.
3870 * @idle: The idle status of this_cpu.
3871 * @cpus: The set of CPUs under consideration for load-balancing.
3872 * @balance: Pointer to a variable indicating if this_cpu
3873 * is the appropriate cpu to perform load balancing at this_level.
3875 * Returns: - the busiest group if imbalance exists.
3876 * - If no imbalance and user has opted for power-savings balance,
3877 * return the least loaded group whose CPUs can be
3878 * put to idle by rebalancing its tasks onto our group.
3880 static struct sched_group *
3881 find_busiest_group(struct sched_domain *sd, int this_cpu,
3882 unsigned long *imbalance, enum cpu_idle_type idle,
3883 const struct cpumask *cpus, int *balance)
3885 struct sd_lb_stats sds;
3887 memset(&sds, 0, sizeof(sds));
3890 * Compute the various statistics relavent for load balancing at
3893 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
3896 * this_cpu is not the appropriate cpu to perform load balancing at
3902 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3903 check_asym_packing(sd, &sds, this_cpu, imbalance))
3906 /* There is no busy sibling group to pull tasks from */
3907 if (!sds.busiest || sds.busiest_nr_running == 0)
3910 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
3913 * If the busiest group is imbalanced the below checks don't
3914 * work because they assumes all things are equal, which typically
3915 * isn't true due to cpus_allowed constraints and the like.
3920 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3921 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3922 !sds.busiest_has_capacity)
3926 * If the local group is more busy than the selected busiest group
3927 * don't try and pull any tasks.
3929 if (sds.this_load >= sds.max_load)
3933 * Don't pull any tasks if this group is already above the domain
3936 if (sds.this_load >= sds.avg_load)
3939 if (idle == CPU_IDLE) {
3941 * This cpu is idle. If the busiest group load doesn't
3942 * have more tasks than the number of available cpu's and
3943 * there is no imbalance between this and busiest group
3944 * wrt to idle cpu's, it is balanced.
3946 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3947 sds.busiest_nr_running <= sds.busiest_group_weight)
3951 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
3952 * imbalance_pct to be conservative.
3954 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3959 /* Looks like there is an imbalance. Compute it */
3960 calculate_imbalance(&sds, this_cpu, imbalance);
3965 * There is no obvious imbalance. But check if we can do some balancing
3968 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3976 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3979 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3980 enum cpu_idle_type idle, unsigned long imbalance,
3981 const struct cpumask *cpus)
3983 struct rq *busiest = NULL, *rq;
3984 unsigned long max_load = 0;
3987 for_each_cpu(i, sched_group_cpus(group)) {
3988 unsigned long power = power_of(i);
3989 unsigned long capacity = DIV_ROUND_CLOSEST(power,
3994 capacity = fix_small_capacity(sd, group);
3996 if (!cpumask_test_cpu(i, cpus))
4000 wl = weighted_cpuload(i);
4003 * When comparing with imbalance, use weighted_cpuload()
4004 * which is not scaled with the cpu power.
4006 if (capacity && rq->nr_running == 1 && wl > imbalance)
4010 * For the load comparisons with the other cpu's, consider
4011 * the weighted_cpuload() scaled with the cpu power, so that
4012 * the load can be moved away from the cpu that is potentially
4013 * running at a lower capacity.
4015 wl = (wl * SCHED_POWER_SCALE) / power;
4017 if (wl > max_load) {
4027 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4028 * so long as it is large enough.
4030 #define MAX_PINNED_INTERVAL 512
4032 /* Working cpumask for load_balance and load_balance_newidle. */
4033 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4035 static int need_active_balance(struct sched_domain *sd, int idle,
4036 int busiest_cpu, int this_cpu)
4038 if (idle == CPU_NEWLY_IDLE) {
4041 * ASYM_PACKING needs to force migrate tasks from busy but
4042 * higher numbered CPUs in order to pack all tasks in the
4043 * lowest numbered CPUs.
