ASoC: Blackfin I2S: fix resuming when device hasn't been used
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / sched_fair.c
blobba7fd6e9556f892dd941ecdd9e77be225929c251
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 s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
271 return se->vruntime - cfs_rq->min_vruntime;
274 static void update_min_vruntime(struct cfs_rq *cfs_rq)
276 u64 vruntime = cfs_rq->min_vruntime;
278 if (cfs_rq->curr)
279 vruntime = cfs_rq->curr->vruntime;
281 if (cfs_rq->rb_leftmost) {
282 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
283 struct sched_entity,
284 run_node);
286 if (!cfs_rq->curr)
287 vruntime = se->vruntime;
288 else
289 vruntime = min_vruntime(vruntime, se->vruntime);
292 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
296 * Enqueue an entity into the rb-tree:
298 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
300 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
301 struct rb_node *parent = NULL;
302 struct sched_entity *entry;
303 s64 key = entity_key(cfs_rq, se);
304 int leftmost = 1;
307 * Find the right place in the rbtree:
309 while (*link) {
310 parent = *link;
311 entry = rb_entry(parent, struct sched_entity, run_node);
313 * We dont care about collisions. Nodes with
314 * the same key stay together.
316 if (key < entity_key(cfs_rq, entry)) {
317 link = &parent->rb_left;
318 } else {
319 link = &parent->rb_right;
320 leftmost = 0;
325 * Maintain a cache of leftmost tree entries (it is frequently
326 * used):
328 if (leftmost)
329 cfs_rq->rb_leftmost = &se->run_node;
331 rb_link_node(&se->run_node, parent, link);
332 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
335 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
337 if (cfs_rq->rb_leftmost == &se->run_node) {
338 struct rb_node *next_node;
340 next_node = rb_next(&se->run_node);
341 cfs_rq->rb_leftmost = next_node;
344 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
347 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
349 struct rb_node *left = cfs_rq->rb_leftmost;
351 if (!left)
352 return NULL;
354 return rb_entry(left, struct sched_entity, run_node);
357 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
359 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
361 if (!last)
362 return NULL;
364 return rb_entry(last, struct sched_entity, run_node);
367 /**************************************************************
368 * Scheduling class statistics methods:
371 #ifdef CONFIG_SCHED_DEBUG
372 int sched_nr_latency_handler(struct ctl_table *table, int write,
373 struct file *filp, void __user *buffer, size_t *lenp,
374 loff_t *ppos)
376 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
378 if (ret || !write)
379 return ret;
381 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
382 sysctl_sched_min_granularity);
384 return 0;
386 #endif
389 * delta /= w
391 static inline unsigned long
392 calc_delta_fair(unsigned long delta, struct sched_entity *se)
394 if (unlikely(se->load.weight != NICE_0_LOAD))
395 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
397 return delta;
401 * The idea is to set a period in which each task runs once.
403 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
404 * this period because otherwise the slices get too small.
406 * p = (nr <= nl) ? l : l*nr/nl
408 static u64 __sched_period(unsigned long nr_running)
410 u64 period = sysctl_sched_latency;
411 unsigned long nr_latency = sched_nr_latency;
413 if (unlikely(nr_running > nr_latency)) {
414 period = sysctl_sched_min_granularity;
415 period *= nr_running;
418 return period;
422 * We calculate the wall-time slice from the period by taking a part
423 * proportional to the weight.
425 * s = p*P[w/rw]
427 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
429 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
431 for_each_sched_entity(se) {
432 struct load_weight *load;
433 struct load_weight lw;
435 cfs_rq = cfs_rq_of(se);
436 load = &cfs_rq->load;
438 if (unlikely(!se->on_rq)) {
439 lw = cfs_rq->load;
441 update_load_add(&lw, se->load.weight);
442 load = &lw;
444 slice = calc_delta_mine(slice, se->load.weight, load);
446 return slice;
450 * We calculate the vruntime slice of a to be inserted task
452 * vs = s/w
454 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
456 return calc_delta_fair(sched_slice(cfs_rq, se), se);
460 * Update the current task's runtime statistics. Skip current tasks that
461 * are not in our scheduling class.
