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);
339 static inline unsigned long
340 calc_delta_weight(unsigned long delta, struct sched_entity *se)
342 for_each_sched_entity(se) {
343 delta = calc_delta_mine(delta,
344 se->load.weight, &cfs_rq_of(se)->load);
353 static inline unsigned long
354 calc_delta_fair(unsigned long delta, struct sched_entity *se)
356 for_each_sched_entity(se) {
357 delta = calc_delta_mine(delta,
358 cfs_rq_of(se)->load.weight, &se->load);
365 * The idea is to set a period in which each task runs once.
367 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
368 * this period because otherwise the slices get too small.
370 * p = (nr <= nl) ? l : l*nr/nl
372 static u64 __sched_period(unsigned long nr_running)
374 u64 period = sysctl_sched_latency;
375 unsigned long nr_latency = sched_nr_latency;
377 if (unlikely(nr_running > nr_latency)) {
378 period = sysctl_sched_min_granularity;
379 period *= nr_running;
386 * We calculate the wall-time slice from the period by taking a part
387 * proportional to the weight.
391 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
393 return calc_delta_weight(__sched_period(cfs_rq->nr_running), se);
397 * We calculate the vruntime slice of a to be inserted task
401 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
403 unsigned long nr_running = cfs_rq->nr_running;
408 return __sched_period(nr_running);
412 * The goal of calc_delta_asym() is to be asymmetrically around NICE_0_LOAD, in
413 * that it favours >=0 over <0.
423 calc_delta_asym(unsigned long delta, struct sched_entity *se)
425 struct load_weight lw = {
426 .weight = NICE_0_LOAD,
427 .inv_weight = 1UL << (WMULT_SHIFT-NICE_0_SHIFT)
430 for_each_sched_entity(se) {
431 struct load_weight *se_lw = &se->load;
433 if (se->load.weight < NICE_0_LOAD)
436 delta = calc_delta_mine(delta,
437 cfs_rq_of(se)->load.weight, se_lw);
444 * Update the current task's runtime statistics. Skip current tasks that
445 * are not in our scheduling class.
448 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
449 unsigned long delta_exec)
451 unsigned long delta_exec_weighted;
453 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
455 curr->sum_exec_runtime += delta_exec;
456 schedstat_add(cfs_rq, exec_clock, delta_exec);
457 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
458 curr->vruntime += delta_exec_weighted;
461 static void update_curr(struct cfs_rq *cfs_rq)
463 struct sched_entity *curr = cfs_rq->curr;
464 u64 now = rq_of(cfs_rq)->clock;
465 unsigned long delta_exec;
471 * Get the amount of time the current task was running
472 * since the last time we changed load (this cannot
473 * overflow on 32 bits):
475 delta_exec = (unsigned long)(now - curr->exec_start);
477 __update_curr(cfs_rq, curr, delta_exec);
478 curr->exec_start = now;
480 if (entity_is_task(curr)) {
481 struct task_struct *curtask = task_of(curr);
483 cpuacct_charge(curtask, delta_exec);
488 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
490 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
494 * Task is being enqueued - update stats:
496 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
499 * Are we enqueueing a waiting task? (for current tasks
500 * a dequeue/enqueue event is a NOP)
502 if (se != cfs_rq->curr)
503 update_stats_wait_start(cfs_rq, se);
507 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
509 schedstat_set(se->wait_max, max(se->wait_max,
510 rq_of(cfs_rq)->clock - se->wait_start));
511 schedstat_set(se->wait_count, se->wait_count + 1);
512 schedstat_set(se->wait_sum, se->wait_sum +
513 rq_of(cfs_rq)->clock - se->wait_start);
514 schedstat_set(se->wait_start, 0);
518 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
521 * Mark the end of the wait period if dequeueing a
524 if (se != cfs_rq->curr)
525 update_stats_wait_end(cfs_rq, se);
529 * We are picking a new current task - update its stats:
532 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
535 * We are starting a new run period:
537 se->exec_start = rq_of(cfs_rq)->clock;
540 /**************************************************
541 * Scheduling class queueing methods:
545 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
547 update_load_add(&cfs_rq->load, se->load.weight);
548 cfs_rq->nr_running++;
550 list_add(&se->group_node, &cfs_rq->tasks);
554 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
556 update_load_sub(&cfs_rq->load, se->load.weight);
557 cfs_rq->nr_running--;
559 list_del_init(&se->group_node);
562 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
564 #ifdef CONFIG_SCHEDSTATS
565 if (se->sleep_start) {
566 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
567 struct task_struct *tsk = task_of(se);
572 if (unlikely(delta > se->sleep_max))
573 se->sleep_max = delta;
576 se->sum_sleep_runtime += delta;
578 account_scheduler_latency(tsk, delta >> 10, 1);
