2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
24 #include <linux/sched.h>
27 * Targeted preemption latency for CPU-bound tasks:
28 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
43 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 enum sched_tunable_scaling sysctl_sched_tunable_scaling
51 = SCHED_TUNABLESCALING_LOG;
54 * Minimal preemption granularity for CPU-bound tasks:
55 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
57 unsigned int sysctl_sched_min_granularity = 750000ULL;
58 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
61 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
63 static unsigned int sched_nr_latency = 8;
66 * After fork, child runs first. If set to 0 (default) then
67 * parent will (try to) run first.
69 unsigned int sysctl_sched_child_runs_first __read_mostly;
72 * sys_sched_yield() compat mode
74 * This option switches the agressive yield implementation of the
75 * old scheduler back on.
77 unsigned int __read_mostly sysctl_sched_compat_yield;
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
87 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
90 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
93 * The exponential sliding window over which load is averaged for shares
97 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
99 static const struct sched_class fair_sched_class;
101 /**************************************************************
102 * CFS operations on generic schedulable entities:
105 #ifdef CONFIG_FAIR_GROUP_SCHED
107 /* cpu runqueue to which this cfs_rq is attached */
108 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
113 /* An entity is a task if it doesn't "own" a runqueue */
114 #define entity_is_task(se) (!se->my_q)
116 static inline struct task_struct *task_of(struct sched_entity *se)
118 #ifdef CONFIG_SCHED_DEBUG
119 WARN_ON_ONCE(!entity_is_task(se));
121 return container_of(se, struct task_struct, se);
124 /* Walk up scheduling entities hierarchy */
125 #define for_each_sched_entity(se) \
126 for (; se; se = se->parent)
128 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
133 /* runqueue on which this entity is (to be) queued */
134 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
139 /* runqueue "owned" by this group */
140 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
145 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
146 * another cpu ('this_cpu')
148 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
150 return cfs_rq->tg->cfs_rq[this_cpu];
153 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
155 if (!cfs_rq->on_list) {
157 * Ensure we either appear before our parent (if already
158 * enqueued) or force our parent to appear after us when it is
159 * enqueued. The fact that we always enqueue bottom-up
160 * reduces this to two cases.
162 if (cfs_rq->tg->parent &&
163 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
164 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
165 &rq_of(cfs_rq)->leaf_cfs_rq_list);
167 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
168 &rq_of(cfs_rq)->leaf_cfs_rq_list);
175 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
177 if (cfs_rq->on_list) {
178 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
183 /* Iterate thr' all leaf cfs_rq's on a runqueue */
184 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
185 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
187 /* Do the two (enqueued) entities belong to the same group ? */
189 is_same_group(struct sched_entity *se, struct sched_entity *pse)
191 if (se->cfs_rq == pse->cfs_rq)
197 static inline struct sched_entity *parent_entity(struct sched_entity *se)
202 /* return depth at which a sched entity is present in the hierarchy */
203 static inline int depth_se(struct sched_entity *se)
207 for_each_sched_entity(se)
214 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
216 int se_depth, pse_depth;
219 * preemption test can be made between sibling entities who are in the
220 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
221 * both tasks until we find their ancestors who are siblings of common
225 /* First walk up until both entities are at same depth */
226 se_depth = depth_se(*se);
227 pse_depth = depth_se(*pse);
229 while (se_depth > pse_depth) {
231 *se = parent_entity(*se);
234 while (pse_depth > se_depth) {
236 *pse = parent_entity(*pse);
239 while (!is_same_group(*se, *pse)) {
240 *se = parent_entity(*se);
241 *pse = parent_entity(*pse);
245 #else /* !CONFIG_FAIR_GROUP_SCHED */
247 static inline struct task_struct *task_of(struct sched_entity *se)
249 return container_of(se, struct task_struct, se);
252 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
254 return container_of(cfs_rq, struct rq, cfs);
257 #define entity_is_task(se) 1
259 #define for_each_sched_entity(se) \
260 for (; se; se = NULL)
262 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
264 return &task_rq(p)->cfs;
267 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
269 struct task_struct *p = task_of(se);
270 struct rq *rq = task_rq(p);
275 /* runqueue "owned" by this group */
276 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
281 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
283 return &cpu_rq(this_cpu)->cfs;
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
294 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
295 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
298 is_same_group(struct sched_entity *se, struct sched_entity *pse)
303 static inline struct sched_entity *parent_entity(struct sched_entity *se)
309 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
313 #endif /* CONFIG_FAIR_GROUP_SCHED */
316 /**************************************************************
317 * Scheduling class tree data structure manipulation methods:
320 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
322 s64 delta = (s64)(vruntime - min_vruntime);
324 min_vruntime = vruntime;
329 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
331 s64 delta = (s64)(vruntime - min_vruntime);
333 min_vruntime = vruntime;
338 static inline int entity_before(struct sched_entity *a,
339 struct sched_entity *b)
341 return (s64)(a->vruntime - b->vruntime) < 0;
344 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
346 return se->vruntime - cfs_rq->min_vruntime;
349 static void update_min_vruntime(struct cfs_rq *cfs_rq)
351 u64 vruntime = cfs_rq->min_vruntime;
354 vruntime = cfs_rq->curr->vruntime;
356 if (cfs_rq->rb_leftmost) {
357 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
362 vruntime = se->vruntime;
364 vruntime = min_vruntime(vruntime, se->vruntime);
367 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
371 * Enqueue an entity into the rb-tree:
373 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
375 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
376 struct rb_node *parent = NULL;
377 struct sched_entity *entry;
378 s64 key = entity_key(cfs_rq, se);
382 * Find the right place in the rbtree:
386 entry = rb_entry(parent, struct sched_entity, run_node);
388 * We dont care about collisions. Nodes with
389 * the same key stay together.
391 if (key < entity_key(cfs_rq, entry)) {
392 link = &parent->rb_left;
394 link = &parent->rb_right;
400 * Maintain a cache of leftmost tree entries (it is frequently
404 cfs_rq->rb_leftmost = &se->run_node;
406 rb_link_node(&se->run_node, parent, link);
407 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
410 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
412 if (cfs_rq->rb_leftmost == &se->run_node) {
413 struct rb_node *next_node;
415 next_node = rb_next(&se->run_node);
416 cfs_rq->rb_leftmost = next_node;
419 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
422 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
424 struct rb_node *left = cfs_rq->rb_leftmost;
429 return rb_entry(left, struct sched_entity, run_node);
432 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
434 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
439 return rb_entry(last, struct sched_entity, run_node);
442 /**************************************************************
443 * Scheduling class statistics methods:
446 #ifdef CONFIG_SCHED_DEBUG
447 int sched_proc_update_handler(struct ctl_table *table, int write,
448 void __user *buffer, size_t *lenp,
451 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
452 int factor = get_update_sysctl_factor();
457 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
458 sysctl_sched_min_granularity);
460 #define WRT_SYSCTL(name) \
461 (normalized_sysctl_##name = sysctl_##name / (factor))
462 WRT_SYSCTL(sched_min_granularity);
463 WRT_SYSCTL(sched_latency);
464 WRT_SYSCTL(sched_wakeup_granularity);
474 static inline unsigned long
475 calc_delta_fair(unsigned long delta, struct sched_entity *se)
477 if (unlikely(se->load.weight != NICE_0_LOAD))
478 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
484 * The idea is to set a period in which each task runs once.
486 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
487 * this period because otherwise the slices get too small.
489 * p = (nr <= nl) ? l : l*nr/nl
491 static u64 __sched_period(unsigned long nr_running)
493 u64 period = sysctl_sched_latency;
494 unsigned long nr_latency = sched_nr_latency;
496 if (unlikely(nr_running > nr_latency)) {
497 period = sysctl_sched_min_granularity;
498 period *= nr_running;
505 * We calculate the wall-time slice from the period by taking a part
506 * proportional to the weight.
510 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
512 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
514 for_each_sched_entity(se) {
515 struct load_weight *load;
516 struct load_weight lw;
518 cfs_rq = cfs_rq_of(se);
519 load = &cfs_rq->load;
521 if (unlikely(!se->on_rq)) {
524 update_load_add(&lw, se->load.weight);
527 slice = calc_delta_mine(slice, se->load.weight, load);
533 * We calculate the vruntime slice of a to be inserted task
537 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
539 return calc_delta_fair(sched_slice(cfs_rq, se), se);
542 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
543 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta);
546 * Update the current task's runtime statistics. Skip current tasks that
547 * are not in our scheduling class.
550 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
551 unsigned long delta_exec)
553 unsigned long delta_exec_weighted;
555 schedstat_set(curr->statistics.exec_max,
556 max((u64)delta_exec, curr->statistics.exec_max));
558 curr->sum_exec_runtime += delta_exec;
559 schedstat_add(cfs_rq, exec_clock, delta_exec);
560 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
562 curr->vruntime += delta_exec_weighted;
563 update_min_vruntime(cfs_rq);
565 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
566 cfs_rq->load_unacc_exec_time += delta_exec;
570 static void update_curr(struct cfs_rq *cfs_rq)
572 struct sched_entity *curr = cfs_rq->curr;
573 u64 now = rq_of(cfs_rq)->clock_task;
574 unsigned long delta_exec;
580 * Get the amount of time the current task was running
581 * since the last time we changed load (this cannot
582 * overflow on 32 bits):
584 delta_exec = (unsigned long)(now - curr->exec_start);
588 __update_curr(cfs_rq, curr, delta_exec);
589 curr->exec_start = now;
591 if (entity_is_task(curr)) {
592 struct task_struct *curtask = task_of(curr);
594 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
595 cpuacct_charge(curtask, delta_exec);
596 account_group_exec_runtime(curtask, delta_exec);
601 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
603 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
607 * Task is being enqueued - update stats:
609 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
612 * Are we enqueueing a waiting task? (for current tasks
613 * a dequeue/enqueue event is a NOP)
615 if (se != cfs_rq->curr)
616 update_stats_wait_start(cfs_rq, se);
620 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
622 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
623 rq_of(cfs_rq)->clock - se->statistics.wait_start));
624 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
625 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
626 rq_of(cfs_rq)->clock - se->statistics.wait_start);
627 #ifdef CONFIG_SCHEDSTATS
628 if (entity_is_task(se)) {
629 trace_sched_stat_wait(task_of(se),
630 rq_of(cfs_rq)->clock - se->statistics.wait_start);
633 schedstat_set(se->statistics.wait_start, 0);
637 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 * Mark the end of the wait period if dequeueing a
643 if (se != cfs_rq->curr)
644 update_stats_wait_end(cfs_rq, se);
648 * We are picking a new current task - update its stats:
651 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
654 * We are starting a new run period:
656 se->exec_start = rq_of(cfs_rq)->clock_task;
659 /**************************************************
660 * Scheduling class queueing methods:
663 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
665 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
667 cfs_rq->task_weight += weight;
671 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
677 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
679 update_load_add(&cfs_rq->load, se->load.weight);
680 if (!parent_entity(se))
681 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
682 if (entity_is_task(se)) {
683 add_cfs_task_weight(cfs_rq, se->load.weight);
684 list_add(&se->group_node, &cfs_rq->tasks);
686 cfs_rq->nr_running++;
690 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
692 update_load_sub(&cfs_rq->load, se->load.weight);
693 if (!parent_entity(se))
694 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
695 if (entity_is_task(se)) {
696 add_cfs_task_weight(cfs_rq, -se->load.weight);
697 list_del_init(&se->group_node);
699 cfs_rq->nr_running--;
702 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
703 static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
706 struct task_group *tg = cfs_rq->tg;
709 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
710 load_avg -= cfs_rq->load_contribution;
712 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
713 atomic_add(load_avg, &tg->load_weight);
714 cfs_rq->load_contribution += load_avg;
718 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
720 u64 period = sysctl_sched_shares_window;
722 unsigned long load = cfs_rq->load.weight;
727 now = rq_of(cfs_rq)->clock;
728 delta = now - cfs_rq->load_stamp;
730 /* truncate load history at 4 idle periods */
731 if (cfs_rq->load_stamp > cfs_rq->load_last &&
732 now - cfs_rq->load_last > 4 * period) {
733 cfs_rq->load_period = 0;
734 cfs_rq->load_avg = 0;
737 cfs_rq->load_stamp = now;
738 cfs_rq->load_unacc_exec_time = 0;
739 cfs_rq->load_period += delta;
741 cfs_rq->load_last = now;
742 cfs_rq->load_avg += delta * load;
745 /* consider updating load contribution on each fold or truncate */
746 if (global_update || cfs_rq->load_period > period
747 || !cfs_rq->load_period)
748 update_cfs_rq_load_contribution(cfs_rq, global_update);
750 while (cfs_rq->load_period > period) {
752 * Inline assembly required to prevent the compiler
753 * optimising this loop into a divmod call.
