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: 10 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 = 10000000UL;
74 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
76 /**************************************************************
77 * CFS operations on generic schedulable entities:
80 static inline struct task_struct *task_of(struct sched_entity *se)
82 return container_of(se, struct task_struct, se);
85 #ifdef CONFIG_FAIR_GROUP_SCHED
87 /* cpu runqueue to which this cfs_rq is attached */
88 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
93 /* An entity is a task if it doesn't "own" a runqueue */
94 #define entity_is_task(se) (!se->my_q)
96 /* Walk up scheduling entities hierarchy */
97 #define for_each_sched_entity(se) \
98 for (; se; se = se->parent)
100 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
105 /* runqueue on which this entity is (to be) queued */
106 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
111 /* runqueue "owned" by this group */
112 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
117 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
118 * another cpu ('this_cpu')
120 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
122 return cfs_rq->tg->cfs_rq[this_cpu];
125 /* Iterate thr' all leaf cfs_rq's on a runqueue */
126 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
127 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
129 /* Do the two (enqueued) entities belong to the same group ? */
131 is_same_group(struct sched_entity *se, struct sched_entity *pse)
133 if (se->cfs_rq == pse->cfs_rq)
139 static inline struct sched_entity *parent_entity(struct sched_entity *se)
144 #else /* CONFIG_FAIR_GROUP_SCHED */
146 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
148 return container_of(cfs_rq, struct rq, cfs);
151 #define entity_is_task(se) 1
153 #define for_each_sched_entity(se) \
154 for (; se; se = NULL)
156 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
158 return &task_rq(p)->cfs;
161 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
163 struct task_struct *p = task_of(se);
164 struct rq *rq = task_rq(p);
169 /* runqueue "owned" by this group */
170 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
175 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
177 return &cpu_rq(this_cpu)->cfs;
180 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
181 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
184 is_same_group(struct sched_entity *se, struct sched_entity *pse)
189 static inline struct sched_entity *parent_entity(struct sched_entity *se)
194 #endif /* CONFIG_FAIR_GROUP_SCHED */
197 /**************************************************************
198 * Scheduling class tree data structure manipulation methods:
201 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
203 s64 delta = (s64)(vruntime - min_vruntime);
205 min_vruntime = vruntime;
210 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
212 s64 delta = (s64)(vruntime - min_vruntime);
214 min_vruntime = vruntime;
219 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
221 return se->vruntime - cfs_rq->min_vruntime;
225 * Enqueue an entity into the rb-tree:
227 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
229 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
230 struct rb_node *parent = NULL;
231 struct sched_entity *entry;
232 s64 key = entity_key(cfs_rq, se);
236 * Find the right place in the rbtree:
240 entry = rb_entry(parent, struct sched_entity, run_node);
242 * We dont care about collisions. Nodes with
243 * the same key stay together.
245 if (key < entity_key(cfs_rq, entry)) {
246 link = &parent->rb_left;
248 link = &parent->rb_right;
254 * Maintain a cache of leftmost tree entries (it is frequently
258 cfs_rq->rb_leftmost = &se->run_node;
260 * maintain cfs_rq->min_vruntime to be a monotonic increasing
261 * value tracking the leftmost vruntime in the tree.
263 cfs_rq->min_vruntime =
264 max_vruntime(cfs_rq->min_vruntime, se->vruntime);
267 rb_link_node(&se->run_node, parent, link);
268 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
271 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
273 if (cfs_rq->rb_leftmost == &se->run_node) {
274 struct rb_node *next_node;
275 struct sched_entity *next;
277 next_node = rb_next(&se->run_node);
278 cfs_rq->rb_leftmost = next_node;
281 next = rb_entry(next_node,
282 struct sched_entity, run_node);
283 cfs_rq->min_vruntime =
284 max_vruntime(cfs_rq->min_vruntime,
289 if (cfs_rq->next == se)
292 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
295 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
297 return cfs_rq->rb_leftmost;
300 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
302 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
305 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
307 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
312 return rb_entry(last, struct sched_entity, run_node);
315 /**************************************************************
316 * Scheduling class statistics methods:
319 #ifdef CONFIG_SCHED_DEBUG
320 int sched_nr_latency_handler(struct ctl_table *table, int write,
321 struct file *filp, void __user *buffer, size_t *lenp,
324 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
329 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
330 sysctl_sched_min_granularity);
337 * The idea is to set a period in which each task runs once.
