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>
24 * Targeted preemption latency for CPU-bound tasks:
25 * (default: 20ms, units: nanoseconds)
27 * NOTE: this latency value is not the same as the concept of
28 * 'timeslice length' - timeslices in CFS are of variable length.
29 * (to see the precise effective timeslice length of your workload,
30 * run vmstat and monitor the context-switches field)
32 * On SMP systems the value of this is multiplied by the log2 of the
33 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
34 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
35 * Targeted preemption latency for CPU-bound tasks:
37 const_debug unsigned int sysctl_sched_latency = 20000000ULL;
40 * After fork, child runs first. (default) If set to 0 then
41 * parent will (try to) run first.
43 const_debug unsigned int sysctl_sched_child_runs_first = 1;
46 * Minimal preemption granularity for CPU-bound tasks:
47 * (default: 2 msec, units: nanoseconds)
49 unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL;
52 * sys_sched_yield() compat mode
54 * This option switches the agressive yield implementation of the
55 * old scheduler back on.
57 unsigned int __read_mostly sysctl_sched_compat_yield;
60 * SCHED_BATCH wake-up granularity.
61 * (default: 25 msec, units: nanoseconds)
63 * This option delays the preemption effects of decoupled workloads
64 * and reduces their over-scheduling. Synchronous workloads will still
65 * have immediate wakeup/sleep latencies.
67 const_debug unsigned int sysctl_sched_batch_wakeup_granularity = 25000000UL;
70 * SCHED_OTHER wake-up granularity.
71 * (default: 1 msec, units: nanoseconds)
73 * This option delays the preemption effects of decoupled workloads
74 * and reduces their over-scheduling. Synchronous workloads will still
75 * have immediate wakeup/sleep latencies.
77 const_debug unsigned int sysctl_sched_wakeup_granularity = 2000000UL;
79 unsigned int sysctl_sched_runtime_limit __read_mostly;
81 extern struct sched_class fair_sched_class;
83 /**************************************************************
84 * CFS operations on generic schedulable entities:
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)
95 /* An entity is a task if it doesn't "own" a runqueue */
96 #define entity_is_task(se) (!se->my_q)
98 #else /* CONFIG_FAIR_GROUP_SCHED */
100 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
102 return container_of(cfs_rq, struct rq, cfs);
105 #define entity_is_task(se) 1
107 #endif /* CONFIG_FAIR_GROUP_SCHED */
109 static inline struct task_struct *task_of(struct sched_entity *se)
111 return container_of(se, struct task_struct, se);
115 /**************************************************************
116 * Scheduling class tree data structure manipulation methods:
120 set_leftmost(struct cfs_rq *cfs_rq, struct rb_node *leftmost)
122 struct sched_entity *se;
124 cfs_rq->rb_leftmost = leftmost;
126 se = rb_entry(leftmost, struct sched_entity, run_node);
127 cfs_rq->min_vruntime = max(se->vruntime,
128 cfs_rq->min_vruntime);
133 * Enqueue an entity into the rb-tree:
136 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
138 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
139 struct rb_node *parent = NULL;
140 struct sched_entity *entry;
141 s64 key = se->fair_key;
145 * Find the right place in the rbtree:
149 entry = rb_entry(parent, struct sched_entity, run_node);
151 * We dont care about collisions. Nodes with
152 * the same key stay together.