754 * See __iter_div_u64_rem() for another example of this.
756 asm("" : "+rm" (cfs_rq->load_period));
757 cfs_rq->load_period /= 2;
758 cfs_rq->load_avg /= 2;
761 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
762 list_del_leaf_cfs_rq(cfs_rq);
765 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
766 unsigned long weight)
769 account_entity_dequeue(cfs_rq, se);
771 update_load_set(&se->load, weight);
774 account_entity_enqueue(cfs_rq, se);
777 static void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
779 struct task_group *tg;
780 struct sched_entity *se;
781 long load_weight, load, shares;
787 se = tg->se[cpu_of(rq_of(cfs_rq))];
791 load = cfs_rq->load.weight + weight_delta;
793 load_weight = atomic_read(&tg->load_weight);
794 load_weight -= cfs_rq->load_contribution;
797 shares = (tg->shares * load);
799 shares /= load_weight;
801 if (shares < MIN_SHARES)
803 if (shares > tg->shares)
806 reweight_entity(cfs_rq_of(se), se, shares);
809 static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
811 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
812 update_cfs_load(cfs_rq, 0);
813 update_cfs_shares(cfs_rq, 0);
816 #else /* CONFIG_FAIR_GROUP_SCHED */
817 static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
821 static inline void update_cfs_shares(struct cfs_rq *cfs_rq, long weight_delta)
825 static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
828 #endif /* CONFIG_FAIR_GROUP_SCHED */
830 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
832 #ifdef CONFIG_SCHEDSTATS
833 struct task_struct *tsk = NULL;
835 if (entity_is_task(se))
838 if (se->statistics.sleep_start) {
839 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
844 if (unlikely(delta > se->statistics.sleep_max))
845 se->statistics.sleep_max = delta;
847 se->statistics.sleep_start = 0;
848 se->statistics.sum_sleep_runtime += delta;
851 account_scheduler_latency(tsk, delta >> 10, 1);
852 trace_sched_stat_sleep(tsk, delta);
855 if (se->statistics.block_start) {
856 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
861 if (unlikely(delta > se->statistics.block_max))
862 se->statistics.block_max = delta;
864 se->statistics.block_start = 0;
865 se->statistics.sum_sleep_runtime += delta;
868 if (tsk->in_iowait) {
869 se->statistics.iowait_sum += delta;
870 se->statistics.iowait_count++;
871 trace_sched_stat_iowait(tsk, delta);
875 * Blocking time is in units of nanosecs, so shift by
876 * 20 to get a milliseconds-range estimation of the
877 * amount of time that the task spent sleeping:
879 if (unlikely(prof_on == SLEEP_PROFILING)) {
880 profile_hits(SLEEP_PROFILING,
881 (void *)get_wchan(tsk),
884 account_scheduler_latency(tsk, delta >> 10, 0);
890 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
892 #ifdef CONFIG_SCHED_DEBUG
893 s64 d = se->vruntime - cfs_rq->min_vruntime;
898 if (d > 3*sysctl_sched_latency)
899 schedstat_inc(cfs_rq, nr_spread_over);
904 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
906 u64 vruntime = cfs_rq->min_vruntime;
909 * The 'current' period is already promised to the current tasks,
910 * however the extra weight of the new task will slow them down a
911 * little, place the new task so that it fits in the slot that
912 * stays open at the end.
914 if (initial && sched_feat(START_DEBIT))
915 vruntime += sched_vslice(cfs_rq, se);
917 /* sleeps up to a single latency don't count. */
919 unsigned long thresh = sysctl_sched_latency;
922 * Halve their sleep time's effect, to allow
923 * for a gentler effect of sleepers:
925 if (sched_feat(GENTLE_FAIR_SLEEPERS))
931 /* ensure we never gain time by being placed backwards. */
932 vruntime = max_vruntime(se->vruntime, vruntime);
934 se->vruntime = vruntime;
938 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
941 * Update the normalized vruntime before updating min_vruntime
942 * through callig update_curr().
944 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
945 se->vruntime += cfs_rq->min_vruntime;
948 * Update run-time statistics of the 'current'.
951 update_cfs_load(cfs_rq, 0);
952 update_cfs_shares(cfs_rq, se->load.weight);
953 account_entity_enqueue(cfs_rq, se);
955 if (flags & ENQUEUE_WAKEUP) {
956 place_entity(cfs_rq, se, 0);
957 enqueue_sleeper(cfs_rq, se);
960 update_stats_enqueue(cfs_rq, se);
961 check_spread(cfs_rq, se);
962 if (se != cfs_rq->curr)
963 __enqueue_entity(cfs_rq, se);
966 if (cfs_rq->nr_running == 1)
967 list_add_leaf_cfs_rq(cfs_rq);
970 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
972 if (!se || cfs_rq->last == se)
975 if (!se || cfs_rq->next == se)
979 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
981 for_each_sched_entity(se)
982 __clear_buddies(cfs_rq_of(se), se);
986 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
989 * Update run-time statistics of the 'current'.
993 update_stats_dequeue(cfs_rq, se);
994 if (flags & DEQUEUE_SLEEP) {
995 #ifdef CONFIG_SCHEDSTATS
996 if (entity_is_task(se)) {
997 struct task_struct *tsk = task_of(se);
999 if (tsk->state & TASK_INTERRUPTIBLE)
1000 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1001 if (tsk->state & TASK_UNINTERRUPTIBLE)
1002 se->statistics.block_start = rq_of(cfs_rq)->clock;
1007 clear_buddies(cfs_rq, se);
1009 if (se != cfs_rq->curr)
1010 __dequeue_entity(cfs_rq, se);
1012 update_cfs_load(cfs_rq, 0);
1013 account_entity_dequeue(cfs_rq, se);
1014 update_min_vruntime(cfs_rq);
1015 update_cfs_shares(cfs_rq, 0);
1018 * Normalize the entity after updating the min_vruntime because the
1019 * update can refer to the ->curr item and we need to reflect this
1020 * movement in our normalized position.
1022 if (!(flags & DEQUEUE_SLEEP))
1023 se->vruntime -= cfs_rq->min_vruntime;
1027 * Preempt the current task with a newly woken task if needed:
1030 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1032 unsigned long ideal_runtime, delta_exec;
1034 ideal_runtime = sched_slice(cfs_rq, curr);
1035 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1036 if (delta_exec > ideal_runtime) {
1037 resched_task(rq_of(cfs_rq)->curr);
1039 * The current task ran long enough, ensure it doesn't get
1040 * re-elected due to buddy favours.
1042 clear_buddies(cfs_rq, curr);
1047 * Ensure that a task that missed wakeup preemption by a
1048 * narrow margin doesn't have to wait for a full slice.
1049 * This also mitigates buddy induced latencies under load.
1051 if (!sched_feat(WAKEUP_PREEMPT))
1054 if (delta_exec < sysctl_sched_min_granularity)
1057 if (cfs_rq->nr_running > 1) {
1058 struct sched_entity *se = __pick_next_entity(cfs_rq);
1059 s64 delta = curr->vruntime - se->vruntime;
1061 if (delta > ideal_runtime)
1062 resched_task(rq_of(cfs_rq)->curr);
1067 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1069 /* 'current' is not kept within the tree. */
1072 * Any task has to be enqueued before it get to execute on
1073 * a CPU. So account for the time it spent waiting on the
1076 update_stats_wait_end(cfs_rq, se);
1077 __dequeue_entity(cfs_rq, se);
1080 update_stats_curr_start(cfs_rq, se);
1082 #ifdef CONFIG_SCHEDSTATS
1084 * Track our maximum slice length, if the CPU's load is at
1085 * least twice that of our own weight (i.e. dont track it
1086 * when there are only lesser-weight tasks around):
1088 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1089 se->statistics.slice_max = max(se->statistics.slice_max,
1090 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1093 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1097 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1099 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1101 struct sched_entity *se = __pick_next_entity(cfs_rq);
1102 struct sched_entity *left = se;
1104 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1108 * Prefer last buddy, try to return the CPU to a preempted task.
1110 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1113 clear_buddies(cfs_rq, se);
1118 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1121 * If still on the runqueue then deactivate_task()
1122 * was not called and update_curr() has to be done:
1125 update_curr(cfs_rq);
1127 check_spread(cfs_rq, prev);
1129 update_stats_wait_start(cfs_rq, prev);
1130 /* Put 'current' back into the tree. */
1131 __enqueue_entity(cfs_rq, prev);
1133 cfs_rq->curr = NULL;
1137 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1140 * Update run-time statistics of the 'current'.
1142 update_curr(cfs_rq);
1145 * Update share accounting for long-running entities.
1147 update_entity_shares_tick(cfs_rq);
1149 #ifdef CONFIG_SCHED_HRTICK
1151 * queued ticks are scheduled to match the slice, so don't bother
1152 * validating it and just reschedule.
1155 resched_task(rq_of(cfs_rq)->curr);
1159 * don't let the period tick interfere with the hrtick preemption
1161 if (!sched_feat(DOUBLE_TICK) &&
1162 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1166 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
1167 check_preempt_tick(cfs_rq, curr);
1170 /**************************************************
1171 * CFS operations on tasks:
1174 #ifdef CONFIG_SCHED_HRTICK
1175 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
1177 struct sched_entity *se = &p->se;
1178 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1180 WARN_ON(task_rq(p) != rq);
1182 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
1183 u64 slice = sched_slice(cfs_rq, se);
1184 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1185 s64 delta = slice - ran;
1194 * Don't schedule slices shorter than 10000ns, that just
1195 * doesn't make sense. Rely on vruntime for fairness.
1198 delta = max_t(s64, 10000LL, delta);
1200 hrtick_start(rq, delta);
1205 * called from enqueue/dequeue and updates the hrtick when the
1206 * current task is from our class and nr_running is low enough
1209 static void hrtick_update(struct rq *rq)
1211 struct task_struct *curr = rq->curr;
1213 if (curr->sched_class != &fair_sched_class)
1216 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
1217 hrtick_start_fair(rq, curr);
1219 #else /* !CONFIG_SCHED_HRTICK */
1221 hrtick_start_fair(struct rq *rq, struct task_struct *p)
1225 static inline void hrtick_update(struct rq *rq)
1231 * The enqueue_task method is called before nr_running is
1232 * increased. Here we update the fair scheduling stats and
1233 * then put the task into the rbtree:
1236 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1238 struct cfs_rq *cfs_rq;
1239 struct sched_entity *se = &p->se;
1241 for_each_sched_entity(se) {
1244 cfs_rq = cfs_rq_of(se);
1245 enqueue_entity(cfs_rq, se, flags);
1246 flags = ENQUEUE_WAKEUP;
1249 for_each_sched_entity(se) {
1250 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1252 update_cfs_load(cfs_rq, 0);
1253 update_cfs_shares(cfs_rq, 0);
1260 * The dequeue_task method is called before nr_running is
1261 * decreased. We remove the task from the rbtree and
1262 * update the fair scheduling stats:
1264 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
1266 struct cfs_rq *cfs_rq;
1267 struct sched_entity *se = &p->se;
1269 for_each_sched_entity(se) {
1270 cfs_rq = cfs_rq_of(se);
1271 dequeue_entity(cfs_rq, se, flags);
1273 /* Don't dequeue parent if it has other entities besides us */
1274 if (cfs_rq->load.weight)
1276 flags |= DEQUEUE_SLEEP;
1279 for_each_sched_entity(se) {
1280 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1282 update_cfs_load(cfs_rq, 0);
1283 update_cfs_shares(cfs_rq, 0);
1290 * sched_yield() support is very simple - we dequeue and enqueue.