339 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
340 * this period because otherwise the slices get too small.
342 * p = (nr <= nl) ? l : l*nr/nl
344 static u64 __sched_period(unsigned long nr_running)
346 u64 period = sysctl_sched_latency;
347 unsigned long nr_latency = sched_nr_latency;
349 if (unlikely(nr_running > nr_latency)) {
350 period = sysctl_sched_min_granularity;
351 period *= nr_running;
358 * We calculate the wall-time slice from the period by taking a part
359 * proportional to the weight.
363 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
365 return calc_delta_mine(__sched_period(cfs_rq->nr_running),
366 se->load.weight, &cfs_rq->load);
370 * We calculate the vruntime slice.
374 static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running)
376 u64 vslice = __sched_period(nr_running);
378 vslice *= NICE_0_LOAD;
379 do_div(vslice, rq_weight);
384 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
386 return __sched_vslice(cfs_rq->load.weight + se->load.weight,
387 cfs_rq->nr_running + 1);
391 * Update the current task's runtime statistics. Skip current tasks that
392 * are not in our scheduling class.
395 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
396 unsigned long delta_exec)
398 unsigned long delta_exec_weighted;
400 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
402 curr->sum_exec_runtime += delta_exec;
403 schedstat_add(cfs_rq, exec_clock, delta_exec);
404 delta_exec_weighted = delta_exec;
405 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
406 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
409 curr->vruntime += delta_exec_weighted;
412 static void update_curr(struct cfs_rq *cfs_rq)
414 struct sched_entity *curr = cfs_rq->curr;
415 u64 now = rq_of(cfs_rq)->clock;
416 unsigned long delta_exec;
422 * Get the amount of time the current task was running
423 * since the last time we changed load (this cannot
424 * overflow on 32 bits):
426 delta_exec = (unsigned long)(now - curr->exec_start);
428 __update_curr(cfs_rq, curr, delta_exec);
429 curr->exec_start = now;
431 if (entity_is_task(curr)) {
432 struct task_struct *curtask = task_of(curr);
434 cpuacct_charge(curtask, delta_exec);
439 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
441 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
445 * Task is being enqueued - update stats:
447 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
450 * Are we enqueueing a waiting task? (for current tasks
451 * a dequeue/enqueue event is a NOP)
453 if (se != cfs_rq->curr)
454 update_stats_wait_start(cfs_rq, se);
458 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
460 schedstat_set(se->wait_max, max(se->wait_max,
461 rq_of(cfs_rq)->clock - se->wait_start));
462 schedstat_set(se->wait_count, se->wait_count + 1);
463 schedstat_set(se->wait_sum, se->wait_sum +
464 rq_of(cfs_rq)->clock - se->wait_start);
465 schedstat_set(se->wait_start, 0);
469 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
472 * Mark the end of the wait period if dequeueing a
475 if (se != cfs_rq->curr)
476 update_stats_wait_end(cfs_rq, se);
480 * We are picking a new current task - update its stats:
483 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
486 * We are starting a new run period:
488 se->exec_start = rq_of(cfs_rq)->clock;
491 /**************************************************
492 * Scheduling class queueing methods:
495 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
