warning removal
[cor_2_6_31.git] / kernel / sched_fair.c
blob9ffb2b2ceba484145b7e2f2a260c311c7eb32921
1 /*
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>
26 * Targeted preemption latency for CPU-bound tasks:
27 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
29 * NOTE: this latency value is not the same as the concept of
30 * 'timeslice length' - timeslices in CFS are of variable length
31 * and have no persistent notion like in traditional, time-slice
32 * based scheduling concepts.
34 * (to see the precise effective timeslice length of your workload,
35 * run vmstat and monitor the context-switches (cs) field)
37 unsigned int sysctl_sched_latency = 20000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity = 4000000ULL;
46 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
48 static unsigned int sched_nr_latency = 5;
51 * After fork, child runs first. (default) If set to 0 then
52 * parent will (try to) run first.
54 const_debug unsigned int sysctl_sched_child_runs_first = 1;
57 * sys_sched_yield() compat mode
59 * This option switches the agressive yield implementation of the
60 * old scheduler back on.
62 unsigned int __read_mostly sysctl_sched_compat_yield;
65 * SCHED_OTHER wake-up granularity.
66 * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 * This option delays the preemption effects of decoupled workloads
69 * and reduces their over-scheduling. Synchronous workloads will still
70 * have immediate wakeup/sleep latencies.
72 unsigned int sysctl_sched_wakeup_granularity = 5000000UL;
74 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
76 static const struct sched_class fair_sched_class;
78 /**************************************************************
79 * CFS operations on generic schedulable entities:
82 static inline struct task_struct *task_of(struct sched_entity *se)
84 return container_of(se, struct task_struct, se);
87 #ifdef CONFIG_FAIR_GROUP_SCHED
89 /* cpu runqueue to which this cfs_rq is attached */
90 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
92 return cfs_rq->rq;
95 /* An entity is a task if it doesn't "own" a runqueue */
96 #define entity_is_task(se) (!se->my_q)
98 /* Walk up scheduling entities hierarchy */
99 #define for_each_sched_entity(se) \
100 for (; se; se = se->parent)
102 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
104 return p->se.cfs_rq;
107 /* runqueue on which this entity is (to be) queued */
108 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
110 return se->cfs_rq;
113 /* runqueue "owned" by this group */
114 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
116 return grp->my_q;
119 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
120 * another cpu ('this_cpu')
122 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
124 return cfs_rq->tg->cfs_rq[this_cpu];
127 /* Iterate thr' all leaf cfs_rq's on a runqueue */
128 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
129 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
131 /* Do the two (enqueued) entities belong to the same group ? */
132 static inline int
133 is_same_group(struct sched_entity *se, struct sched_entity *pse)
135 if (se->cfs_rq == pse->cfs_rq)
136 return 1;
138 return 0;
141 static inline struct sched_entity *parent_entity(struct sched_entity *se)
143 return se->parent;
146 /* return depth at which a sched entity is present in the hierarchy */
147 static inline int depth_se(struct sched_entity *se)
149 int depth = 0;
151 for_each_sched_entity(se)
152 depth++;
154 return depth;
157 static void
158 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
160 int se_depth, pse_depth;
163 * preemption test can be made between sibling entities who are in the
164 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
165 * both tasks until we find their ancestors who are siblings of common
166 * parent.
169 /* First walk up until both entities are at same depth */
170 se_depth = depth_se(*se);
171 pse_depth = depth_se(*pse);
173 while (se_depth > pse_depth) {
174 se_depth--;
175 *se = parent_entity(*se);
178 while (pse_depth > se_depth) {
179 pse_depth--;
180 *pse = parent_entity(*pse);
183 while (!is_same_group(*se, *pse)) {
184 *se = parent_entity(*se);
185 *pse = parent_entity(*pse);
189 #else /* CONFIG_FAIR_GROUP_SCHED */
191 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
193 return container_of(cfs_rq, struct rq, cfs);
196 #define entity_is_task(se) 1
198 #define for_each_sched_entity(se) \
199 for (; se; se = NULL)
201 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
203 return &task_rq(p)->cfs;
206 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
208 struct task_struct *p = task_of(se);
209 struct rq *rq = task_rq(p);
211 return &rq->cfs;
214 /* runqueue "owned" by this group */
215 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
217 return NULL;
220 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
222 return &cpu_rq(this_cpu)->cfs;
225 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
226 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
228 static inline int
229 is_same_group(struct sched_entity *se, struct sched_entity *pse)
231 return 1;
234 static inline struct sched_entity *parent_entity(struct sched_entity *se)
236 return NULL;
239 static inline void
240 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
244 #endif /* CONFIG_FAIR_GROUP_SCHED */
247 /**************************************************************
248 * Scheduling class tree data structure manipulation methods:
251 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
253 s64 delta = (s64)(vruntime - min_vruntime);
254 if (delta > 0)
255 min_vruntime = vruntime;
257 return min_vruntime;
260 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
262 s64 delta = (s64)(vruntime - min_vruntime);
263 if (delta < 0)
264 min_vruntime = vruntime;
266 return min_vruntime;
269 static inline int entity_before(struct sched_entity *a,
270 struct sched_entity *b)
272 return (s64)(a->vruntime - b->vruntime) < 0;
275 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
277 return se->vruntime - cfs_rq->min_vruntime;
280 static void update_min_vruntime(struct cfs_rq *cfs_rq)
282 u64 vruntime = cfs_rq->min_vruntime;
284 if (cfs_rq->curr)
285 vruntime = cfs_rq->curr->vruntime;
287 if (cfs_rq->rb_leftmost) {
288 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
289 struct sched_entity,
290 run_node);
292 if (!cfs_rq->curr)
293 vruntime = se->vruntime;
294 else
295 vruntime = min_vruntime(vruntime, se->vruntime);
298 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
302 * Enqueue an entity into the rb-tree:
304 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
306 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
307 struct rb_node *parent = NULL;
308 struct sched_entity *entry;
309 s64 key = entity_key(cfs_rq, se);
310 int leftmost = 1;
313 * Find the right place in the rbtree:
315 while (*link) {
316 parent = *link;
317 entry = rb_entry(parent, struct sched_entity, run_node);
319 * We dont care about collisions. Nodes with
320 * the same key stay together.