4045 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
4049 * The only task running in a non-idle cpu can be moved to this
4050 * cpu in an attempt to completely freeup the other CPU
4053 * The package power saving logic comes from
4054 * find_busiest_group(). If there are no imbalance, then
4055 * f_b_g() will return NULL. However when sched_mc={1,2} then
4056 * f_b_g() will select a group from which a running task may be
4057 * pulled to this cpu in order to make the other package idle.
4058 * If there is no opportunity to make a package idle and if
4059 * there are no imbalance, then f_b_g() will return NULL and no
4060 * action will be taken in load_balance_newidle().
4062 * Under normal task pull operation due to imbalance, there
4063 * will be more than one task in the source run queue and
4064 * move_tasks() will succeed. ld_moved will be true and this
4065 * active balance code will not be triggered.
4067 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4071 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4074 static int active_load_balance_cpu_stop(void *data);
4077 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4078 * tasks if there is an imbalance.
4080 static int load_balance(int this_cpu, struct rq *this_rq,
4081 struct sched_domain *sd, enum cpu_idle_type idle,
4084 int ld_moved, all_pinned = 0, active_balance = 0;
4085 struct sched_group *group;
4086 unsigned long imbalance;
4088 unsigned long flags;
4089 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4091 cpumask_copy(cpus, cpu_active_mask);
4093 schedstat_inc(sd, lb_count[idle]);
4096 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
4103 schedstat_inc(sd, lb_nobusyg[idle]);
4107 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
4109 schedstat_inc(sd, lb_nobusyq[idle]);
4113 BUG_ON(busiest == this_rq);
4115 schedstat_add(sd, lb_imbalance[idle], imbalance);
4118 if (busiest->nr_running > 1) {
4120 * Attempt to move tasks. If find_busiest_group has found
4121 * an imbalance but busiest->nr_running <= 1, the group is
4122 * still unbalanced. ld_moved simply stays zero, so it is
4123 * correctly treated as an imbalance.
4126 local_irq_save(flags);
4127 double_rq_lock(this_rq, busiest);
4128 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4129 imbalance, sd, idle, &all_pinned);
4130 double_rq_unlock(this_rq, busiest);
4131 local_irq_restore(flags);
4134 * some other cpu did the load balance for us.
4136 if (ld_moved && this_cpu != smp_processor_id())
4137 resched_cpu(this_cpu);
4139 /* All tasks on this runqueue were pinned by CPU affinity */
4140 if (unlikely(all_pinned)) {
4141 cpumask_clear_cpu(cpu_of(busiest), cpus);
4142 if (!cpumask_empty(cpus))
4149 schedstat_inc(sd, lb_failed[idle]);
4151 * Increment the failure counter only on periodic balance.
4152 * We do not want newidle balance, which can be very
4153 * frequent, pollute the failure counter causing
4154 * excessive cache_hot migrations and active balances.
4156 if (idle != CPU_NEWLY_IDLE)
4157 sd->nr_balance_failed++;
4159 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
4160 raw_spin_lock_irqsave(&busiest->lock, flags);
4162 /* don't kick the active_load_balance_cpu_stop,
4163 * if the curr task on busiest cpu can't be
4166 if (!cpumask_test_cpu(this_cpu,
4167 tsk_cpus_allowed(busiest->curr))) {
4168 raw_spin_unlock_irqrestore(&busiest->lock,
4171 goto out_one_pinned;
4175 * ->active_balance synchronizes accesses to
4176 * ->active_balance_work. Once set, it's cleared
4177 * only after active load balance is finished.
4179 if (!busiest->active_balance) {
4180 busiest->active_balance = 1;
4181 busiest->push_cpu = this_cpu;
4184 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4187 stop_one_cpu_nowait(cpu_of(busiest),
4188 active_load_balance_cpu_stop, busiest,
4189 &busiest->active_balance_work);
4192 * We've kicked active balancing, reset the failure
4195 sd->nr_balance_failed = sd->cache_nice_tries+1;
4198 sd->nr_balance_failed = 0;
4200 if (likely(!active_balance)) {
4201 /* We were unbalanced, so reset the balancing interval */
4202 sd->balance_interval = sd->min_interval;
4205 * If we've begun active balancing, start to back off. This
4206 * case may not be covered by the all_pinned logic if there
4207 * is only 1 task on the busy runqueue (because we don't call
4210 if (sd->balance_interval < sd->max_interval)
4211 sd->balance_interval *= 2;
4217 schedstat_inc(sd, lb_balanced[idle]);
4219 sd->nr_balance_failed = 0;
4222 /* tune up the balancing interval */
4223 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4224 (sd->balance_interval < sd->max_interval))
4225 sd->balance_interval *= 2;
4233 * idle_balance is called by schedule() if this_cpu is about to become
4234 * idle. Attempts to pull tasks from other CPUs.