1292 * If compat_yield is turned on then we requeue to the end of the tree.
1294 static void yield_task_fair(struct rq *rq)
1296 struct task_struct *curr = rq->curr;
1297 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1298 struct sched_entity *rightmost, *se = &curr->se;
1301 * Are we the only task in the tree?
1303 if (unlikely(cfs_rq->nr_running == 1))
1306 clear_buddies(cfs_rq, se);
1308 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1309 update_rq_clock(rq);
1311 * Update run-time statistics of the 'current'.
1313 update_curr(cfs_rq);
1318 * Find the rightmost entry in the rbtree:
1320 rightmost = __pick_last_entity(cfs_rq);
1322 * Already in the rightmost position?
1324 if (unlikely(!rightmost || entity_before(rightmost, se)))
1328 * Minimally necessary key value to be last in the tree:
1329 * Upon rescheduling, sched_class::put_prev_task() will place
1330 * 'current' within the tree based on its new key value.
1332 se->vruntime = rightmost->vruntime + 1;
1337 static void task_waking_fair(struct rq *rq, struct task_struct *p)
1339 struct sched_entity *se = &p->se;
1340 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1342 se->vruntime -= cfs_rq->min_vruntime;
1345 #ifdef CONFIG_FAIR_GROUP_SCHED
1347 * effective_load() calculates the load change as seen from the root_task_group
1349 * Adding load to a group doesn't make a group heavier, but can cause movement
1350 * of group shares between cpus. Assuming the shares were perfectly aligned one
1351 * can calculate the shift in shares.
1353 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
1355 struct sched_entity *se = tg->se[cpu];
1360 for_each_sched_entity(se) {
1361 long S, rw, s, a, b;
1363 S = se->my_q->tg->shares;
1364 s = se->load.weight;
1365 rw = se->my_q->load.weight;
1376 * Assume the group is already running and will
1377 * thus already be accounted for in the weight.
1379 * That is, moving shares between CPUs, does not
1380 * alter the group weight.
1390 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1391 unsigned long wl, unsigned long wg)
1398 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
1400 unsigned long this_load, load;
1401 int idx, this_cpu, prev_cpu;
1402 unsigned long tl_per_task;
1403 struct task_group *tg;
1404 unsigned long weight;
1408 this_cpu = smp_processor_id();
1409 prev_cpu = task_cpu(p);
1410 load = source_load(prev_cpu, idx);
1411 this_load = target_load(this_cpu, idx);
1414 * If sync wakeup then subtract the (maximum possible)
1415 * effect of the currently running task from the load
1416 * of the current CPU:
1420 tg = task_group(current);
1421 weight = current->se.load.weight;
1423 this_load += effective_load(tg, this_cpu, -weight, -weight);
1424 load += effective_load(tg, prev_cpu, 0, -weight);
1428 weight = p->se.load.weight;
1431 * In low-load situations, where prev_cpu is idle and this_cpu is idle
1432 * due to the sync cause above having dropped this_load to 0, we'll
1433 * always have an imbalance, but there's really nothing you can do
1434 * about that, so that's good too.
1436 * Otherwise check if either cpus are near enough in load to allow this
1437 * task to be woken on this_cpu.
1440 unsigned long this_eff_load, prev_eff_load;
1442 this_eff_load = 100;
1443 this_eff_load *= power_of(prev_cpu);
1444 this_eff_load *= this_load +
1445 effective_load(tg, this_cpu, weight, weight);
1447 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
1448 prev_eff_load *= power_of(this_cpu);
1449 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
1451 balanced = this_eff_load <= prev_eff_load;
1457 * If the currently running task will sleep within
1458 * a reasonable amount of time then attract this newly
1461 if (sync && balanced)
1464 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
1465 tl_per_task = cpu_avg_load_per_task(this_cpu);
1468 (this_load <= load &&
1469 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
1471 * This domain has SD_WAKE_AFFINE and
1472 * p is cache cold in this domain, and
1473 * there is no bad imbalance.
1475 schedstat_inc(sd, ttwu_move_affine);
1476 schedstat_inc(p, se.statistics.nr_wakeups_affine);
1484 * find_idlest_group finds and returns the least busy CPU group within the
1487 static struct sched_group *
1488 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
1489 int this_cpu, int load_idx)
1491 struct sched_group *idlest = NULL, *group = sd->groups;
1492 unsigned long min_load = ULONG_MAX, this_load = 0;
1493 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1496 unsigned long load, avg_load;
1500 /* Skip over this group if it has no CPUs allowed */
1501 if (!cpumask_intersects(sched_group_cpus(group),
1505 local_group = cpumask_test_cpu(this_cpu,
1506 sched_group_cpus(group));
1508 /* Tally up the load of all CPUs in the group */
1511 for_each_cpu(i, sched_group_cpus(group)) {
1512 /* Bias balancing toward cpus of our domain */
1514 load = source_load(i, load_idx);
1516 load = target_load(i, load_idx);
1521 /* Adjust by relative CPU power of the group */
1522 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1525 this_load = avg_load;
1526 } else if (avg_load < min_load) {
1527 min_load = avg_load;
1530 } while (group = group->next, group != sd->groups);
1532 if (!idlest || 100*this_load < imbalance*min_load)
1538 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1541 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
1543 unsigned long load, min_load = ULONG_MAX;
1547 /* Traverse only the allowed CPUs */
1548 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
1549 load = weighted_cpuload(i);
1551 if (load < min_load || (load == min_load && i == this_cpu)) {
1561 * Try and locate an idle CPU in the sched_domain.
1563 static int select_idle_sibling(struct task_struct *p, int target)
1565 int cpu = smp_processor_id();
1566 int prev_cpu = task_cpu(p);
1567 struct sched_domain *sd;
1571 * If the task is going to be woken-up on this cpu and if it is
1572 * already idle, then it is the right target.
1574 if (target == cpu && idle_cpu(cpu))
1578 * If the task is going to be woken-up on the cpu where it previously
1579 * ran and if it is currently idle, then it the right target.
1581 if (target == prev_cpu && idle_cpu(prev_cpu))
1585 * Otherwise, iterate the domains and find an elegible idle cpu.
1587 for_each_domain(target, sd) {
1588 if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
1591 for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
1599 * Lets stop looking for an idle sibling when we reached
1600 * the domain that spans the current cpu and prev_cpu.
1602 if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
1603 cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
1611 * sched_balance_self: balance the current task (running on cpu) in domains
1612 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1615 * Balance, ie. select the least loaded group.
1617 * Returns the target CPU number, or the same CPU if no balancing is needed.
1619 * preempt must be disabled.
1622 select_task_rq_fair(struct rq *rq, struct task_struct *p, int sd_flag, int wake_flags)
1624 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
1625 int cpu = smp_processor_id();
1626 int prev_cpu = task_cpu(p);
1628 int want_affine = 0;
1630 int sync = wake_flags & WF_SYNC;
1632 if (sd_flag & SD_BALANCE_WAKE) {
1633 if (cpumask_test_cpu(cpu, &p->cpus_allowed))
1638 for_each_domain(cpu, tmp) {
1639 if (!(tmp->flags & SD_LOAD_BALANCE))
1643 * If power savings logic is enabled for a domain, see if we
1644 * are not overloaded, if so, don't balance wider.
1646 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
1647 unsigned long power = 0;
1648 unsigned long nr_running = 0;
1649 unsigned long capacity;
1652 for_each_cpu(i, sched_domain_span(tmp)) {
1653 power += power_of(i);
1654 nr_running += cpu_rq(i)->cfs.nr_running;
1657 capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
1659 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
1662 if (nr_running < capacity)
1667 * If both cpu and prev_cpu are part of this domain,
1668 * cpu is a valid SD_WAKE_AFFINE target.
1670 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
1671 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
1676 if (!want_sd && !want_affine)
1679 if (!(tmp->flags & sd_flag))
1687 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
1688 return select_idle_sibling(p, cpu);
1690 return select_idle_sibling(p, prev_cpu);
1694 int load_idx = sd->forkexec_idx;
1695 struct sched_group *group;
1698 if (!(sd->flags & sd_flag)) {
1703 if (sd_flag & SD_BALANCE_WAKE)
1704 load_idx = sd->wake_idx;
1706 group = find_idlest_group(sd, p, cpu, load_idx);
1712 new_cpu = find_idlest_cpu(group, p, cpu);
1713 if (new_cpu == -1 || new_cpu == cpu) {
1714 /* Now try balancing at a lower domain level of cpu */
1719 /* Now try balancing at a lower domain level of new_cpu */
1721 weight = sd->span_weight;
1723 for_each_domain(cpu, tmp) {
1724 if (weight <= tmp->span_weight)
1726 if (tmp->flags & sd_flag)
1729 /* while loop will break here if sd == NULL */
1734 #endif /* CONFIG_SMP */
1736 static unsigned long
1737 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1739 unsigned long gran = sysctl_sched_wakeup_granularity;
1742 * Since its curr running now, convert the gran from real-time
1743 * to virtual-time in his units.
1745 * By using 'se' instead of 'curr' we penalize light tasks, so
1746 * they get preempted easier. That is, if 'se' < 'curr' then
1747 * the resulting gran will be larger, therefore penalizing the
1748 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1749 * be smaller, again penalizing the lighter task.
1751 * This is especially important for buddies when the leftmost
1752 * task is higher priority than the buddy.
1754 if (unlikely(se->load.weight != NICE_0_LOAD))
1755 gran = calc_delta_fair(gran, se);
1761 * Should 'se' preempt 'curr'.
1775 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1777 s64 gran, vdiff = curr->vruntime - se->vruntime;
1782 gran = wakeup_gran(curr, se);
1789 static void set_last_buddy(struct sched_entity *se)
1791 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1792 for_each_sched_entity(se)
1793 cfs_rq_of(se)->last = se;
1797 static void set_next_buddy(struct sched_entity *se)
1799 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1800 for_each_sched_entity(se)
1801 cfs_rq_of(se)->next = se;
1806 * Preempt the current task with a newly woken task if needed:
1808 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1810 struct task_struct *curr = rq->curr;
1811 struct sched_entity *se = &curr->se, *pse = &p->se;
1812 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1813 int scale = cfs_rq->nr_running >= sched_nr_latency;
1815 if (unlikely(se == pse))
1818 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK))
1819 set_next_buddy(pse);
1822 * We can come here with TIF_NEED_RESCHED already set from new task
1825 if (test_tsk_need_resched(curr))
1829 * Batch and idle tasks do not preempt (their preemption is driven by
1832 if (unlikely(p->policy != SCHED_NORMAL))
1835 /* Idle tasks are by definition preempted by everybody. */
1836 if (unlikely(curr->policy == SCHED_IDLE))
1839 if (!sched_feat(WAKEUP_PREEMPT))
1842 update_curr(cfs_rq);
1843 find_matching_se(&se, &pse);
1845 if (wakeup_preempt_entity(se, pse) == 1)
1853 * Only set the backward buddy when the current task is still
1854 * on the rq. This can happen when a wakeup gets interleaved
1855 * with schedule on the ->pre_schedule() or idle_balance()
1856 * point, either of which can * drop the rq lock.