322 if (key < entity_key(cfs_rq, entry)) {
323 link = &parent->rb_left;
324 } else {
325 link = &parent->rb_right;
326 leftmost = 0;
331 * Maintain a cache of leftmost tree entries (it is frequently
332 * used):
334 if (leftmost)
335 cfs_rq->rb_leftmost = &se->run_node;
337 rb_link_node(&se->run_node, parent, link);
338 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
341 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
343 if (cfs_rq->rb_leftmost == &se->run_node) {
344 struct rb_node *next_node;
346 next_node = rb_next(&se->run_node);
347 cfs_rq->rb_leftmost = next_node;
350 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
353 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
355 struct rb_node *left = cfs_rq->rb_leftmost;
357 if (!left)
358 return NULL;
360 return rb_entry(left, struct sched_entity, run_node);
363 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
365 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
367 if (!last)
368 return NULL;
370 return rb_entry(last, struct sched_entity, run_node);
373 /**************************************************************
374 * Scheduling class statistics methods:
377 #ifdef CONFIG_SCHED_DEBUG
378 int sched_nr_latency_handler(struct ctl_table *table, int write,
379 struct file *filp, void __user *buffer, size_t *lenp,
380 loff_t *ppos)
382 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
384 if (ret || !write)
385 return ret;
387 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
388 sysctl_sched_min_granularity);
390 return 0;
392 #endif
395 * delta /= w
397 static inline unsigned long
398 calc_delta_fair(unsigned long delta, struct sched_entity *se)
400 if (unlikely(se->load.weight != NICE_0_LOAD))
401 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
403 return delta;
407 * The idea is to set a period in which each task runs once.
409 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
410 * this period because otherwise the slices get too small.
412 * p = (nr <= nl) ? l : l*nr/nl
414 static u64 __sched_period(unsigned long nr_running)
416 u64 period = sysctl_sched_latency;
417 unsigned long nr_latency = sched_nr_latency;
419 if (unlikely(nr_running > nr_latency)) {
420 period = sysctl_sched_min_granularity;
421 period *= nr_running;
424 return period;
428 * We calculate the wall-time slice from the period by taking a part
429 * proportional to the weight.
431 * s = p*P[w/rw]
433 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
435 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
437 for_each_sched_entity(se) {
438 struct load_weight *load;
439 struct load_weight lw;
441 cfs_rq = cfs_rq_of(se);
442 load = &cfs_rq->load;
444 if (unlikely(!se->on_rq)) {
445 lw = cfs_rq->load;
447 update_load_add(&lw, se->load.weight);
448 load = &lw;
450 slice = calc_delta_mine(slice, se->load.weight, load);
452 return slice;
456 * We calculate the vruntime slice of a to be inserted task
458 * vs = s/w
460 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
462 return calc_delta_fair(sched_slice(cfs_rq, se), se);
466 * Update the current task's runtime statistics. Skip current tasks that
467 * are not in our scheduling class.
469 static inline void
470 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
471 unsigned long delta_exec)
473 unsigned long delta_exec_weighted;
475 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
477 curr->sum_exec_runtime += delta_exec;
478 schedstat_add(cfs_rq, exec_clock, delta_exec);
479 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
480 curr->vruntime += delta_exec_weighted;
481 update_min_vruntime(cfs_rq);
484 static void update_curr(struct cfs_rq *cfs_rq)
486 struct sched_entity *curr = cfs_rq->curr;
487 u64 now = rq_of(cfs_rq)->clock;
488 unsigned long delta_exec;
490 if (unlikely(!curr))
491 return;
494 * Get the amount of time the current task was running
495 * since the last time we changed load (this cannot
496 * overflow on 32 bits):
498 delta_exec = (unsigned long)(now - curr->exec_start);
499 if (!delta_exec)
500 return;
502 __update_curr(cfs_rq, curr, delta_exec);
503 curr->exec_start = now;
505 if (entity_is_task(curr)) {
506 struct task_struct *curtask = task_of(curr);
508 cpuacct_charge(curtask, delta_exec);
509 account_group_exec_runtime(curtask, delta_exec);
513 static inline void
514 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
516 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
520 * Task is being enqueued - update stats:
522 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 * Are we enqueueing a waiting task? (for current tasks
526 * a dequeue/enqueue event is a NOP)
528 if (se != cfs_rq->curr)
529 update_stats_wait_start(cfs_rq, se);
532 static void
533 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
535 schedstat_set(se->wait_max, max(se->wait_max,
536 rq_of(cfs_rq)->clock - se->wait_start));
537 schedstat_set(se->wait_count, se->wait_count + 1);
538 schedstat_set(se->wait_sum, se->wait_sum +
539 rq_of(cfs_rq)->clock - se->wait_start);
540 schedstat_set(se->wait_start, 0);
543 static inline void
544 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
547 * Mark the end of the wait period if dequeueing a
548 * waiting task:
550 if (se != cfs_rq->curr)
551 update_stats_wait_end(cfs_rq, se);
555 * We are picking a new current task - update its stats:
557 static inline void
558 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
561 * We are starting a new run period:
563 se->exec_start = rq_of(cfs_rq)->clock;
566 /**************************************************
567 * Scheduling class queueing methods:
570 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
571 static void
572 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
574 cfs_rq->task_weight += weight;
576 #else
577 static inline void
578 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
581 #endif
583 static void
584 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
586 update_load_add(&cfs_rq->load, se->load.weight);
587 if (!parent_entity(se))
588 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
589 if (entity_is_task(se)) {
590 add_cfs_task_weight(cfs_rq, se->load.weight);
591 list_add(&se->group_node, &cfs_rq->tasks);
593 cfs_rq->nr_running++;
594 se->on_rq = 1;
597 static void
598 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
600 update_load_sub(&cfs_rq->load, se->load.weight);
601 if (!parent_entity(se))
602 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
603 if (entity_is_task(se)) {
604 add_cfs_task_weight(cfs_rq, -se->load.weight);
605 list_del_init(&se->group_node);
607 cfs_rq->nr_running--;
608 se->on_rq = 0;
611 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
613 #ifdef CONFIG_SCHEDSTATS
614 if (se->sleep_start) {
615 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
616 struct task_struct *tsk = task_of(se);
618 if ((s64)delta < 0)
619 delta = 0;
621 if (unlikely(delta > se->sleep_max))
622 se->sleep_max = delta;
624 se->sleep_start = 0;
625 se->sum_sleep_runtime += delta;
627 account_scheduler_latency(tsk, delta >> 10, 1);
629 if (se->block_start) {
630 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
631 struct task_struct *tsk = task_of(se);
633 if ((s64)delta < 0)
634 delta = 0;
636 if (unlikely(delta > se->block_max))
637 se->block_max = delta;
639 se->block_start = 0;
640 se->sum_sleep_runtime += delta;
643 * Blocking time is in units of nanosecs, so shift by 20 to
644 * get a milliseconds-range estimation of the amount of
645 * time that the task spent sleeping:
647 if (unlikely(prof_on == SLEEP_PROFILING)) {
649 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
650 delta >> 20);
652 account_scheduler_latency(tsk, delta >> 10, 0);
654 #endif
657 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
659 #ifdef CONFIG_SCHED_DEBUG
660 s64 d = se->vruntime - cfs_rq->min_vruntime;
662 if (d < 0)
663 d = -d;
665 if (d > 3*sysctl_sched_latency)
666 schedstat_inc(cfs_rq, nr_spread_over);
667 #endif
670 static void
671 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
673 u64 vruntime = cfs_rq->min_vruntime;
676 * The 'current' period is already promised to the current tasks,
677 * however the extra weight of the new task will slow them down a
678 * little, place the new task so that it fits in the slot that
679 * stays open at the end.
681 if (initial && sched_feat(START_DEBIT))
682 vruntime += sched_vslice(cfs_rq, se);
684 if (!initial) {
685 /* sleeps upto a single latency don't count. */
686 if (sched_feat(NEW_FAIR_SLEEPERS)) {
687 unsigned long thresh = sysctl_sched_latency;
690 * Convert the sleeper threshold into virtual time.
691 * SCHED_IDLE is a special sub-class. We care about
692 * fairness only relative to other SCHED_IDLE tasks,
693 * all of which have the same weight.
695 if (sched_feat(NORMALIZED_SLEEPER) &&
696 (!entity_is_task(se) ||
697 task_of(se)->policy != SCHED_IDLE))
698 thresh = calc_delta_fair(thresh, se);
700 vruntime -= thresh;
703 /* ensure we never gain time by being placed backwards. */
704 vruntime = max_vruntime(se->vruntime, vruntime);
707 se->vruntime = vruntime;
710 static void
711 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
714 * Update run-time statistics of the 'current'.
716 update_curr(cfs_rq);
717 account_entity_enqueue(cfs_rq, se);
719 if (wakeup) {
720 place_entity(cfs_rq, se, 0);
721 enqueue_sleeper(cfs_rq, se);
724 update_stats_enqueue(cfs_rq, se);
725 check_spread(cfs_rq, se);
726 if (se != cfs_rq->curr)
727 __enqueue_entity(cfs_rq, se);
730 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
732 if (cfs_rq->last == se)
733 cfs_rq->last = NULL;
735 if (cfs_rq->next == se)
736 cfs_rq->next = NULL;
739 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
741 for_each_sched_entity(se)
742 __clear_buddies(cfs_rq_of(se), se);
745 static void
746 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
749 * Update run-time statistics of the 'current'.
751 update_curr(cfs_rq);
753 update_stats_dequeue(cfs_rq, se);
754 if (sleep) {
755 #ifdef CONFIG_SCHEDSTATS
756 if (entity_is_task(se)) {
757 struct task_struct *tsk = task_of(se);
759 if (tsk->state & TASK_INTERRUPTIBLE)
760 se->sleep_start = rq_of(cfs_rq)->clock;
761 if (tsk->state & TASK_UNINTERRUPTIBLE)
762 se->block_start = rq_of(cfs_rq)->clock;
764 #endif
767 clear_buddies(cfs_rq, se);
769 if (se != cfs_rq->curr)
770 __dequeue_entity(cfs_rq, se);
771 account_entity_dequeue(cfs_rq, se);
772 update_min_vruntime(cfs_rq);
776 * Preempt the current task with a newly woken task if needed:
778 static void
779 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
781 unsigned long ideal_runtime, delta_exec;
783 ideal_runtime = sched_slice(cfs_rq, curr);
784 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
785 if (delta_exec > ideal_runtime) {
786 resched_task(rq_of(cfs_rq)->curr);
788 * The current task ran long enough, ensure it doesn't get
789 * re-elected due to buddy favours.