497 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
499 cfs_rq->task_weight += weight;
503 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
509 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
511 update_load_add(&cfs_rq->load, se->load.weight);
512 if (!parent_entity(se))
513 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
514 if (entity_is_task(se))
515 add_cfs_task_weight(cfs_rq, se->load.weight);
516 cfs_rq->nr_running++;
521 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
523 update_load_sub(&cfs_rq->load, se->load.weight);
524 if (!parent_entity(se))
525 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
526 if (entity_is_task(se))
527 add_cfs_task_weight(cfs_rq, -se->load.weight);
528 cfs_rq->nr_running--;
532 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
534 #ifdef CONFIG_SCHEDSTATS
535 if (se->sleep_start) {
536 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
537 struct task_struct *tsk = task_of(se);
542 if (unlikely(delta > se->sleep_max))
543 se->sleep_max = delta;
546 se->sum_sleep_runtime += delta;
548 account_scheduler_latency(tsk, delta >> 10, 1);
550 if (se->block_start) {
551 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
552 struct task_struct *tsk = task_of(se);
557 if (unlikely(delta > se->block_max))
558 se->block_max = delta;
561 se->sum_sleep_runtime += delta;
564 * Blocking time is in units of nanosecs, so shift by 20 to
565 * get a milliseconds-range estimation of the amount of
566 * time that the task spent sleeping:
568 if (unlikely(prof_on == SLEEP_PROFILING)) {
570 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
573 account_scheduler_latency(tsk, delta >> 10, 0);
578 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
580 #ifdef CONFIG_SCHED_DEBUG
581 s64 d = se->vruntime - cfs_rq->min_vruntime;
586 if (d > 3*sysctl_sched_latency)
587 schedstat_inc(cfs_rq, nr_spread_over);
592 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
596 if (first_fair(cfs_rq)) {
597 vruntime = min_vruntime(cfs_rq->min_vruntime,
598 __pick_next_entity(cfs_rq)->vruntime);
600 vruntime = cfs_rq->min_vruntime;
603 * The 'current' period is already promised to the current tasks,
604 * however the extra weight of the new task will slow them down a
605 * little, place the new task so that it fits in the slot that
606 * stays open at the end.
608 if (initial && sched_feat(START_DEBIT))
609 vruntime += sched_vslice_add(cfs_rq, se);
612 /* sleeps upto a single latency don't count. */
613 if (sched_feat(NEW_FAIR_SLEEPERS)) {
614 if (sched_feat(NORMALIZED_SLEEPER))
615 vruntime -= calc_delta_fair(sysctl_sched_latency,
618 vruntime -= sysctl_sched_latency;
621 /* ensure we never gain time by being placed backwards. */
622 vruntime = max_vruntime(se->vruntime, vruntime);
625 se->vruntime = vruntime;
629 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
632 * Update run-time statistics of the 'current'.
637 place_entity(cfs_rq, se, 0);
638 enqueue_sleeper(cfs_rq, se);
641 update_stats_enqueue(cfs_rq, se);
642 check_spread(cfs_rq, se);
643 if (se != cfs_rq->curr)
644 __enqueue_entity(cfs_rq, se);
645 account_entity_enqueue(cfs_rq, se);
648 static void update_avg(u64 *avg, u64 sample)
650 s64 diff = sample - *avg;
654 static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 if (!se->last_wakeup)
659 update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup);
664 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
667 * Update run-time statistics of the 'current'.