463 static inline void
464 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
465 unsigned long delta_exec)
467 unsigned long delta_exec_weighted;
469 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
471 curr->sum_exec_runtime += delta_exec;
472 schedstat_add(cfs_rq, exec_clock, delta_exec);
473 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
474 curr->vruntime += delta_exec_weighted;
475 update_min_vruntime(cfs_rq);
478 static void update_curr(struct cfs_rq *cfs_rq)
480 struct sched_entity *curr = cfs_rq->curr;
481 u64 now = rq_of(cfs_rq)->clock;
482 unsigned long delta_exec;
484 if (unlikely(!curr))
485 return;
488 * Get the amount of time the current task was running
489 * since the last time we changed load (this cannot
490 * overflow on 32 bits):
492 delta_exec = (unsigned long)(now - curr->exec_start);
493 if (!delta_exec)
494 return;
496 __update_curr(cfs_rq, curr, delta_exec);
497 curr->exec_start = now;
499 if (entity_is_task(curr)) {
500 struct task_struct *curtask = task_of(curr);
502 cpuacct_charge(curtask, delta_exec);
503 account_group_exec_runtime(curtask, delta_exec);
507 static inline void
508 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
510 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
514 * Task is being enqueued - update stats:
516 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
519 * Are we enqueueing a waiting task? (for current tasks
520 * a dequeue/enqueue event is a NOP)
522 if (se != cfs_rq->curr)
523 update_stats_wait_start(cfs_rq, se);
526 static void
527 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
529 schedstat_set(se->wait_max, max(se->wait_max,
530 rq_of(cfs_rq)->clock - se->wait_start));
531 schedstat_set(se->wait_count, se->wait_count + 1);
532 schedstat_set(se->wait_sum, se->wait_sum +
533 rq_of(cfs_rq)->clock - se->wait_start);
534 schedstat_set(se->wait_start, 0);
537 static inline void
538 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
541 * Mark the end of the wait period if dequeueing a
542 * waiting task:
544 if (se != cfs_rq->curr)
545 update_stats_wait_end(cfs_rq, se);
549 * We are picking a new current task - update its stats:
551 static inline void
552 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
555 * We are starting a new run period:
557 se->exec_start = rq_of(cfs_rq)->clock;
560 /**************************************************
561 * Scheduling class queueing methods:
564 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
565 static void
566 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
568 cfs_rq->task_weight += weight;
570 #else
571 static inline void
572 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
575 #endif
577 static void
578 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
580 update_load_add(&cfs_rq->load, se->load.weight);
581 if (!parent_entity(se))
582 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
583 if (entity_is_task(se)) {
584 add_cfs_task_weight(cfs_rq, se->load.weight);
585 list_add(&se->group_node, &cfs_rq->tasks);
587 cfs_rq->nr_running++;
588 se->on_rq = 1;
591 static void
592 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
594 update_load_sub(&cfs_rq->load, se->load.weight);
595 if (!parent_entity(se))
596 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
597 if (entity_is_task(se)) {
598 add_cfs_task_weight(cfs_rq, -se->load.weight);
599 list_del_init(&se->group_node);
601 cfs_rq->nr_running--;
602 se->on_rq = 0;
605 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
607 #ifdef CONFIG_SCHEDSTATS
608 if (se->sleep_start) {
609 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
610 struct task_struct *tsk = task_of(se);
612 if ((s64)delta < 0)
613 delta = 0;
615 if (unlikely(delta > se->sleep_max))
616 se->sleep_max = delta;
618 se->sleep_start = 0;
619 se->sum_sleep_runtime += delta;
621 account_scheduler_latency(tsk, delta >> 10, 1);
623 if (se->block_start) {
624 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
625 struct task_struct *tsk = task_of(se);
627 if ((s64)delta < 0)
628 delta = 0;
630 if (unlikely(delta > se->block_max))
631 se->block_max = delta;
633 se->block_start = 0;
634 se->sum_sleep_runtime += delta;
637 * Blocking time is in units of nanosecs, so shift by 20 to
638 * get a milliseconds-range estimation of the amount of
639 * time that the task spent sleeping:
641 if (unlikely(prof_on == SLEEP_PROFILING)) {
643 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
644 delta >> 20);
646 account_scheduler_latency(tsk, delta >> 10, 0);
648 #endif
651 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
653 #ifdef CONFIG_SCHED_DEBUG
654 s64 d = se->vruntime - cfs_rq->min_vruntime;
656 if (d < 0)
657 d = -d;
659 if (d > 3*sysctl_sched_latency)
660 schedstat_inc(cfs_rq, nr_spread_over);
661 #endif
664 static void
665 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
667 u64 vruntime = cfs_rq->min_vruntime;
670 * The 'current' period is already promised to the current tasks,
671 * however the extra weight of the new task will slow them down a
672 * little, place the new task so that it fits in the slot that
673 * stays open at the end.
675 if (initial && sched_feat(START_DEBIT))
676 vruntime += sched_vslice(cfs_rq, se);
678 if (!initial) {
679 /* sleeps upto a single latency don't count. */
680 if (sched_feat(NEW_FAIR_SLEEPERS)) {
681 unsigned long thresh = sysctl_sched_latency;
684 * Convert the sleeper threshold into virtual time.
685 * SCHED_IDLE is a special sub-class. We care about
686 * fairness only relative to other SCHED_IDLE tasks,
687 * all of which have the same weight.