4236 static void idle_balance(int this_cpu, struct rq *this_rq)
4238 struct sched_domain *sd;
4239 int pulled_task = 0;
4240 unsigned long next_balance = jiffies + HZ;
4242 this_rq->idle_stamp = this_rq->clock;
4244 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4248 * Drop the rq->lock, but keep IRQ/preempt disabled.
4250 raw_spin_unlock(&this_rq->lock);
4252 update_shares(this_cpu);
4254 for_each_domain(this_cpu, sd) {
4255 unsigned long interval;
4258 if (!(sd->flags & SD_LOAD_BALANCE))
4261 if (sd->flags & SD_BALANCE_NEWIDLE) {
4262 /* If we've pulled tasks over stop searching: */
4263 pulled_task = load_balance(this_cpu, this_rq,
4264 sd, CPU_NEWLY_IDLE, &balance);
4267 interval = msecs_to_jiffies(sd->balance_interval);
4268 if (time_after(next_balance, sd->last_balance + interval))
4269 next_balance = sd->last_balance + interval;
4271 this_rq->idle_stamp = 0;
4277 raw_spin_lock(&this_rq->lock);
4279 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4281 * We are going idle. next_balance may be set based on
4282 * a busy processor. So reset next_balance.
4284 this_rq->next_balance = next_balance;
4289 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4290 * running tasks off the busiest CPU onto idle CPUs. It requires at
4291 * least 1 task to be running on each physical CPU where possible, and
4292 * avoids physical / logical imbalances.
4294 static int active_load_balance_cpu_stop(void *data)
4296 struct rq *busiest_rq = data;
4297 int busiest_cpu = cpu_of(busiest_rq);
4298 int target_cpu = busiest_rq->push_cpu;
4299 struct rq *target_rq = cpu_rq(target_cpu);
4300 struct sched_domain *sd;
4302 raw_spin_lock_irq(&busiest_rq->lock);
4304 /* make sure the requested cpu hasn't gone down in the meantime */
4305 if (unlikely(busiest_cpu != smp_processor_id() ||
4306 !busiest_rq->active_balance))
4309 /* Is there any task to move? */
4310 if (busiest_rq->nr_running <= 1)
4314 * This condition is "impossible", if it occurs
4315 * we need to fix it. Originally reported by
4316 * Bjorn Helgaas on a 128-cpu setup.
4318 BUG_ON(busiest_rq == target_rq);
4320 /* move a task from busiest_rq to target_rq */
4321 double_lock_balance(busiest_rq, target_rq);
4323 /* Search for an sd spanning us and the target CPU. */
4325 for_each_domain(target_cpu, sd) {
4326 if ((sd->flags & SD_LOAD_BALANCE) &&
4327 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4332 schedstat_inc(sd, alb_count);
4334 if (move_one_task(target_rq, target_cpu, busiest_rq,
4336 schedstat_inc(sd, alb_pushed);
4338 schedstat_inc(sd, alb_failed);
4341 double_unlock_balance(busiest_rq, target_rq);
4343 busiest_rq->active_balance = 0;
4344 raw_spin_unlock_irq(&busiest_rq->lock);
4350 * idle load balancing details
4351 * - One of the idle CPUs nominates itself as idle load_balancer, while
4353 * - This idle load balancer CPU will also go into tickless mode when
4354 * it is idle, just like all other idle CPUs
4355 * - When one of the busy CPUs notice that there may be an idle rebalancing
4356 * needed, they will kick the idle load balancer, which then does idle
4357 * load balancing for all the idle CPUs.