671 update_stats_dequeue(cfs_rq, se);
673 update_avg_stats(cfs_rq, se);
674 #ifdef CONFIG_SCHEDSTATS
675 if (entity_is_task(se)) {
676 struct task_struct *tsk = task_of(se);
678 if (tsk->state & TASK_INTERRUPTIBLE)
679 se->sleep_start = rq_of(cfs_rq)->clock;
680 if (tsk->state & TASK_UNINTERRUPTIBLE)
681 se->block_start = rq_of(cfs_rq)->clock;
686 if (se != cfs_rq->curr)
687 __dequeue_entity(cfs_rq, se);
688 account_entity_dequeue(cfs_rq, se);
692 * Preempt the current task with a newly woken task if needed:
695 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
697 unsigned long ideal_runtime, delta_exec;
699 ideal_runtime = sched_slice(cfs_rq, curr);
700 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
701 if (delta_exec > ideal_runtime)
702 resched_task(rq_of(cfs_rq)->curr);
706 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
708 /* 'current' is not kept within the tree. */
711 * Any task has to be enqueued before it get to execute on
712 * a CPU. So account for the time it spent waiting on the
715 update_stats_wait_end(cfs_rq, se);
716 __dequeue_entity(cfs_rq, se);
719 update_stats_curr_start(cfs_rq, se);
721 #ifdef CONFIG_SCHEDSTATS
723 * Track our maximum slice length, if the CPU's load is at
724 * least twice that of our own weight (i.e. dont track it
725 * when there are only lesser-weight tasks around):
727 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
728 se->slice_max = max(se->slice_max,
729 se->sum_exec_runtime - se->prev_sum_exec_runtime);
732 se->prev_sum_exec_runtime = se->sum_exec_runtime;
736 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
738 static struct sched_entity *
739 pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se)
744 if (wakeup_preempt_entity(cfs_rq->next, se) != 0)
750 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
752 struct sched_entity *se = NULL;
754 if (first_fair(cfs_rq)) {
755 se = __pick_next_entity(cfs_rq);
756 se = pick_next(cfs_rq, se);
757 set_next_entity(cfs_rq, se);
763 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
766 * If still on the runqueue then deactivate_task()
767 * was not called and update_curr() has to be done:
772 check_spread(cfs_rq, prev);
774 update_stats_wait_start(cfs_rq, prev);
775 /* Put 'current' back into the tree. */
776 __enqueue_entity(cfs_rq, prev);
782 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
785 * Update run-time statistics of the 'current'.
789 #ifdef CONFIG_SCHED_HRTICK
791 * queued ticks are scheduled to match the slice, so don't bother
792 * validating it and just reschedule.
795 return resched_task(rq_of(cfs_rq)->curr);
797 * don't let the period tick interfere with the hrtick preemption
799 if (!sched_feat(DOUBLE_TICK) &&
800 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
804 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
805 check_preempt_tick(cfs_rq, curr);
808 /**************************************************
809 * CFS operations on tasks:
812 #ifdef CONFIG_SCHED_HRTICK
813 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
815 int requeue = rq->curr == p;
816 struct sched_entity *se = &p->se;
817 struct cfs_rq *cfs_rq = cfs_rq_of(se);
819 WARN_ON(task_rq(p) != rq);
821 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
822 u64 slice = sched_slice(cfs_rq, se);
823 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
824 s64 delta = slice - ran;
833 * Don't schedule slices shorter than 10000ns, that just
834 * doesn't make sense. Rely on vruntime for fairness.
837 delta = max(10000LL, delta);
839 hrtick_start(rq, delta, requeue);
844 hrtick_start_fair(struct rq *rq, struct task_struct *p)
850 * The enqueue_task method is called before nr_running is
851 * increased. Here we update the fair scheduling stats and
852 * then put the task into the rbtree:
854 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
856 struct cfs_rq *cfs_rq;
857 struct sched_entity *se = &p->se;
859 for_each_sched_entity(se) {
862 cfs_rq = cfs_rq_of(se);
863 enqueue_entity(cfs_rq, se, wakeup);
867 hrtick_start_fair(rq, rq->curr);
871 * The dequeue_task method is called before nr_running is
872 * decreased. We remove the task from the rbtree and
873 * update the fair scheduling stats:
875 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
877 struct cfs_rq *cfs_rq;
878 struct sched_entity *se = &p->se;
880 for_each_sched_entity(se) {
881 cfs_rq = cfs_rq_of(se);
882 dequeue_entity(cfs_rq, se, sleep);
883 /* Don't dequeue parent if it has other entities besides us */
884 if (cfs_rq->load.weight)
889 hrtick_start_fair(rq, rq->curr);
893 * sched_yield() support is very simple - we dequeue and enqueue.