1858 * Also, during early boot the idle thread is in the fair class,
1859 * for obvious reasons its a bad idea to schedule back to it.
1861 if (unlikely(!se->on_rq || curr == rq->idle))
1864 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
1868 static struct task_struct *pick_next_task_fair(struct rq *rq)
1870 struct task_struct *p;
1871 struct cfs_rq *cfs_rq = &rq->cfs;
1872 struct sched_entity *se;
1874 if (!cfs_rq->nr_running)
1878 se = pick_next_entity(cfs_rq);
1879 set_next_entity(cfs_rq, se);
1880 cfs_rq = group_cfs_rq(se);
1884 hrtick_start_fair(rq, p);
1890 * Account for a descheduled task:
1892 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1894 struct sched_entity *se = &prev->se;
1895 struct cfs_rq *cfs_rq;
1897 for_each_sched_entity(se) {
1898 cfs_rq = cfs_rq_of(se);
1899 put_prev_entity(cfs_rq, se);
1904 /**************************************************
1905 * Fair scheduling class load-balancing methods:
1909 * pull_task - move a task from a remote runqueue to the local runqueue.
1910 * Both runqueues must be locked.
1912 static void pull_task(struct rq *src_rq, struct task_struct *p,
1913 struct rq *this_rq, int this_cpu)
1915 deactivate_task(src_rq, p, 0);
1916 set_task_cpu(p, this_cpu);
1917 activate_task(this_rq, p, 0);
1918 check_preempt_curr(this_rq, p, 0);
1922 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1925 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
1926 struct sched_domain *sd, enum cpu_idle_type idle,
1929 int tsk_cache_hot = 0;
1931 * We do not migrate tasks that are:
1932 * 1) running (obviously), or
1933 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1934 * 3) are cache-hot on their current CPU.
1936 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
1937 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
1942 if (task_running(rq, p)) {
1943 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
1948 * Aggressive migration if:
1949 * 1) task is cache cold, or
1950 * 2) too many balance attempts have failed.
1953 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
1954 if (!tsk_cache_hot ||
1955 sd->nr_balance_failed > sd->cache_nice_tries) {
1956 #ifdef CONFIG_SCHEDSTATS
1957 if (tsk_cache_hot) {
1958 schedstat_inc(sd, lb_hot_gained[idle]);
1959 schedstat_inc(p, se.statistics.nr_forced_migrations);
1965 if (tsk_cache_hot) {
1966 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
1973 * move_one_task tries to move exactly one task from busiest to this_rq, as
1974 * part of active balancing operations within "domain".
1975 * Returns 1 if successful and 0 otherwise.
1977 * Called with both runqueues locked.
1980 move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1981 struct sched_domain *sd, enum cpu_idle_type idle)
1983 struct task_struct *p, *n;
1984 struct cfs_rq *cfs_rq;
1987 for_each_leaf_cfs_rq(busiest, cfs_rq) {
1988 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
1990 if (!can_migrate_task(p, busiest, this_cpu,
1994 pull_task(busiest, p, this_rq, this_cpu);
1996 * Right now, this is only the second place pull_task()
1997 * is called, so we can safely collect pull_task()
1998 * stats here rather than inside pull_task().
2000 schedstat_inc(sd, lb_gained[idle]);
2008 static unsigned long
2009 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2010 unsigned long max_load_move, struct sched_domain *sd,
2011 enum cpu_idle_type idle, int *all_pinned,
2012 int *this_best_prio, struct cfs_rq *busiest_cfs_rq)
2014 int loops = 0, pulled = 0, pinned = 0;
2015 long rem_load_move = max_load_move;
2016 struct task_struct *p, *n;
2018 if (max_load_move == 0)
2023 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
2024 if (loops++ > sysctl_sched_nr_migrate)
2027 if ((p->se.load.weight >> 1) > rem_load_move ||
2028 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned))
2031 pull_task(busiest, p, this_rq, this_cpu);
2033 rem_load_move -= p->se.load.weight;
2035 #ifdef CONFIG_PREEMPT
2037 * NEWIDLE balancing is a source of latency, so preemptible
2038 * kernels will stop after the first task is pulled to minimize
2039 * the critical section.
2041 if (idle == CPU_NEWLY_IDLE)
2046 * We only want to steal up to the prescribed amount of
2049 if (rem_load_move <= 0)
2052 if (p->prio < *this_best_prio)
2053 *this_best_prio = p->prio;
2057 * Right now, this is one of only two places pull_task() is called,
2058 * so we can safely collect pull_task() stats here rather than
2059 * inside pull_task().
2061 schedstat_add(sd, lb_gained[idle], pulled);
2064 *all_pinned = pinned;
2066 return max_load_move - rem_load_move;
2069 #ifdef CONFIG_FAIR_GROUP_SCHED
2071 * update tg->load_weight by folding this cpu's load_avg
2073 static int update_shares_cpu(struct task_group *tg, int cpu)
2075 struct cfs_rq *cfs_rq;
2076 unsigned long flags;
2083 cfs_rq = tg->cfs_rq[cpu];
2085 raw_spin_lock_irqsave(&rq->lock, flags);
2087 update_rq_clock(rq);
2088 update_cfs_load(cfs_rq, 1);
2091 * We need to update shares after updating tg->load_weight in
2092 * order to adjust the weight of groups with long running tasks.
2094 update_cfs_shares(cfs_rq, 0);
2096 raw_spin_unlock_irqrestore(&rq->lock, flags);
2101 static void update_shares(int cpu)
2103 struct cfs_rq *cfs_rq;
2104 struct rq *rq = cpu_rq(cpu);
2107 for_each_leaf_cfs_rq(rq, cfs_rq)
2108 update_shares_cpu(cfs_rq->tg, cpu);
2112 static unsigned long
2113 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2114 unsigned long max_load_move,
2115 struct sched_domain *sd, enum cpu_idle_type idle,
2116 int *all_pinned, int *this_best_prio)
2118 long rem_load_move = max_load_move;
2119 int busiest_cpu = cpu_of(busiest);
2120 struct task_group *tg;
2123 update_h_load(busiest_cpu);
2125 list_for_each_entry_rcu(tg, &task_groups, list) {
2126 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
2127 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
2128 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
2129 u64 rem_load, moved_load;
2134 if (!busiest_cfs_rq->task_weight)
2137 rem_load = (u64)rem_load_move * busiest_weight;
2138 rem_load = div_u64(rem_load, busiest_h_load + 1);
2140 moved_load = balance_tasks(this_rq, this_cpu, busiest,
2141 rem_load, sd, idle, all_pinned, this_best_prio,
2147 moved_load *= busiest_h_load;
2148 moved_load = div_u64(moved_load, busiest_weight + 1);
2150 rem_load_move -= moved_load;
2151 if (rem_load_move < 0)
2156 return max_load_move - rem_load_move;
2159 static inline void update_shares(int cpu)
2163 static unsigned long
2164 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
2165 unsigned long max_load_move,
2166 struct sched_domain *sd, enum cpu_idle_type idle,
2167 int *all_pinned, int *this_best_prio)
2169 return balance_tasks(this_rq, this_cpu, busiest,
2170 max_load_move, sd, idle, all_pinned,
2171 this_best_prio, &busiest->cfs);
2176 * move_tasks tries to move up to max_load_move weighted load from busiest to
2177 * this_rq, as part of a balancing operation within domain "sd".
2178 * Returns 1 if successful and 0 otherwise.
2180 * Called with both runqueues locked.
2182 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2183 unsigned long max_load_move,
2184 struct sched_domain *sd, enum cpu_idle_type idle,
2187 unsigned long total_load_moved = 0, load_moved;
2188 int this_best_prio = this_rq->curr->prio;
2191 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
2192 max_load_move - total_load_moved,
2193 sd, idle, all_pinned, &this_best_prio);
2195 total_load_moved += load_moved;
2197 #ifdef CONFIG_PREEMPT
2199 * NEWIDLE balancing is a source of latency, so preemptible
2200 * kernels will stop after the first task is pulled to minimize
2201 * the critical section.
2203 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
2206 if (raw_spin_is_contended(&this_rq->lock) ||
2207 raw_spin_is_contended(&busiest->lock))
2210 } while (load_moved && max_load_move > total_load_moved);
2212 return total_load_moved > 0;
2215 /********** Helpers for find_busiest_group ************************/
2217 * sd_lb_stats - Structure to store the statistics of a sched_domain
2218 * during load balancing.
2220 struct sd_lb_stats {
2221 struct sched_group *busiest; /* Busiest group in this sd */
2222 struct sched_group *this; /* Local group in this sd */
2223 unsigned long total_load; /* Total load of all groups in sd */
2224 unsigned long total_pwr; /* Total power of all groups in sd */
2225 unsigned long avg_load; /* Average load across all groups in sd */
2227 /** Statistics of this group */
2228 unsigned long this_load;
2229 unsigned long this_load_per_task;
2230 unsigned long this_nr_running;
2231 unsigned long this_has_capacity;
2232 unsigned int this_idle_cpus;
2234 /* Statistics of the busiest group */
2235 unsigned int busiest_idle_cpus;
2236 unsigned long max_load;
2237 unsigned long busiest_load_per_task;
2238 unsigned long busiest_nr_running;
2239 unsigned long busiest_group_capacity;
2240 unsigned long busiest_has_capacity;
2241 unsigned int busiest_group_weight;
2243 int group_imb; /* Is there imbalance in this sd */
2244 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2245 int power_savings_balance; /* Is powersave balance needed for this sd */
2246 struct sched_group *group_min; /* Least loaded group in sd */
2247 struct sched_group *group_leader; /* Group which relieves group_min */
2248 unsigned long min_load_per_task; /* load_per_task in group_min */
2249 unsigned long leader_nr_running; /* Nr running of group_leader */
2250 unsigned long min_nr_running; /* Nr running of group_min */
2255 * sg_lb_stats - stats of a sched_group required for load_balancing
2257 struct sg_lb_stats {
2258 unsigned long avg_load; /*Avg load across the CPUs of the group */
2259 unsigned long group_load; /* Total load over the CPUs of the group */
2260 unsigned long sum_nr_running; /* Nr tasks running in the group */
2261 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
2262 unsigned long group_capacity;
2263 unsigned long idle_cpus;
2264 unsigned long group_weight;
2265 int group_imb; /* Is there an imbalance in the group ? */
2266 int group_has_capacity; /* Is there extra capacity in the group? */
2270 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
2271 * @group: The group whose first cpu is to be returned.
2273 static inline unsigned int group_first_cpu(struct sched_group *group)
2275 return cpumask_first(sched_group_cpus(group));
2279 * get_sd_load_idx - Obtain the load index for a given sched domain.
2280 * @sd: The sched_domain whose load_idx is to be obtained.
2281 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
2283 static inline int get_sd_load_idx(struct sched_domain *sd,
2284 enum cpu_idle_type idle)
2290 load_idx = sd->busy_idx;
2293 case CPU_NEWLY_IDLE:
2294 load_idx = sd->newidle_idx;
2297 load_idx = sd->idle_idx;
2305 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2307 * init_sd_power_savings_stats - Initialize power savings statistics for
2308 * the given sched_domain, during load balancing.
2310 * @sd: Sched domain whose power-savings statistics are to be initialized.
2311 * @sds: Variable containing the statistics for sd.
2312 * @idle: Idle status of the CPU at which we're performing load-balancing.
2314 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2315 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2318 * Busy processors will not participate in power savings
2321 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
2322 sds->power_savings_balance = 0;
2324 sds->power_savings_balance = 1;
2325 sds->min_nr_running = ULONG_MAX;
2326 sds->leader_nr_running = 0;
2331 * update_sd_power_savings_stats - Update the power saving stats for a
2332 * sched_domain while performing load balancing.