4360 atomic_t load_balancer;
4361 atomic_t first_pick_cpu;
4362 atomic_t second_pick_cpu;
4363 cpumask_var_t idle_cpus_mask;
4364 cpumask_var_t grp_idle_mask;
4365 unsigned long next_balance; /* in jiffy units */
4366 } nohz ____cacheline_aligned;
4368 int get_nohz_load_balancer(void)
4370 return atomic_read(&nohz.load_balancer);
4373 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4375 * lowest_flag_domain - Return lowest sched_domain containing flag.
4376 * @cpu: The cpu whose lowest level of sched domain is to
4378 * @flag: The flag to check for the lowest sched_domain
4379 * for the given cpu.
4381 * Returns the lowest sched_domain of a cpu which contains the given flag.
4383 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4385 struct sched_domain *sd;
4387 for_each_domain(cpu, sd)
4388 if (sd->flags & flag)
4395 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4396 * @cpu: The cpu whose domains we're iterating over.
4397 * @sd: variable holding the value of the power_savings_sd
4399 * @flag: The flag to filter the sched_domains to be iterated.
4401 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4402 * set, starting from the lowest sched_domain to the highest.
4404 #define for_each_flag_domain(cpu, sd, flag) \
4405 for (sd = lowest_flag_domain(cpu, flag); \
4406 (sd && (sd->flags & flag)); sd = sd->parent)
4409 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4410 * @ilb_group: group to be checked for semi-idleness
4412 * Returns: 1 if the group is semi-idle. 0 otherwise.
4414 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4415 * and atleast one non-idle CPU. This helper function checks if the given
4416 * sched_group is semi-idle or not.
4418 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4420 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
4421 sched_group_cpus(ilb_group));
4424 * A sched_group is semi-idle when it has atleast one busy cpu
4425 * and atleast one idle cpu.
4427 if (cpumask_empty(nohz.grp_idle_mask))
4430 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
4436 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4437 * @cpu: The cpu which is nominating a new idle_load_balancer.
4439 * Returns: Returns the id of the idle load balancer if it exists,
4440 * Else, returns >= nr_cpu_ids.
4442 * This algorithm picks the idle load balancer such that it belongs to a
4443 * semi-idle powersavings sched_domain. The idea is to try and avoid
4444 * completely idle packages/cores just for the purpose of idle load balancing
4445 * when there are other idle cpu's which are better suited for that job.
4447 static int find_new_ilb(int cpu)
4449 struct sched_domain *sd;
4450 struct sched_group *ilb_group;
4451 int ilb = nr_cpu_ids;
4454 * Have idle load balancer selection from semi-idle packages only
4455 * when power-aware load balancing is enabled
4457 if (!(sched_smt_power_savings || sched_mc_power_savings))
4461 * Optimize for the case when we have no idle CPUs or only one
4462 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4464 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4468 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4469 ilb_group = sd->groups;
4472 if (is_semi_idle_group(ilb_group)) {
4473 ilb = cpumask_first(nohz.grp_idle_mask);
4477 ilb_group = ilb_group->next;
4479 } while (ilb_group != sd->groups);
4487 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4488 static inline int find_new_ilb(int call_cpu)
4495 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4496 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4497 * CPU (if there is one).
4499 static void nohz_balancer_kick(int cpu)
4503 nohz.next_balance++;
4505 ilb_cpu = get_nohz_load_balancer();
4507 if (ilb_cpu >= nr_cpu_ids) {
4508 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4509 if (ilb_cpu >= nr_cpu_ids)
4513 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4514 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4518 * Use smp_send_reschedule() instead of resched_cpu().
4519 * This way we generate a sched IPI on the target cpu which
4520 * is idle. And the softirq performing nohz idle load balance
4521 * will be run before returning from the IPI.
4523 smp_send_reschedule(ilb_cpu);
4529 * This routine will try to nominate the ilb (idle load balancing)
4530 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4531 * load balancing on behalf of all those cpus.