895 * If compat_yield is turned on then we requeue to the end of the tree.
897 static void yield_task_fair(struct rq *rq)
899 struct task_struct *curr = rq->curr;
900 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
901 struct sched_entity *rightmost, *se = &curr->se;
904 * Are we the only task in the tree?
906 if (unlikely(cfs_rq->nr_running == 1))
909 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
910 __update_rq_clock(rq);
912 * Update run-time statistics of the 'current'.
919 * Find the rightmost entry in the rbtree:
921 rightmost = __pick_last_entity(cfs_rq);
923 * Already in the rightmost position?
925 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
929 * Minimally necessary key value to be last in the tree:
930 * Upon rescheduling, sched_class::put_prev_task() will place
931 * 'current' within the tree based on its new key value.
933 se->vruntime = rightmost->vruntime + 1;
937 * wake_idle() will wake a task on an idle cpu if task->cpu is
938 * not idle and an idle cpu is available. The span of cpus to
939 * search starts with cpus closest then further out as needed,
940 * so we always favor a closer, idle cpu.
942 * Returns the CPU we should wake onto.
944 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
945 static int wake_idle(int cpu, struct task_struct *p)
948 struct sched_domain *sd;
952 * If it is idle, then it is the best cpu to run this task.
954 * This cpu is also the best, if it has more than one task already.
955 * Siblings must be also busy(in most cases) as they didn't already
956 * pickup the extra load from this cpu and hence we need not check
957 * sibling runqueue info. This will avoid the checks and cache miss
958 * penalities associated with that.
960 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
963 for_each_domain(cpu, sd) {
964 if ((sd->flags & SD_WAKE_IDLE)
965 || ((sd->flags & SD_WAKE_IDLE_FAR)
966 && !task_hot(p, task_rq(p)->clock, sd))) {
967 cpus_and(tmp, sd->span, p->cpus_allowed);
968 for_each_cpu_mask(i, tmp) {
970 if (i != task_cpu(p)) {
984 static inline int wake_idle(int cpu, struct task_struct *p)
992 static const struct sched_class fair_sched_class;
995 wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq,
996 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
997 int idx, unsigned long load, unsigned long this_load,
998 unsigned int imbalance)
1000 struct task_struct *curr = this_rq->curr;
1001 unsigned long tl = this_load;
1002 unsigned long tl_per_task;
1004 if (!(this_sd->flags & SD_WAKE_AFFINE))
1008 * If the currently running task will sleep within
1009 * a reasonable amount of time then attract this newly
1012 if (sync && curr->sched_class == &fair_sched_class) {
1013 if (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1014 p->se.avg_overlap < sysctl_sched_migration_cost)
1018 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1019 tl_per_task = cpu_avg_load_per_task(this_cpu);
1022 * If sync wakeup then subtract the (maximum possible)
1023 * effect of the currently running task from the load
1024 * of the current CPU:
1027 tl -= current->se.load.weight;
1029 if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) ||
1030 100*(tl + p->se.load.weight) <= imbalance*load) {
1032 * This domain has SD_WAKE_AFFINE and
1033 * p is cache cold in this domain, and
1034 * there is no bad imbalance.
1036 schedstat_inc(this_sd, ttwu_move_affine);
1037 schedstat_inc(p, se.nr_wakeups_affine);
1044 static int select_task_rq_fair(struct task_struct *p, int sync)
1046 struct sched_domain *sd, *this_sd = NULL;
1047 int prev_cpu, this_cpu, new_cpu;
1048 unsigned long load, this_load;
1049 struct rq *rq, *this_rq;
1050 unsigned int imbalance;
1053 prev_cpu = task_cpu(p);
1055 this_cpu = smp_processor_id();
1056 this_rq = cpu_rq(this_cpu);
1060 * 'this_sd' is the first domain that both
1061 * this_cpu and prev_cpu are present in:
1063 for_each_domain(this_cpu, sd) {
1064 if (cpu_isset(prev_cpu, sd->span)) {
1070 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1074 * Check for affine wakeup and passive balancing possibilities.