580 if (se->block_start) {
581 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
582 struct task_struct *tsk = task_of(se);
587 if (unlikely(delta > se->block_max))
588 se->block_max = delta;
591 se->sum_sleep_runtime += delta;
594 * Blocking time is in units of nanosecs, so shift by 20 to
595 * get a milliseconds-range estimation of the amount of
596 * time that the task spent sleeping:
598 if (unlikely(prof_on == SLEEP_PROFILING)) {
600 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
603 account_scheduler_latency(tsk, delta >> 10, 0);
608 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
610 #ifdef CONFIG_SCHED_DEBUG
611 s64 d = se->vruntime - cfs_rq->min_vruntime;
616 if (d > 3*sysctl_sched_latency)
617 schedstat_inc(cfs_rq, nr_spread_over);
622 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
626 if (first_fair(cfs_rq)) {
627 vruntime = min_vruntime(cfs_rq->min_vruntime,
628 __pick_next_entity(cfs_rq)->vruntime);
630 vruntime = cfs_rq->min_vruntime;
633 * The 'current' period is already promised to the current tasks,
634 * however the extra weight of the new task will slow them down a
635 * little, place the new task so that it fits in the slot that
636 * stays open at the end.
638 if (initial && sched_feat(START_DEBIT))
639 vruntime += sched_vslice_add(cfs_rq, se);
642 /* sleeps upto a single latency don't count. */
643 if (sched_feat(NEW_FAIR_SLEEPERS)) {
644 unsigned long thresh = sysctl_sched_latency;
647 * convert the sleeper threshold into virtual time
649 if (sched_feat(NORMALIZED_SLEEPER))
650 thresh = calc_delta_fair(thresh, se);
655 /* ensure we never gain time by being placed backwards. */
656 vruntime = max_vruntime(se->vruntime, vruntime);
659 se->vruntime = vruntime;
663 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
666 * Update run-time statistics of the 'current'.
669 account_entity_enqueue(cfs_rq, se);
672 place_entity(cfs_rq, se, 0);
673 enqueue_sleeper(cfs_rq, se);
676 update_stats_enqueue(cfs_rq, se);
677 check_spread(cfs_rq, se);
678 if (se != cfs_rq->curr)
679 __enqueue_entity(cfs_rq, se);
682 static void update_avg(u64 *avg, u64 sample)
684 s64 diff = sample - *avg;
688 static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se)
690 if (!se->last_wakeup)
693 update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup);
698 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
701 * Update run-time statistics of the 'current'.
705 update_stats_dequeue(cfs_rq, se);
707 update_avg_stats(cfs_rq, se);
708 #ifdef CONFIG_SCHEDSTATS
709 if (entity_is_task(se)) {
710 struct task_struct *tsk = task_of(se);
712 if (tsk->state & TASK_INTERRUPTIBLE)
713 se->sleep_start = rq_of(cfs_rq)->clock;
714 if (tsk->state & TASK_UNINTERRUPTIBLE)
715 se->block_start = rq_of(cfs_rq)->clock;
720 if (se != cfs_rq->curr)
721 __dequeue_entity(cfs_rq, se);
722 account_entity_dequeue(cfs_rq, se);
726 * Preempt the current task with a newly woken task if needed:
729 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
731 unsigned long ideal_runtime, delta_exec;
733 ideal_runtime = sched_slice(cfs_rq, curr);
734 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
735 if (delta_exec > ideal_runtime)
736 resched_task(rq_of(cfs_rq)->curr);
740 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
742 /* 'current' is not kept within the tree. */
745 * Any task has to be enqueued before it get to execute on
746 * a CPU. So account for the time it spent waiting on the
749 update_stats_wait_end(cfs_rq, se);
750 __dequeue_entity(cfs_rq, se);
753 update_stats_curr_start(cfs_rq, se);
755 #ifdef CONFIG_SCHEDSTATS
757 * Track our maximum slice length, if the CPU's load is at
758 * least twice that of our own weight (i.e. dont track it
759 * when there are only lesser-weight tasks around):
761 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
762 se->slice_max = max(se->slice_max,
763 se->sum_exec_runtime - se->prev_sum_exec_runtime);
766 se->prev_sum_exec_runtime = se->sum_exec_runtime;
770 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
772 static struct sched_entity *
773 pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se)
778 if (wakeup_preempt_entity(cfs_rq->next, se) != 0)
784 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
786 struct sched_entity *se = NULL;
788 if (first_fair(cfs_rq)) {
789 se = __pick_next_entity(cfs_rq);
790 se = pick_next(cfs_rq, se);
791 set_next_entity(cfs_rq, se);
797 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
800 * If still on the runqueue then deactivate_task()
801 * was not called and update_curr() has to be done:
806 check_spread(cfs_rq, prev);
808 update_stats_wait_start(cfs_rq, prev);
809 /* Put 'current' back into the tree. */
810 __enqueue_entity(cfs_rq, prev);
816 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
819 * Update run-time statistics of the 'current'.
