2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 8;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
99 static const struct sched_class fair_sched_class;
101 /**************************************************************
102 * CFS operations on generic schedulable entities:
105 #ifdef CONFIG_FAIR_GROUP_SCHED
107 /* cpu runqueue to which this cfs_rq is attached */
108 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
113 /* An entity is a task if it doesn't "own" a runqueue */
114 #define entity_is_task(se) (!se->my_q)
116 static inline struct task_struct *task_of(struct sched_entity *se)
118 #ifdef CONFIG_SCHED_DEBUG
119 WARN_ON_ONCE(!entity_is_task(se));
121 return container_of(se, struct task_struct, se);
124 /* Walk up scheduling entities hierarchy */
125 #define for_each_sched_entity(se) \
126 for (; se; se = se->parent)
128 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
133 /* runqueue on which this entity is (to be) queued */
134 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
139 /* runqueue "owned" by this group */
140 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
145 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
146 * another cpu ('this_cpu')
148 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
150 return cfs_rq->tg->cfs_rq[this_cpu];
153 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
155 if (!cfs_rq->on_list) {
157 * Ensure we either appear before our parent (if already
158 * enqueued) or force our parent to appear after us when it is
159 * enqueued. The fact that we always enqueue bottom-up
160 * reduces this to two cases.
162 if (cfs_rq->tg->parent &&
163 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
164 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
165 &rq_of(cfs_rq)->leaf_cfs_rq_list);
167 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
168 &rq_of(cfs_rq)->leaf_cfs_rq_list);
175 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
177 if (cfs_rq->on_list) {
178 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
183 /* Iterate thr' all leaf cfs_rq's on a runqueue */
184 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
185 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
187 /* Do the two (enqueued) entities belong to the same group ? */
189 is_same_group(struct sched_entity *se, struct sched_entity *pse)
191 if (se->cfs_rq == pse->cfs_rq)
197 static inline struct sched_entity *parent_entity(struct sched_entity *se)
202 /* return depth at which a sched entity is present in the hierarchy */
203 static inline int depth_se(struct sched_entity *se)
207 for_each_sched_entity(se)
214 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
216 int se_depth, pse_depth;
219 * preemption test can be made between sibling entities who are in the
220 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
221 * both tasks until we find their ancestors who are siblings of common
225 /* First walk up until both entities are at same depth */
226 se_depth = depth_se(*se);
227 pse_depth = depth_se(*pse);
229 while (se_depth > pse_depth) {
231 *se = parent_entity(*se);
234 while (pse_depth > se_depth) {
236 *pse = parent_entity(*pse);
239 while (!is_same_group(*se, *pse)) {
240 *se = parent_entity(*se);
241 *pse = parent_entity(*pse);
245 #else /* !CONFIG_FAIR_GROUP_SCHED */
247 static inline struct task_struct *task_of(struct sched_entity *se)
249 return container_of(se, struct task_struct, se);
252 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 return container_of(cfs_rq, struct rq, cfs);
257 #define entity_is_task(se) 1
259 #define for_each_sched_entity(se) \
260 for (; se; se = NULL)
262 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
264 return &task_rq(p)->cfs;
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
269 struct task_struct *p = task_of(se);
270 struct rq *rq = task_rq(p);
275 /* runqueue "owned" by this group */
276 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
283 return &cpu_rq(this_cpu)->cfs;
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
295 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
298 is_same_group(struct sched_entity *se, struct sched_entity *pse)
303 static inline struct sched_entity *parent_entity(struct sched_entity *se)
309 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
313 #endif /* CONFIG_FAIR_GROUP_SCHED */
316 /**************************************************************
317 * Scheduling class tree data structure manipulation methods:
320 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
322 s64 delta = (s64)(vruntime - min_vruntime);
324 min_vruntime = vruntime;
329 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
331 s64 delta = (s64)(vruntime - min_vruntime);
333 min_vruntime = vruntime;
338 static inline int entity_before(struct sched_entity *a,
339 struct sched_entity *b)
341 return (s64)(a->vruntime - b->vruntime) < 0;
344 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
346 return se->vruntime - cfs_rq->min_vruntime;
349 static void update_min_vruntime(struct cfs_rq *cfs_rq)
351 u64 vruntime = cfs_rq->min_vruntime;
354 vruntime = cfs_rq->curr->vruntime;
356 if (cfs_rq->rb_leftmost) {
357 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
362 vruntime = se->vruntime;
364 vruntime = min_vruntime(vruntime, se->vruntime);
367 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
371 * Enqueue an entity into the rb-tree:
373 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
375 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
376 struct rb_node *parent = NULL;
377 struct sched_entity *entry;
378 s64 key = entity_key(cfs_rq, se);
382 * Find the right place in the rbtree:
386 entry = rb_entry(parent, struct sched_entity, run_node);
388 * We dont care about collisions. Nodes with
389 * the same key stay together.
391 if (key < entity_key(cfs_rq, entry)) {
392 link = &parent->rb_left;
394 link = &parent->rb_right;
400 * Maintain a cache of leftmost tree entries (it is frequently
404 cfs_rq->rb_leftmost = &se->run_node;
406 rb_link_node(&se->run_node, parent, link);
407 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
410 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
412 if (cfs_rq->rb_leftmost == &se->run_node) {
413 struct rb_node *next_node;
415 next_node = rb_next(&se->run_node);
416 cfs_rq->rb_leftmost = next_node;
419 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
422 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
424 struct rb_node *left = cfs_rq->rb_leftmost;
429 return rb_entry(left, struct sched_entity, run_node);
432 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
434 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
439 return rb_entry(last, struct sched_entity, run_node);
442 /**************************************************************
443 * Scheduling class statistics methods:
446 #ifdef CONFIG_SCHED_DEBUG
447 int sched_proc_update_handler(struct ctl_table *table, int write,
448 void __user *buffer, size_t *lenp,
451 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
452 int factor = get_update_sysctl_factor();
457 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
458 sysctl_sched_min_granularity);
460 #define WRT_SYSCTL(name) \
461 (normalized_sysctl_##name = sysctl_##name / (factor))
462 WRT_SYSCTL(sched_min_granularity);
463 WRT_SYSCTL(sched_latency);
464 WRT_SYSCTL(sched_wakeup_granularity);
474 static inline unsigned long
475 calc_delta_fair(unsigned long delta, struct sched_entity *se)
477 if (unlikely(se->load.weight != NICE_0_LOAD))
478 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
484 * The idea is to set a period in which each task runs once.
486 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
487 * this period because otherwise the slices get too small.
489 * p = (nr <= nl) ? l : l*nr/nl
491 static u64 __sched_period(unsigned long nr_running)
493 u64 period = sysctl_sched_latency;
494 unsigned long nr_latency = sched_nr_latency;
496 if (unlikely(nr_running > nr_latency)) {
497 period = sysctl_sched_min_granularity;
498 period *= nr_running;
505 * We calculate the wall-time slice from the period by taking a part
506 * proportional to the weight.
510 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
512 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
514 for_each_sched_entity(se) {
515 struct load_weight *load;
516 struct load_weight lw;
518 cfs_rq = cfs_rq_of(se);
519 load = &cfs_rq->load;
521 if (unlikely(!se->on_rq)) {
524 update_load_add(&lw, se->load.weight);
527 slice = calc_delta_mine(slice, se->load.weight, load);
533 * We calculate the vruntime slice of a to be inserted task
537 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
539 return calc_delta_fair(sched_slice(cfs_rq, se), se);
543 * Update the current task's runtime statistics. Skip current tasks that
544 * are not in our scheduling class.
547 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
548 unsigned long delta_exec)
550 unsigned long delta_exec_weighted;
552 schedstat_set(curr->statistics.exec_max,
553 max((u64)delta_exec, curr->statistics.exec_max));
555 curr->sum_exec_runtime += delta_exec;
556 schedstat_add(cfs_rq, exec_clock, delta_exec);
557 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
559 curr->vruntime += delta_exec_weighted;
560 update_min_vruntime(cfs_rq);
563 static void update_curr(struct cfs_rq *cfs_rq)
565 struct sched_entity *curr = cfs_rq->curr;
566 u64 now = rq_of(cfs_rq)->clock_task;
567 unsigned long delta_exec;
573 * Get the amount of time the current task was running
574 * since the last time we changed load (this cannot
575 * overflow on 32 bits):
577 delta_exec = (unsigned long)(now - curr->exec_start);
581 __update_curr(cfs_rq, curr, delta_exec);
582 curr->exec_start = now;
584 if (entity_is_task(curr)) {
585 struct task_struct *curtask = task_of(curr);
587 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
588 cpuacct_charge(curtask, delta_exec);
589 account_group_exec_runtime(curtask, delta_exec);
594 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
596 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
600 * Task is being enqueued - update stats:
602 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
605 * Are we enqueueing a waiting task? (for current tasks
606 * a dequeue/enqueue event is a NOP)
608 if (se != cfs_rq->curr)
609 update_stats_wait_start(cfs_rq, se);
613 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
615 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
616 rq_of(cfs_rq)->clock - se->statistics.wait_start));
617 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
618 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
619 rq_of(cfs_rq)->clock - se->statistics.wait_start);
620 #ifdef CONFIG_SCHEDSTATS
621 if (entity_is_task(se)) {
622 trace_sched_stat_wait(task_of(se),
623 rq_of(cfs_rq)->clock - se->statistics.wait_start);
626 schedstat_set(se->statistics.wait_start, 0);
630 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
633 * Mark the end of the wait period if dequeueing a
636 if (se != cfs_rq->curr)
637 update_stats_wait_end(cfs_rq, se);
641 * We are picking a new current task - update its stats:
644 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
647 * We are starting a new run period:
649 se->exec_start = rq_of(cfs_rq)->clock_task;
652 /**************************************************
653 * Scheduling class queueing methods:
656 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
658 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
660 cfs_rq->task_weight += weight;
664 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
670 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
672 update_load_add(&cfs_rq->load, se->load.weight);
673 if (!parent_entity(se))
674 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
675 if (entity_is_task(se)) {
676 add_cfs_task_weight(cfs_rq, se->load.weight);
677 list_add(&se->group_node, &cfs_rq->tasks);
679 cfs_rq->nr_running++;
683 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
685 update_load_sub(&cfs_rq->load, se->load.weight);
686 if (!parent_entity(se))
687 dec_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_del_init(&se->group_node);
692 cfs_rq->nr_running--;
695 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
696 static void update_cfs_load(struct cfs_rq *cfs_rq)
698 u64 period = sysctl_sched_shares_window;
700 unsigned long load = cfs_rq->load.weight;
705 now = rq_of(cfs_rq)->clock;
706 delta = now - cfs_rq->load_stamp;
708 /* truncate load history at 4 idle periods */
709 if (cfs_rq->load_stamp > cfs_rq->load_last &&
710 now - cfs_rq->load_last > 4 * period) {
711 cfs_rq->load_period = 0;
712 cfs_rq->load_avg = 0;
715 cfs_rq->load_stamp = now;
716 cfs_rq->load_period += delta;
718 cfs_rq->load_last = now;
719 cfs_rq->load_avg += delta * load;
722 while (cfs_rq->load_period > period) {
724 * Inline assembly required to prevent the compiler
725 * optimising this loop into a divmod call.