791 clear_buddies(cfs_rq, curr);
795 static void
796 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
798 /* 'current' is not kept within the tree. */
799 if (se->on_rq) {
801 * Any task has to be enqueued before it get to execute on
802 * a CPU. So account for the time it spent waiting on the
803 * runqueue.
805 update_stats_wait_end(cfs_rq, se);
806 __dequeue_entity(cfs_rq, se);
809 update_stats_curr_start(cfs_rq, se);
810 cfs_rq->curr = se;
811 #ifdef CONFIG_SCHEDSTATS
813 * Track our maximum slice length, if the CPU's load is at
814 * least twice that of our own weight (i.e. dont track it
815 * when there are only lesser-weight tasks around):
817 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
818 se->slice_max = max(se->slice_max,
819 se->sum_exec_runtime - se->prev_sum_exec_runtime);
821 #endif
822 se->prev_sum_exec_runtime = se->sum_exec_runtime;
825 static int
826 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
828 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
830 struct sched_entity *se = __pick_next_entity(cfs_rq);
832 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
833 return cfs_rq->next;
835 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
836 return cfs_rq->last;
838 return se;
841 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
844 * If still on the runqueue then deactivate_task()
845 * was not called and update_curr() has to be done:
847 if (prev->on_rq)
848 update_curr(cfs_rq);
850 check_spread(cfs_rq, prev);
851 if (prev->on_rq) {
852 update_stats_wait_start(cfs_rq, prev);
853 /* Put 'current' back into the tree. */
854 __enqueue_entity(cfs_rq, prev);
856 cfs_rq->curr = NULL;
859 static void
860 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
863 * Update run-time statistics of the 'current'.
865 update_curr(cfs_rq);
867 #ifdef CONFIG_SCHED_HRTICK
869 * queued ticks are scheduled to match the slice, so don't bother
870 * validating it and just reschedule.
872 if (queued) {
873 resched_task(rq_of(cfs_rq)->curr);
874 return;
877 * don't let the period tick interfere with the hrtick preemption
879 if (!sched_feat(DOUBLE_TICK) &&
880 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
881 return;
882 #endif
884 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
885 check_preempt_tick(cfs_rq, curr);
888 /**************************************************
889 * CFS operations on tasks:
892 #ifdef CONFIG_SCHED_HRTICK
893 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
895 struct sched_entity *se = &p->se;
896 struct cfs_rq *cfs_rq = cfs_rq_of(se);
898 WARN_ON(task_rq(p) != rq);
900 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
901 u64 slice = sched_slice(cfs_rq, se);
902 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
903 s64 delta = slice - ran;
905 if (delta < 0) {
906 if (rq->curr == p)
907 resched_task(p);
908 return;
912 * Don't schedule slices shorter than 10000ns, that just
913 * doesn't make sense. Rely on vruntime for fairness.
915 if (rq->curr != p)
916 delta = max_t(s64, 10000LL, delta);
918 hrtick_start(rq, delta);
923 * called from enqueue/dequeue and updates the hrtick when the
924 * current task is from our class and nr_running is low enough
925 * to matter.
927 static void hrtick_update(struct rq *rq)
929 struct task_struct *curr = rq->curr;
931 if (curr->sched_class != &fair_sched_class)
932 return;
934 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
935 hrtick_start_fair(rq, curr);
937 #else /* !CONFIG_SCHED_HRTICK */
938 static inline void
939 hrtick_start_fair(struct rq *rq, struct task_struct *p)
943 static inline void hrtick_update(struct rq *rq)
946 #endif
949 * The enqueue_task method is called before nr_running is
950 * increased. Here we update the fair scheduling stats and
951 * then put the task into the rbtree:
953 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
955 struct cfs_rq *cfs_rq;
956 struct sched_entity *se = &p->se;
958 for_each_sched_entity(se) {
959 if (se->on_rq)
960 break;
961 cfs_rq = cfs_rq_of(se);
962 enqueue_entity(cfs_rq, se, wakeup);
963 wakeup = 1;
966 hrtick_update(rq);
970 * The dequeue_task method is called before nr_running is
971 * decreased. We remove the task from the rbtree and
972 * update the fair scheduling stats:
974 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
976 struct cfs_rq *cfs_rq;
977 struct sched_entity *se = &p->se;
979 for_each_sched_entity(se) {
980 cfs_rq = cfs_rq_of(se);
981 dequeue_entity(cfs_rq, se, sleep);
982 /* Don't dequeue parent if it has other entities besides us */
983 if (cfs_rq->load.weight)
984 break;
985 sleep = 1;
988 hrtick_update(rq);
992 * sched_yield() support is very simple - we dequeue and enqueue.
994 * If compat_yield is turned on then we requeue to the end of the tree.
996 static void yield_task_fair(struct rq *rq)
998 struct task_struct *curr = rq->curr;
999 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1000 struct sched_entity *rightmost, *se = &curr->se;
1003 * Are we the only task in the tree?