4533 * When the ilb owner becomes busy, we will not have new ilb owner until some
4534 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4535 * idle load balancing by kicking one of the idle CPUs.
4537 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4538 * ilb owner CPU in future (when there is a need for idle load balancing on
4539 * behalf of all idle CPUs).
4541 void select_nohz_load_balancer(int stop_tick)
4543 int cpu = smp_processor_id();
4546 if (!cpu_active(cpu)) {
4547 if (atomic_read(&nohz.load_balancer) != cpu)
4551 * If we are going offline and still the leader,
4554 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4561 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4563 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4564 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4565 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4566 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4568 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4571 /* make me the ilb owner */
4572 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4577 * Check to see if there is a more power-efficient
4580 new_ilb = find_new_ilb(cpu);
4581 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4582 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4583 resched_cpu(new_ilb);
4589 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4592 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4594 if (atomic_read(&nohz.load_balancer) == cpu)
4595 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4603 static DEFINE_SPINLOCK(balancing);
4605 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4608 * Scale the max load_balance interval with the number of CPUs in the system.
4609 * This trades load-balance latency on larger machines for less cross talk.
4611 static void update_max_interval(void)
4613 max_load_balance_interval = HZ*num_online_cpus()/10;
4617 * It checks each scheduling domain to see if it is due to be balanced,
4618 * and initiates a balancing operation if so.
4620 * Balancing parameters are set up in arch_init_sched_domains.
4622 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4625 struct rq *rq = cpu_rq(cpu);
4626 unsigned long interval;
4627 struct sched_domain *sd;
4628 /* Earliest time when we have to do rebalance again */
4629 unsigned long next_balance = jiffies + 60*HZ;
4630 int update_next_balance = 0;
4636 for_each_domain(cpu, sd) {
4637 if (!(sd->flags & SD_LOAD_BALANCE))
4640 interval = sd->balance_interval;
4641 if (idle != CPU_IDLE)
4642 interval *= sd->busy_factor;
4644 /* scale ms to jiffies */
4645 interval = msecs_to_jiffies(interval);
4646 interval = clamp(interval, 1UL, max_load_balance_interval);
4648 need_serialize = sd->flags & SD_SERIALIZE;
4650 if (need_serialize) {
4651 if (!spin_trylock(&balancing))
4655 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4656 if (load_balance(cpu, rq, sd, idle, &balance)) {
4658 * We've pulled tasks over so either we're no
4661 idle = CPU_NOT_IDLE;
4663 sd->last_balance = jiffies;
4666 spin_unlock(&balancing);
4668 if (time_after(next_balance, sd->last_balance + interval)) {
4669 next_balance = sd->last_balance + interval;
4670 update_next_balance = 1;
4674 * Stop the load balance at this level. There is another
4675 * CPU in our sched group which is doing load balancing more
4684 * next_balance will be updated only when there is a need.
4685 * When the cpu is attached to null domain for ex, it will not be
4688 if (likely(update_next_balance))
4689 rq->next_balance = next_balance;
4694 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4695 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4697 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4699 struct rq *this_rq = cpu_rq(this_cpu);
4703 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
4706 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4707 if (balance_cpu == this_cpu)
4711 * If this cpu gets work to do, stop the load balancing
4712 * work being done for other cpus. Next load
4713 * balancing owner will pick it up.
4715 if (need_resched()) {
4716 this_rq->nohz_balance_kick = 0;
4720 raw_spin_lock_irq(&this_rq->lock);
4721 update_rq_clock(this_rq);
4722 update_cpu_load(this_rq);
4723 raw_spin_unlock_irq(&this_rq->lock);
4725 rebalance_domains(balance_cpu, CPU_IDLE);
4727 rq = cpu_rq(balance_cpu);
4728 if (time_after(this_rq->next_balance, rq->next_balance))
4729 this_rq->next_balance = rq->next_balance;
4731 nohz.next_balance = this_rq->next_balance;
4732 this_rq->nohz_balance_kick = 0;
4736 * Current heuristic for kicking the idle load balancer
4737 * - first_pick_cpu is the one of the busy CPUs. It will kick
4738 * idle load balancer when it has more than one process active. This
4739 * eliminates the need for idle load balancing altogether when we have
4740 * only one running process in the system (common case).