689 if (sched_feat(NORMALIZED_SLEEPER) &&
690 task_of(se)->policy != SCHED_IDLE)
691 thresh = calc_delta_fair(thresh, se);
693 vruntime -= thresh;
696 /* ensure we never gain time by being placed backwards. */
697 vruntime = max_vruntime(se->vruntime, vruntime);
700 se->vruntime = vruntime;
703 static void
704 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
707 * Update run-time statistics of the 'current'.
709 update_curr(cfs_rq);
710 account_entity_enqueue(cfs_rq, se);
712 if (wakeup) {
713 place_entity(cfs_rq, se, 0);
714 enqueue_sleeper(cfs_rq, se);
717 update_stats_enqueue(cfs_rq, se);
718 check_spread(cfs_rq, se);
719 if (se != cfs_rq->curr)
720 __enqueue_entity(cfs_rq, se);
723 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
725 if (cfs_rq->last == se)
726 cfs_rq->last = NULL;
728 if (cfs_rq->next == se)
729 cfs_rq->next = NULL;
732 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
734 for_each_sched_entity(se)
735 __clear_buddies(cfs_rq_of(se), se);
738 static void
739 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
742 * Update run-time statistics of the 'current'.
744 update_curr(cfs_rq);
746 update_stats_dequeue(cfs_rq, se);
747 if (sleep) {
748 #ifdef CONFIG_SCHEDSTATS
749 if (entity_is_task(se)) {
750 struct task_struct *tsk = task_of(se);
752 if (tsk->state & TASK_INTERRUPTIBLE)
753 se->sleep_start = rq_of(cfs_rq)->clock;
754 if (tsk->state & TASK_UNINTERRUPTIBLE)
755 se->block_start = rq_of(cfs_rq)->clock;
757 #endif
760 clear_buddies(cfs_rq, se);
762 if (se != cfs_rq->curr)
763 __dequeue_entity(cfs_rq, se);
764 account_entity_dequeue(cfs_rq, se);
765 update_min_vruntime(cfs_rq);
769 * Preempt the current task with a newly woken task if needed:
771 static void
772 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
774 unsigned long ideal_runtime, delta_exec;
776 ideal_runtime = sched_slice(cfs_rq, curr);
777 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
778 if (delta_exec > ideal_runtime) {
779 resched_task(rq_of(cfs_rq)->curr);
781 * The current task ran long enough, ensure it doesn't get
782 * re-elected due to buddy favours.
784 clear_buddies(cfs_rq, curr);
788 static void
789 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
791 /* 'current' is not kept within the tree. */
792 if (se->on_rq) {
794 * Any task has to be enqueued before it get to execute on
795 * a CPU. So account for the time it spent waiting on the
796 * runqueue.
798 update_stats_wait_end(cfs_rq, se);
799 __dequeue_entity(cfs_rq, se);
802 update_stats_curr_start(cfs_rq, se);
803 cfs_rq->curr = se;
804 #ifdef CONFIG_SCHEDSTATS
806 * Track our maximum slice length, if the CPU's load is at
807 * least twice that of our own weight (i.e. dont track it
808 * when there are only lesser-weight tasks around):
810 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
811 se->slice_max = max(se->slice_max,
812 se->sum_exec_runtime - se->prev_sum_exec_runtime);
814 #endif
815 se->prev_sum_exec_runtime = se->sum_exec_runtime;
818 static int
819 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
821 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
823 struct sched_entity *se = __pick_next_entity(cfs_rq);
825 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
826 return cfs_rq->next;
828 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
829 return cfs_rq->last;
831 return se;
834 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
837 * If still on the runqueue then deactivate_task()
838 * was not called and update_curr() has to be done:
840 if (prev->on_rq)
841 update_curr(cfs_rq);
843 check_spread(cfs_rq, prev);
844 if (prev->on_rq) {
845 update_stats_wait_start(cfs_rq, prev);
846 /* Put 'current' back into the tree. */
847 __enqueue_entity(cfs_rq, prev);
849 cfs_rq->curr = NULL;
852 static void
853 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
856 * Update run-time statistics of the 'current'.
858 update_curr(cfs_rq);
860 #ifdef CONFIG_SCHED_HRTICK
862 * queued ticks are scheduled to match the slice, so don't bother
863 * validating it and just reschedule.
865 if (queued) {
866 resched_task(rq_of(cfs_rq)->curr);
867 return;
870 * don't let the period tick interfere with the hrtick preemption
872 if (!sched_feat(DOUBLE_TICK) &&
873 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
874 return;
875 #endif
877 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
878 check_preempt_tick(cfs_rq, curr);
881 /**************************************************
882 * CFS operations on tasks:
885 #ifdef CONFIG_SCHED_HRTICK
886 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
888 struct sched_entity *se = &p->se;
889 struct cfs_rq *cfs_rq = cfs_rq_of(se);
891 WARN_ON(task_rq(p) != rq);
893 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
894 u64 slice = sched_slice(cfs_rq, se);
895 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
896 s64 delta = slice - ran;
898 if (delta < 0) {
899 if (rq->curr == p)
900 resched_task(p);
901 return;
905 * Don't schedule slices shorter than 10000ns, that just
906 * doesn't make sense. Rely on vruntime for fairness.