1005 if (unlikely(cfs_rq->nr_running == 1))
1006 return;
1008 clear_buddies(cfs_rq, se);
1010 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1011 update_rq_clock(rq);
1013 * Update run-time statistics of the 'current'.
1015 update_curr(cfs_rq);
1017 return;
1020 * Find the rightmost entry in the rbtree:
1022 rightmost = __pick_last_entity(cfs_rq);
1024 * Already in the rightmost position?
1026 if (unlikely(!rightmost || entity_before(rightmost, se)))
1027 return;
1030 * Minimally necessary key value to be last in the tree:
1031 * Upon rescheduling, sched_class::put_prev_task() will place
1032 * 'current' within the tree based on its new key value.
1034 se->vruntime = rightmost->vruntime + 1;
1038 * wake_idle() will wake a task on an idle cpu if task->cpu is
1039 * not idle and an idle cpu is available. The span of cpus to
1040 * search starts with cpus closest then further out as needed,
1041 * so we always favor a closer, idle cpu.
1042 * Domains may include CPUs that are not usable for migration,
1043 * hence we need to mask them out (cpu_active_mask)
1045 * Returns the CPU we should wake onto.
1047 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1048 static int wake_idle(int cpu, struct task_struct *p)
1050 struct sched_domain *sd;
1051 int i;
1052 unsigned int chosen_wakeup_cpu;
1053 int this_cpu;
1056 * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu
1057 * are idle and this is not a kernel thread and this task's affinity
1058 * allows it to be moved to preferred cpu, then just move!
1061 this_cpu = smp_processor_id();
1062 chosen_wakeup_cpu =
1063 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu;
1065 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP &&
1066 idle_cpu(cpu) && idle_cpu(this_cpu) &&
1067 p->mm && !(p->flags & PF_KTHREAD) &&
1068 cpu_isset(chosen_wakeup_cpu, p->cpus_allowed))
1069 return chosen_wakeup_cpu;
1072 * If it is idle, then it is the best cpu to run this task.
1074 * This cpu is also the best, if it has more than one task already.
1075 * Siblings must be also busy(in most cases) as they didn't already
1076 * pickup the extra load from this cpu and hence we need not check
1077 * sibling runqueue info. This will avoid the checks and cache miss
1078 * penalities associated with that.
1080 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
1081 return cpu;
1083 for_each_domain(cpu, sd) {
1084 if ((sd->flags & SD_WAKE_IDLE)
1085 || ((sd->flags & SD_WAKE_IDLE_FAR)
1086 && !task_hot(p, task_rq(p)->clock, sd))) {
1087 for_each_cpu_and(i, sched_domain_span(sd),
1088 &p->cpus_allowed) {
1089 if (cpu_active(i) && idle_cpu(i)) {
1090 if (i != task_cpu(p)) {
1091 schedstat_inc(p,
1092 se.nr_wakeups_idle);
1094 return i;
1097 } else {
1098 break;
1101 return cpu;
1103 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1104 static inline int wake_idle(int cpu, struct task_struct *p)
1106 return cpu;
1108 #endif
1110 #ifdef CONFIG_SMP
1112 #ifdef CONFIG_FAIR_GROUP_SCHED
1114 * effective_load() calculates the load change as seen from the root_task_group
1116 * Adding load to a group doesn't make a group heavier, but can cause movement
1117 * of group shares between cpus. Assuming the shares were perfectly aligned one
1118 * can calculate the shift in shares.
1120 * The problem is that perfectly aligning the shares is rather expensive, hence
1121 * we try to avoid doing that too often - see update_shares(), which ratelimits
1122 * this change.
1124 * We compensate this by not only taking the current delta into account, but
1125 * also considering the delta between when the shares were last adjusted and
1126 * now.
1128 * We still saw a performance dip, some tracing learned us that between
1129 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1130 * significantly. Therefore try to bias the error in direction of failing
1131 * the affine wakeup.
1134 static long effective_load(struct task_group *tg, int cpu,
1135 long wl, long wg)
1137 struct sched_entity *se = tg->se[cpu];
1139 if (!tg->parent)
1140 return wl;
1143 * By not taking the decrease of shares on the other cpu into
1144 * account our error leans towards reducing the affine wakeups.
1146 if (!wl && sched_feat(ASYM_EFF_LOAD))
1147 return wl;
1149 for_each_sched_entity(se) {
1150 long S, rw, s, a, b;
1151 long more_w;
1154 * Instead of using this increment, also add the difference
1155 * between when the shares were last updated and now.
1157 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1158 wl += more_w;
1159 wg += more_w;
1161 S = se->my_q->tg->shares;
1162 s = se->my_q->shares;
1163 rw = se->my_q->rq_weight;
1165 a = S*(rw + wl);
1166 b = S*rw + s*wg;
1168 wl = s*(a-b);
1170 if (likely(b))
1171 wl /= b;
1174 * Assume the group is already running and will
1175 * thus already be accounted for in the weight.
1177 * That is, moving shares between CPUs, does not
1178 * alter the group weight.