823 #ifdef CONFIG_SCHED_HRTICK
825 * queued ticks are scheduled to match the slice, so don't bother
826 * validating it and just reschedule.
829 resched_task(rq_of(cfs_rq)->curr);
833 * don't let the period tick interfere with the hrtick preemption
835 if (!sched_feat(DOUBLE_TICK) &&
836 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
840 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
841 check_preempt_tick(cfs_rq, curr);
844 /**************************************************
845 * CFS operations on tasks:
848 #ifdef CONFIG_SCHED_HRTICK
849 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
851 int requeue = rq->curr == p;
852 struct sched_entity *se = &p->se;
853 struct cfs_rq *cfs_rq = cfs_rq_of(se);
855 WARN_ON(task_rq(p) != rq);
857 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
858 u64 slice = sched_slice(cfs_rq, se);
859 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
860 s64 delta = slice - ran;
869 * Don't schedule slices shorter than 10000ns, that just
870 * doesn't make sense. Rely on vruntime for fairness.
873 delta = max(10000LL, delta);
875 hrtick_start(rq, delta, requeue);
880 hrtick_start_fair(struct rq *rq, struct task_struct *p)
886 * The enqueue_task method is called before nr_running is
887 * increased. Here we update the fair scheduling stats and
888 * then put the task into the rbtree:
890 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
892 struct cfs_rq *cfs_rq;
893 struct sched_entity *se = &p->se;
895 for_each_sched_entity(se) {
898 cfs_rq = cfs_rq_of(se);
899 enqueue_entity(cfs_rq, se, wakeup);
903 hrtick_start_fair(rq, rq->curr);
907 * The dequeue_task method is called before nr_running is
908 * decreased. We remove the task from the rbtree and
909 * update the fair scheduling stats:
911 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
913 struct cfs_rq *cfs_rq;
914 struct sched_entity *se = &p->se;
916 for_each_sched_entity(se) {
917 cfs_rq = cfs_rq_of(se);
918 dequeue_entity(cfs_rq, se, sleep);
919 /* Don't dequeue parent if it has other entities besides us */
920 if (cfs_rq->load.weight)
925 hrtick_start_fair(rq, rq->curr);
929 * sched_yield() support is very simple - we dequeue and enqueue.
931 * If compat_yield is turned on then we requeue to the end of the tree.
933 static void yield_task_fair(struct rq *rq)
935 struct task_struct *curr = rq->curr;
936 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
937 struct sched_entity *rightmost, *se = &curr->se;
940 * Are we the only task in the tree?
942 if (unlikely(cfs_rq->nr_running == 1))
945 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
948 * Update run-time statistics of the 'current'.
955 * Find the rightmost entry in the rbtree:
957 rightmost = __pick_last_entity(cfs_rq);
959 * Already in the rightmost position?
961 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
965 * Minimally necessary key value to be last in the tree:
966 * Upon rescheduling, sched_class::put_prev_task() will place
967 * 'current' within the tree based on its new key value.
969 se->vruntime = rightmost->vruntime + 1;
973 * wake_idle() will wake a task on an idle cpu if task->cpu is
974 * not idle and an idle cpu is available. The span of cpus to
975 * search starts with cpus closest then further out as needed,
976 * so we always favor a closer, idle cpu.
978 * Returns the CPU we should wake onto.
980 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
981 static int wake_idle(int cpu, struct task_struct *p)
984 struct sched_domain *sd;
988 * If it is idle, then it is the best cpu to run this task.