908 if (rq->curr != p)
909 delta = max_t(s64, 10000LL, delta);
911 hrtick_start(rq, delta);
916 * called from enqueue/dequeue and updates the hrtick when the
917 * current task is from our class and nr_running is low enough
918 * to matter.
920 static void hrtick_update(struct rq *rq)
922 struct task_struct *curr = rq->curr;
924 if (curr->sched_class != &fair_sched_class)
925 return;
927 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
928 hrtick_start_fair(rq, curr);
930 #else /* !CONFIG_SCHED_HRTICK */
931 static inline void
932 hrtick_start_fair(struct rq *rq, struct task_struct *p)
936 static inline void hrtick_update(struct rq *rq)
939 #endif
942 * The enqueue_task method is called before nr_running is
943 * increased. Here we update the fair scheduling stats and
944 * then put the task into the rbtree:
946 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
948 struct cfs_rq *cfs_rq;
949 struct sched_entity *se = &p->se;
951 for_each_sched_entity(se) {
952 if (se->on_rq)
953 break;
954 cfs_rq = cfs_rq_of(se);
955 enqueue_entity(cfs_rq, se, wakeup);
956 wakeup = 1;
959 hrtick_update(rq);
963 * The dequeue_task method is called before nr_running is
964 * decreased. We remove the task from the rbtree and
965 * update the fair scheduling stats:
967 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
969 struct cfs_rq *cfs_rq;
970 struct sched_entity *se = &p->se;
972 for_each_sched_entity(se) {
973 cfs_rq = cfs_rq_of(se);
974 dequeue_entity(cfs_rq, se, sleep);
975 /* Don't dequeue parent if it has other entities besides us */
976 if (cfs_rq->load.weight)
977 break;
978 sleep = 1;
981 hrtick_update(rq);
985 * sched_yield() support is very simple - we dequeue and enqueue.
987 * If compat_yield is turned on then we requeue to the end of the tree.
989 static void yield_task_fair(struct rq *rq)
991 struct task_struct *curr = rq->curr;
992 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
993 struct sched_entity *rightmost, *se = &curr->se;
996 * Are we the only task in the tree?
998 if (unlikely(cfs_rq->nr_running == 1))
999 return;
1001 clear_buddies(cfs_rq, se);
1003 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1004 update_rq_clock(rq);
1006 * Update run-time statistics of the 'current'.
1008 update_curr(cfs_rq);
1010 return;
1013 * Find the rightmost entry in the rbtree:
1015 rightmost = __pick_last_entity(cfs_rq);
1017 * Already in the rightmost position?
1019 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
1020 return;
1023 * Minimally necessary key value to be last in the tree:
1024 * Upon rescheduling, sched_class::put_prev_task() will place
1025 * 'current' within the tree based on its new key value.
1027 se->vruntime = rightmost->vruntime + 1;
1031 * wake_idle() will wake a task on an idle cpu if task->cpu is
1032 * not idle and an idle cpu is available. The span of cpus to
1033 * search starts with cpus closest then further out as needed,
1034 * so we always favor a closer, idle cpu.
1035 * Domains may include CPUs that are not usable for migration,
1036 * hence we need to mask them out (cpu_active_mask)
1038 * Returns the CPU we should wake onto.
1040 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1041 static int wake_idle(int cpu, struct task_struct *p)
1043 struct sched_domain *sd;
1044 int i;
1045 unsigned int chosen_wakeup_cpu;
1046 int this_cpu;
1049 * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu
1050 * are idle and this is not a kernel thread and this task's affinity
1051 * allows it to be moved to preferred cpu, then just move!
1054 this_cpu = smp_processor_id();
1055 chosen_wakeup_cpu =
1056 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu;
1058 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP &&
1059 idle_cpu(cpu) && idle_cpu(this_cpu) &&
1060 p->mm && !(p->flags & PF_KTHREAD) &&
1061 cpu_isset(chosen_wakeup_cpu, p->cpus_allowed))
1062 return chosen_wakeup_cpu;
1065 * If it is idle, then it is the best cpu to run this task.
1067 * This cpu is also the best, if it has more than one task already.
1068 * Siblings must be also busy(in most cases) as they didn't already
1069 * pickup the extra load from this cpu and hence we need not check
1070 * sibling runqueue info. This will avoid the checks and cache miss
1071 * penalities associated with that.