1180 wg = 0;
1183 return wl;
1186 #else
1188 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1189 unsigned long wl, unsigned long wg)
1191 return wl;
1194 #endif
1196 static int
1197 wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
1198 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1199 int idx, unsigned long load, unsigned long this_load,
1200 unsigned int imbalance)
1202 struct task_struct *curr = this_rq->curr;
1203 struct task_group *tg;
1204 unsigned long tl = this_load;
1205 unsigned long tl_per_task;
1206 unsigned long weight;
1207 int balanced;
1209 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1210 return 0;
1212 if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1213 p->se.avg_overlap > sysctl_sched_migration_cost))
1214 sync = 0;
1217 * If sync wakeup then subtract the (maximum possible)
1218 * effect of the currently running task from the load
1219 * of the current CPU:
1221 if (sync) {
1222 tg = task_group(current);
1223 weight = current->se.load.weight;
1225 tl += effective_load(tg, this_cpu, -weight, -weight);
1226 load += effective_load(tg, prev_cpu, 0, -weight);
1229 tg = task_group(p);
1230 weight = p->se.load.weight;
1232 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1233 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1236 * If the currently running task will sleep within
1237 * a reasonable amount of time then attract this newly
1238 * woken task:
1240 if (sync && balanced)
1241 return 1;
1243 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1244 tl_per_task = cpu_avg_load_per_task(this_cpu);
1246 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1247 tl_per_task)) {
1249 * This domain has SD_WAKE_AFFINE and
1250 * p is cache cold in this domain, and
1251 * there is no bad imbalance.
1253 schedstat_inc(this_sd, ttwu_move_affine);
1254 schedstat_inc(p, se.nr_wakeups_affine);
1256 return 1;
1258 return 0;
1261 static int select_task_rq_fair(struct task_struct *p, int sync)
1263 struct sched_domain *sd, *this_sd = NULL;
1264 int prev_cpu, this_cpu, new_cpu;
1265 unsigned long load, this_load;
1266 struct rq *this_rq;
1267 unsigned int imbalance;
1268 int idx;
1270 prev_cpu = task_cpu(p);
1271 this_cpu = smp_processor_id();
1272 this_rq = cpu_rq(this_cpu);
1273 new_cpu = prev_cpu;
1275 if (prev_cpu == this_cpu)
1276 goto out;
1278 * 'this_sd' is the first domain that both
1279 * this_cpu and prev_cpu are present in:
1281 for_each_domain(this_cpu, sd) {
1282 if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) {
1283 this_sd = sd;
1284 break;
1288 if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed)))
1289 goto out;
1292 * Check for affine wakeup and passive balancing possibilities.
1294 if (!this_sd)
1295 goto out;
1297 idx = this_sd->wake_idx;
1299 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1301 load = source_load(prev_cpu, idx);
1302 this_load = target_load(this_cpu, idx);
1304 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1305 load, this_load, imbalance))
1306 return this_cpu;
1309 * Start passive balancing when half the imbalance_pct
1310 * limit is reached.
1312 if (this_sd->flags & SD_WAKE_BALANCE) {
1313 if (imbalance*this_load <= 100*load) {
1314 schedstat_inc(this_sd, ttwu_move_balance);
1315 schedstat_inc(p, se.nr_wakeups_passive);
1316 return this_cpu;
1320 out:
1321 return wake_idle(new_cpu, p);
1323 #endif /* CONFIG_SMP */
1326 * Adaptive granularity
1328 * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1329 * with the limit of wakeup_gran -- when it never does a wakeup.
1331 * So the smaller avg_wakeup is the faster we want this task to preempt,
1332 * but we don't want to treat the preemptee unfairly and therefore allow it
1333 * to run for at least the amount of time we'd like to run.
1335 * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1337 * NOTE: we use *nr_running to scale with load, this nicely matches the
1338 * degrading latency on load.
1340 static unsigned long
1341 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1343 u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1344 u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1345 u64 gran = 0;
1347 if (this_run < expected_wakeup)
1348 gran = expected_wakeup - this_run;
1350 return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1353 static unsigned long
1354 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1356 unsigned long gran = sysctl_sched_wakeup_granularity;
1358 if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1359 gran = adaptive_gran(curr, se);
1362 * Since its curr running now, convert the gran from real-time
1363 * to virtual-time in his units.
1365 if (sched_feat(ASYM_GRAN)) {
1367 * By using 'se' instead of 'curr' we penalize light tasks, so
1368 * they get preempted easier. That is, if 'se' < 'curr' then
1369 * the resulting gran will be larger, therefore penalizing the
1370 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1371 * be smaller, again penalizing the lighter task.
1373 * This is especially important for buddies when the leftmost
1374 * task is higher priority than the buddy.
1376 if (unlikely(se->load.weight != NICE_0_LOAD))
1377 gran = calc_delta_fair(gran, se);
1378 } else {
1379 if (unlikely(curr->load.weight != NICE_0_LOAD))
1380 gran = calc_delta_fair(gran, curr);
1383 return gran;
1387 * Should 'se' preempt 'curr'.
1389 * |s1
1390 * |s2
1391 * |s3
1393 * |<--->|c
1395 * w(c, s1) = -1
1396 * w(c, s2) = 0
1397 * w(c, s3) = 1
1400 static int
1401 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1403 s64 gran, vdiff = curr->vruntime - se->vruntime;
1405 if (vdiff <= 0)
1406 return -1;
1408 gran = wakeup_gran(curr, se);
1409 if (vdiff > gran)
1410 return 1;
1412 return 0;
1415 static void set_last_buddy(struct sched_entity *se)
1417 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1418 for_each_sched_entity(se)
1419 cfs_rq_of(se)->last = se;
1423 static void set_next_buddy(struct sched_entity *se)
1425 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1426 for_each_sched_entity(se)
1427 cfs_rq_of(se)->next = se;
1432 * Preempt the current task with a newly woken task if needed:
1434 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
1436 struct task_struct *curr = rq->curr;
1437 struct sched_entity *se = &curr->se, *pse = &p->se;
1438 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1440 update_curr(cfs_rq);
1442 if (unlikely(rt_prio(p->prio))) {
1443 resched_task(curr);
1444 return;
1447 if (unlikely(p->sched_class != &fair_sched_class))
1448 return;
1450 if (unlikely(se == pse))
1451 return;
1454 * Only set the backward buddy when the current task is still on the
1455 * rq. This can happen when a wakeup gets interleaved with schedule on
1456 * the ->pre_schedule() or idle_balance() point, either of which can
1457 * drop the rq lock.