990 * This cpu is also the best, if it has more than one task already.
991 * Siblings must be also busy(in most cases) as they didn't already
992 * pickup the extra load from this cpu and hence we need not check
993 * sibling runqueue info. This will avoid the checks and cache miss
994 * penalities associated with that.
996 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
999 for_each_domain(cpu, sd) {
1000 if ((sd->flags & SD_WAKE_IDLE)
1001 || ((sd->flags & SD_WAKE_IDLE_FAR)
1002 && !task_hot(p, task_rq(p)->clock, sd))) {
1003 cpus_and(tmp, sd->span, p->cpus_allowed);
1004 for_each_cpu_mask(i, tmp) {
1006 if (i != task_cpu(p)) {
1008 se.nr_wakeups_idle);
1020 static inline int wake_idle(int cpu, struct task_struct *p)
1028 static const struct sched_class fair_sched_class;
1031 wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq,
1032 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1033 int idx, unsigned long load, unsigned long this_load,
1034 unsigned int imbalance)
1036 struct task_struct *curr = this_rq->curr;
1037 unsigned long tl = this_load;
1038 unsigned long tl_per_task;
1041 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1045 * If sync wakeup then subtract the (maximum possible)
1046 * effect of the currently running task from the load
1047 * of the current CPU:
1050 tl -= current->se.load.weight;
1052 balanced = 100*(tl + p->se.load.weight) <= imbalance*load;
1055 * If the currently running task will sleep within
1056 * a reasonable amount of time then attract this newly
1059 if (sync && balanced && curr->sched_class == &fair_sched_class) {
1060 if (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1061 p->se.avg_overlap < sysctl_sched_migration_cost)
1065 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1066 tl_per_task = cpu_avg_load_per_task(this_cpu);
1068 if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) ||
1071 * This domain has SD_WAKE_AFFINE and
1072 * p is cache cold in this domain, and
1073 * there is no bad imbalance.
1075 schedstat_inc(this_sd, ttwu_move_affine);
1076 schedstat_inc(p, se.nr_wakeups_affine);
1083 static int select_task_rq_fair(struct task_struct *p, int sync)
1085 struct sched_domain *sd, *this_sd = NULL;
1086 int prev_cpu, this_cpu, new_cpu;
1087 unsigned long load, this_load;
1088 struct rq *rq, *this_rq;
1089 unsigned int imbalance;
1092 prev_cpu = task_cpu(p);
1094 this_cpu = smp_processor_id();
1095 this_rq = cpu_rq(this_cpu);
1099 * 'this_sd' is the first domain that both
1100 * this_cpu and prev_cpu are present in:
1102 for_each_domain(this_cpu, sd) {
1103 if (cpu_isset(prev_cpu, sd->span)) {
1109 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1113 * Check for affine wakeup and passive balancing possibilities.
1118 idx = this_sd->wake_idx;
1120 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1122 load = source_load(prev_cpu, idx);
1123 this_load = target_load(this_cpu, idx);
1125 if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1126 load, this_load, imbalance))
1129 if (prev_cpu == this_cpu)
1133 * Start passive balancing when half the imbalance_pct
1136 if (this_sd->flags & SD_WAKE_BALANCE) {
1137 if (imbalance*this_load <= 100*load) {
1138 schedstat_inc(this_sd, ttwu_move_balance);
1139 schedstat_inc(p, se.nr_wakeups_passive);
1145 return wake_idle(new_cpu, p);
1147 #endif /* CONFIG_SMP */
1149 static unsigned long wakeup_gran(struct sched_entity *se)
1151 unsigned long gran = sysctl_sched_wakeup_granularity;
1154 * More easily preempt - nice tasks, while not making it harder for
1157 gran = calc_delta_asym(sysctl_sched_wakeup_granularity, se);
1163 * Should 'se' preempt 'curr'.