1073 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
1074 return cpu;
1076 for_each_domain(cpu, sd) {
1077 if ((sd->flags & SD_WAKE_IDLE)
1078 || ((sd->flags & SD_WAKE_IDLE_FAR)
1079 && !task_hot(p, task_rq(p)->clock, sd))) {
1080 for_each_cpu_and(i, sched_domain_span(sd),
1081 &p->cpus_allowed) {
1082 if (cpu_active(i) && idle_cpu(i)) {
1083 if (i != task_cpu(p)) {
1084 schedstat_inc(p,
1085 se.nr_wakeups_idle);
1087 return i;
1090 } else {
1091 break;
1094 return cpu;
1096 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1097 static inline int wake_idle(int cpu, struct task_struct *p)
1099 return cpu;
1101 #endif
1103 #ifdef CONFIG_SMP
1105 #ifdef CONFIG_FAIR_GROUP_SCHED
1107 * effective_load() calculates the load change as seen from the root_task_group
1109 * Adding load to a group doesn't make a group heavier, but can cause movement
1110 * of group shares between cpus. Assuming the shares were perfectly aligned one
1111 * can calculate the shift in shares.
1113 * The problem is that perfectly aligning the shares is rather expensive, hence
1114 * we try to avoid doing that too often - see update_shares(), which ratelimits
1115 * this change.
1117 * We compensate this by not only taking the current delta into account, but
1118 * also considering the delta between when the shares were last adjusted and
1119 * now.
1121 * We still saw a performance dip, some tracing learned us that between
1122 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1123 * significantly. Therefore try to bias the error in direction of failing
1124 * the affine wakeup.
1127 static long effective_load(struct task_group *tg, int cpu,
1128 long wl, long wg)
1130 struct sched_entity *se = tg->se[cpu];
1132 if (!tg->parent)
1133 return wl;
1136 * By not taking the decrease of shares on the other cpu into
1137 * account our error leans towards reducing the affine wakeups.
1139 if (!wl && sched_feat(ASYM_EFF_LOAD))
1140 return wl;
1142 for_each_sched_entity(se) {
1143 long S, rw, s, a, b;
1144 long more_w;
1147 * Instead of using this increment, also add the difference
1148 * between when the shares were last updated and now.
1150 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1151 wl += more_w;
1152 wg += more_w;
1154 S = se->my_q->tg->shares;
1155 s = se->my_q->shares;
1156 rw = se->my_q->rq_weight;
1158 a = S*(rw + wl);
1159 b = S*rw + s*wg;
1161 wl = s*(a-b);
1163 if (likely(b))
1164 wl /= b;
1167 * Assume the group is already running and will
1168 * thus already be accounted for in the weight.
1170 * That is, moving shares between CPUs, does not
1171 * alter the group weight.
1173 wg = 0;
1176 return wl;
1179 #else
1181 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1182 unsigned long wl, unsigned long wg)
1184 return wl;
1187 #endif
1189 static int
1190 wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
1191 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1192 int idx, unsigned long load, unsigned long this_load,
1193 unsigned int imbalance)
1195 struct task_struct *curr = this_rq->curr;
1196 struct task_group *tg;
1197 unsigned long tl = this_load;
1198 unsigned long tl_per_task;
1199 unsigned long weight;
1200 int balanced;
1202 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1203 return 0;
1205 if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1206 p->se.avg_overlap > sysctl_sched_migration_cost))
1207 sync = 0;
1210 * If sync wakeup then subtract the (maximum possible)
1211 * effect of the currently running task from the load
1212 * of the current CPU:
1214 if (sync) {
1215 tg = task_group(current);
1216 weight = current->se.load.weight;
1218 tl += effective_load(tg, this_cpu, -weight, -weight);
1219 load += effective_load(tg, prev_cpu, 0, -weight);
1222 tg = task_group(p);
1223 weight = p->se.load.weight;
1225 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1226 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1229 * If the currently running task will sleep within
1230 * a reasonable amount of time then attract this newly
1231 * woken task:
1233 if (sync && balanced)
1234 return 1;
1236 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1237 tl_per_task = cpu_avg_load_per_task(this_cpu);
1239 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1240 tl_per_task)) {
1242 * This domain has SD_WAKE_AFFINE and
1243 * p is cache cold in this domain, and
1244 * there is no bad imbalance.
1246 schedstat_inc(this_sd, ttwu_move_affine);
1247 schedstat_inc(p, se.nr_wakeups_affine);
1249 return 1;
1251 return 0;
1254 static int select_task_rq_fair(struct task_struct *p, int sync)
1256 struct sched_domain *sd, *this_sd = NULL;
1257 int prev_cpu, this_cpu, new_cpu;
1258 unsigned long load, this_load;
1259 struct rq *this_rq;
1260 unsigned int imbalance;
1261 int idx;
1263 prev_cpu = task_cpu(p);
1264 this_cpu = smp_processor_id();
1265 this_rq = cpu_rq(this_cpu);
1266 new_cpu = prev_cpu;
1268 if (prev_cpu == this_cpu)
1269 goto out;
1271 * 'this_sd' is the first domain that both
1272 * this_cpu and prev_cpu are present in:
1274 for_each_domain(this_cpu, sd) {
1275 if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) {
1276 this_sd = sd;
1277 break;
1281 if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed)))
1282 goto out;
1285 * Check for affine wakeup and passive balancing possibilities.