726 * See __iter_div_u64_rem() for another example of this.
728 asm("" : "+rm" (cfs_rq->load_period));
729 cfs_rq->load_period /= 2;
730 cfs_rq->load_avg /= 2;
733 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
734 list_del_leaf_cfs_rq(cfs_rq);
737 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
738 unsigned long weight)
741 account_entity_dequeue(cfs_rq, se);
743 update_load_set(&se->load, weight);
746 account_entity_enqueue(cfs_rq, se);
749 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
751 struct task_group *tg;
752 struct sched_entity *se;
753 long load_weight, load, shares;
759 se = tg->se[cpu_of(rq_of(cfs_rq))];
763 load = cfs_rq->load.weight + weight_delta;
765 load_weight = atomic_read(&tg->load_weight);
766 load_weight -= cfs_rq->load_contribution;
769 shares = (tg->shares * load);
771 shares /= load_weight;
773 if (shares < MIN_SHARES)
775 if (shares > tg->shares)
778 reweight_entity(cfs_rq_of(se), se, shares);
780 #else /* CONFIG_FAIR_GROUP_SCHED */
781 static inline void update_cfs_load(struct cfs_rq *cfs_rq)
785 static inline void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
788 #endif /* CONFIG_FAIR_GROUP_SCHED */
790 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
792 #ifdef CONFIG_SCHEDSTATS
793 struct task_struct *tsk = NULL;
795 if (entity_is_task(se))
798 if (se->statistics.sleep_start) {
799 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
804 if (unlikely(delta > se->statistics.sleep_max))
805 se->statistics.sleep_max = delta;
807 se->statistics.sleep_start = 0;
808 se->statistics.sum_sleep_runtime += delta;
811 account_scheduler_latency(tsk, delta >> 10, 1);
812 trace_sched_stat_sleep(tsk, delta);
815 if (se->statistics.block_start) {
816 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
821 if (unlikely(delta > se->statistics.block_max))
822 se->statistics.block_max = delta;
824 se->statistics.block_start = 0;
825 se->statistics.sum_sleep_runtime += delta;
828 if (tsk->in_iowait) {
829 se->statistics.iowait_sum += delta;
830 se->statistics.iowait_count++;
831 trace_sched_stat_iowait(tsk, delta);
835 * Blocking time is in units of nanosecs, so shift by
836 * 20 to get a milliseconds-range estimation of the
837 * amount of time that the task spent sleeping:
839 if (unlikely(prof_on == SLEEP_PROFILING)) {
840 profile_hits(SLEEP_PROFILING,
841 (void *)get_wchan(tsk),
844 account_scheduler_latency(tsk, delta >> 10, 0);
850 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
852 #ifdef CONFIG_SCHED_DEBUG
853 s64 d = se->vruntime - cfs_rq->min_vruntime;
858 if (d > 3*sysctl_sched_latency)
859 schedstat_inc(cfs_rq, nr_spread_over);
864 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
866 u64 vruntime = cfs_rq->min_vruntime;
869 * The 'current' period is already promised to the current tasks,
870 * however the extra weight of the new task will slow them down a
871 * little, place the new task so that it fits in the slot that
872 * stays open at the end.
874 if (initial && sched_feat(START_DEBIT))
875 vruntime += sched_vslice(cfs_rq, se);
877 /* sleeps up to a single latency don't count. */
879 unsigned long thresh = sysctl_sched_latency;
882 * Halve their sleep time's effect, to allow
883 * for a gentler effect of sleepers:
885 if (sched_feat(GENTLE_FAIR_SLEEPERS))
891 /* ensure we never gain time by being placed backwards. */
892 vruntime = max_vruntime(se->vruntime, vruntime);
894 se->vruntime = vruntime;
898 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
901 * Update the normalized vruntime before updating min_vruntime
902 * through callig update_curr().
904 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
905 se->vruntime += cfs_rq->min_vruntime;
908 * Update run-time statistics of the 'current'.
911 update_cfs_load(cfs_rq);
912 update_cfs_shares(cfs_rq, se->load.weight);
913 account_entity_enqueue(cfs_rq, se);
915 if (flags & ENQUEUE_WAKEUP) {
916 place_entity(cfs_rq, se, 0);
917 enqueue_sleeper(cfs_rq, se);
920 update_stats_enqueue(cfs_rq, se);
921 check_spread(cfs_rq, se);
922 if (se != cfs_rq->curr)
923 __enqueue_entity(cfs_rq, se);
926 if (cfs_rq->nr_running == 1)
927 list_add_leaf_cfs_rq(cfs_rq);
930 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
932 if (!se || cfs_rq->last == se)
935 if (!se || cfs_rq->next == se)
939 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
941 for_each_sched_entity(se)
942 __clear_buddies(cfs_rq_of(se), se);
946 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
949 * Update run-time statistics of the 'current'.
953 update_stats_dequeue(cfs_rq, se);
954 if (flags & DEQUEUE_SLEEP) {
955 #ifdef CONFIG_SCHEDSTATS
956 if (entity_is_task(se)) {
957 struct task_struct *tsk = task_of(se);
959 if (tsk->state & TASK_INTERRUPTIBLE)
960 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
961 if (tsk->state & TASK_UNINTERRUPTIBLE)
962 se->statistics.block_start = rq_of(cfs_rq)->clock;
967 clear_buddies(cfs_rq, se);
969 if (se != cfs_rq->curr)
970 __dequeue_entity(cfs_rq, se);
972 update_cfs_load(cfs_rq);
973 account_entity_dequeue(cfs_rq, se);
974 update_min_vruntime(cfs_rq);
975 update_cfs_shares(cfs_rq, 0);
978 * Normalize the entity after updating the min_vruntime because the
979 * update can refer to the ->curr item and we need to reflect this
980 * movement in our normalized position.
982 if (!(flags & DEQUEUE_SLEEP))
983 se->vruntime -= cfs_rq->min_vruntime;
987 * Preempt the current task with a newly woken task if needed:
990 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
992 unsigned long ideal_runtime, delta_exec;
994 ideal_runtime = sched_slice(cfs_rq, curr);
995 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
996 if (delta_exec > ideal_runtime) {
997 resched_task(rq_of(cfs_rq)->curr);
999 * The current task ran long enough, ensure it doesn't get
1000 * re-elected due to buddy favours.
1002 clear_buddies(cfs_rq, curr);
1007 * Ensure that a task that missed wakeup preemption by a
1008 * narrow margin doesn't have to wait for a full slice.
1009 * This also mitigates buddy induced latencies under load.
1011 if (!sched_feat(WAKEUP_PREEMPT))
1014 if (delta_exec < sysctl_sched_min_granularity)
1017 if (cfs_rq->nr_running > 1) {
1018 struct sched_entity *se = __pick_next_entity(cfs_rq);
1019 s64 delta = curr->vruntime - se->vruntime;
1021 if (delta > ideal_runtime)
1022 resched_task(rq_of(cfs_rq)->curr);
1027 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1029 /* 'current' is not kept within the tree. */
1032 * Any task has to be enqueued before it get to execute on
1033 * a CPU. So account for the time it spent waiting on the
1036 update_stats_wait_end(cfs_rq, se);
1037 __dequeue_entity(cfs_rq, se);
1040 update_stats_curr_start(cfs_rq, se);
1042 #ifdef CONFIG_SCHEDSTATS
1044 * Track our maximum slice length, if the CPU's load is at
1045 * least twice that of our own weight (i.e. dont track it
1046 * when there are only lesser-weight tasks around):
1048 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1049 se->statistics.slice_max = max(se->statistics.slice_max,
1050 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1053 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1057 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1059 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1061 struct sched_entity *se = __pick_next_entity(cfs_rq);
1062 struct sched_entity *left = se;
1064 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1068 * Prefer last buddy, try to return the CPU to a preempted task.
1070 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1073 clear_buddies(cfs_rq, se);
1078 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1081 * If still on the runqueue then deactivate_task()
1082 * was not called and update_curr() has to be done:
1085 update_curr(cfs_rq);
1087 check_spread(cfs_rq, prev);
1089 update_stats_wait_start(cfs_rq, prev);
1090 /* Put 'current' back into the tree. */
1091 __enqueue_entity(cfs_rq, prev);
1093 cfs_rq->curr = NULL;
1097 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1100 * Update run-time statistics of the 'current'.
1102 update_curr(cfs_rq);
1104 #ifdef CONFIG_SCHED_HRTICK
1106 * queued ticks are scheduled to match the slice, so don't bother
1107 * validating it and just reschedule.
1110 resched_task(rq_of(cfs_rq)->curr);
1114 * don't let the period tick interfere with the hrtick preemption
1116 if (!sched_feat(DOUBLE_TICK) &&
1117 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1121 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1122 check_preempt_tick(cfs_rq, curr);
1125 /**************************************************
1126 * CFS operations on tasks:
1129 #ifdef CONFIG_SCHED_HRTICK
1130 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1132 struct sched_entity *se = &p->se;
1133 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1135 WARN_ON(task_rq(p) != rq);
1137 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1138 u64 slice = sched_slice(cfs_rq, se);
1139 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1140 s64 delta = slice - ran;
1149 * Don't schedule slices shorter than 10000ns, that just
1150 * doesn't make sense. Rely on vruntime for fairness.
1153 delta = max_t(s64, 10000LL, delta);
1155 hrtick_start(rq, delta);
1160 * called from enqueue/dequeue and updates the hrtick when the
1161 * current task is from our class and nr_running is low enough
1164 static void hrtick_update(struct rq *rq)
1166 struct task_struct *curr = rq->curr;
1168 if (curr->sched_class != &fair_sched_class)
1171 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1172 hrtick_start_fair(rq, curr);
1174 #else /* !CONFIG_SCHED_HRTICK */
1176 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1180 static inline void hrtick_update(struct rq *rq)
1186 * The enqueue_task method is called before nr_running is
1187 * increased. Here we update the fair scheduling stats and
1188 * then put the task into the rbtree:
1191 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1193 struct cfs_rq *cfs_rq;
1194 struct sched_entity *se = &p->se;
1196 for_each_sched_entity(se) {
1199 cfs_rq = cfs_rq_of(se);
1200 enqueue_entity(cfs_rq, se, flags);
1201 flags = ENQUEUE_WAKEUP;
1204 for_each_sched_entity(se) {
1205 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1207 update_cfs_load(cfs_rq);
1208 update_cfs_shares(cfs_rq, 0);
1215 * The dequeue_task method is called before nr_running is
1216 * decreased. We remove the task from the rbtree and
1217 * update the fair scheduling stats:
1219 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1221 struct cfs_rq *cfs_rq;
1222 struct sched_entity *se = &p->se;
1224 for_each_sched_entity(se) {
1225 cfs_rq = cfs_rq_of(se);
1226 dequeue_entity(cfs_rq, se, flags);
1228 /* Don't dequeue parent if it has other entities besides us */
1229 if (cfs_rq->load.weight)
1231 flags |= DEQUEUE_SLEEP;
1234 for_each_sched_entity(se) {
1235 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1237 update_cfs_load(cfs_rq);
1238 update_cfs_shares(cfs_rq, 0);
1245 * sched_yield() support is very simple - we dequeue and enqueue.