1079 idx = this_sd->wake_idx;
1081 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1083 load = source_load(prev_cpu, idx);
1084 this_load = target_load(this_cpu, idx);
1086 if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1087 load, this_load, imbalance))
1090 if (prev_cpu == this_cpu)
1094 * Start passive balancing when half the imbalance_pct
1097 if (this_sd->flags & SD_WAKE_BALANCE) {
1098 if (imbalance*this_load <= 100*load) {
1099 schedstat_inc(this_sd, ttwu_move_balance);
1100 schedstat_inc(p, se.nr_wakeups_passive);
1106 return wake_idle(new_cpu, p);
1108 #endif /* CONFIG_SMP */
1110 static unsigned long wakeup_gran(struct sched_entity *se)
1112 unsigned long gran = sysctl_sched_wakeup_granularity;
1115 * More easily preempt - nice tasks, while not making
1116 * it harder for + nice tasks.
1118 if (unlikely(se->load.weight > NICE_0_LOAD))
1119 gran = calc_delta_fair(gran, &se->load);
1125 * Should 'se' preempt 'curr'.
1139 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1141 s64 gran, vdiff = curr->vruntime - se->vruntime;
1146 gran = wakeup_gran(curr);
1153 /* return depth at which a sched entity is present in the hierarchy */
1154 static inline int depth_se(struct sched_entity *se)
1158 for_each_sched_entity(se)
1165 * Preempt the current task with a newly woken task if needed:
1167 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
1169 struct task_struct *curr = rq->curr;
1170 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1171 struct sched_entity *se = &curr->se, *pse = &p->se;
1172 int se_depth, pse_depth;
1174 if (unlikely(rt_prio(p->prio))) {
1175 update_rq_clock(rq);
1176 update_curr(cfs_rq);
1181 se->last_wakeup = se->sum_exec_runtime;
1182 if (unlikely(se == pse))
1185 cfs_rq_of(pse)->next = pse;
1188 * Batch tasks do not preempt (their preemption is driven by
1191 if (unlikely(p->policy == SCHED_BATCH))
1194 if (!sched_feat(WAKEUP_PREEMPT))
1198 * preemption test can be made between sibling entities who are in the
1199 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
1200 * both tasks until we find their ancestors who are siblings of common
1204 /* First walk up until both entities are at same depth */
1205 se_depth = depth_se(se);
1206 pse_depth = depth_se(pse);
1208 while (se_depth > pse_depth) {
1210 se = parent_entity(se);
1213 while (pse_depth > se_depth) {
1215 pse = parent_entity(pse);
1218 while (!is_same_group(se, pse)) {
1219 se = parent_entity(se);
1220 pse = parent_entity(pse);
1223 if (wakeup_preempt_entity(se, pse) == 1)
1227 static struct task_struct *pick_next_task_fair(struct rq *rq)
1229 struct task_struct *p;
1230 struct cfs_rq *cfs_rq = &rq->cfs;
1231 struct sched_entity *se;
1233 if (unlikely(!cfs_rq->nr_running))
1237 se = pick_next_entity(cfs_rq);
1238 cfs_rq = group_cfs_rq(se);
1242 hrtick_start_fair(rq, p);
1248 * Account for a descheduled task:
1250 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1252 struct sched_entity *se = &prev->se;
1253 struct cfs_rq *cfs_rq;
1255 for_each_sched_entity(se) {
1256 cfs_rq = cfs_rq_of(se);
1257 put_prev_entity(cfs_rq, se);
1262 /**************************************************
1263 * Fair scheduling class load-balancing methods:
1267 * Load-balancing iterator. Note: while the runqueue stays locked
1268 * during the whole iteration, the current task might be
1269 * dequeued so the iterator has to be dequeue-safe. Here we
1270 * achieve that by always pre-iterating before returning
1273 static struct task_struct *
1274 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1276 struct task_struct *p = NULL;
1277 struct sched_entity *se;
1282 /* Skip over entities that are not tasks */
1284 se = rb_entry(curr, struct sched_entity, run_node);
1285 curr = rb_next(curr);
1286 } while (curr && !entity_is_task(se));