1287 if (!this_sd)
1288 goto out;
1290 idx = this_sd->wake_idx;
1292 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1294 load = source_load(prev_cpu, idx);
1295 this_load = target_load(this_cpu, idx);
1297 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1298 load, this_load, imbalance))
1299 return this_cpu;
1302 * Start passive balancing when half the imbalance_pct
1303 * limit is reached.
1305 if (this_sd->flags & SD_WAKE_BALANCE) {
1306 if (imbalance*this_load <= 100*load) {
1307 schedstat_inc(this_sd, ttwu_move_balance);
1308 schedstat_inc(p, se.nr_wakeups_passive);
1309 return this_cpu;
1313 out:
1314 return wake_idle(new_cpu, p);
1316 #endif /* CONFIG_SMP */
1319 * Adaptive granularity
1321 * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1322 * with the limit of wakeup_gran -- when it never does a wakeup.
1324 * So the smaller avg_wakeup is the faster we want this task to preempt,
1325 * but we don't want to treat the preemptee unfairly and therefore allow it
1326 * to run for at least the amount of time we'd like to run.
1328 * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1330 * NOTE: we use *nr_running to scale with load, this nicely matches the
1331 * degrading latency on load.
1333 static unsigned long
1334 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1336 u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1337 u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1338 u64 gran = 0;
1340 if (this_run < expected_wakeup)
1341 gran = expected_wakeup - this_run;
1343 return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1346 static unsigned long
1347 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1349 unsigned long gran = sysctl_sched_wakeup_granularity;
1351 if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1352 gran = adaptive_gran(curr, se);
1355 * Since its curr running now, convert the gran from real-time
1356 * to virtual-time in his units.
1358 if (sched_feat(ASYM_GRAN)) {
1360 * By using 'se' instead of 'curr' we penalize light tasks, so
1361 * they get preempted easier. That is, if 'se' < 'curr' then
1362 * the resulting gran will be larger, therefore penalizing the
1363 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1364 * be smaller, again penalizing the lighter task.
1366 * This is especially important for buddies when the leftmost
1367 * task is higher priority than the buddy.
1369 if (unlikely(se->load.weight != NICE_0_LOAD))
1370 gran = calc_delta_fair(gran, se);
1371 } else {
1372 if (unlikely(curr->load.weight != NICE_0_LOAD))
1373 gran = calc_delta_fair(gran, curr);
1376 return gran;
1380 * Should 'se' preempt 'curr'.
1382 * |s1
1383 * |s2
1384 * |s3
1386 * |<--->|c
1388 * w(c, s1) = -1
1389 * w(c, s2) = 0
1390 * w(c, s3) = 1
1393 static int
1394 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1396 s64 gran, vdiff = curr->vruntime - se->vruntime;
1398 if (vdiff <= 0)
1399 return -1;
1401 gran = wakeup_gran(curr, se);
1402 if (vdiff > gran)
1403 return 1;
1405 return 0;
1408 static void set_last_buddy(struct sched_entity *se)
1410 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1411 for_each_sched_entity(se)
1412 cfs_rq_of(se)->last = se;
1416 static void set_next_buddy(struct sched_entity *se)
1418 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1419 for_each_sched_entity(se)
1420 cfs_rq_of(se)->next = se;
1425 * Preempt the current task with a newly woken task if needed:
1427 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
1429 struct task_struct *curr = rq->curr;
1430 struct sched_entity *se = &curr->se, *pse = &p->se;
1431 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1433 update_curr(cfs_rq);
1435 if (unlikely(rt_prio(p->prio))) {
1436 resched_task(curr);
1437 return;
1440 if (unlikely(p->sched_class != &fair_sched_class))
1441 return;
1443 if (unlikely(se == pse))
1444 return;
1447 * Only set the backward buddy when the current task is still on the
1448 * rq. This can happen when a wakeup gets interleaved with schedule on
1449 * the ->pre_schedule() or idle_balance() point, either of which can
1450 * drop the rq lock.
1452 * Also, during early boot the idle thread is in the fair class, for
1453 * obvious reasons its a bad idea to schedule back to the idle thread.
1455 if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1456 set_last_buddy(se);
1457 set_next_buddy(pse);
1460 * We can come here with TIF_NEED_RESCHED already set from new task
1461 * wake up path.