1247 * If compat_yield is turned on then we requeue to the end of the tree.
1249 static void yield_task_fair(struct rq *rq)
1251 struct task_struct *curr = rq->curr;
1252 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1253 struct sched_entity *rightmost, *se = &curr->se;
1256 * Are we the only task in the tree?
1258 if (unlikely(cfs_rq->nr_running == 1))
1261 clear_buddies(cfs_rq, se);
1263 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1264 update_rq_clock(rq);
1266 * Update run-time statistics of the 'current'.
1268 update_curr(cfs_rq);
1273 * Find the rightmost entry in the rbtree:
1275 rightmost = __pick_last_entity(cfs_rq);
1277 * Already in the rightmost position?
1279 if (unlikely(!rightmost || entity_before(rightmost, se)))
1283 * Minimally necessary key value to be last in the tree:
1284 * Upon rescheduling, sched_class::put_prev_task() will place
1285 * 'current' within the tree based on its new key value.
1287 se->vruntime = rightmost->vruntime + 1;
1292 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1294 struct sched_entity *se = &p->se;
1295 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1297 se->vruntime -= cfs_rq->min_vruntime;
1300 #ifdef CONFIG_FAIR_GROUP_SCHED
1302 * effective_load() calculates the load change as seen from the root_task_group
1304 * Adding load to a group doesn't make a group heavier, but can cause movement
1305 * of group shares between cpus. Assuming the shares were perfectly aligned one
1306 * can calculate the shift in shares.
1308 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1310 struct sched_entity *se = tg->se[cpu];
1315 for_each_sched_entity(se) {
1316 long S, rw, s, a, b;
1318 S = se->my_q->tg->shares;
1319 s = se->load.weight;
1320 rw = se->my_q->load.weight;
1331 * Assume the group is already running and will
1332 * thus already be accounted for in the weight.
1334 * That is, moving shares between CPUs, does not
1335 * alter the group weight.
1345 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1346 unsigned long wl, unsigned long wg)
1353 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1355 unsigned long this_load, load;
1356 int idx, this_cpu, prev_cpu;
1357 unsigned long tl_per_task;
1358 struct task_group *tg;
1359 unsigned long weight;
1363 this_cpu = smp_processor_id();
1364 prev_cpu = task_cpu(p);
1365 load = source_load(prev_cpu, idx);
1366 this_load = target_load(this_cpu, idx);
1369 * If sync wakeup then subtract the (maximum possible)
1370 * effect of the currently running task from the load
1371 * of the current CPU:
1375 tg = task_group(current);
1376 weight = current->se.load.weight;
1378 this_load += effective_load(tg, this_cpu, -weight, -weight);
1379 load += effective_load(tg, prev_cpu, 0, -weight);
1383 weight = p->se.load.weight;
1386 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1387 * due to the sync cause above having dropped this_load to 0, we'll
1388 * always have an imbalance, but there's really nothing you can do
1389 * about that, so that's good too.
1391 * Otherwise check if either cpus are near enough in load to allow this
1392 * task to be woken on this_cpu.
1395 unsigned long this_eff_load, prev_eff_load;
1397 this_eff_load = 100;
1398 this_eff_load *= power_of(prev_cpu);
1399 this_eff_load *= this_load +
1400 effective_load(tg, this_cpu, weight, weight);
1402 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1403 prev_eff_load *= power_of(this_cpu);
1404 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1406 balanced = this_eff_load <= prev_eff_load;
1412 * If the currently running task will sleep within
1413 * a reasonable amount of time then attract this newly
1416 if (sync && balanced)
1419 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1420 tl_per_task = cpu_avg_load_per_task(this_cpu);
1423 (this_load <= load &&
1424 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1426 * This domain has SD_WAKE_AFFINE and
1427 * p is cache cold in this domain, and
1428 * there is no bad imbalance.
1430 schedstat_inc(sd, ttwu_move_affine);
1431 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1439 * find_idlest_group finds and returns the least busy CPU group within the
1442 static struct sched_group *
1443 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1444 int this_cpu, int load_idx)
1446 struct sched_group *idlest = NULL, *group = sd->groups;
1447 unsigned long min_load = ULONG_MAX, this_load = 0;
1448 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1451 unsigned long load, avg_load;
1455 /* Skip over this group if it has no CPUs allowed */
1456 if (!cpumask_intersects(sched_group_cpus(group),
1460 local_group = cpumask_test_cpu(this_cpu,
1461 sched_group_cpus(group));
1463 /* Tally up the load of all CPUs in the group */
1466 for_each_cpu(i, sched_group_cpus(group)) {
1467 /* Bias balancing toward cpus of our domain */
1469 load = source_load(i, load_idx);
1471 load = target_load(i, load_idx);
1476 /* Adjust by relative CPU power of the group */
1477 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1480 this_load = avg_load;
1481 } else if (avg_load < min_load) {
1482 min_load = avg_load;
1485 } while (group = group->next, group != sd->groups);
1487 if (!idlest || 100*this_load < imbalance*min_load)
1493 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1496 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1498 unsigned long load, min_load = ULONG_MAX;
1502 /* Traverse only the allowed CPUs */
1503 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1504 load = weighted_cpuload(i);
1506 if (load < min_load || (load == min_load && i == this_cpu)) {
1516 * Try and locate an idle CPU in the sched_domain.
1518 static int select_idle_sibling(struct task_struct *p, int target)
1520 int cpu = smp_processor_id();
1521 int prev_cpu = task_cpu(p);
1522 struct sched_domain *sd;
1526 * If the task is going to be woken-up on this cpu and if it is
1527 * already idle, then it is the right target.
1529 if (target == cpu && idle_cpu(cpu))
1533 * If the task is going to be woken-up on the cpu where it previously
1534 * ran and if it is currently idle, then it the right target.
1536 if (target == prev_cpu && idle_cpu(prev_cpu))
1540 * Otherwise, iterate the domains and find an elegible idle cpu.
1542 for_each_domain(target, sd) {
1543 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1546 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1554 * Lets stop looking for an idle sibling when we reached
1555 * the domain that spans the current cpu and prev_cpu.
1557 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1558 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1566 * sched_balance_self: balance the current task (running on cpu) in domains
1567 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1570 * Balance, ie. select the least loaded group.
1572 * Returns the target CPU number, or the same CPU if no balancing is needed.
1574 * preempt must be disabled.
1577 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1579 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1580 int cpu = smp_processor_id();
1581 int prev_cpu = task_cpu(p);
1583 int want_affine = 0;
1585 int sync = wake_flags & WF_SYNC;
1587 if (sd_flag & SD_BALANCE_WAKE) {
1588 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1593 for_each_domain(cpu, tmp) {
1594 if (!(tmp->flags & SD_LOAD_BALANCE))
1598 * If power savings logic is enabled for a domain, see if we
1599 * are not overloaded, if so, don't balance wider.
1601 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1602 unsigned long power = 0;
1603 unsigned long nr_running = 0;
1604 unsigned long capacity;
1607 for_each_cpu(i, sched_domain_span(tmp)) {
1608 power += power_of(i);
1609 nr_running += cpu_rq(i)->cfs.nr_running;
1612 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1614 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1617 if (nr_running < capacity)
1622 * If both cpu and prev_cpu are part of this domain,
1623 * cpu is a valid SD_WAKE_AFFINE target.
1625 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1626 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1631 if (!want_sd && !want_affine)
1634 if (!(tmp->flags & sd_flag))
1642 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1643 return select_idle_sibling(p, cpu);
1645 return select_idle_sibling(p, prev_cpu);
1649 int load_idx = sd->forkexec_idx;
1650 struct sched_group *group;
1653 if (!(sd->flags & sd_flag)) {
1658 if (sd_flag & SD_BALANCE_WAKE)
1659 load_idx = sd->wake_idx;
1661 group = find_idlest_group(sd, p, cpu, load_idx);
1667 new_cpu = find_idlest_cpu(group, p, cpu);
1668 if (new_cpu == -1 || new_cpu == cpu) {
1669 /* Now try balancing at a lower domain level of cpu */
1674 /* Now try balancing at a lower domain level of new_cpu */
1676 weight = sd->span_weight;
1678 for_each_domain(cpu, tmp) {
1679 if (weight <= tmp->span_weight)
1681 if (tmp->flags & sd_flag)
1684 /* while loop will break here if sd == NULL */
1689 #endif /* CONFIG_SMP */
1691 static unsigned long
1692 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1694 unsigned long gran = sysctl_sched_wakeup_granularity;
1697 * Since its curr running now, convert the gran from real-time
1698 * to virtual-time in his units.
1700 * By using 'se' instead of 'curr' we penalize light tasks, so
1701 * they get preempted easier. That is, if 'se' < 'curr' then
1702 * the resulting gran will be larger, therefore penalizing the
1703 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1704 * be smaller, again penalizing the lighter task.
1706 * This is especially important for buddies when the leftmost
1707 * task is higher priority than the buddy.
1709 if (unlikely(se->load.weight != NICE_0_LOAD))
1710 gran = calc_delta_fair(gran, se);
1716 * Should 'se' preempt 'curr'.
1730 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1732 s64 gran, vdiff = curr->vruntime - se->vruntime;
1737 gran = wakeup_gran(curr, se);
1744 static void set_last_buddy(struct sched_entity *se)
1746 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1747 for_each_sched_entity(se)
1748 cfs_rq_of(se)->last = se;
1752 static void set_next_buddy(struct sched_entity *se)
1754 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1755 for_each_sched_entity(se)
1756 cfs_rq_of(se)->next = se;
1761 * Preempt the current task with a newly woken task if needed:
1763 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1765 struct task_struct *curr = rq->curr;
1766 struct sched_entity *se = &curr->se, *pse = &p->se;
1767 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1768 int scale = cfs_rq->nr_running >= sched_nr_latency;
1770 if (unlikely(rt_prio(p->prio)))
1773 if (unlikely(p->sched_class != &fair_sched_class))
1776 if (unlikely(se == pse))
1779 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1780 set_next_buddy(pse);
1783 * We can come here with TIF_NEED_RESCHED already set from new task
1786 if (test_tsk_need_resched(curr))
1790 * Batch and idle tasks do not preempt (their preemption is driven by
1793 if (unlikely(p->policy != SCHED_NORMAL))
1796 /* Idle tasks are by definition preempted by everybody. */
1797 if (unlikely(curr->policy == SCHED_IDLE))
1800 if (!sched_feat(WAKEUP_PREEMPT))
1803 update_curr(cfs_rq);
1804 find_matching_se(&se, &pse);
1806 if (wakeup_preempt_entity(se, pse) == 1)
1814 * Only set the backward buddy when the current task is still
1815 * on the rq. This can happen when a wakeup gets interleaved
1816 * with schedule on the ->pre_schedule() or idle_balance()
1817 * point, either of which can * drop the rq lock.