2334 * @group: sched_group belonging to the sched_domain under consideration.
2335 * @sds: Variable containing the statistics of the sched_domain
2336 * @local_group: Does group contain the CPU for which we're performing
2338 * @sgs: Variable containing the statistics of the group.
2340 static inline void update_sd_power_savings_stats(struct sched_group *group,
2341 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2344 if (!sds->power_savings_balance)
2348 * If the local group is idle or completely loaded
2349 * no need to do power savings balance at this domain
2351 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
2352 !sds->this_nr_running))
2353 sds->power_savings_balance = 0;
2356 * If a group is already running at full capacity or idle,
2357 * don't include that group in power savings calculations
2359 if (!sds->power_savings_balance ||
2360 sgs->sum_nr_running >= sgs->group_capacity ||
2361 !sgs->sum_nr_running)
2365 * Calculate the group which has the least non-idle load.
2366 * This is the group from where we need to pick up the load
2369 if ((sgs->sum_nr_running < sds->min_nr_running) ||
2370 (sgs->sum_nr_running == sds->min_nr_running &&
2371 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
2372 sds->group_min = group;
2373 sds->min_nr_running = sgs->sum_nr_running;
2374 sds->min_load_per_task = sgs->sum_weighted_load /
2375 sgs->sum_nr_running;
2379 * Calculate the group which is almost near its
2380 * capacity but still has some space to pick up some load
2381 * from other group and save more power
2383 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
2386 if (sgs->sum_nr_running > sds->leader_nr_running ||
2387 (sgs->sum_nr_running == sds->leader_nr_running &&
2388 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
2389 sds->group_leader = group;
2390 sds->leader_nr_running = sgs->sum_nr_running;
2395 * check_power_save_busiest_group - see if there is potential for some power-savings balance
2396 * @sds: Variable containing the statistics of the sched_domain
2397 * under consideration.
2398 * @this_cpu: Cpu at which we're currently performing load-balancing.
2399 * @imbalance: Variable to store the imbalance.
2402 * Check if we have potential to perform some power-savings balance.
2403 * If yes, set the busiest group to be the least loaded group in the
2404 * sched_domain, so that it's CPUs can be put to idle.
2406 * Returns 1 if there is potential to perform power-savings balance.
2409 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2410 int this_cpu, unsigned long *imbalance)
2412 if (!sds->power_savings_balance)
2415 if (sds->this != sds->group_leader ||
2416 sds->group_leader == sds->group_min)
2419 *imbalance = sds->min_load_per_task;
2420 sds->busiest = sds->group_min;
2425 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2426 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
2427 struct sd_lb_stats *sds, enum cpu_idle_type idle)
2432 static inline void update_sd_power_savings_stats(struct sched_group *group,
2433 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
2438 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
2439 int this_cpu, unsigned long *imbalance)
2443 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
2446 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
2448 return SCHED_LOAD_SCALE;
2451 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
2453 return default_scale_freq_power(sd, cpu);
2456 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
2458 unsigned long weight = sd->span_weight;
2459 unsigned long smt_gain = sd->smt_gain;
2466 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
2468 return default_scale_smt_power(sd, cpu);
2471 unsigned long scale_rt_power(int cpu)
2473 struct rq *rq = cpu_rq(cpu);
2474 u64 total, available;
2476 total = sched_avg_period() + (rq->clock - rq->age_stamp);
2478 if (unlikely(total < rq->rt_avg)) {
2479 /* Ensures that power won't end up being negative */
2482 available = total - rq->rt_avg;
2485 if (unlikely((s64)total < SCHED_LOAD_SCALE))
2486 total = SCHED_LOAD_SCALE;
2488 total >>= SCHED_LOAD_SHIFT;
2490 return div_u64(available, total);
2493 static void update_cpu_power(struct sched_domain *sd, int cpu)
2495 unsigned long weight = sd->span_weight;
2496 unsigned long power = SCHED_LOAD_SCALE;
2497 struct sched_group *sdg = sd->groups;
2499 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
2500 if (sched_feat(ARCH_POWER))
2501 power *= arch_scale_smt_power(sd, cpu);
2503 power *= default_scale_smt_power(sd, cpu);
2505 power >>= SCHED_LOAD_SHIFT;
2508 sdg->cpu_power_orig = power;
2510 if (sched_feat(ARCH_POWER))
2511 power *= arch_scale_freq_power(sd, cpu);
2513 power *= default_scale_freq_power(sd, cpu);
2515 power >>= SCHED_LOAD_SHIFT;
2517 power *= scale_rt_power(cpu);
2518 power >>= SCHED_LOAD_SHIFT;
2523 cpu_rq(cpu)->cpu_power = power;
2524 sdg->cpu_power = power;
2527 static void update_group_power(struct sched_domain *sd, int cpu)
2529 struct sched_domain *child = sd->child;
2530 struct sched_group *group, *sdg = sd->groups;
2531 unsigned long power;
2534 update_cpu_power(sd, cpu);
2540 group = child->groups;
2542 power += group->cpu_power;
2543 group = group->next;
2544 } while (group != child->groups);
2546 sdg->cpu_power = power;
2550 * Try and fix up capacity for tiny siblings, this is needed when
2551 * things like SD_ASYM_PACKING need f_b_g to select another sibling
2552 * which on its own isn't powerful enough.
2554 * See update_sd_pick_busiest() and check_asym_packing().
2557 fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
2560 * Only siblings can have significantly less than SCHED_LOAD_SCALE
2562 if (sd->level != SD_LV_SIBLING)
2566 * If ~90% of the cpu_power is still there, we're good.
2568 if (group->cpu_power * 32 > group->cpu_power_orig * 29)
2575 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
2576 * @sd: The sched_domain whose statistics are to be updated.
2577 * @group: sched_group whose statistics are to be updated.
2578 * @this_cpu: Cpu for which load balance is currently performed.
2579 * @idle: Idle status of this_cpu
2580 * @load_idx: Load index of sched_domain of this_cpu for load calc.
2581 * @sd_idle: Idle status of the sched_domain containing group.
2582 * @local_group: Does group contain this_cpu.
2583 * @cpus: Set of cpus considered for load balancing.
2584 * @balance: Should we balance.
2585 * @sgs: variable to hold the statistics for this group.
2587 static inline void update_sg_lb_stats(struct sched_domain *sd,
2588 struct sched_group *group, int this_cpu,
2589 enum cpu_idle_type idle, int load_idx, int *sd_idle,
2590 int local_group, const struct cpumask *cpus,
2591 int *balance, struct sg_lb_stats *sgs)
2593 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
2595 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2596 unsigned long avg_load_per_task = 0;
2599 balance_cpu = group_first_cpu(group);
2601 /* Tally up the load of all CPUs in the group */
2603 min_cpu_load = ~0UL;
2606 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
2607 struct rq *rq = cpu_rq(i);
2609 if (*sd_idle && rq->nr_running)
2612 /* Bias balancing toward cpus of our domain */
2614 if (idle_cpu(i) && !first_idle_cpu) {
2619 load = target_load(i, load_idx);
2621 load = source_load(i, load_idx);
2622 if (load > max_cpu_load) {
2623 max_cpu_load = load;
2624 max_nr_running = rq->nr_running;
2626 if (min_cpu_load > load)
2627 min_cpu_load = load;
2630 sgs->group_load += load;
2631 sgs->sum_nr_running += rq->nr_running;
2632 sgs->sum_weighted_load += weighted_cpuload(i);
2638 * First idle cpu or the first cpu(busiest) in this sched group
2639 * is eligible for doing load balancing at this and above
2640 * domains. In the newly idle case, we will allow all the cpu's
2641 * to do the newly idle load balance.
2643 if (idle != CPU_NEWLY_IDLE && local_group) {
2644 if (balance_cpu != this_cpu) {
2648 update_group_power(sd, this_cpu);
2651 /* Adjust by relative CPU power of the group */
2652 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
2655 * Consider the group unbalanced when the imbalance is larger
2656 * than the average weight of two tasks.
2658 * APZ: with cgroup the avg task weight can vary wildly and
2659 * might not be a suitable number - should we keep a
2660 * normalized nr_running number somewhere that negates
2663 if (sgs->sum_nr_running)
2664 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
2666 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
2669 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
2670 if (!sgs->group_capacity)
2671 sgs->group_capacity = fix_small_capacity(sd, group);
2672 sgs->group_weight = group->group_weight;
2674 if (sgs->group_capacity > sgs->sum_nr_running)
2675 sgs->group_has_capacity = 1;
2679 * update_sd_pick_busiest - return 1 on busiest group
2680 * @sd: sched_domain whose statistics are to be checked
2681 * @sds: sched_domain statistics
2682 * @sg: sched_group candidate to be checked for being the busiest
2683 * @sgs: sched_group statistics
2684 * @this_cpu: the current cpu
2686 * Determine if @sg is a busier group than the previously selected
2689 static bool update_sd_pick_busiest(struct sched_domain *sd,
2690 struct sd_lb_stats *sds,
2691 struct sched_group *sg,
2692 struct sg_lb_stats *sgs,
2695 if (sgs->avg_load <= sds->max_load)
2698 if (sgs->sum_nr_running > sgs->group_capacity)
2705 * ASYM_PACKING needs to move all the work to the lowest
2706 * numbered CPUs in the group, therefore mark all groups
2707 * higher than ourself as busy.
2709 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
2710 this_cpu < group_first_cpu(sg)) {
2714 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
2722 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
2723 * @sd: sched_domain whose statistics are to be updated.
2724 * @this_cpu: Cpu for which load balance is currently performed.
2725 * @idle: Idle status of this_cpu
2726 * @sd_idle: Idle status of the sched_domain containing sg.
2727 * @cpus: Set of cpus considered for load balancing.
2728 * @balance: Should we balance.
2729 * @sds: variable to hold the statistics for this sched_domain.
2731 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
2732 enum cpu_idle_type idle, int *sd_idle,
2733 const struct cpumask *cpus, int *balance,
2734 struct sd_lb_stats *sds)
2736 struct sched_domain *child = sd->child;
2737 struct sched_group *sg = sd->groups;
2738 struct sg_lb_stats sgs;
2739 int load_idx, prefer_sibling = 0;
2741 if (child && child->flags & SD_PREFER_SIBLING)
2744 init_sd_power_savings_stats(sd, sds, idle);
2745 load_idx = get_sd_load_idx(sd, idle);
2750 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
2751 memset(&sgs, 0, sizeof(sgs));
2752 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx, sd_idle,
2753 local_group, cpus, balance, &sgs);
2755 if (local_group && !(*balance))
2758 sds->total_load += sgs.group_load;
2759 sds->total_pwr += sg->cpu_power;
2762 * In case the child domain prefers tasks go to siblings
2763 * first, lower the sg capacity to one so that we'll try
2764 * and move all the excess tasks away. We lower the capacity
2765 * of a group only if the local group has the capacity to fit
2766 * these excess tasks, i.e. nr_running < group_capacity. The
2767 * extra check prevents the case where you always pull from the
2768 * heaviest group when it is already under-utilized (possible
2769 * with a large weight task outweighs the tasks on the system).