1177 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1179 s64 gran, vdiff = curr->vruntime - se->vruntime;
1184 gran = wakeup_gran(curr);
1191 /* return depth at which a sched entity is present in the hierarchy */
1192 static inline int depth_se(struct sched_entity *se)
1196 for_each_sched_entity(se)
1203 * Preempt the current task with a newly woken task if needed:
1205 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
1207 struct task_struct *curr = rq->curr;
1208 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1209 struct sched_entity *se = &curr->se, *pse = &p->se;
1210 int se_depth, pse_depth;
1212 if (unlikely(rt_prio(p->prio))) {
1213 update_rq_clock(rq);
1214 update_curr(cfs_rq);
1219 se->last_wakeup = se->sum_exec_runtime;
1220 if (unlikely(se == pse))
1223 cfs_rq_of(pse)->next = pse;
1226 * Batch tasks do not preempt (their preemption is driven by
1229 if (unlikely(p->policy == SCHED_BATCH))
1232 if (!sched_feat(WAKEUP_PREEMPT))
1236 * preemption test can be made between sibling entities who are in the
1237 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
1238 * both tasks until we find their ancestors who are siblings of common
1242 /* First walk up until both entities are at same depth */
1243 se_depth = depth_se(se);
1244 pse_depth = depth_se(pse);
1246 while (se_depth > pse_depth) {
1248 se = parent_entity(se);
1251 while (pse_depth > se_depth) {
1253 pse = parent_entity(pse);
1256 while (!is_same_group(se, pse)) {
1257 se = parent_entity(se);
1258 pse = parent_entity(pse);
1261 if (wakeup_preempt_entity(se, pse) == 1)
1265 static struct task_struct *pick_next_task_fair(struct rq *rq)
1267 struct task_struct *p;
1268 struct cfs_rq *cfs_rq = &rq->cfs;
1269 struct sched_entity *se;
1271 if (unlikely(!cfs_rq->nr_running))
1275 se = pick_next_entity(cfs_rq);
1276 cfs_rq = group_cfs_rq(se);
1280 hrtick_start_fair(rq, p);
1286 * Account for a descheduled task:
1288 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1290 struct sched_entity *se = &prev->se;
1291 struct cfs_rq *cfs_rq;
1293 for_each_sched_entity(se) {
1294 cfs_rq = cfs_rq_of(se);
1295 put_prev_entity(cfs_rq, se);
1300 /**************************************************
1301 * Fair scheduling class load-balancing methods:
1305 * Load-balancing iterator. Note: while the runqueue stays locked
1306 * during the whole iteration, the current task might be
1307 * dequeued so the iterator has to be dequeue-safe. Here we
1308 * achieve that by always pre-iterating before returning
1311 static struct task_struct *
1312 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1314 struct task_struct *p = NULL;
1315 struct sched_entity *se;
1317 while (next != &cfs_rq->tasks) {
1318 se = list_entry(next, struct sched_entity, group_node);
1321 /* Skip over entities that are not tasks */
1322 if (entity_is_task(se)) {
1328 cfs_rq->balance_iterator = next;
1332 static struct task_struct *load_balance_start_fair(void *arg)
1334 struct cfs_rq *cfs_rq = arg;
1336 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1339 static struct task_struct *load_balance_next_fair(void *arg)
1341 struct cfs_rq *cfs_rq = arg;
1343 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1346 #ifdef CONFIG_FAIR_GROUP_SCHED
1347 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1349 struct sched_entity *curr;
1350 struct task_struct *p;
1352 if (!cfs_rq->nr_running || !first_fair(cfs_rq))
1355 curr = cfs_rq->curr;
1357 curr = __pick_next_entity(cfs_rq);
1365 static unsigned long
1366 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1367 unsigned long max_load_move,
1368 struct sched_domain *sd, enum cpu_idle_type idle,
1369 int *all_pinned, int *this_best_prio)
1371 struct cfs_rq *busy_cfs_rq;
1372 long rem_load_move = max_load_move;
1373 struct rq_iterator cfs_rq_iterator;
1375 cfs_rq_iterator.start = load_balance_start_fair;
1376 cfs_rq_iterator.next = load_balance_next_fair;
1378 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1379 #ifdef CONFIG_FAIR_GROUP_SCHED
1380 struct cfs_rq *this_cfs_rq;
1382 unsigned long maxload;
1384 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1386 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1387 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1391 /* Don't pull more than imbalance/2 */
1393 maxload = min(rem_load_move, imbalance);
1395 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1397 # define maxload rem_load_move
1400 * pass busy_cfs_rq argument into
1401 * load_balance_[start|next]_fair iterators
1403 cfs_rq_iterator.arg = busy_cfs_rq;
1404 rem_load_move -= balance_tasks(this_rq, this_cpu, busiest,
1405 maxload, sd, idle, all_pinned,
1409 if (rem_load_move <= 0)
1413 return max_load_move - rem_load_move;
1417 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 struct sched_domain *sd, enum cpu_idle_type idle)
1420 struct cfs_rq *busy_cfs_rq;
1421 struct rq_iterator cfs_rq_iterator;
1423 cfs_rq_iterator.start = load_balance_start_fair;
1424 cfs_rq_iterator.next = load_balance_next_fair;
1426 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1428 * pass busy_cfs_rq argument into
1429 * load_balance_[start|next]_fair iterators
1431 cfs_rq_iterator.arg = busy_cfs_rq;
1432 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1442 * scheduler tick hitting a task of our scheduling class:
1444 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1446 struct cfs_rq *cfs_rq;
1447 struct sched_entity *se = &curr->se;
1449 for_each_sched_entity(se) {
1450 cfs_rq = cfs_rq_of(se);
1451 entity_tick(cfs_rq, se, queued);
1455 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1458 * Share the fairness runtime between parent and child, thus the
1459 * total amount of pressure for CPU stays equal - new tasks
1460 * get a chance to run but frequent forkers are not allowed to
1461 * monopolize the CPU. Note: the parent runqueue is locked,
1462 * the child is not running yet.