1819 * Also, during early boot the idle thread is in the fair class,
1820 * for obvious reasons its a bad idea to schedule back to it.
1822 if (unlikely(!se->on_rq || curr == rq->idle))
1825 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1829 static struct task_struct *pick_next_task_fair(struct rq *rq)
1831 struct task_struct *p;
1832 struct cfs_rq *cfs_rq = &rq->cfs;
1833 struct sched_entity *se;
1835 if (!cfs_rq->nr_running)
1839 se = pick_next_entity(cfs_rq);
1840 set_next_entity(cfs_rq, se);
1841 cfs_rq = group_cfs_rq(se);
1845 hrtick_start_fair(rq, p);
1851 * Account for a descheduled task:
1853 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1855 struct sched_entity *se = &prev->se;
1856 struct cfs_rq *cfs_rq;
1858 for_each_sched_entity(se) {
1859 cfs_rq = cfs_rq_of(se);
1860 put_prev_entity(cfs_rq, se);
1865 /**************************************************
1866 * Fair scheduling class load-balancing methods:
1870 * pull_task - move a task from a remote runqueue to the local runqueue.
1871 * Both runqueues must be locked.
1873 static void pull_task(struct rq *src_rq, struct task_struct *p,
1874 struct rq *this_rq, int this_cpu)
1876 deactivate_task(src_rq, p, 0);
1877 set_task_cpu(p, this_cpu);
1878 activate_task(this_rq, p, 0);
1879 check_preempt_curr(this_rq, p, 0);
1881 /* re-arm NEWIDLE balancing when moving tasks */
1882 src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost;
1883 this_rq->idle_stamp = 0;
1887 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1890 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1891 struct sched_domain *sd, enum cpu_idle_type idle,
1894 int tsk_cache_hot = 0;
1896 * We do not migrate tasks that are:
1897 * 1) running (obviously), or
1898 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1899 * 3) are cache-hot on their current CPU.
1901 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1902 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1907 if (task_running(rq, p)) {
1908 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1913 * Aggressive migration if:
1914 * 1) task is cache cold, or
1915 * 2) too many balance attempts have failed.
1918 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
1919 if (!tsk_cache_hot ||
1920 sd->nr_balance_failed > sd->cache_nice_tries) {
1921 #ifdef CONFIG_SCHEDSTATS
1922 if (tsk_cache_hot) {
1923 schedstat_inc(sd, lb_hot_gained[idle]);
1924 schedstat_inc(p, se.statistics.nr_forced_migrations);
1930 if (tsk_cache_hot) {
1931 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
1938 * move_one_task tries to move exactly one task from busiest to this_rq, as
1939 * part of active balancing operations within "domain".
1940 * Returns 1 if successful and 0 otherwise.
1942 * Called with both runqueues locked.
1945 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1946 struct sched_domain *sd, enum cpu_idle_type idle)
1948 struct task_struct *p, *n;
1949 struct cfs_rq *cfs_rq;
1952 for_each_leaf_cfs_rq(busiest, cfs_rq) {
1953 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1955 if (!can_migrate_task(p, busiest, this_cpu,
1959 pull_task(busiest, p, this_rq, this_cpu);
1961 * Right now, this is only the second place pull_task()
1962 * is called, so we can safely collect pull_task()
1963 * stats here rather than inside pull_task().
1965 schedstat_inc(sd, lb_gained[idle]);
1973 static unsigned long
1974 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1975 unsigned long max_load_move, struct sched_domain *sd,
1976 enum cpu_idle_type idle, int *all_pinned,
1977 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
1979 int loops = 0, pulled = 0, pinned = 0;
1980 long rem_load_move = max_load_move;
1981 struct task_struct *p, *n;
1983 if (max_load_move == 0)
1988 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
1989 if (loops++ > sysctl_sched_nr_migrate)
1992 if ((p->se.load.weight >> 1) > rem_load_move ||
1993 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
1996 pull_task(busiest, p, this_rq, this_cpu);
1998 rem_load_move -= p->se.load.weight;
2000 #ifdef CONFIG_PREEMPT
2002 * NEWIDLE balancing is a source of latency, so preemptible
2003 * kernels will stop after the first task is pulled to minimize
2004 * the critical section.
2006 if (idle == CPU_NEWLY_IDLE)
2011 * We only want to steal up to the prescribed amount of
2014 if (rem_load_move <= 0)
2017 if (p->prio < *this_best_prio)
2018 *this_best_prio = p->prio;
2022 * Right now, this is one of only two places pull_task() is called,
2023 * so we can safely collect pull_task() stats here rather than
2024 * inside pull_task().
2026 schedstat_add(sd, lb_gained[idle], pulled);
2029 *all_pinned = pinned;
2031 return max_load_move - rem_load_move;
2034 #ifdef CONFIG_FAIR_GROUP_SCHED
2036 * update tg->load_weight by folding this cpu's load_avg
2038 static int update_shares_cpu(struct task_group *tg, int cpu)
2040 struct cfs_rq *cfs_rq;
2041 unsigned long flags;
2049 cfs_rq = tg->cfs_rq[cpu];
2051 raw_spin_lock_irqsave(&rq->lock, flags);
2053 update_rq_clock(rq);
2054 update_cfs_load(cfs_rq);
2056 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
2057 load_avg -= cfs_rq->load_contribution;
2058 atomic_add(load_avg, &tg->load_weight);
2059 cfs_rq->load_contribution += load_avg;
2062 * We need to update shares after updating tg->load_weight in
2063 * order to adjust the weight of groups with long running tasks.
2065 update_cfs_shares(cfs_rq, 0);
2067 raw_spin_unlock_irqrestore(&rq->lock, flags);
2072 static void update_shares(int cpu)
2074 struct cfs_rq *cfs_rq;
2075 struct rq *rq = cpu_rq(cpu);
2078 for_each_leaf_cfs_rq(rq, cfs_rq)
2079 update_shares_cpu(cfs_rq->tg, cpu);
2083 static unsigned long
2084 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2085 unsigned long max_load_move,
2086 struct sched_domain *sd, enum cpu_idle_type idle,
2087 int *all_pinned, int *this_best_prio)
2089 long rem_load_move = max_load_move;
2090 int busiest_cpu = cpu_of(busiest);
2091 struct task_group *tg;
2094 update_h_load(busiest_cpu);
2096 list_for_each_entry_rcu(tg, &task_groups, list) {
2097 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2098 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2099 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2100 u64 rem_load, moved_load;
2105 if (!busiest_cfs_rq->task_weight)
2108 rem_load = (u64)rem_load_move * busiest_weight;
2109 rem_load = div_u64(rem_load, busiest_h_load + 1);
2111 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2112 rem_load, sd, idle, all_pinned, this_best_prio,
2118 moved_load *= busiest_h_load;
2119 moved_load = div_u64(moved_load, busiest_weight + 1);
2121 rem_load_move -= moved_load;
2122 if (rem_load_move < 0)
2127 return max_load_move - rem_load_move;
2130 static inline void update_shares(int cpu)
2134 static unsigned long
2135 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2136 unsigned long max_load_move,
2137 struct sched_domain *sd, enum cpu_idle_type idle,
2138 int *all_pinned, int *this_best_prio)
2140 return balance_tasks(this_rq, this_cpu, busiest,
2141 max_load_move, sd, idle, all_pinned,
2142 this_best_prio, &busiest->cfs);
2147 * move_tasks tries to move up to max_load_move weighted load from busiest to
2148 * this_rq, as part of a balancing operation within domain "sd".
2149 * Returns 1 if successful and 0 otherwise.
2151 * Called with both runqueues locked.
2153 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2154 unsigned long max_load_move,
2155 struct sched_domain *sd, enum cpu_idle_type idle,
2158 unsigned long total_load_moved = 0, load_moved;
2159 int this_best_prio = this_rq->curr->prio;
2162 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2163 max_load_move - total_load_moved,
2164 sd, idle, all_pinned, &this_best_prio);
2166 total_load_moved += load_moved;
2168 #ifdef CONFIG_PREEMPT
2170 * NEWIDLE balancing is a source of latency, so preemptible
2171 * kernels will stop after the first task is pulled to minimize
2172 * the critical section.
2174 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2177 if (raw_spin_is_contended(&this_rq->lock) ||
2178 raw_spin_is_contended(&busiest->lock))
2181 } while (load_moved && max_load_move > total_load_moved);
2183 return total_load_moved > 0;
2186 /********** Helpers for find_busiest_group ************************/
2188 * sd_lb_stats - Structure to store the statistics of a sched_domain
2189 * during load balancing.
2191 struct sd_lb_stats {
2192 struct sched_group *busiest; /* Busiest group in this sd */
2193 struct sched_group *this; /* Local group in this sd */
2194 unsigned long total_load; /* Total load of all groups in sd */
2195 unsigned long total_pwr; /* Total power of all groups in sd */
2196 unsigned long avg_load; /* Average load across all groups in sd */
2198 /** Statistics of this group */
2199 unsigned long this_load;
2200 unsigned long this_load_per_task;
2201 unsigned long this_nr_running;
2202 unsigned long this_has_capacity;
2204 /* Statistics of the busiest group */
2205 unsigned long max_load;
2206 unsigned long busiest_load_per_task;
2207 unsigned long busiest_nr_running;
2208 unsigned long busiest_group_capacity;
2209 unsigned long busiest_has_capacity;
2211 int group_imb; /* Is there imbalance in this sd */
2212 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2213 int power_savings_balance; /* Is powersave balance needed for this sd */
2214 struct sched_group *group_min; /* Least loaded group in sd */
2215 struct sched_group *group_leader; /* Group which relieves group_min */
2216 unsigned long min_load_per_task; /* load_per_task in group_min */
2217 unsigned long leader_nr_running; /* Nr running of group_leader */
2218 unsigned long min_nr_running; /* Nr running of group_min */
2223 * sg_lb_stats - stats of a sched_group required for load_balancing
2225 struct sg_lb_stats {
2226 unsigned long avg_load; /*Avg load across the CPUs of the group */
2227 unsigned long group_load; /* Total load over the CPUs of the group */
2228 unsigned long sum_nr_running; /* Nr tasks running in the group */
2229 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2230 unsigned long group_capacity;
2231 int group_imb; /* Is there an imbalance in the group ? */
2232 int group_has_capacity; /* Is there extra capacity in the group? */
2236 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2237 * @group: The group whose first cpu is to be returned.
2239 static inline unsigned int group_first_cpu(struct sched_group *group)
2241 return cpumask_first(sched_group_cpus(group));
2245 * get_sd_load_idx - Obtain the load index for a given sched domain.
2246 * @sd: The sched_domain whose load_idx is to be obtained.
2247 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2249 static inline int get_sd_load_idx(struct sched_domain *sd,
2250 enum cpu_idle_type idle)
2256 load_idx = sd->busy_idx;
2259 case CPU_NEWLY_IDLE:
2260 load_idx = sd->newidle_idx;
2263 load_idx = sd->idle_idx;
2271 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2273 * init_sd_power_savings_stats - Initialize power savings statistics for
2274 * the given sched_domain, during load balancing.
2276 * @sd: Sched domain whose power-savings statistics are to be initialized.