2771 if (prefer_sibling && !local_group && sds->this_has_capacity)
2772 sgs.group_capacity = min(sgs.group_capacity, 1UL);
2775 sds->this_load = sgs.avg_load;
2777 sds->this_nr_running = sgs.sum_nr_running;
2778 sds->this_load_per_task = sgs.sum_weighted_load;
2779 sds->this_has_capacity = sgs.group_has_capacity;
2780 sds->this_idle_cpus = sgs.idle_cpus;
2781 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
2782 sds->max_load = sgs.avg_load;
2784 sds->busiest_nr_running = sgs.sum_nr_running;
2785 sds->busiest_idle_cpus = sgs.idle_cpus;
2786 sds->busiest_group_capacity = sgs.group_capacity;
2787 sds->busiest_load_per_task = sgs.sum_weighted_load;
2788 sds->busiest_has_capacity = sgs.group_has_capacity;
2789 sds->busiest_group_weight = sgs.group_weight;
2790 sds->group_imb = sgs.group_imb;
2793 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
2795 } while (sg != sd->groups);
2798 int __weak arch_sd_sibling_asym_packing(void)
2800 return 0*SD_ASYM_PACKING;
2804 * check_asym_packing - Check to see if the group is packed into the
2807 * This is primarily intended to used at the sibling level. Some
2808 * cores like POWER7 prefer to use lower numbered SMT threads. In the
2809 * case of POWER7, it can move to lower SMT modes only when higher
2810 * threads are idle. When in lower SMT modes, the threads will
2811 * perform better since they share less core resources. Hence when we
2812 * have idle threads, we want them to be the higher ones.
2814 * This packing function is run on idle threads. It checks to see if
2815 * the busiest CPU in this domain (core in the P7 case) has a higher
2816 * CPU number than the packing function is being run on. Here we are
2817 * assuming lower CPU number will be equivalent to lower a SMT thread
2820 * Returns 1 when packing is required and a task should be moved to
2821 * this CPU. The amount of the imbalance is returned in *imbalance.
2823 * @sd: The sched_domain whose packing is to be checked.
2824 * @sds: Statistics of the sched_domain which is to be packed
2825 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2826 * @imbalance: returns amount of imbalanced due to packing.
2828 static int check_asym_packing(struct sched_domain *sd,
2829 struct sd_lb_stats *sds,
2830 int this_cpu, unsigned long *imbalance)
2834 if (!(sd->flags & SD_ASYM_PACKING))
2840 busiest_cpu = group_first_cpu(sds->busiest);
2841 if (this_cpu > busiest_cpu)
2844 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->cpu_power,
2850 * fix_small_imbalance - Calculate the minor imbalance that exists
2851 * amongst the groups of a sched_domain, during
2853 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
2854 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
2855 * @imbalance: Variable to store the imbalance.
2857 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
2858 int this_cpu, unsigned long *imbalance)
2860 unsigned long tmp, pwr_now = 0, pwr_move = 0;
2861 unsigned int imbn = 2;
2862 unsigned long scaled_busy_load_per_task;
2864 if (sds->this_nr_running) {
2865 sds->this_load_per_task /= sds->this_nr_running;
2866 if (sds->busiest_load_per_task >
2867 sds->this_load_per_task)
2870 sds->this_load_per_task =
2871 cpu_avg_load_per_task(this_cpu);
2873 scaled_busy_load_per_task = sds->busiest_load_per_task
2875 scaled_busy_load_per_task /= sds->busiest->cpu_power;
2877 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
2878 (scaled_busy_load_per_task * imbn)) {
2879 *imbalance = sds->busiest_load_per_task;
2884 * OK, we don't have enough imbalance to justify moving tasks,
2885 * however we may be able to increase total CPU power used by
2889 pwr_now += sds->busiest->cpu_power *
2890 min(sds->busiest_load_per_task, sds->max_load);
2891 pwr_now += sds->this->cpu_power *
2892 min(sds->this_load_per_task, sds->this_load);
2893 pwr_now /= SCHED_LOAD_SCALE;
2895 /* Amount of load we'd subtract */
2896 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2897 sds->busiest->cpu_power;
2898 if (sds->max_load > tmp)
2899 pwr_move += sds->busiest->cpu_power *
2900 min(sds->busiest_load_per_task, sds->max_load - tmp);
2902 /* Amount of load we'd add */
2903 if (sds->max_load * sds->busiest->cpu_power <
2904 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
2905 tmp = (sds->max_load * sds->busiest->cpu_power) /
2906 sds->this->cpu_power;
2908 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
2909 sds->this->cpu_power;
2910 pwr_move += sds->this->cpu_power *
2911 min(sds->this_load_per_task, sds->this_load + tmp);
2912 pwr_move /= SCHED_LOAD_SCALE;
2914 /* Move if we gain throughput */
2915 if (pwr_move > pwr_now)
2916 *imbalance = sds->busiest_load_per_task;
2920 * calculate_imbalance - Calculate the amount of imbalance present within the
2921 * groups of a given sched_domain during load balance.
2922 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
2923 * @this_cpu: Cpu for which currently load balance is being performed.
2924 * @imbalance: The variable to store the imbalance.
2926 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
2927 unsigned long *imbalance)
2929 unsigned long max_pull, load_above_capacity = ~0UL;
2931 sds->busiest_load_per_task /= sds->busiest_nr_running;
2932 if (sds->group_imb) {
2933 sds->busiest_load_per_task =
2934 min(sds->busiest_load_per_task, sds->avg_load);
2938 * In the presence of smp nice balancing, certain scenarios can have
2939 * max load less than avg load(as we skip the groups at or below
2940 * its cpu_power, while calculating max_load..)
2942 if (sds->max_load < sds->avg_load) {
2944 return fix_small_imbalance(sds, this_cpu, imbalance);
2947 if (!sds->group_imb) {
2949 * Don't want to pull so many tasks that a group would go idle.
2951 load_above_capacity = (sds->busiest_nr_running -
2952 sds->busiest_group_capacity);
2954 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
2956 load_above_capacity /= sds->busiest->cpu_power;
2960 * We're trying to get all the cpus to the average_load, so we don't
2961 * want to push ourselves above the average load, nor do we wish to
2962 * reduce the max loaded cpu below the average load. At the same time,
2963 * we also don't want to reduce the group load below the group capacity
2964 * (so that we can implement power-savings policies etc). Thus we look
2965 * for the minimum possible imbalance.
2966 * Be careful of negative numbers as they'll appear as very large values
2967 * with unsigned longs.
2969 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
2971 /* How much load to actually move to equalise the imbalance */
2972 *imbalance = min(max_pull * sds->busiest->cpu_power,
2973 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
2977 * if *imbalance is less than the average load per runnable task
2978 * there is no gaurantee that any tasks will be moved so we'll have
2979 * a think about bumping its value to force at least one task to be
2982 if (*imbalance < sds->busiest_load_per_task)
2983 return fix_small_imbalance(sds, this_cpu, imbalance);
2987 /******* find_busiest_group() helpers end here *********************/
2990 * find_busiest_group - Returns the busiest group within the sched_domain
2991 * if there is an imbalance. If there isn't an imbalance, and
2992 * the user has opted for power-savings, it returns a group whose
2993 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
2994 * such a group exists.
2996 * Also calculates the amount of weighted load which should be moved
2997 * to restore balance.
2999 * @sd: The sched_domain whose busiest group is to be returned.
3000 * @this_cpu: The cpu for which load balancing is currently being performed.
3001 * @imbalance: Variable which stores amount of weighted load which should
3002 * be moved to restore balance/put a group to idle.
3003 * @idle: The idle status of this_cpu.
3004 * @sd_idle: The idleness of sd
3005 * @cpus: The set of CPUs under consideration for load-balancing.
3006 * @balance: Pointer to a variable indicating if this_cpu
3007 * is the appropriate cpu to perform load balancing at this_level.
3009 * Returns: - the busiest group if imbalance exists.
3010 * - If no imbalance and user has opted for power-savings balance,
3011 * return the least loaded group whose CPUs can be
3012 * put to idle by rebalancing its tasks onto our group.
3014 static struct sched_group *
3015 find_busiest_group(struct sched_domain *sd, int this_cpu,
3016 unsigned long *imbalance, enum cpu_idle_type idle,
3017 int *sd_idle, const struct cpumask *cpus, int *balance)
3019 struct sd_lb_stats sds;
3021 memset(&sds, 0, sizeof(sds));
3024 * Compute the various statistics relavent for load balancing at
3027 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3030 /* Cases where imbalance does not exist from POV of this_cpu */
3031 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3033 * 2) There is no busy sibling group to pull from.
3034 * 3) This group is the busiest group.
3035 * 4) This group is more busy than the avg busieness at this
3037 * 5) The imbalance is within the specified limit.
3039 * Note: when doing newidle balance, if the local group has excess
3040 * capacity (i.e. nr_running < group_capacity) and the busiest group
3041 * does not have any capacity, we force a load balance to pull tasks
3042 * to the local group. In this case, we skip past checks 3, 4 and 5.
3047 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
3048 check_asym_packing(sd, &sds, this_cpu, imbalance))
3051 if (!sds.busiest || sds.busiest_nr_running == 0)
3054 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
3055 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
3056 !sds.busiest_has_capacity)
3059 if (sds.this_load >= sds.max_load)
3062 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3064 if (sds.this_load >= sds.avg_load)
3068 * In the CPU_NEWLY_IDLE, use imbalance_pct to be conservative.
3069 * And to check for busy balance use !idle_cpu instead of
3070 * CPU_NOT_IDLE. This is because HT siblings will use CPU_NOT_IDLE
3071 * even when they are idle.
3073 if (idle == CPU_NEWLY_IDLE || !idle_cpu(this_cpu)) {
3074 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3078 * This cpu is idle. If the busiest group load doesn't
3079 * have more tasks than the number of available cpu's and
3080 * there is no imbalance between this and busiest group
3081 * wrt to idle cpu's, it is balanced.
3083 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
3084 sds.busiest_nr_running <= sds.busiest_group_weight)
3089 /* Looks like there is an imbalance. Compute it */
3090 calculate_imbalance(&sds, this_cpu, imbalance);
3095 * There is no obvious imbalance. But check if we can do some balancing
3098 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3106 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3109 find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
3110 enum cpu_idle_type idle, unsigned long imbalance,
3111 const struct cpumask *cpus)
3113 struct rq *busiest = NULL, *rq;
3114 unsigned long max_load = 0;
3117 for_each_cpu(i, sched_group_cpus(group)) {
3118 unsigned long power = power_of(i);
3119 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
3123 capacity = fix_small_capacity(sd, group);
3125 if (!cpumask_test_cpu(i, cpus))
3129 wl = weighted_cpuload(i);
3132 * When comparing with imbalance, use weighted_cpuload()
3133 * which is not scaled with the cpu power.
3135 if (capacity && rq->nr_running == 1 && wl > imbalance)
3139 * For the load comparisons with the other cpu's, consider
3140 * the weighted_cpuload() scaled with the cpu power, so that
3141 * the load can be moved away from the cpu that is potentially
3142 * running at a lower capacity.
3144 wl = (wl * SCHED_LOAD_SCALE) / power;
3146 if (wl > max_load) {
3156 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3157 * so long as it is large enough.
3159 #define MAX_PINNED_INTERVAL 512
3161 /* Working cpumask for load_balance and load_balance_newidle. */
3162 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3164 static int need_active_balance(struct sched_domain *sd, int sd_idle, int idle,
3165 int busiest_cpu, int this_cpu)
3167 if (idle == CPU_NEWLY_IDLE) {
3170 * ASYM_PACKING needs to force migrate tasks from busy but
3171 * higher numbered CPUs in order to pack all tasks in the
3172 * lowest numbered CPUs.
3174 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
3178 * The only task running in a non-idle cpu can be moved to this
3179 * cpu in an attempt to completely freeup the other CPU
3182 * The package power saving logic comes from
3183 * find_busiest_group(). If there are no imbalance, then
3184 * f_b_g() will return NULL. However when sched_mc={1,2} then
3185 * f_b_g() will select a group from which a running task may be
3186 * pulled to this cpu in order to make the other package idle.
3187 * If there is no opportunity to make a package idle and if
3188 * there are no imbalance, then f_b_g() will return NULL and no
3189 * action will be taken in load_balance_newidle().