4741 * - If there are more than one busy CPU, idle load balancer may have
4742 * to run for active_load_balance to happen (i.e., two busy CPUs are
4743 * SMT or core siblings and can run better if they move to different
4744 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
4745 * which will kick idle load balancer as soon as it has any load.
4747 static inline int nohz_kick_needed(struct rq *rq, int cpu)
4749 unsigned long now = jiffies;
4751 int first_pick_cpu, second_pick_cpu;
4753 if (time_before(now, nohz.next_balance))
4759 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
4760 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
4762 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
4763 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
4766 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
4767 if (ret == nr_cpu_ids || ret == cpu) {
4768 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4769 if (rq->nr_running > 1)
4772 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
4773 if (ret == nr_cpu_ids || ret == cpu) {
4781 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4785 * run_rebalance_domains is triggered when needed from the scheduler tick.
4786 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4788 static void run_rebalance_domains(struct softirq_action *h)
4790 int this_cpu = smp_processor_id();
4791 struct rq *this_rq = cpu_rq(this_cpu);
4792 enum cpu_idle_type idle = this_rq->idle_balance ?
4793 CPU_IDLE : CPU_NOT_IDLE;
4795 rebalance_domains(this_cpu, idle);
4798 * If this cpu has a pending nohz_balance_kick, then do the
4799 * balancing on behalf of the other idle cpus whose ticks are
4802 nohz_idle_balance(this_cpu, idle);
4805 static inline int on_null_domain(int cpu)
4807 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4811 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4813 static inline void trigger_load_balance(struct rq *rq, int cpu)
4815 /* Don't need to rebalance while attached to NULL domain */
4816 if (time_after_eq(jiffies, rq->next_balance) &&
4817 likely(!on_null_domain(cpu)))
4818 raise_softirq(SCHED_SOFTIRQ);
4820 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4821 nohz_balancer_kick(cpu);
4825 static void rq_online_fair(struct rq *rq)
4830 static void rq_offline_fair(struct rq *rq)
4835 #else /* CONFIG_SMP */
4838 * on UP we do not need to balance between CPUs:
4840 static inline void idle_balance(int cpu, struct rq *rq)
4844 #endif /* CONFIG_SMP */
4847 * scheduler tick hitting a task of our scheduling class:
4849 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4851 struct cfs_rq *cfs_rq;
4852 struct sched_entity *se = &curr->se;
4854 for_each_sched_entity(se) {
4855 cfs_rq = cfs_rq_of(se);
4856 entity_tick(cfs_rq, se, queued);
4861 * called on fork with the child task as argument from the parent's context
4862 * - child not yet on the tasklist
4863 * - preemption disabled
4865 static void task_fork_fair(struct task_struct *p)
4867 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4868 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4869 int this_cpu = smp_processor_id();
4870 struct rq *rq = this_rq();
4871 unsigned long flags;
4873 raw_spin_lock_irqsave(&rq->lock, flags);
4875 update_rq_clock(rq);
4877 if (unlikely(task_cpu(p) != this_cpu)) {
4879 __set_task_cpu(p, this_cpu);
4883 update_curr(cfs_rq);
4886 se->vruntime = curr->vruntime;
4887 place_entity(cfs_rq, se, 1);
4889 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4891 * Upon rescheduling, sched_class::put_prev_task() will place
4892 * 'current' within the tree based on its new key value.
4894 swap(curr->vruntime, se->vruntime);
4895 resched_task(rq->curr);
4898 se->vruntime -= cfs_rq->min_vruntime;
4900 raw_spin_unlock_irqrestore(&rq->lock, flags);
4904 * Priority of the task has changed. Check to see if we preempt
4908 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4914 * Reschedule if we are currently running on this runqueue and
4915 * our priority decreased, or if we are not currently running on
4916 * this runqueue and our priority is higher than the current's
4918 if (rq->curr == p) {
4919 if (p->prio > oldprio)
4920 resched_task(rq->curr);
4922 check_preempt_curr(rq, p, 0);
4925 static void switched_from_fair(struct rq *rq, struct task_struct *p)
4927 struct sched_entity *se = &p->se;
4928 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4931 * Ensure the task's vruntime is normalized, so that when its
4932 * switched back to the fair class the enqueue_entity(.flags=0) will
4933 * do the right thing.