2277 * @sds: Variable containing the statistics for sd.
2278 * @idle: Idle status of the CPU at which we're performing load-balancing.
2280 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2281 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2284 * Busy processors will not participate in power savings
2287 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2288 sds->power_savings_balance = 0;
2290 sds->power_savings_balance = 1;
2291 sds->min_nr_running = ULONG_MAX;
2292 sds->leader_nr_running = 0;
2297 * update_sd_power_savings_stats - Update the power saving stats for a
2298 * sched_domain while performing load balancing.
2300 * @group: sched_group belonging to the sched_domain under consideration.
2301 * @sds: Variable containing the statistics of the sched_domain
2302 * @local_group: Does group contain the CPU for which we're performing
2304 * @sgs: Variable containing the statistics of the group.
2306 static inline void update_sd_power_savings_stats(struct sched_group *group,
2307 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2310 if (!sds->power_savings_balance)
2314 * If the local group is idle or completely loaded
2315 * no need to do power savings balance at this domain
2317 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2318 !sds->this_nr_running))
2319 sds->power_savings_balance = 0;
2322 * If a group is already running at full capacity or idle,
2323 * don't include that group in power savings calculations
2325 if (!sds->power_savings_balance ||
2326 sgs->sum_nr_running >= sgs->group_capacity ||
2327 !sgs->sum_nr_running)
2331 * Calculate the group which has the least non-idle load.
2332 * This is the group from where we need to pick up the load
2335 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2336 (sgs->sum_nr_running == sds->min_nr_running &&
2337 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2338 sds->group_min = group;
2339 sds->min_nr_running = sgs->sum_nr_running;
2340 sds->min_load_per_task = sgs->sum_weighted_load /
2341 sgs->sum_nr_running;
2345 * Calculate the group which is almost near its
2346 * capacity but still has some space to pick up some load
2347 * from other group and save more power
2349 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2352 if (sgs->sum_nr_running > sds->leader_nr_running ||
2353 (sgs->sum_nr_running == sds->leader_nr_running &&
2354 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2355 sds->group_leader = group;
2356 sds->leader_nr_running = sgs->sum_nr_running;
2361 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2362 * @sds: Variable containing the statistics of the sched_domain
2363 * under consideration.
2364 * @this_cpu: Cpu at which we're currently performing load-balancing.
2365 * @imbalance: Variable to store the imbalance.
2368 * Check if we have potential to perform some power-savings balance.
2369 * If yes, set the busiest group to be the least loaded group in the
2370 * sched_domain, so that it's CPUs can be put to idle.
2372 * Returns 1 if there is potential to perform power-savings balance.
2375 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2376 int this_cpu, unsigned long *imbalance)
2378 if (!sds->power_savings_balance)
2381 if (sds->this != sds->group_leader ||
2382 sds->group_leader == sds->group_min)
2385 *imbalance = sds->min_load_per_task;
2386 sds->busiest = sds->group_min;
2391 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2392 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2393 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2398 static inline void update_sd_power_savings_stats(struct sched_group *group,
2399 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2404 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2405 int this_cpu, unsigned long *imbalance)
2409 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2412 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2414 return SCHED_LOAD_SCALE;
2417 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2419 return default_scale_freq_power(sd, cpu);
2422 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2424 unsigned long weight = sd->span_weight;
2425 unsigned long smt_gain = sd->smt_gain;
2432 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2434 return default_scale_smt_power(sd, cpu);
2437 unsigned long scale_rt_power(int cpu)
2439 struct rq *rq = cpu_rq(cpu);
2440 u64 total, available;
2442 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2444 if (unlikely(total < rq->rt_avg)) {
2445 /* Ensures that power won't end up being negative */
2448 available = total - rq->rt_avg;
2451 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2452 total = SCHED_LOAD_SCALE;
2454 total >>= SCHED_LOAD_SHIFT;
2456 return div_u64(available, total);
2459 static void update_cpu_power(struct sched_domain *sd, int cpu)
2461 unsigned long weight = sd->span_weight;
2462 unsigned long power = SCHED_LOAD_SCALE;
2463 struct sched_group *sdg = sd->groups;
2465 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2466 if (sched_feat(ARCH_POWER))
2467 power *= arch_scale_smt_power(sd, cpu);
2469 power *= default_scale_smt_power(sd, cpu);
2471 power >>= SCHED_LOAD_SHIFT;
2474 sdg->cpu_power_orig = power;
2476 if (sched_feat(ARCH_POWER))
2477 power *= arch_scale_freq_power(sd, cpu);
2479 power *= default_scale_freq_power(sd, cpu);
2481 power >>= SCHED_LOAD_SHIFT;
2483 power *= scale_rt_power(cpu);
2484 power >>= SCHED_LOAD_SHIFT;
2489 cpu_rq(cpu)->cpu_power = power;
2490 sdg->cpu_power = power;
2493 static void update_group_power(struct sched_domain *sd, int cpu)
2495 struct sched_domain *child = sd->child;
2496 struct sched_group *group, *sdg = sd->groups;
2497 unsigned long power;
2500 update_cpu_power(sd, cpu);
2506 group = child->groups;
2508 power += group->cpu_power;
2509 group = group->next;
2510 } while (group != child->groups);
2512 sdg->cpu_power = power;
2516 * Try and fix up capacity for tiny siblings, this is needed when
2517 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2518 * which on its own isn't powerful enough.
2520 * See update_sd_pick_busiest() and check_asym_packing().
2523 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2526 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2528 if (sd->level != SD_LV_SIBLING)
2532 * If ~90% of the cpu_power is still there, we're good.
2534 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2541 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2542 * @sd: The sched_domain whose statistics are to be updated.
2543 * @group: sched_group whose statistics are to be updated.
2544 * @this_cpu: Cpu for which load balance is currently performed.
2545 * @idle: Idle status of this_cpu
2546 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2547 * @sd_idle: Idle status of the sched_domain containing group.
2548 * @local_group: Does group contain this_cpu.
2549 * @cpus: Set of cpus considered for load balancing.
2550 * @balance: Should we balance.
2551 * @sgs: variable to hold the statistics for this group.
2553 static inline void update_sg_lb_stats(struct sched_domain *sd,
2554 struct sched_group *group, int this_cpu,
2555 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2556 int local_group, const struct cpumask *cpus,
2557 int *balance, struct sg_lb_stats *sgs)
2559 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2561 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2562 unsigned long avg_load_per_task = 0;
2565 balance_cpu = group_first_cpu(group);
2567 /* Tally up the load of all CPUs in the group */
2569 min_cpu_load = ~0UL;
2572 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2573 struct rq *rq = cpu_rq(i);
2575 if (*sd_idle && rq->nr_running)
2578 /* Bias balancing toward cpus of our domain */
2580 if (idle_cpu(i) && !first_idle_cpu) {
2585 load = target_load(i, load_idx);
2587 load = source_load(i, load_idx);
2588 if (load > max_cpu_load) {
2589 max_cpu_load = load;
2590 max_nr_running = rq->nr_running;
2592 if (min_cpu_load > load)
2593 min_cpu_load = load;
2596 sgs->group_load += load;
2597 sgs->sum_nr_running += rq->nr_running;
2598 sgs->sum_weighted_load += weighted_cpuload(i);
2603 * First idle cpu or the first cpu(busiest) in this sched group
2604 * is eligible for doing load balancing at this and above
2605 * domains. In the newly idle case, we will allow all the cpu's
2606 * to do the newly idle load balance.
2608 if (idle != CPU_NEWLY_IDLE && local_group) {
2609 if (balance_cpu != this_cpu) {
2613 update_group_power(sd, this_cpu);
2616 /* Adjust by relative CPU power of the group */
2617 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2620 * Consider the group unbalanced when the imbalance is larger
2621 * than the average weight of two tasks.
2623 * APZ: with cgroup the avg task weight can vary wildly and
2624 * might not be a suitable number - should we keep a
2625 * normalized nr_running number somewhere that negates
2628 if (sgs->sum_nr_running)
2629 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2631 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2634 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2635 if (!sgs->group_capacity)
2636 sgs->group_capacity = fix_small_capacity(sd, group);
2638 if (sgs->group_capacity > sgs->sum_nr_running)
2639 sgs->group_has_capacity = 1;
2643 * update_sd_pick_busiest - return 1 on busiest group
2644 * @sd: sched_domain whose statistics are to be checked
2645 * @sds: sched_domain statistics
2646 * @sg: sched_group candidate to be checked for being the busiest
2647 * @sgs: sched_group statistics
2648 * @this_cpu: the current cpu
2650 * Determine if @sg is a busier group than the previously selected
2653 static bool update_sd_pick_busiest(struct sched_domain *sd,
2654 struct sd_lb_stats *sds,
2655 struct sched_group *sg,
2656 struct sg_lb_stats *sgs,
2659 if (sgs->avg_load <= sds->max_load)
2662 if (sgs->sum_nr_running > sgs->group_capacity)
2669 * ASYM_PACKING needs to move all the work to the lowest
2670 * numbered CPUs in the group, therefore mark all groups
2671 * higher than ourself as busy.
2673 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2674 this_cpu < group_first_cpu(sg)) {
2678 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2686 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2687 * @sd: sched_domain whose statistics are to be updated.
2688 * @this_cpu: Cpu for which load balance is currently performed.
2689 * @idle: Idle status of this_cpu
2690 * @sd_idle: Idle status of the sched_domain containing sg.
2691 * @cpus: Set of cpus considered for load balancing.
2692 * @balance: Should we balance.
2693 * @sds: variable to hold the statistics for this sched_domain.
2695 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2696 enum cpu_idle_type idle, int *sd_idle,
2697 const struct cpumask *cpus, int *balance,
2698 struct sd_lb_stats *sds)
2700 struct sched_domain *child = sd->child;
2701 struct sched_group *sg = sd->groups;
2702 struct sg_lb_stats sgs;
2703 int load_idx, prefer_sibling = 0;
2705 if (child && child->flags & SD_PREFER_SIBLING)
2708 init_sd_power_savings_stats(sd, sds, idle);
2709 load_idx = get_sd_load_idx(sd, idle);
2714 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2715 memset(&sgs, 0, sizeof(sgs));
2716 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2717 local_group, cpus, balance, &sgs);
2719 if (local_group && !(*balance))
2722 sds->total_load += sgs.group_load;
2723 sds->total_pwr += sg->cpu_power;
2726 * In case the child domain prefers tasks go to siblings
2727 * first, lower the sg capacity to one so that we'll try
2728 * and move all the excess tasks away. We lower the capacity
2729 * of a group only if the local group has the capacity to fit
2730 * these excess tasks, i.e. nr_running < group_capacity. The
2731 * extra check prevents the case where you always pull from the
2732 * heaviest group when it is already under-utilized (possible
2733 * with a large weight task outweighs the tasks on the system).