3191 * Under normal task pull operation due to imbalance, there
3192 * will be more than one task in the source run queue and
3193 * move_tasks() will succeed. ld_moved will be true and this
3194 * active balance code will not be triggered.
3196 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3197 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3200 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3204 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
3207 static int active_load_balance_cpu_stop(void *data);
3210 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3211 * tasks if there is an imbalance.
3213 static int load_balance(int this_cpu, struct rq *this_rq,
3214 struct sched_domain *sd, enum cpu_idle_type idle,
3217 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3218 struct sched_group *group;
3219 unsigned long imbalance;
3221 unsigned long flags;
3222 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3224 cpumask_copy(cpus, cpu_active_mask);
3227 * When power savings policy is enabled for the parent domain, idle
3228 * sibling can pick up load irrespective of busy siblings. In this case,
3229 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3230 * portraying it as CPU_NOT_IDLE.
3232 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3233 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3236 schedstat_inc(sd, lb_count[idle]);
3239 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3246 schedstat_inc(sd, lb_nobusyg[idle]);
3250 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
3252 schedstat_inc(sd, lb_nobusyq[idle]);
3256 BUG_ON(busiest == this_rq);
3258 schedstat_add(sd, lb_imbalance[idle], imbalance);
3261 if (busiest->nr_running > 1) {
3263 * Attempt to move tasks. If find_busiest_group has found
3264 * an imbalance but busiest->nr_running <= 1, the group is
3265 * still unbalanced. ld_moved simply stays zero, so it is
3266 * correctly treated as an imbalance.
3268 local_irq_save(flags);
3269 double_rq_lock(this_rq, busiest);
3270 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3271 imbalance, sd, idle, &all_pinned);
3272 double_rq_unlock(this_rq, busiest);
3273 local_irq_restore(flags);
3276 * some other cpu did the load balance for us.
3278 if (ld_moved && this_cpu != smp_processor_id())
3279 resched_cpu(this_cpu);
3281 /* All tasks on this runqueue were pinned by CPU affinity */
3282 if (unlikely(all_pinned)) {
3283 cpumask_clear_cpu(cpu_of(busiest), cpus);
3284 if (!cpumask_empty(cpus))
3291 schedstat_inc(sd, lb_failed[idle]);
3293 * Increment the failure counter only on periodic balance.
3294 * We do not want newidle balance, which can be very
3295 * frequent, pollute the failure counter causing
3296 * excessive cache_hot migrations and active balances.
3298 if (idle != CPU_NEWLY_IDLE)
3299 sd->nr_balance_failed++;
3301 if (need_active_balance(sd, sd_idle, idle, cpu_of(busiest),
3303 raw_spin_lock_irqsave(&busiest->lock, flags);
3305 /* don't kick the active_load_balance_cpu_stop,
3306 * if the curr task on busiest cpu can't be
3309 if (!cpumask_test_cpu(this_cpu,
3310 &busiest->curr->cpus_allowed)) {
3311 raw_spin_unlock_irqrestore(&busiest->lock,
3314 goto out_one_pinned;
3318 * ->active_balance synchronizes accesses to
3319 * ->active_balance_work. Once set, it's cleared
3320 * only after active load balance is finished.
3322 if (!busiest->active_balance) {
3323 busiest->active_balance = 1;
3324 busiest->push_cpu = this_cpu;
3327 raw_spin_unlock_irqrestore(&busiest->lock, flags);
3330 stop_one_cpu_nowait(cpu_of(busiest),
3331 active_load_balance_cpu_stop, busiest,
3332 &busiest->active_balance_work);
3335 * We've kicked active balancing, reset the failure
3338 sd->nr_balance_failed = sd->cache_nice_tries+1;
3341 sd->nr_balance_failed = 0;
3343 if (likely(!active_balance)) {
3344 /* We were unbalanced, so reset the balancing interval */
3345 sd->balance_interval = sd->min_interval;
3348 * If we've begun active balancing, start to back off. This
3349 * case may not be covered by the all_pinned logic if there
3350 * is only 1 task on the busy runqueue (because we don't call
3353 if (sd->balance_interval < sd->max_interval)
3354 sd->balance_interval *= 2;
3357 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3358 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3364 schedstat_inc(sd, lb_balanced[idle]);
3366 sd->nr_balance_failed = 0;
3369 /* tune up the balancing interval */
3370 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3371 (sd->balance_interval < sd->max_interval))
3372 sd->balance_interval *= 2;
3374 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3375 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3384 * idle_balance is called by schedule() if this_cpu is about to become
3385 * idle. Attempts to pull tasks from other CPUs.
3387 static void idle_balance(int this_cpu, struct rq *this_rq)
3389 struct sched_domain *sd;
3390 int pulled_task = 0;
3391 unsigned long next_balance = jiffies + HZ;
3393 this_rq->idle_stamp = this_rq->clock;
3395 if (this_rq->avg_idle < sysctl_sched_migration_cost)
3399 * Drop the rq->lock, but keep IRQ/preempt disabled.
3401 raw_spin_unlock(&this_rq->lock);
3403 update_shares(this_cpu);
3404 for_each_domain(this_cpu, sd) {
3405 unsigned long interval;
3408 if (!(sd->flags & SD_LOAD_BALANCE))
3411 if (sd->flags & SD_BALANCE_NEWIDLE) {
3412 /* If we've pulled tasks over stop searching: */
3413 pulled_task = load_balance(this_cpu, this_rq,
3414 sd, CPU_NEWLY_IDLE, &balance);
3417 interval = msecs_to_jiffies(sd->balance_interval);
3418 if (time_after(next_balance, sd->last_balance + interval))
3419 next_balance = sd->last_balance + interval;
3421 this_rq->idle_stamp = 0;
3426 raw_spin_lock(&this_rq->lock);
3428 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3430 * We are going idle. next_balance may be set based on
3431 * a busy processor. So reset next_balance.
3433 this_rq->next_balance = next_balance;
3438 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
3439 * running tasks off the busiest CPU onto idle CPUs. It requires at
3440 * least 1 task to be running on each physical CPU where possible, and
3441 * avoids physical / logical imbalances.
3443 static int active_load_balance_cpu_stop(void *data)
3445 struct rq *busiest_rq = data;
3446 int busiest_cpu = cpu_of(busiest_rq);
3447 int target_cpu = busiest_rq->push_cpu;
3448 struct rq *target_rq = cpu_rq(target_cpu);
3449 struct sched_domain *sd;
3451 raw_spin_lock_irq(&busiest_rq->lock);
3453 /* make sure the requested cpu hasn't gone down in the meantime */
3454 if (unlikely(busiest_cpu != smp_processor_id() ||
3455 !busiest_rq->active_balance))
3458 /* Is there any task to move? */
3459 if (busiest_rq->nr_running <= 1)
3463 * This condition is "impossible", if it occurs
3464 * we need to fix it. Originally reported by
3465 * Bjorn Helgaas on a 128-cpu setup.
3467 BUG_ON(busiest_rq == target_rq);
3469 /* move a task from busiest_rq to target_rq */
3470 double_lock_balance(busiest_rq, target_rq);
3472 /* Search for an sd spanning us and the target CPU. */
3473 for_each_domain(target_cpu, sd) {
3474 if ((sd->flags & SD_LOAD_BALANCE) &&
3475 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3480 schedstat_inc(sd, alb_count);
3482 if (move_one_task(target_rq, target_cpu, busiest_rq,
3484 schedstat_inc(sd, alb_pushed);
3486 schedstat_inc(sd, alb_failed);
3488 double_unlock_balance(busiest_rq, target_rq);
3490 busiest_rq->active_balance = 0;
3491 raw_spin_unlock_irq(&busiest_rq->lock);
3497 static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
3499 static void trigger_sched_softirq(void *data)
3501 raise_softirq_irqoff(SCHED_SOFTIRQ);
3504 static inline void init_sched_softirq_csd(struct call_single_data *csd)
3506 csd->func = trigger_sched_softirq;
3513 * idle load balancing details
3514 * - One of the idle CPUs nominates itself as idle load_balancer, while
3516 * - This idle load balancer CPU will also go into tickless mode when
3517 * it is idle, just like all other idle CPUs
3518 * - When one of the busy CPUs notice that there may be an idle rebalancing
3519 * needed, they will kick the idle load balancer, which then does idle
3520 * load balancing for all the idle CPUs.
3523 atomic_t load_balancer;
3524 atomic_t first_pick_cpu;
3525 atomic_t second_pick_cpu;
3526 cpumask_var_t idle_cpus_mask;
3527 cpumask_var_t grp_idle_mask;
3528 unsigned long next_balance; /* in jiffy units */
3529 } nohz ____cacheline_aligned;
3531 int get_nohz_load_balancer(void)
3533 return atomic_read(&nohz.load_balancer);
3536 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3538 * lowest_flag_domain - Return lowest sched_domain containing flag.
3539 * @cpu: The cpu whose lowest level of sched domain is to
3541 * @flag: The flag to check for the lowest sched_domain
3542 * for the given cpu.
3544 * Returns the lowest sched_domain of a cpu which contains the given flag.
3546 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
3548 struct sched_domain *sd;
3550 for_each_domain(cpu, sd)
3551 if (sd && (sd->flags & flag))
3558 * for_each_flag_domain - Iterates over sched_domains containing the flag.
3559 * @cpu: The cpu whose domains we're iterating over.
3560 * @sd: variable holding the value of the power_savings_sd
3562 * @flag: The flag to filter the sched_domains to be iterated.
3564 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
3565 * set, starting from the lowest sched_domain to the highest.
3567 #define for_each_flag_domain(cpu, sd, flag) \
3568 for (sd = lowest_flag_domain(cpu, flag); \
3569 (sd && (sd->flags & flag)); sd = sd->parent)
3572 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
3573 * @ilb_group: group to be checked for semi-idleness
3575 * Returns: 1 if the group is semi-idle. 0 otherwise.
3577 * We define a sched_group to be semi idle if it has atleast one idle-CPU
3578 * and atleast one non-idle CPU. This helper function checks if the given
3579 * sched_group is semi-idle or not.
3581 static inline int is_semi_idle_group(struct sched_group *ilb_group)
3583 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
3584 sched_group_cpus(ilb_group));
3587 * A sched_group is semi-idle when it has atleast one busy cpu
3588 * and atleast one idle cpu.
3590 if (cpumask_empty(nohz.grp_idle_mask))
3593 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
3599 * find_new_ilb - Finds the optimum idle load balancer for nomination.
3600 * @cpu: The cpu which is nominating a new idle_load_balancer.
3602 * Returns: Returns the id of the idle load balancer if it exists,
3603 * Else, returns >= nr_cpu_ids.
3605 * This algorithm picks the idle load balancer such that it belongs to a
3606 * semi-idle powersavings sched_domain. The idea is to try and avoid
3607 * completely idle packages/cores just for the purpose of idle load balancing
3608 * when there are other idle cpu's which are better suited for that job.
3610 static int find_new_ilb(int cpu)
3612 struct sched_domain *sd;
3613 struct sched_group *ilb_group;
3616 * Have idle load balancer selection from semi-idle packages only
3617 * when power-aware load balancing is enabled
3619 if (!(sched_smt_power_savings || sched_mc_power_savings))
3623 * Optimize for the case when we have no idle CPUs or only one
3624 * idle CPU. Don't walk the sched_domain hierarchy in such cases
3626 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
3629 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
3630 ilb_group = sd->groups;
3633 if (is_semi_idle_group(ilb_group))
3634 return cpumask_first(nohz.grp_idle_mask);
3636 ilb_group = ilb_group->next;
3638 } while (ilb_group != sd->groups);
3644 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
3645 static inline int find_new_ilb(int call_cpu)
3652 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
3653 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
3654 * CPU (if there is one).