4935 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4936 * have normalized the vruntime, if it was !on_rq, then only when
4937 * the task is sleeping will it still have non-normalized vruntime.
4939 if (!se->on_rq && p->state != TASK_RUNNING) {
4941 * Fix up our vruntime so that the current sleep doesn't
4942 * cause 'unlimited' sleep bonus.
4944 place_entity(cfs_rq, se, 0);
4945 se->vruntime -= cfs_rq->min_vruntime;
4950 * We switched to the sched_fair class.
4952 static void switched_to_fair(struct rq *rq, struct task_struct *p)
4958 * We were most likely switched from sched_rt, so
4959 * kick off the schedule if running, otherwise just see
4960 * if we can still preempt the current task.
4963 resched_task(rq->curr);
4965 check_preempt_curr(rq, p, 0);
4968 /* Account for a task changing its policy or group.
4970 * This routine is mostly called to set cfs_rq->curr field when a task
4971 * migrates between groups/classes.
4973 static void set_curr_task_fair(struct rq *rq)
4975 struct sched_entity *se = &rq->curr->se;
4977 for_each_sched_entity(se) {
4978 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4980 set_next_entity(cfs_rq, se);
4981 /* ensure bandwidth has been allocated on our new cfs_rq */
4982 account_cfs_rq_runtime(cfs_rq, 0);
4986 #ifdef CONFIG_FAIR_GROUP_SCHED
4987 static void task_move_group_fair(struct task_struct *p, int on_rq)
4990 * If the task was not on the rq at the time of this cgroup movement
4991 * it must have been asleep, sleeping tasks keep their ->vruntime
4992 * absolute on their old rq until wakeup (needed for the fair sleeper
4993 * bonus in place_entity()).
4995 * If it was on the rq, we've just 'preempted' it, which does convert
4996 * ->vruntime to a relative base.
4998 * Make sure both cases convert their relative position when migrating
4999 * to another cgroup's rq. This does somewhat interfere with the
5000 * fair sleeper stuff for the first placement, but who cares.
5003 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5004 set_task_rq(p, task_cpu(p));
5006 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5010 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5012 struct sched_entity *se = &task->se;
5013 unsigned int rr_interval = 0;
5016 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5019 if (rq->cfs.load.weight)
5020 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5026 * All the scheduling class methods:
5028 static const struct sched_class fair_sched_class = {
5029 .next = &idle_sched_class,
5030 .enqueue_task = enqueue_task_fair,
5031 .dequeue_task = dequeue_task_fair,
5032 .yield_task = yield_task_fair,
5033 .yield_to_task = yield_to_task_fair,
5035 .check_preempt_curr = check_preempt_wakeup,
5037 .pick_next_task = pick_next_task_fair,
5038 .put_prev_task = put_prev_task_fair,
5041 .select_task_rq = select_task_rq_fair,
5043 .rq_online = rq_online_fair,
5044 .rq_offline = rq_offline_fair,
5046 .task_waking = task_waking_fair,
5049 .set_curr_task = set_curr_task_fair,
5050 .task_tick = task_tick_fair,
5051 .task_fork = task_fork_fair,
5053 .prio_changed = prio_changed_fair,
5054 .switched_from = switched_from_fair,
5055 .switched_to = switched_to_fair,
5057 .get_rr_interval = get_rr_interval_fair,
5059 #ifdef CONFIG_FAIR_GROUP_SCHED
5060 .task_move_group = task_move_group_fair,
5064 #ifdef CONFIG_SCHED_DEBUG
5065 static void print_cfs_stats(struct seq_file *m, int cpu)
5067 struct cfs_rq *cfs_rq;
5070 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5071 print_cfs_rq(m, cpu, cfs_rq);