2735 if (prefer_sibling && !local_group && sds->this_has_capacity)
2736 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2739 sds->this_load = sgs.avg_load;
2741 sds->this_nr_running = sgs.sum_nr_running;
2742 sds->this_load_per_task = sgs.sum_weighted_load;
2743 sds->this_has_capacity = sgs.group_has_capacity;
2744 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2745 sds->max_load = sgs.avg_load;
2747 sds->busiest_nr_running = sgs.sum_nr_running;
2748 sds->busiest_group_capacity = sgs.group_capacity;
2749 sds->busiest_load_per_task = sgs.sum_weighted_load;
2750 sds->busiest_has_capacity = sgs.group_has_capacity;
2751 sds->group_imb = sgs.group_imb;
2754 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2756 } while (sg != sd->groups);
2759 int __weak arch_sd_sibling_asym_packing(void)
2761 return 0*SD_ASYM_PACKING;
2765 * check_asym_packing - Check to see if the group is packed into the
2768 * This is primarily intended to used at the sibling level. Some
2769 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2770 * case of POWER7, it can move to lower SMT modes only when higher
2771 * threads are idle. When in lower SMT modes, the threads will
2772 * perform better since they share less core resources. Hence when we
2773 * have idle threads, we want them to be the higher ones.
2775 * This packing function is run on idle threads. It checks to see if
2776 * the busiest CPU in this domain (core in the P7 case) has a higher
2777 * CPU number than the packing function is being run on. Here we are
2778 * assuming lower CPU number will be equivalent to lower a SMT thread
2781 * Returns 1 when packing is required and a task should be moved to
2782 * this CPU. The amount of the imbalance is returned in *imbalance.
2784 * @sd: The sched_domain whose packing is to be checked.
2785 * @sds: Statistics of the sched_domain which is to be packed
2786 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2787 * @imbalance: returns amount of imbalanced due to packing.
2789 static int check_asym_packing(struct sched_domain *sd,
2790 struct sd_lb_stats *sds,
2791 int this_cpu, unsigned long *imbalance)
2795 if (!(sd->flags & SD_ASYM_PACKING))
2801 busiest_cpu = group_first_cpu(sds->busiest);
2802 if (this_cpu > busiest_cpu)
2805 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2811 * fix_small_imbalance - Calculate the minor imbalance that exists
2812 * amongst the groups of a sched_domain, during
2814 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2815 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2816 * @imbalance: Variable to store the imbalance.
2818 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2819 int this_cpu, unsigned long *imbalance)
2821 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2822 unsigned int imbn = 2;
2823 unsigned long scaled_busy_load_per_task;
2825 if (sds->this_nr_running) {
2826 sds->this_load_per_task /= sds->this_nr_running;
2827 if (sds->busiest_load_per_task >
2828 sds->this_load_per_task)
2831 sds->this_load_per_task =
2832 cpu_avg_load_per_task(this_cpu);
2834 scaled_busy_load_per_task = sds->busiest_load_per_task
2836 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2838 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2839 (scaled_busy_load_per_task * imbn)) {
2840 *imbalance = sds->busiest_load_per_task;
2845 * OK, we don't have enough imbalance to justify moving tasks,
2846 * however we may be able to increase total CPU power used by
2850 pwr_now += sds->busiest->cpu_power *
2851 min(sds->busiest_load_per_task, sds->max_load);
2852 pwr_now += sds->this->cpu_power *
2853 min(sds->this_load_per_task, sds->this_load);
2854 pwr_now /= SCHED_LOAD_SCALE;
2856 /* Amount of load we'd subtract */
2857 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2858 sds->busiest->cpu_power;
2859 if (sds->max_load > tmp)
2860 pwr_move += sds->busiest->cpu_power *
2861 min(sds->busiest_load_per_task, sds->max_load - tmp);
2863 /* Amount of load we'd add */
2864 if (sds->max_load * sds->busiest->cpu_power <
2865 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2866 tmp = (sds->max_load * sds->busiest->cpu_power) /
2867 sds->this->cpu_power;
2869 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2870 sds->this->cpu_power;
2871 pwr_move += sds->this->cpu_power *
2872 min(sds->this_load_per_task, sds->this_load + tmp);
2873 pwr_move /= SCHED_LOAD_SCALE;
2875 /* Move if we gain throughput */
2876 if (pwr_move > pwr_now)
2877 *imbalance = sds->busiest_load_per_task;
2881 * calculate_imbalance - Calculate the amount of imbalance present within the
2882 * groups of a given sched_domain during load balance.
2883 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2884 * @this_cpu: Cpu for which currently load balance is being performed.
2885 * @imbalance: The variable to store the imbalance.
2887 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2888 unsigned long *imbalance)
2890 unsigned long max_pull, load_above_capacity = ~0UL;
2892 sds->busiest_load_per_task /= sds->busiest_nr_running;
2893 if (sds->group_imb) {
2894 sds->busiest_load_per_task =
2895 min(sds->busiest_load_per_task, sds->avg_load);
2899 * In the presence of smp nice balancing, certain scenarios can have
2900 * max load less than avg load(as we skip the groups at or below
2901 * its cpu_power, while calculating max_load..)
2903 if (sds->max_load < sds->avg_load) {
2905 return fix_small_imbalance(sds, this_cpu, imbalance);
2908 if (!sds->group_imb) {
2910 * Don't want to pull so many tasks that a group would go idle.
2912 load_above_capacity = (sds->busiest_nr_running -
2913 sds->busiest_group_capacity);
2915 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2917 load_above_capacity /= sds->busiest->cpu_power;
2921 * We're trying to get all the cpus to the average_load, so we don't
2922 * want to push ourselves above the average load, nor do we wish to
2923 * reduce the max loaded cpu below the average load. At the same time,
2924 * we also don't want to reduce the group load below the group capacity
2925 * (so that we can implement power-savings policies etc). Thus we look
2926 * for the minimum possible imbalance.
2927 * Be careful of negative numbers as they'll appear as very large values
2928 * with unsigned longs.
2930 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2932 /* How much load to actually move to equalise the imbalance */
2933 *imbalance = min(max_pull * sds->busiest->cpu_power,
2934 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
2938 * if *imbalance is less than the average load per runnable task
2939 * there is no gaurantee that any tasks will be moved so we'll have
2940 * a think about bumping its value to force at least one task to be
2943 if (*imbalance < sds->busiest_load_per_task)
2944 return fix_small_imbalance(sds, this_cpu, imbalance);
2948 /******* find_busiest_group() helpers end here *********************/
2951 * find_busiest_group - Returns the busiest group within the sched_domain
2952 * if there is an imbalance. If there isn't an imbalance, and
2953 * the user has opted for power-savings, it returns a group whose
2954 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2955 * such a group exists.
2957 * Also calculates the amount of weighted load which should be moved
2958 * to restore balance.
2960 * @sd: The sched_domain whose busiest group is to be returned.
2961 * @this_cpu: The cpu for which load balancing is currently being performed.
2962 * @imbalance: Variable which stores amount of weighted load which should
2963 * be moved to restore balance/put a group to idle.
2964 * @idle: The idle status of this_cpu.
2965 * @sd_idle: The idleness of sd
2966 * @cpus: The set of CPUs under consideration for load-balancing.
2967 * @balance: Pointer to a variable indicating if this_cpu
2968 * is the appropriate cpu to perform load balancing at this_level.
2970 * Returns: - the busiest group if imbalance exists.
2971 * - If no imbalance and user has opted for power-savings balance,
2972 * return the least loaded group whose CPUs can be
2973 * put to idle by rebalancing its tasks onto our group.
2975 static struct sched_group *
2976 find_busiest_group(struct sched_domain *sd, int this_cpu,
2977 unsigned long *imbalance, enum cpu_idle_type idle,
2978 int *sd_idle, const struct cpumask *cpus, int *balance)
2980 struct sd_lb_stats sds;
2982 memset(&sds, 0, sizeof(sds));
2985 * Compute the various statistics relavent for load balancing at
2988 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
2991 /* Cases where imbalance does not exist from POV of this_cpu */
2992 /* 1) this_cpu is not the appropriate cpu to perform load balancing
2994 * 2) There is no busy sibling group to pull from.
2995 * 3) This group is the busiest group.
2996 * 4) This group is more busy than the avg busieness at this
2998 * 5) The imbalance is within the specified limit.
3000 * Note: when doing newidle balance, if the local group has excess
3001 * capacity (i.e. nr_running < group_capacity) and the busiest group
3002 * does not have any capacity, we force a load balance to pull tasks
3003 * to the local group. In this case, we skip past checks 3, 4 and 5.
3008 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3009 check_asym_packing(sd, &sds, this_cpu, imbalance))
3012 if (!sds.busiest || sds.busiest_nr_running == 0)
3015 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3016 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3017 !sds.busiest_has_capacity)
3020 if (sds.this_load >= sds.max_load)
3023 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3025 if (sds.this_load >= sds.avg_load)
3028 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3032 /* Looks like there is an imbalance. Compute it */
3033 calculate_imbalance(&sds, this_cpu, imbalance);
3038 * There is no obvious imbalance. But check if we can do some balancing
3041 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3049 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3052 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3053 enum cpu_idle_type idle, unsigned long imbalance,
3054 const struct cpumask *cpus)
3056 struct rq *busiest = NULL, *rq;
3057 unsigned long max_load = 0;
3060 for_each_cpu(i, sched_group_cpus(group)) {
3061 unsigned long power = power_of(i);
3062 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3066 capacity = fix_small_capacity(sd, group);
3068 if (!cpumask_test_cpu(i, cpus))
3072 wl = weighted_cpuload(i);
3075 * When comparing with imbalance, use weighted_cpuload()
3076 * which is not scaled with the cpu power.
3078 if (capacity && rq->nr_running == 1 && wl > imbalance)
3082 * For the load comparisons with the other cpu's, consider
3083 * the weighted_cpuload() scaled with the cpu power, so that
3084 * the load can be moved away from the cpu that is potentially
3085 * running at a lower capacity.
3087 wl = (wl * SCHED_LOAD_SCALE) / power;
3089 if (wl > max_load) {
3099 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3100 * so long as it is large enough.
3102 #define MAX_PINNED_INTERVAL 512
3104 /* Working cpumask for load_balance and load_balance_newidle. */
3105 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3107 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
3108 int busiest_cpu, int this_cpu)
3110 if (idle == CPU_NEWLY_IDLE) {
3113 * ASYM_PACKING needs to force migrate tasks from busy but
3114 * higher numbered CPUs in order to pack all tasks in the
3115 * lowest numbered CPUs.
3117 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3121 * The only task running in a non-idle cpu can be moved to this
3122 * cpu in an attempt to completely freeup the other CPU
3125 * The package power saving logic comes from
3126 * find_busiest_group(). If there are no imbalance, then
3127 * f_b_g() will return NULL. However when sched_mc={1,2} then
3128 * f_b_g() will select a group from which a running task may be
3129 * pulled to this cpu in order to make the other package idle.
3130 * If there is no opportunity to make a package idle and if
3131 * there are no imbalance, then f_b_g() will return NULL and no
3132 * action will be taken in load_balance_newidle().
3134 * Under normal task pull operation due to imbalance, there
3135 * will be more than one task in the source run queue and
3136 * move_tasks() will succeed. ld_moved will be true and this
3137 * active balance code will not be triggered.