1464 static void task_new_fair(struct rq *rq, struct task_struct *p)
1466 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1467 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1468 int this_cpu = smp_processor_id();
1470 sched_info_queued(p);
1472 update_curr(cfs_rq);
1473 place_entity(cfs_rq, se, 1);
1475 /* 'curr' will be NULL if the child belongs to a different group */
1476 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1477 curr && curr->vruntime < se->vruntime) {
1479 * Upon rescheduling, sched_class::put_prev_task() will place
1480 * 'current' within the tree based on its new key value.
1482 swap(curr->vruntime, se->vruntime);
1485 enqueue_task_fair(rq, p, 0);
1486 resched_task(rq->curr);
1490 * Priority of the task has changed. Check to see if we preempt
1493 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1494 int oldprio, int running)
1497 * Reschedule if we are currently running on this runqueue and
1498 * our priority decreased, or if we are not currently running on
1499 * this runqueue and our priority is higher than the current's
1502 if (p->prio > oldprio)
1503 resched_task(rq->curr);
1505 check_preempt_curr(rq, p);
1509 * We switched to the sched_fair class.
1511 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1515 * We were most likely switched from sched_rt, so
1516 * kick off the schedule if running, otherwise just see
1517 * if we can still preempt the current task.
1520 resched_task(rq->curr);
1522 check_preempt_curr(rq, p);
1525 /* Account for a task changing its policy or group.
1527 * This routine is mostly called to set cfs_rq->curr field when a task
1528 * migrates between groups/classes.
1530 static void set_curr_task_fair(struct rq *rq)
1532 struct sched_entity *se = &rq->curr->se;
1534 for_each_sched_entity(se)
1535 set_next_entity(cfs_rq_of(se), se);
1538 #ifdef CONFIG_FAIR_GROUP_SCHED
1539 static void moved_group_fair(struct task_struct *p)
1541 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1543 update_curr(cfs_rq);
1544 place_entity(cfs_rq, &p->se, 1);
1549 * All the scheduling class methods:
1551 static const struct sched_class fair_sched_class = {
1552 .next = &idle_sched_class,
1553 .enqueue_task = enqueue_task_fair,
1554 .dequeue_task = dequeue_task_fair,
1555 .yield_task = yield_task_fair,
1557 .select_task_rq = select_task_rq_fair,
1558 #endif /* CONFIG_SMP */
1560 .check_preempt_curr = check_preempt_wakeup,
1562 .pick_next_task = pick_next_task_fair,
1563 .put_prev_task = put_prev_task_fair,
1566 .load_balance = load_balance_fair,
1567 .move_one_task = move_one_task_fair,
1570 .set_curr_task = set_curr_task_fair,
1571 .task_tick = task_tick_fair,
1572 .task_new = task_new_fair,
1574 .prio_changed = prio_changed_fair,
1575 .switched_to = switched_to_fair,
1577 #ifdef CONFIG_FAIR_GROUP_SCHED
1578 .moved_group = moved_group_fair,
1582 #ifdef CONFIG_SCHED_DEBUG
1583 static void print_cfs_stats(struct seq_file *m, int cpu)
1585 struct cfs_rq *cfs_rq;
1588 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1589 print_cfs_rq(m, cpu, cfs_rq);