1459 * Also, during early boot the idle thread is in the fair class, for
1460 * obvious reasons its a bad idea to schedule back to the idle thread.
1462 if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1463 set_last_buddy(se);
1464 set_next_buddy(pse);
1467 * We can come here with TIF_NEED_RESCHED already set from new task
1468 * wake up path.
1470 if (test_tsk_need_resched(curr))
1471 return;
1474 * Batch and idle tasks do not preempt (their preemption is driven by
1475 * the tick):
1477 if (unlikely(p->policy != SCHED_NORMAL))
1478 return;
1480 /* Idle tasks are by definition preempted by everybody. */
1481 if (unlikely(curr->policy == SCHED_IDLE)) {
1482 resched_task(curr);
1483 return;
1486 if (!sched_feat(WAKEUP_PREEMPT))
1487 return;
1489 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1490 (se->avg_overlap < sysctl_sched_migration_cost &&
1491 pse->avg_overlap < sysctl_sched_migration_cost))) {
1492 resched_task(curr);
1493 return;
1496 find_matching_se(&se, &pse);
1498 BUG_ON(!pse);
1500 if (wakeup_preempt_entity(se, pse) == 1)
1501 resched_task(curr);
1504 static struct task_struct *pick_next_task_fair(struct rq *rq)
1506 struct task_struct *p;
1507 struct cfs_rq *cfs_rq = &rq->cfs;
1508 struct sched_entity *se;
1510 if (unlikely(!cfs_rq->nr_running))
1511 return NULL;
1513 do {
1514 se = pick_next_entity(cfs_rq);
1516 * If se was a buddy, clear it so that it will have to earn
1517 * the favour again.
1519 __clear_buddies(cfs_rq, se);
1520 set_next_entity(cfs_rq, se);
1521 cfs_rq = group_cfs_rq(se);
1522 } while (cfs_rq);
1524 p = task_of(se);
1525 hrtick_start_fair(rq, p);
1527 return p;
1531 * Account for a descheduled task:
1533 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1535 struct sched_entity *se = &prev->se;
1536 struct cfs_rq *cfs_rq;
1538 for_each_sched_entity(se) {
1539 cfs_rq = cfs_rq_of(se);
1540 put_prev_entity(cfs_rq, se);
1544 #ifdef CONFIG_SMP
1545 /**************************************************
1546 * Fair scheduling class load-balancing methods:
1550 * Load-balancing iterator. Note: while the runqueue stays locked
1551 * during the whole iteration, the current task might be
1552 * dequeued so the iterator has to be dequeue-safe. Here we
1553 * achieve that by always pre-iterating before returning
1554 * the current task:
1556 static struct task_struct *
1557 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1559 struct task_struct *p = NULL;
1560 struct sched_entity *se;
1562 if (next == &cfs_rq->tasks)
1563 return NULL;
1565 se = list_entry(next, struct sched_entity, group_node);
1566 p = task_of(se);
1567 cfs_rq->balance_iterator = next->next;
1569 return p;
1572 static struct task_struct *load_balance_start_fair(void *arg)
1574 struct cfs_rq *cfs_rq = arg;
1576 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1579 static struct task_struct *load_balance_next_fair(void *arg)
1581 struct cfs_rq *cfs_rq = arg;
1583 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1586 static unsigned long
1587 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1588 unsigned long max_load_move, struct sched_domain *sd,
1589 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1590 struct cfs_rq *cfs_rq)
1592 struct rq_iterator cfs_rq_iterator;
1594 cfs_rq_iterator.start = load_balance_start_fair;
1595 cfs_rq_iterator.next = load_balance_next_fair;
1596 cfs_rq_iterator.arg = cfs_rq;
1598 return balance_tasks(this_rq, this_cpu, busiest,
1599 max_load_move, sd, idle, all_pinned,
1600 this_best_prio, &cfs_rq_iterator);
1603 #ifdef CONFIG_FAIR_GROUP_SCHED
1604 static unsigned long
1605 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1606 unsigned long max_load_move,
1607 struct sched_domain *sd, enum cpu_idle_type idle,
1608 int *all_pinned, int *this_best_prio)
1610 long rem_load_move = max_load_move;
1611 int busiest_cpu = cpu_of(busiest);
1612 struct task_group *tg;
1614 rcu_read_lock();
1615 update_h_load(busiest_cpu);
1617 list_for_each_entry_rcu(tg, &task_groups, list) {
1618 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1619 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1620 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1621 u64 rem_load, moved_load;
1624 * empty group
1626 if (!busiest_cfs_rq->task_weight)
1627 continue;
1629 rem_load = (u64)rem_load_move * busiest_weight;
1630 rem_load = div_u64(rem_load, busiest_h_load + 1);
1632 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1633 rem_load, sd, idle, all_pinned, this_best_prio,
1634 tg->cfs_rq[busiest_cpu]);
1636 if (!moved_load)
1637 continue;
1639 moved_load *= busiest_h_load;
1640 moved_load = div_u64(moved_load, busiest_weight + 1);
1642 rem_load_move -= moved_load;
1643 if (rem_load_move < 0)
1644 break;
1646 rcu_read_unlock();
1648 return max_load_move - rem_load_move;
1650 #else
1651 static unsigned long