1288 cfs_rq->rb_load_balance_curr = curr;
1290 if (entity_is_task(se))
1296 static struct task_struct *load_balance_start_fair(void *arg)
1298 struct cfs_rq *cfs_rq = arg;
1300 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1303 static struct task_struct *load_balance_next_fair(void *arg)
1305 struct cfs_rq *cfs_rq = arg;
1307 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1310 static unsigned long
1311 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1312 unsigned long max_load_move, struct sched_domain *sd,
1313 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1314 struct cfs_rq *cfs_rq)
1316 struct rq_iterator cfs_rq_iterator;
1318 cfs_rq_iterator.start = load_balance_start_fair;
1319 cfs_rq_iterator.next = load_balance_next_fair;
1320 cfs_rq_iterator.arg = cfs_rq;
1322 return balance_tasks(this_rq, this_cpu, busiest,
1323 max_load_move, sd, idle, all_pinned,
1324 this_best_prio, &cfs_rq_iterator);
1327 #ifdef CONFIG_FAIR_GROUP_SCHED
1328 static unsigned long
1329 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1330 unsigned long max_load_move,
1331 struct sched_domain *sd, enum cpu_idle_type idle,
1332 int *all_pinned, int *this_best_prio)
1334 long rem_load_move = max_load_move;
1335 int busiest_cpu = cpu_of(busiest);
1336 struct task_group *tg;
1339 list_for_each_entry(tg, &task_groups, list) {
1341 unsigned long this_weight, busiest_weight;
1342 long rem_load, max_load, moved_load;
1347 if (!aggregate(tg, sd)->task_weight)
1350 rem_load = rem_load_move * aggregate(tg, sd)->rq_weight;
1351 rem_load /= aggregate(tg, sd)->load + 1;
1353 this_weight = tg->cfs_rq[this_cpu]->task_weight;
1354 busiest_weight = tg->cfs_rq[busiest_cpu]->task_weight;
1356 imbalance = (busiest_weight - this_weight) / 2;
1359 imbalance = busiest_weight;
1361 max_load = max(rem_load, imbalance);
1362 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1363 max_load, sd, idle, all_pinned, this_best_prio,
1364 tg->cfs_rq[busiest_cpu]);
1369 move_group_shares(tg, sd, busiest_cpu, this_cpu);
1371 moved_load *= aggregate(tg, sd)->load;
1372 moved_load /= aggregate(tg, sd)->rq_weight + 1;
1374 rem_load_move -= moved_load;
1375 if (rem_load_move < 0)
1380 return max_load_move - rem_load_move;
1383 static unsigned long
1384 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1385 unsigned long max_load_move,
1386 struct sched_domain *sd, enum cpu_idle_type idle,
1387 int *all_pinned, int *this_best_prio)
1389 return __load_balance_fair(this_rq, this_cpu, busiest,
1390 max_load_move, sd, idle, all_pinned,
1391 this_best_prio, &busiest->cfs);
1396 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1397 struct sched_domain *sd, enum cpu_idle_type idle)
1399 struct cfs_rq *busy_cfs_rq;
1400 struct rq_iterator cfs_rq_iterator;
1402 cfs_rq_iterator.start = load_balance_start_fair;
1403 cfs_rq_iterator.next = load_balance_next_fair;
1405 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1407 * pass busy_cfs_rq argument into
1408 * load_balance_[start|next]_fair iterators
1410 cfs_rq_iterator.arg = busy_cfs_rq;
1411 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1421 * scheduler tick hitting a task of our scheduling class:
1423 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1425 struct cfs_rq *cfs_rq;
1426 struct sched_entity *se = &curr->se;
1428 for_each_sched_entity(se) {
1429 cfs_rq = cfs_rq_of(se);
1430 entity_tick(cfs_rq, se, queued);
1434 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1437 * Share the fairness runtime between parent and child, thus the
1438 * total amount of pressure for CPU stays equal - new tasks
1439 * get a chance to run but frequent forkers are not allowed to
1440 * monopolize the CPU. Note: the parent runqueue is locked,
1441 * the child is not running yet.