3139 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3140 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3143 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3147 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3150 static int active_load_balance_cpu_stop(void *data);
3153 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3154 * tasks if there is an imbalance.
3156 static int load_balance(int this_cpu, struct rq *this_rq,
3157 struct sched_domain *sd, enum cpu_idle_type idle,
3160 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3161 struct sched_group *group;
3162 unsigned long imbalance;
3164 unsigned long flags;
3165 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3167 cpumask_copy(cpus, cpu_active_mask);
3170 * When power savings policy is enabled for the parent domain, idle
3171 * sibling can pick up load irrespective of busy siblings. In this case,
3172 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3173 * portraying it as CPU_NOT_IDLE.
3175 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3176 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3179 schedstat_inc(sd, lb_count[idle]);
3182 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3189 schedstat_inc(sd, lb_nobusyg[idle]);
3193 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3195 schedstat_inc(sd, lb_nobusyq[idle]);
3199 BUG_ON(busiest == this_rq);
3201 schedstat_add(sd, lb_imbalance[idle], imbalance);
3204 if (busiest->nr_running > 1) {
3206 * Attempt to move tasks. If find_busiest_group has found
3207 * an imbalance but busiest->nr_running <= 1, the group is
3208 * still unbalanced. ld_moved simply stays zero, so it is
3209 * correctly treated as an imbalance.
3211 local_irq_save(flags);
3212 double_rq_lock(this_rq, busiest);
3213 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3214 imbalance, sd, idle, &all_pinned);
3215 double_rq_unlock(this_rq, busiest);
3216 local_irq_restore(flags);
3219 * some other cpu did the load balance for us.
3221 if (ld_moved && this_cpu != smp_processor_id())
3222 resched_cpu(this_cpu);
3224 /* All tasks on this runqueue were pinned by CPU affinity */
3225 if (unlikely(all_pinned)) {
3226 cpumask_clear_cpu(cpu_of(busiest), cpus);
3227 if (!cpumask_empty(cpus))
3234 schedstat_inc(sd, lb_failed[idle]);
3236 * Increment the failure counter only on periodic balance.
3237 * We do not want newidle balance, which can be very
3238 * frequent, pollute the failure counter causing
3239 * excessive cache_hot migrations and active balances.
3241 if (idle != CPU_NEWLY_IDLE)
3242 sd->nr_balance_failed++;
3244 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3246 raw_spin_lock_irqsave(&busiest->lock, flags);
3248 /* don't kick the active_load_balance_cpu_stop,
3249 * if the curr task on busiest cpu can't be
3252 if (!cpumask_test_cpu(this_cpu,
3253 &busiest->curr->cpus_allowed)) {
3254 raw_spin_unlock_irqrestore(&busiest->lock,
3257 goto out_one_pinned;
3261 * ->active_balance synchronizes accesses to
3262 * ->active_balance_work. Once set, it's cleared
3263 * only after active load balance is finished.
3265 if (!busiest->active_balance) {
3266 busiest->active_balance = 1;
3267 busiest->push_cpu = this_cpu;
3270 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3273 stop_one_cpu_nowait(cpu_of(busiest),
3274 active_load_balance_cpu_stop, busiest,
3275 &busiest->active_balance_work);
3278 * We've kicked active balancing, reset the failure
3281 sd->nr_balance_failed = sd->cache_nice_tries+1;
3284 sd->nr_balance_failed = 0;
3286 if (likely(!active_balance)) {
3287 /* We were unbalanced, so reset the balancing interval */
3288 sd->balance_interval = sd->min_interval;
3291 * If we've begun active balancing, start to back off. This
3292 * case may not be covered by the all_pinned logic if there
3293 * is only 1 task on the busy runqueue (because we don't call
3296 if (sd->balance_interval < sd->max_interval)
3297 sd->balance_interval *= 2;
3300 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3301 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3307 schedstat_inc(sd, lb_balanced[idle]);
3309 sd->nr_balance_failed = 0;
3312 /* tune up the balancing interval */
3313 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3314 (sd->balance_interval < sd->max_interval))
3315 sd->balance_interval *= 2;
3317 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3318 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3327 * idle_balance is called by schedule() if this_cpu is about to become
3328 * idle. Attempts to pull tasks from other CPUs.
3330 static void idle_balance(int this_cpu, struct rq *this_rq)
3332 struct sched_domain *sd;
3333 int pulled_task = 0;
3334 unsigned long next_balance = jiffies + HZ;
3336 this_rq->idle_stamp = this_rq->clock;
3338 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3342 * Drop the rq->lock, but keep IRQ/preempt disabled.
3344 raw_spin_unlock(&this_rq->lock);
3346 update_shares(this_cpu);
3347 for_each_domain(this_cpu, sd) {
3348 unsigned long interval;
3351 if (!(sd->flags & SD_LOAD_BALANCE))
3354 if (sd->flags & SD_BALANCE_NEWIDLE) {
3355 /* If we've pulled tasks over stop searching: */
3356 pulled_task = load_balance(this_cpu, this_rq,
3357 sd, CPU_NEWLY_IDLE, &balance);
3360 interval = msecs_to_jiffies(sd->balance_interval);
3361 if (time_after(next_balance, sd->last_balance + interval))
3362 next_balance = sd->last_balance + interval;
3367 raw_spin_lock(&this_rq->lock);
3369 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3371 * We are going idle. next_balance may be set based on
3372 * a busy processor. So reset next_balance.
3374 this_rq->next_balance = next_balance;
3379 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3380 * running tasks off the busiest CPU onto idle CPUs. It requires at
3381 * least 1 task to be running on each physical CPU where possible, and
3382 * avoids physical / logical imbalances.
3384 static int active_load_balance_cpu_stop(void *data)
3386 struct rq *busiest_rq = data;
3387 int busiest_cpu = cpu_of(busiest_rq);
3388 int target_cpu = busiest_rq->push_cpu;
3389 struct rq *target_rq = cpu_rq(target_cpu);
3390 struct sched_domain *sd;
3392 raw_spin_lock_irq(&busiest_rq->lock);
3394 /* make sure the requested cpu hasn't gone down in the meantime */
3395 if (unlikely(busiest_cpu != smp_processor_id() ||
3396 !busiest_rq->active_balance))
3399 /* Is there any task to move? */
3400 if (busiest_rq->nr_running <= 1)
3404 * This condition is "impossible", if it occurs
3405 * we need to fix it. Originally reported by
3406 * Bjorn Helgaas on a 128-cpu setup.
3408 BUG_ON(busiest_rq == target_rq);
3410 /* move a task from busiest_rq to target_rq */
3411 double_lock_balance(busiest_rq, target_rq);
3413 /* Search for an sd spanning us and the target CPU. */
3414 for_each_domain(target_cpu, sd) {
3415 if ((sd->flags & SD_LOAD_BALANCE) &&
3416 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3421 schedstat_inc(sd, alb_count);
3423 if (move_one_task(target_rq, target_cpu, busiest_rq,
3425 schedstat_inc(sd, alb_pushed);
3427 schedstat_inc(sd, alb_failed);
3429 double_unlock_balance(busiest_rq, target_rq);
3431 busiest_rq->active_balance = 0;
3432 raw_spin_unlock_irq(&busiest_rq->lock);
3438 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3440 static void trigger_sched_softirq(void *data)
3442 raise_softirq_irqoff(SCHED_SOFTIRQ);
3445 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3447 csd->func = trigger_sched_softirq;
3454 * idle load balancing details
3455 * - One of the idle CPUs nominates itself as idle load_balancer, while
3457 * - This idle load balancer CPU will also go into tickless mode when
3458 * it is idle, just like all other idle CPUs
3459 * - When one of the busy CPUs notice that there may be an idle rebalancing
3460 * needed, they will kick the idle load balancer, which then does idle
3461 * load balancing for all the idle CPUs.
3464 atomic_t load_balancer;
3465 atomic_t first_pick_cpu;
3466 atomic_t second_pick_cpu;
3467 cpumask_var_t idle_cpus_mask;
3468 cpumask_var_t grp_idle_mask;
3469 unsigned long next_balance; /* in jiffy units */
3470 } nohz ____cacheline_aligned;
3472 int get_nohz_load_balancer(void)
3474 return atomic_read(&nohz.load_balancer);
3477 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3479 * lowest_flag_domain - Return lowest sched_domain containing flag.
3480 * @cpu: The cpu whose lowest level of sched domain is to
3482 * @flag: The flag to check for the lowest sched_domain
3483 * for the given cpu.
3485 * Returns the lowest sched_domain of a cpu which contains the given flag.
3487 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3489 struct sched_domain *sd;
3491 for_each_domain(cpu, sd)
3492 if (sd && (sd->flags & flag))
3499 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3500 * @cpu: The cpu whose domains we're iterating over.
3501 * @sd: variable holding the value of the power_savings_sd
3503 * @flag: The flag to filter the sched_domains to be iterated.
3505 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3506 * set, starting from the lowest sched_domain to the highest.
3508 #define for_each_flag_domain(cpu, sd, flag) \
3509 for (sd = lowest_flag_domain(cpu, flag); \
3510 (sd && (sd->flags & flag)); sd = sd->parent)
3513 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3514 * @ilb_group: group to be checked for semi-idleness
3516 * Returns: 1 if the group is semi-idle. 0 otherwise.
3518 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3519 * and atleast one non-idle CPU. This helper function checks if the given
3520 * sched_group is semi-idle or not.
3522 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3524 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3525 sched_group_cpus(ilb_group));
3528 * A sched_group is semi-idle when it has atleast one busy cpu
3529 * and atleast one idle cpu.
3531 if (cpumask_empty(nohz.grp_idle_mask))
3534 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3540 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3541 * @cpu: The cpu which is nominating a new idle_load_balancer.
3543 * Returns: Returns the id of the idle load balancer if it exists,
3544 * Else, returns >= nr_cpu_ids.
3546 * This algorithm picks the idle load balancer such that it belongs to a
3547 * semi-idle powersavings sched_domain. The idea is to try and avoid
3548 * completely idle packages/cores just for the purpose of idle load balancing
3549 * when there are other idle cpu's which are better suited for that job.
3551 static int find_new_ilb(int cpu)
3553 struct sched_domain *sd;
3554 struct sched_group *ilb_group;
3557 * Have idle load balancer selection from semi-idle packages only
3558 * when power-aware load balancing is enabled
3560 if (!(sched_smt_power_savings || sched_mc_power_savings))
3564 * Optimize for the case when we have no idle CPUs or only one
3565 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3567 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3570 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3571 ilb_group = sd->groups;
3574 if (is_semi_idle_group(ilb_group))
3575 return cpumask_first(nohz.grp_idle_mask);
3577 ilb_group = ilb_group->next;
3579 } while (ilb_group != sd->groups);
3585 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3586 static inline int find_new_ilb(int call_cpu)
3593 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3594 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3595 * CPU (if there is one).
3597 static void nohz_balancer_kick(int cpu)
3601 nohz.next_balance++;
3603 ilb_cpu = get_nohz_load_balancer();
3605 if (ilb_cpu >= nr_cpu_ids) {
3606 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3607 if (ilb_cpu >= nr_cpu_ids)
3611 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3612 struct call_single_data *cp;
3614 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3615 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3616 __smp_call_function_single(ilb_cpu, cp, 0);
3622 * This routine will try to nominate the ilb (idle load balancing)
3623 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3624 * load balancing on behalf of all those cpus.