1652 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1653 unsigned long max_load_move,
1654 struct sched_domain *sd, enum cpu_idle_type idle,
1655 int *all_pinned, int *this_best_prio)
1657 return __load_balance_fair(this_rq, this_cpu, busiest,
1658 max_load_move, sd, idle, all_pinned,
1659 this_best_prio, &busiest->cfs);
1661 #endif
1663 static int
1664 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1665 struct sched_domain *sd, enum cpu_idle_type idle)
1667 struct cfs_rq *busy_cfs_rq;
1668 struct rq_iterator cfs_rq_iterator;
1670 cfs_rq_iterator.start = load_balance_start_fair;
1671 cfs_rq_iterator.next = load_balance_next_fair;
1673 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1675 * pass busy_cfs_rq argument into
1676 * load_balance_[start|next]_fair iterators
1678 cfs_rq_iterator.arg = busy_cfs_rq;
1679 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1680 &cfs_rq_iterator))
1681 return 1;
1684 return 0;
1686 #endif /* CONFIG_SMP */
1689 * scheduler tick hitting a task of our scheduling class:
1691 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1693 struct cfs_rq *cfs_rq;
1694 struct sched_entity *se = &curr->se;
1696 for_each_sched_entity(se) {
1697 cfs_rq = cfs_rq_of(se);
1698 entity_tick(cfs_rq, se, queued);
1703 * Share the fairness runtime between parent and child, thus the
1704 * total amount of pressure for CPU stays equal - new tasks
1705 * get a chance to run but frequent forkers are not allowed to
1706 * monopolize the CPU. Note: the parent runqueue is locked,
1707 * the child is not running yet.
1709 static void task_new_fair(struct rq *rq, struct task_struct *p)
1711 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1712 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1713 int this_cpu = smp_processor_id();
1715 sched_info_queued(p);
1717 update_curr(cfs_rq);
1718 place_entity(cfs_rq, se, 1);
1720 /* 'curr' will be NULL if the child belongs to a different group */
1721 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1722 curr && entity_before(curr, se)) {
1724 * Upon rescheduling, sched_class::put_prev_task() will place
1725 * 'current' within the tree based on its new key value.
1727 swap(curr->vruntime, se->vruntime);
1728 resched_task(rq->curr);
1731 enqueue_task_fair(rq, p, 0);
1735 * Priority of the task has changed. Check to see if we preempt
1736 * the current task.
1738 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1739 int oldprio, int running)
1742 * Reschedule if we are currently running on this runqueue and
1743 * our priority decreased, or if we are not currently running on
1744 * this runqueue and our priority is higher than the current's
1746 if (running) {
1747 if (p->prio > oldprio)
1748 resched_task(rq->curr);
1749 } else
1750 check_preempt_curr(rq, p, 0);
1754 * We switched to the sched_fair class.
1756 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1757 int running)
1760 * We were most likely switched from sched_rt, so
1761 * kick off the schedule if running, otherwise just see
1762 * if we can still preempt the current task.
1764 if (running)
1765 resched_task(rq->curr);
1766 else
1767 check_preempt_curr(rq, p, 0);
1770 /* Account for a task changing its policy or group.
1772 * This routine is mostly called to set cfs_rq->curr field when a task
1773 * migrates between groups/classes.
1775 static void set_curr_task_fair(struct rq *rq)
1777 struct sched_entity *se = &rq->curr->se;
1779 for_each_sched_entity(se)
1780 set_next_entity(cfs_rq_of(se), se);
1783 #ifdef CONFIG_FAIR_GROUP_SCHED
1784 static void moved_group_fair(struct task_struct *p)
1786 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1788 update_curr(cfs_rq);
1789 place_entity(cfs_rq, &p->se, 1);
1791 #endif
1794 * All the scheduling class methods:
1796 static const struct sched_class fair_sched_class = {
1797 .next = &idle_sched_class,
1798 .enqueue_task = enqueue_task_fair,
1799 .dequeue_task = dequeue_task_fair,
1800 .yield_task = yield_task_fair,
1802 .check_preempt_curr = check_preempt_wakeup,
1804 .pick_next_task = pick_next_task_fair,
1805 .put_prev_task = put_prev_task_fair,
1807 #ifdef CONFIG_SMP
1808 .select_task_rq = select_task_rq_fair,
1810 .load_balance = load_balance_fair,
1811 .move_one_task = move_one_task_fair,
1812 #endif
1814 .set_curr_task = set_curr_task_fair,
1815 .task_tick = task_tick_fair,
1816 .task_new = task_new_fair,
1818 .prio_changed = prio_changed_fair,
1819 .switched_to = switched_to_fair,
1821 #ifdef CONFIG_FAIR_GROUP_SCHED
1822 .moved_group = moved_group_fair,
1823 #endif
1826 #ifdef CONFIG_SCHED_DEBUG
1827 static void print_cfs_stats(struct seq_file *m, int cpu)
1829 struct cfs_rq *cfs_rq;
1831 rcu_read_lock();
1832 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1833 print_cfs_rq(m, cpu, cfs_rq);
1834 rcu_read_unlock();
1836 #endif