1443 static void task_new_fair(struct rq *rq, struct task_struct *p)
1445 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1446 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1447 int this_cpu = smp_processor_id();
1449 sched_info_queued(p);
1451 update_curr(cfs_rq);
1452 place_entity(cfs_rq, se, 1);
1454 /* 'curr' will be NULL if the child belongs to a different group */
1455 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1456 curr && curr->vruntime < se->vruntime) {
1458 * Upon rescheduling, sched_class::put_prev_task() will place
1459 * 'current' within the tree based on its new key value.
1461 swap(curr->vruntime, se->vruntime);
1464 enqueue_task_fair(rq, p, 0);
1465 resched_task(rq->curr);
1469 * Priority of the task has changed. Check to see if we preempt
1472 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1473 int oldprio, int running)
1476 * Reschedule if we are currently running on this runqueue and
1477 * our priority decreased, or if we are not currently running on
1478 * this runqueue and our priority is higher than the current's
1481 if (p->prio > oldprio)
1482 resched_task(rq->curr);
1484 check_preempt_curr(rq, p);
1488 * We switched to the sched_fair class.
1490 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1494 * We were most likely switched from sched_rt, so
1495 * kick off the schedule if running, otherwise just see
1496 * if we can still preempt the current task.
1499 resched_task(rq->curr);
1501 check_preempt_curr(rq, p);
1504 /* Account for a task changing its policy or group.
1506 * This routine is mostly called to set cfs_rq->curr field when a task
1507 * migrates between groups/classes.
1509 static void set_curr_task_fair(struct rq *rq)
1511 struct sched_entity *se = &rq->curr->se;
1513 for_each_sched_entity(se)
1514 set_next_entity(cfs_rq_of(se), se);
1517 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 static void moved_group_fair(struct task_struct *p)
1520 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1522 update_curr(cfs_rq);
1523 place_entity(cfs_rq, &p->se, 1);
1528 * All the scheduling class methods:
1530 static const struct sched_class fair_sched_class = {
1531 .next = &idle_sched_class,
1532 .enqueue_task = enqueue_task_fair,
1533 .dequeue_task = dequeue_task_fair,
1534 .yield_task = yield_task_fair,
1536 .select_task_rq = select_task_rq_fair,
1537 #endif /* CONFIG_SMP */
1539 .check_preempt_curr = check_preempt_wakeup,
1541 .pick_next_task = pick_next_task_fair,
1542 .put_prev_task = put_prev_task_fair,
1545 .load_balance = load_balance_fair,
1546 .move_one_task = move_one_task_fair,
1549 .set_curr_task = set_curr_task_fair,
1550 .task_tick = task_tick_fair,
1551 .task_new = task_new_fair,
1553 .prio_changed = prio_changed_fair,
1554 .switched_to = switched_to_fair,
1556 #ifdef CONFIG_FAIR_GROUP_SCHED
1557 .moved_group = moved_group_fair,
1561 #ifdef CONFIG_SCHED_DEBUG
1562 static void print_cfs_stats(struct seq_file *m, int cpu)
1564 struct cfs_rq *cfs_rq;
1567 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1568 print_cfs_rq(m, cpu, cfs_rq);