1463 if (test_tsk_need_resched(curr))
1464 return;
1467 * Batch and idle tasks do not preempt (their preemption is driven by
1468 * the tick):
1470 if (unlikely(p->policy != SCHED_NORMAL))
1471 return;
1473 /* Idle tasks are by definition preempted by everybody. */
1474 if (unlikely(curr->policy == SCHED_IDLE)) {
1475 resched_task(curr);
1476 return;
1479 if (!sched_feat(WAKEUP_PREEMPT))
1480 return;
1482 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1483 (se->avg_overlap < sysctl_sched_migration_cost &&
1484 pse->avg_overlap < sysctl_sched_migration_cost))) {
1485 resched_task(curr);
1486 return;
1489 find_matching_se(&se, &pse);
1491 BUG_ON(!pse);
1493 if (wakeup_preempt_entity(se, pse) == 1)
1494 resched_task(curr);
1497 static struct task_struct *pick_next_task_fair(struct rq *rq)
1499 struct task_struct *p;
1500 struct cfs_rq *cfs_rq = &rq->cfs;
1501 struct sched_entity *se;
1503 if (unlikely(!cfs_rq->nr_running))
1504 return NULL;
1506 do {
1507 se = pick_next_entity(cfs_rq);
1509 * If se was a buddy, clear it so that it will have to earn
1510 * the favour again.
1512 __clear_buddies(cfs_rq, se);
1513 set_next_entity(cfs_rq, se);
1514 cfs_rq = group_cfs_rq(se);
1515 } while (cfs_rq);
1517 p = task_of(se);
1518 hrtick_start_fair(rq, p);
1520 return p;
1524 * Account for a descheduled task:
1526 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1528 struct sched_entity *se = &prev->se;
1529 struct cfs_rq *cfs_rq;
1531 for_each_sched_entity(se) {
1532 cfs_rq = cfs_rq_of(se);
1533 put_prev_entity(cfs_rq, se);
1537 #ifdef CONFIG_SMP
1538 /**************************************************
1539 * Fair scheduling class load-balancing methods:
1543 * Load-balancing iterator. Note: while the runqueue stays locked
1544 * during the whole iteration, the current task might be
1545 * dequeued so the iterator has to be dequeue-safe. Here we
1546 * achieve that by always pre-iterating before returning
1547 * the current task:
1549 static struct task_struct *
1550 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1552 struct task_struct *p = NULL;
1553 struct sched_entity *se;
1555 if (next == &cfs_rq->tasks)
1556 return NULL;
1558 se = list_entry(next, struct sched_entity, group_node);
1559 p = task_of(se);
1560 cfs_rq->balance_iterator = next->next;
1562 return p;
1565 static struct task_struct *load_balance_start_fair(void *arg)
1567 struct cfs_rq *cfs_rq = arg;
1569 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1572 static struct task_struct *load_balance_next_fair(void *arg)
1574 struct cfs_rq *cfs_rq = arg;
1576 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1579 static unsigned long
1580 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1581 unsigned long max_load_move, struct sched_domain *sd,
1582 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1583 struct cfs_rq *cfs_rq)
1585 struct rq_iterator cfs_rq_iterator;
1587 cfs_rq_iterator.start = load_balance_start_fair;
1588 cfs_rq_iterator.next = load_balance_next_fair;
1589 cfs_rq_iterator.arg = cfs_rq;
1591 return balance_tasks(this_rq, this_cpu, busiest,
1592 max_load_move, sd, idle, all_pinned,
1593 this_best_prio, &cfs_rq_iterator);
1596 #ifdef CONFIG_FAIR_GROUP_SCHED
1597 static unsigned long
1598 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1599 unsigned long max_load_move,
1600 struct sched_domain *sd, enum cpu_idle_type idle,
1601 int *all_pinned, int *this_best_prio)
1603 long rem_load_move = max_load_move;
1604 int busiest_cpu = cpu_of(busiest);
1605 struct task_group *tg;
1607 rcu_read_lock();
1608 update_h_load(busiest_cpu);
1610 list_for_each_entry_rcu(tg, &task_groups, list) {
1611 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1612 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1613 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1614 u64 rem_load, moved_load;
1617 * empty group
1619 if (!busiest_cfs_rq->task_weight)
1620 continue;
1622 rem_load = (u64)rem_load_move * busiest_weight;
1623 rem_load = div_u64(rem_load, busiest_h_load + 1);
1625 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1626 rem_load, sd, idle, all_pinned, this_best_prio,
1627 tg->cfs_rq[busiest_cpu]);
1629 if (!moved_load)
1630 continue;
1632 moved_load *= busiest_h_load;
1633 moved_load = div_u64(moved_load, busiest_weight + 1);
1635 rem_load_move -= moved_load;
1636 if (rem_load_move < 0)
1637 break;
1639 rcu_read_unlock();
1641 return max_load_move - rem_load_move;
1643 #else
1644 static unsigned long
1645 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1646 unsigned long max_load_move,
1647 struct sched_domain *sd, enum cpu_idle_type idle,
1648 int *all_pinned, int *this_best_prio)
1650 return __load_balance_fair(this_rq, this_cpu, busiest,
1651 max_load_move, sd, idle, all_pinned,
1652 this_best_prio, &busiest->cfs);
1654 #endif
1656 static int
1657 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1658 struct sched_domain *sd, enum cpu_idle_type idle)
1660 struct cfs_rq *busy_cfs_rq;
1661 struct rq_iterator cfs_rq_iterator;
1663 cfs_rq_iterator.start = load_balance_start_fair;
1664 cfs_rq_iterator.next = load_balance_next_fair;
1666 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1668 * pass busy_cfs_rq argument into
1669 * load_balance_[start|next]_fair iterators
1671 cfs_rq_iterator.arg = busy_cfs_rq;
1672 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1673 &cfs_rq_iterator))
1674 return 1;
1677 return 0;
1679 #endif /* CONFIG_SMP */
1682 * scheduler tick hitting a task of our scheduling class:
1684 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1686 struct cfs_rq *cfs_rq;
1687 struct sched_entity *se = &curr->se;
1689 for_each_sched_entity(se) {
1690 cfs_rq = cfs_rq_of(se);
1691 entity_tick(cfs_rq, se, queued);
1696 * Share the fairness runtime between parent and child, thus the
1697 * total amount of pressure for CPU stays equal - new tasks
1698 * get a chance to run but frequent forkers are not allowed to
1699 * monopolize the CPU. Note: the parent runqueue is locked,
1700 * the child is not running yet.