3626 * When the ilb owner becomes busy, we will not have new ilb owner until some
3627 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3628 * idle load balancing by kicking one of the idle CPUs.
3630 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3631 * ilb owner CPU in future (when there is a need for idle load balancing on
3632 * behalf of all idle CPUs).
3634 void select_nohz_load_balancer(int stop_tick)
3636 int cpu = smp_processor_id();
3639 if (!cpu_active(cpu)) {
3640 if (atomic_read(&nohz.load_balancer) != cpu)
3644 * If we are going offline and still the leader,
3647 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3654 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3656 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3657 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3658 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3659 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3661 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3664 /* make me the ilb owner */
3665 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3670 * Check to see if there is a more power-efficient
3673 new_ilb = find_new_ilb(cpu);
3674 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3675 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3676 resched_cpu(new_ilb);
3682 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3685 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3687 if (atomic_read(&nohz.load_balancer) == cpu)
3688 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3696 static DEFINE_SPINLOCK(balancing);
3699 * It checks each scheduling domain to see if it is due to be balanced,
3700 * and initiates a balancing operation if so.
3702 * Balancing parameters are set up in arch_init_sched_domains.
3704 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3707 struct rq *rq = cpu_rq(cpu);
3708 unsigned long interval;
3709 struct sched_domain *sd;
3710 /* Earliest time when we have to do rebalance again */
3711 unsigned long next_balance = jiffies + 60*HZ;
3712 int update_next_balance = 0;
3717 for_each_domain(cpu, sd) {
3718 if (!(sd->flags & SD_LOAD_BALANCE))
3721 interval = sd->balance_interval;
3722 if (idle != CPU_IDLE)
3723 interval *= sd->busy_factor;
3725 /* scale ms to jiffies */
3726 interval = msecs_to_jiffies(interval);
3727 if (unlikely(!interval))
3729 if (interval > HZ*NR_CPUS/10)
3730 interval = HZ*NR_CPUS/10;
3732 need_serialize = sd->flags & SD_SERIALIZE;
3734 if (need_serialize) {
3735 if (!spin_trylock(&balancing))
3739 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3740 if (load_balance(cpu, rq, sd, idle, &balance)) {
3742 * We've pulled tasks over so either we're no
3743 * longer idle, or one of our SMT siblings is
3746 idle = CPU_NOT_IDLE;
3748 sd->last_balance = jiffies;
3751 spin_unlock(&balancing);
3753 if (time_after(next_balance, sd->last_balance + interval)) {
3754 next_balance = sd->last_balance + interval;
3755 update_next_balance = 1;
3759 * Stop the load balance at this level. There is another
3760 * CPU in our sched group which is doing load balancing more
3768 * next_balance will be updated only when there is a need.
3769 * When the cpu is attached to null domain for ex, it will not be
3772 if (likely(update_next_balance))
3773 rq->next_balance = next_balance;
3778 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3779 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3781 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3783 struct rq *this_rq = cpu_rq(this_cpu);
3787 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3790 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3791 if (balance_cpu == this_cpu)
3795 * If this cpu gets work to do, stop the load balancing
3796 * work being done for other cpus. Next load
3797 * balancing owner will pick it up.
3799 if (need_resched()) {
3800 this_rq->nohz_balance_kick = 0;
3804 raw_spin_lock_irq(&this_rq->lock);
3805 update_rq_clock(this_rq);
3806 update_cpu_load(this_rq);
3807 raw_spin_unlock_irq(&this_rq->lock);
3809 rebalance_domains(balance_cpu, CPU_IDLE);
3811 rq = cpu_rq(balance_cpu);
3812 if (time_after(this_rq->next_balance, rq->next_balance))
3813 this_rq->next_balance = rq->next_balance;
3815 nohz.next_balance = this_rq->next_balance;
3816 this_rq->nohz_balance_kick = 0;
3820 * Current heuristic for kicking the idle load balancer
3821 * - first_pick_cpu is the one of the busy CPUs. It will kick
3822 * idle load balancer when it has more than one process active. This
3823 * eliminates the need for idle load balancing altogether when we have
3824 * only one running process in the system (common case).
3825 * - If there are more than one busy CPU, idle load balancer may have
3826 * to run for active_load_balance to happen (i.e., two busy CPUs are
3827 * SMT or core siblings and can run better if they move to different
3828 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3829 * which will kick idle load balancer as soon as it has any load.
3831 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3833 unsigned long now = jiffies;
3835 int first_pick_cpu, second_pick_cpu;
3837 if (time_before(now, nohz.next_balance))
3840 if (rq->idle_at_tick)
3843 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3844 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3846 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3847 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3850 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3851 if (ret == nr_cpu_ids || ret == cpu) {
3852 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3853 if (rq->nr_running > 1)
3856 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3857 if (ret == nr_cpu_ids || ret == cpu) {
3865 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3869 * run_rebalance_domains is triggered when needed from the scheduler tick.
3870 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3872 static void run_rebalance_domains(struct softirq_action *h)
3874 int this_cpu = smp_processor_id();
3875 struct rq *this_rq = cpu_rq(this_cpu);
3876 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3877 CPU_IDLE : CPU_NOT_IDLE;
3879 rebalance_domains(this_cpu, idle);
3882 * If this cpu has a pending nohz_balance_kick, then do the
3883 * balancing on behalf of the other idle cpus whose ticks are
3886 nohz_idle_balance(this_cpu, idle);
3889 static inline int on_null_domain(int cpu)
3891 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3895 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3897 static inline void trigger_load_balance(struct rq *rq, int cpu)
3899 /* Don't need to rebalance while attached to NULL domain */
3900 if (time_after_eq(jiffies, rq->next_balance) &&
3901 likely(!on_null_domain(cpu)))
3902 raise_softirq(SCHED_SOFTIRQ);
3904 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
3905 nohz_balancer_kick(cpu);
3909 static void rq_online_fair(struct rq *rq)
3914 static void rq_offline_fair(struct rq *rq)
3919 #else /* CONFIG_SMP */
3922 * on UP we do not need to balance between CPUs:
3924 static inline void idle_balance(int cpu, struct rq *rq)
3928 #endif /* CONFIG_SMP */
3931 * scheduler tick hitting a task of our scheduling class:
3933 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
3935 struct cfs_rq *cfs_rq;
3936 struct sched_entity *se = &curr->se;
3938 for_each_sched_entity(se) {
3939 cfs_rq = cfs_rq_of(se);
3940 entity_tick(cfs_rq, se, queued);
3945 * called on fork with the child task as argument from the parent's context
3946 * - child not yet on the tasklist
3947 * - preemption disabled
3949 static void task_fork_fair(struct task_struct *p)
3951 struct cfs_rq *cfs_rq = task_cfs_rq(current);
3952 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
3953 int this_cpu = smp_processor_id();
3954 struct rq *rq = this_rq();
3955 unsigned long flags;
3957 raw_spin_lock_irqsave(&rq->lock, flags);
3959 update_rq_clock(rq);
3961 if (unlikely(task_cpu(p) != this_cpu)) {
3963 __set_task_cpu(p, this_cpu);
3967 update_curr(cfs_rq);
3970 se->vruntime = curr->vruntime;
3971 place_entity(cfs_rq, se, 1);
3973 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
3975 * Upon rescheduling, sched_class::put_prev_task() will place
3976 * 'current' within the tree based on its new key value.
3978 swap(curr->vruntime, se->vruntime);
3979 resched_task(rq->curr);
3982 se->vruntime -= cfs_rq->min_vruntime;
3984 raw_spin_unlock_irqrestore(&rq->lock, flags);
3988 * Priority of the task has changed. Check to see if we preempt
3991 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
3992 int oldprio, int running)
3995 * Reschedule if we are currently running on this runqueue and
3996 * our priority decreased, or if we are not currently running on
3997 * this runqueue and our priority is higher than the current's
4000 if (p->prio > oldprio)
4001 resched_task(rq->curr);
4003 check_preempt_curr(rq, p, 0);
4007 * We switched to the sched_fair class.
4009 static void switched_to_fair(struct rq *rq, struct task_struct *p,
4013 * We were most likely switched from sched_rt, so
4014 * kick off the schedule if running, otherwise just see
4015 * if we can still preempt the current task.
4018 resched_task(rq->curr);
4020 check_preempt_curr(rq, p, 0);
4023 /* Account for a task changing its policy or group.
4025 * This routine is mostly called to set cfs_rq->curr field when a task
4026 * migrates between groups/classes.
4028 static void set_curr_task_fair(struct rq *rq)
4030 struct sched_entity *se = &rq->curr->se;
4032 for_each_sched_entity(se)
4033 set_next_entity(cfs_rq_of(se), se);
4036 #ifdef CONFIG_FAIR_GROUP_SCHED
4037 static void task_move_group_fair(struct task_struct *p, int on_rq)
4040 * If the task was not on the rq at the time of this cgroup movement
4041 * it must have been asleep, sleeping tasks keep their ->vruntime
4042 * absolute on their old rq until wakeup (needed for the fair sleeper
4043 * bonus in place_entity()).
4045 * If it was on the rq, we've just 'preempted' it, which does convert
4046 * ->vruntime to a relative base.
4048 * Make sure both cases convert their relative position when migrating
4049 * to another cgroup's rq. This does somewhat interfere with the
4050 * fair sleeper stuff for the first placement, but who cares.
4053 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4054 set_task_rq(p, task_cpu(p));
4056 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4060 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4062 struct sched_entity *se = &task->se;
4063 unsigned int rr_interval = 0;
4066 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4069 if (rq->cfs.load.weight)
4070 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4076 * All the scheduling class methods:
4078 static const struct sched_class fair_sched_class = {
4079 .next = &idle_sched_class,
4080 .enqueue_task = enqueue_task_fair,
4081 .dequeue_task = dequeue_task_fair,
4082 .yield_task = yield_task_fair,
4084 .check_preempt_curr = check_preempt_wakeup,
4086 .pick_next_task = pick_next_task_fair,
4087 .put_prev_task = put_prev_task_fair,
4090 .select_task_rq = select_task_rq_fair,
4092 .rq_online = rq_online_fair,
4093 .rq_offline = rq_offline_fair,
4095 .task_waking = task_waking_fair,
4098 .set_curr_task = set_curr_task_fair,
4099 .task_tick = task_tick_fair,
4100 .task_fork = task_fork_fair,
4102 .prio_changed = prio_changed_fair,
4103 .switched_to = switched_to_fair,
4105 .get_rr_interval = get_rr_interval_fair,
4107 #ifdef CONFIG_FAIR_GROUP_SCHED
4108 .task_move_group = task_move_group_fair,
4112 #ifdef CONFIG_SCHED_DEBUG
4113 static void print_cfs_stats(struct seq_file *m, int cpu)
4115 struct cfs_rq *cfs_rq;
4118 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4119 print_cfs_rq(m, cpu, cfs_rq);