3656 static void nohz_balancer_kick(int cpu)
3660 nohz.next_balance++;
3662 ilb_cpu = get_nohz_load_balancer();
3664 if (ilb_cpu >= nr_cpu_ids) {
3665 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
3666 if (ilb_cpu >= nr_cpu_ids)
3670 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
3671 struct call_single_data *cp;
3673 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
3674 cp = &per_cpu(remote_sched_softirq_cb, cpu);
3675 __smp_call_function_single(ilb_cpu, cp, 0);
3681 * This routine will try to nominate the ilb (idle load balancing)
3682 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3683 * load balancing on behalf of all those cpus.
3685 * When the ilb owner becomes busy, we will not have new ilb owner until some
3686 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
3687 * idle load balancing by kicking one of the idle CPUs.
3689 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
3690 * ilb owner CPU in future (when there is a need for idle load balancing on
3691 * behalf of all idle CPUs).
3693 void select_nohz_load_balancer(int stop_tick)
3695 int cpu = smp_processor_id();
3698 if (!cpu_active(cpu)) {
3699 if (atomic_read(&nohz.load_balancer) != cpu)
3703 * If we are going offline and still the leader,
3706 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3713 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
3715 if (atomic_read(&nohz.first_pick_cpu) == cpu)
3716 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
3717 if (atomic_read(&nohz.second_pick_cpu) == cpu)
3718 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3720 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
3723 /* make me the ilb owner */
3724 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
3729 * Check to see if there is a more power-efficient
3732 new_ilb = find_new_ilb(cpu);
3733 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
3734 atomic_set(&nohz.load_balancer, nr_cpu_ids);
3735 resched_cpu(new_ilb);
3741 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
3744 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
3746 if (atomic_read(&nohz.load_balancer) == cpu)
3747 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
3755 static DEFINE_SPINLOCK(balancing);
3758 * It checks each scheduling domain to see if it is due to be balanced,
3759 * and initiates a balancing operation if so.
3761 * Balancing parameters are set up in arch_init_sched_domains.
3763 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3766 struct rq *rq = cpu_rq(cpu);
3767 unsigned long interval;
3768 struct sched_domain *sd;
3769 /* Earliest time when we have to do rebalance again */
3770 unsigned long next_balance = jiffies + 60*HZ;
3771 int update_next_balance = 0;
3776 for_each_domain(cpu, sd) {
3777 if (!(sd->flags & SD_LOAD_BALANCE))
3780 interval = sd->balance_interval;
3781 if (idle != CPU_IDLE)
3782 interval *= sd->busy_factor;
3784 /* scale ms to jiffies */
3785 interval = msecs_to_jiffies(interval);
3786 if (unlikely(!interval))
3788 if (interval > HZ*NR_CPUS/10)
3789 interval = HZ*NR_CPUS/10;
3791 need_serialize = sd->flags & SD_SERIALIZE;
3793 if (need_serialize) {
3794 if (!spin_trylock(&balancing))
3798 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3799 if (load_balance(cpu, rq, sd, idle, &balance)) {
3801 * We've pulled tasks over so either we're no
3802 * longer idle, or one of our SMT siblings is
3805 idle = CPU_NOT_IDLE;
3807 sd->last_balance = jiffies;
3810 spin_unlock(&balancing);
3812 if (time_after(next_balance, sd->last_balance + interval)) {
3813 next_balance = sd->last_balance + interval;
3814 update_next_balance = 1;
3818 * Stop the load balance at this level. There is another
3819 * CPU in our sched group which is doing load balancing more
3827 * next_balance will be updated only when there is a need.
3828 * When the cpu is attached to null domain for ex, it will not be
3831 if (likely(update_next_balance))
3832 rq->next_balance = next_balance;
3837 * In CONFIG_NO_HZ case, the idle balance kickee will do the
3838 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3840 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
3842 struct rq *this_rq = cpu_rq(this_cpu);
3846 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
3849 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
3850 if (balance_cpu == this_cpu)
3854 * If this cpu gets work to do, stop the load balancing
3855 * work being done for other cpus. Next load
3856 * balancing owner will pick it up.
3858 if (need_resched()) {
3859 this_rq->nohz_balance_kick = 0;
3863 raw_spin_lock_irq(&this_rq->lock);
3864 update_rq_clock(this_rq);
3865 update_cpu_load(this_rq);
3866 raw_spin_unlock_irq(&this_rq->lock);
3868 rebalance_domains(balance_cpu, CPU_IDLE);
3870 rq = cpu_rq(balance_cpu);
3871 if (time_after(this_rq->next_balance, rq->next_balance))
3872 this_rq->next_balance = rq->next_balance;
3874 nohz.next_balance = this_rq->next_balance;
3875 this_rq->nohz_balance_kick = 0;
3879 * Current heuristic for kicking the idle load balancer
3880 * - first_pick_cpu is the one of the busy CPUs. It will kick
3881 * idle load balancer when it has more than one process active. This
3882 * eliminates the need for idle load balancing altogether when we have
3883 * only one running process in the system (common case).
3884 * - If there are more than one busy CPU, idle load balancer may have
3885 * to run for active_load_balance to happen (i.e., two busy CPUs are
3886 * SMT or core siblings and can run better if they move to different
3887 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
3888 * which will kick idle load balancer as soon as it has any load.
3890 static inline int nohz_kick_needed(struct rq *rq, int cpu)
3892 unsigned long now = jiffies;
3894 int first_pick_cpu, second_pick_cpu;
3896 if (time_before(now, nohz.next_balance))
3899 if (rq->idle_at_tick)
3902 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
3903 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
3905 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
3906 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
3909 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
3910 if (ret == nr_cpu_ids || ret == cpu) {
3911 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
3912 if (rq->nr_running > 1)
3915 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
3916 if (ret == nr_cpu_ids || ret == cpu) {
3924 static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
3928 * run_rebalance_domains is triggered when needed from the scheduler tick.
3929 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
3931 static void run_rebalance_domains(struct softirq_action *h)
3933 int this_cpu = smp_processor_id();
3934 struct rq *this_rq = cpu_rq(this_cpu);
3935 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3936 CPU_IDLE : CPU_NOT_IDLE;
3938 rebalance_domains(this_cpu, idle);
3941 * If this cpu has a pending nohz_balance_kick, then do the
3942 * balancing on behalf of the other idle cpus whose ticks are
3945 nohz_idle_balance(this_cpu, idle);
3948 static inline int on_null_domain(int cpu)
3950 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
3954 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3956 static inline void trigger_load_balance(struct rq *rq, int cpu)
3958 /* Don't need to rebalance while attached to NULL domain */
3959 if (time_after_eq(jiffies, rq->next_balance) &&
3960 likely(!on_null_domain(cpu)))
3961 raise_softirq(SCHED_SOFTIRQ);
3963 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
3964 nohz_balancer_kick(cpu);
3968 static void rq_online_fair(struct rq *rq)
3973 static void rq_offline_fair(struct rq *rq)
3978 #else /* CONFIG_SMP */
3981 * on UP we do not need to balance between CPUs:
3983 static inline void idle_balance(int cpu, struct rq *rq)
3987 #endif /* CONFIG_SMP */
3990 * scheduler tick hitting a task of our scheduling class:
3992 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
3994 struct cfs_rq *cfs_rq;
3995 struct sched_entity *se = &curr->se;
3997 for_each_sched_entity(se) {
3998 cfs_rq = cfs_rq_of(se);
3999 entity_tick(cfs_rq, se, queued);
4004 * called on fork with the child task as argument from the parent's context
4005 * - child not yet on the tasklist
4006 * - preemption disabled
4008 static void task_fork_fair(struct task_struct *p)
4010 struct cfs_rq *cfs_rq = task_cfs_rq(current);
4011 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
4012 int this_cpu = smp_processor_id();
4013 struct rq *rq = this_rq();
4014 unsigned long flags;
4016 raw_spin_lock_irqsave(&rq->lock, flags);
4018 update_rq_clock(rq);
4020 if (unlikely(task_cpu(p) != this_cpu)) {
4022 __set_task_cpu(p, this_cpu);
4026 update_curr(cfs_rq);
4029 se->vruntime = curr->vruntime;
4030 place_entity(cfs_rq, se, 1);
4032 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4034 * Upon rescheduling, sched_class::put_prev_task() will place
4035 * 'current' within the tree based on its new key value.
4037 swap(curr->vruntime, se->vruntime);
4038 resched_task(rq->curr);
4041 se->vruntime -= cfs_rq->min_vruntime;
4043 raw_spin_unlock_irqrestore(&rq->lock, flags);
4047 * Priority of the task has changed. Check to see if we preempt
4050 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
4051 int oldprio, int running)
4054 * Reschedule if we are currently running on this runqueue and
4055 * our priority decreased, or if we are not currently running on
4056 * this runqueue and our priority is higher than the current's
4059 if (p->prio > oldprio)
4060 resched_task(rq->curr);
4062 check_preempt_curr(rq, p, 0);
4066 * We switched to the sched_fair class.
4068 static void switched_to_fair(struct rq *rq, struct task_struct *p,
4072 * We were most likely switched from sched_rt, so
4073 * kick off the schedule if running, otherwise just see
4074 * if we can still preempt the current task.
4077 resched_task(rq->curr);
4079 check_preempt_curr(rq, p, 0);
4082 /* Account for a task changing its policy or group.
4084 * This routine is mostly called to set cfs_rq->curr field when a task
4085 * migrates between groups/classes.
4087 static void set_curr_task_fair(struct rq *rq)
4089 struct sched_entity *se = &rq->curr->se;
4091 for_each_sched_entity(se)
4092 set_next_entity(cfs_rq_of(se), se);
4095 #ifdef CONFIG_FAIR_GROUP_SCHED
4096 static void task_move_group_fair(struct task_struct *p, int on_rq)
4099 * If the task was not on the rq at the time of this cgroup movement
4100 * it must have been asleep, sleeping tasks keep their ->vruntime
4101 * absolute on their old rq until wakeup (needed for the fair sleeper
4102 * bonus in place_entity()).
4104 * If it was on the rq, we've just 'preempted' it, which does convert
4105 * ->vruntime to a relative base.
4107 * Make sure both cases convert their relative position when migrating
4108 * to another cgroup's rq. This does somewhat interfere with the
4109 * fair sleeper stuff for the first placement, but who cares.
4112 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
4113 set_task_rq(p, task_cpu(p));
4115 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
4119 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
4121 struct sched_entity *se = &task->se;
4122 unsigned int rr_interval = 0;
4125 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
4128 if (rq->cfs.load.weight)
4129 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
4135 * All the scheduling class methods:
4137 static const struct sched_class fair_sched_class = {
4138 .next = &idle_sched_class,
4139 .enqueue_task = enqueue_task_fair,
4140 .dequeue_task = dequeue_task_fair,
4141 .yield_task = yield_task_fair,
4143 .check_preempt_curr = check_preempt_wakeup,
4145 .pick_next_task = pick_next_task_fair,
4146 .put_prev_task = put_prev_task_fair,
4149 .select_task_rq = select_task_rq_fair,
4151 .rq_online = rq_online_fair,
4152 .rq_offline = rq_offline_fair,
4154 .task_waking = task_waking_fair,
4157 .set_curr_task = set_curr_task_fair,
4158 .task_tick = task_tick_fair,
4159 .task_fork = task_fork_fair,
4161 .prio_changed = prio_changed_fair,
4162 .switched_to = switched_to_fair,
4164 .get_rr_interval = get_rr_interval_fair,
4166 #ifdef CONFIG_FAIR_GROUP_SCHED
4167 .task_move_group = task_move_group_fair,
4171 #ifdef CONFIG_SCHED_DEBUG
4172 static void print_cfs_stats(struct seq_file *m, int cpu)
4174 struct cfs_rq *cfs_rq;
4177 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
4178 print_cfs_rq(m, cpu, cfs_rq);