1702 static void task_new_fair(struct rq *rq, struct task_struct *p)
1704 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1705 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1706 int this_cpu = smp_processor_id();
1708 sched_info_queued(p);
1710 update_curr(cfs_rq);
1711 place_entity(cfs_rq, se, 1);
1713 /* 'curr' will be NULL if the child belongs to a different group */
1714 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1715 curr && curr->vruntime < se->vruntime) {
1717 * Upon rescheduling, sched_class::put_prev_task() will place
1718 * 'current' within the tree based on its new key value.
1720 swap(curr->vruntime, se->vruntime);
1721 resched_task(rq->curr);
1724 enqueue_task_fair(rq, p, 0);
1728 * Priority of the task has changed. Check to see if we preempt
1729 * the current task.
1731 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1732 int oldprio, int running)
1735 * Reschedule if we are currently running on this runqueue and
1736 * our priority decreased, or if we are not currently running on
1737 * this runqueue and our priority is higher than the current's
1739 if (running) {
1740 if (p->prio > oldprio)
1741 resched_task(rq->curr);
1742 } else
1743 check_preempt_curr(rq, p, 0);
1747 * We switched to the sched_fair class.
1749 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1750 int running)
1753 * We were most likely switched from sched_rt, so
1754 * kick off the schedule if running, otherwise just see
1755 * if we can still preempt the current task.
1757 if (running)
1758 resched_task(rq->curr);
1759 else
1760 check_preempt_curr(rq, p, 0);
1763 /* Account for a task changing its policy or group.
1765 * This routine is mostly called to set cfs_rq->curr field when a task
1766 * migrates between groups/classes.
1768 static void set_curr_task_fair(struct rq *rq)
1770 struct sched_entity *se = &rq->curr->se;
1772 for_each_sched_entity(se)
1773 set_next_entity(cfs_rq_of(se), se);
1776 #ifdef CONFIG_FAIR_GROUP_SCHED
1777 static void moved_group_fair(struct task_struct *p)
1779 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1781 update_curr(cfs_rq);
1782 place_entity(cfs_rq, &p->se, 1);
1784 #endif
1787 * All the scheduling class methods:
1789 static const struct sched_class fair_sched_class = {
1790 .next = &idle_sched_class,
1791 .enqueue_task = enqueue_task_fair,
1792 .dequeue_task = dequeue_task_fair,
1793 .yield_task = yield_task_fair,
1795 .check_preempt_curr = check_preempt_wakeup,
1797 .pick_next_task = pick_next_task_fair,
1798 .put_prev_task = put_prev_task_fair,
1800 #ifdef CONFIG_SMP
1801 .select_task_rq = select_task_rq_fair,
1803 .load_balance = load_balance_fair,
1804 .move_one_task = move_one_task_fair,
1805 #endif
1807 .set_curr_task = set_curr_task_fair,
1808 .task_tick = task_tick_fair,
1809 .task_new = task_new_fair,
1811 .prio_changed = prio_changed_fair,
1812 .switched_to = switched_to_fair,
1814 #ifdef CONFIG_FAIR_GROUP_SCHED
1815 .moved_group = moved_group_fair,
1816 #endif
1819 #ifdef CONFIG_SCHED_DEBUG
1820 static void print_cfs_stats(struct seq_file *m, int cpu)
1822 struct cfs_rq *cfs_rq;
1824 rcu_read_lock();
1825 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1826 print_cfs_rq(m, cpu, cfs_rq);
1827 rcu_read_unlock();
1829 #endif