154 if (key - entry->fair_key < 0) {
155 link = &parent->rb_left;
157 link = &parent->rb_right;
163 * Maintain a cache of leftmost tree entries (it is frequently
167 set_leftmost(cfs_rq, &se->run_node);
169 rb_link_node(&se->run_node, parent, link);
170 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
171 update_load_add(&cfs_rq->load, se->load.weight);
172 cfs_rq->nr_running++;
175 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
179 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
181 if (cfs_rq->rb_leftmost == &se->run_node)
182 set_leftmost(cfs_rq, rb_next(&se->run_node));
184 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
185 update_load_sub(&cfs_rq->load, se->load.weight);
186 cfs_rq->nr_running--;
189 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
192 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
194 return cfs_rq->rb_leftmost;
197 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
199 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
202 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
204 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
205 struct sched_entity *se = NULL;
206 struct rb_node *parent;
210 se = rb_entry(parent, struct sched_entity, run_node);
211 link = &parent->rb_right;
217 /**************************************************************
218 * Scheduling class statistics methods:
221 static u64 __sched_period(unsigned long nr_running)
223 u64 period = sysctl_sched_latency;
224 unsigned long nr_latency =
225 sysctl_sched_latency / sysctl_sched_min_granularity;
227 if (unlikely(nr_running > nr_latency)) {
228 period *= nr_running;
229 do_div(period, nr_latency);
235 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
237 u64 period = __sched_period(cfs_rq->nr_running);
239 period *= se->load.weight;
240 do_div(period, cfs_rq->load.weight);
246 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
248 long limit = sysctl_sched_runtime_limit;
251 * Niced tasks have the same history dynamic range as
254 if (unlikely(se->wait_runtime > limit)) {
255 se->wait_runtime = limit;
256 schedstat_inc(se, wait_runtime_overruns);
257 schedstat_inc(cfs_rq, wait_runtime_overruns);
259 if (unlikely(se->wait_runtime < -limit)) {
260 se->wait_runtime = -limit;
261 schedstat_inc(se, wait_runtime_underruns);
262 schedstat_inc(cfs_rq, wait_runtime_underruns);
267 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
269 se->wait_runtime += delta;
270 schedstat_add(se, sum_wait_runtime, delta);
271 limit_wait_runtime(cfs_rq, se);
275 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
277 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
278 __add_wait_runtime(cfs_rq, se, delta);
279 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
283 * Update the current task's runtime statistics. Skip current tasks that
284 * are not in our scheduling class.
287 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
288 unsigned long delta_exec)
290 unsigned long delta, delta_fair, delta_mine, delta_exec_weighted;
291 struct load_weight *lw = &cfs_rq->load;
292 unsigned long load = lw->weight;
294 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
296 curr->sum_exec_runtime += delta_exec;
297 cfs_rq->exec_clock += delta_exec;
298 delta_exec_weighted = delta_exec;
299 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
300 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
303 curr->vruntime += delta_exec_weighted;
305 if (!sched_feat(FAIR_SLEEPERS))
311 delta_fair = calc_delta_fair(delta_exec, lw);
312 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
314 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
315 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
316 delta = min(delta, (unsigned long)(
317 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
318 cfs_rq->sleeper_bonus -= delta;
322 cfs_rq->fair_clock += delta_fair;
324 * We executed delta_exec amount of time on the CPU,
325 * but we were only entitled to delta_mine amount of
326 * time during that period (if nr_running == 1 then
327 * the two values are equal)
328 * [Note: delta_mine - delta_exec is negative]:
330 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
333 static void update_curr(struct cfs_rq *cfs_rq)
335 struct sched_entity *curr = cfs_rq->curr;
336 u64 now = rq_of(cfs_rq)->clock;
337 unsigned long delta_exec;
343 * Get the amount of time the current task was running
344 * since the last time we changed load (this cannot
345 * overflow on 32 bits):
347 delta_exec = (unsigned long)(now - curr->exec_start);
349 __update_curr(cfs_rq, curr, delta_exec);
350 curr->exec_start = now;
354 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
356 se->wait_start_fair = cfs_rq->fair_clock;
357 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
360 static inline unsigned long
361 calc_weighted(unsigned long delta, struct sched_entity *se)
363 unsigned long weight = se->load.weight;
365 if (unlikely(weight != NICE_0_LOAD))
366 return (u64)delta * se->load.weight >> NICE_0_SHIFT;
372 * Task is being enqueued - update stats:
374 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
377 * Are we enqueueing a waiting task? (for current tasks
378 * a dequeue/enqueue event is a NOP)
380 if (se != cfs_rq->curr)
381 update_stats_wait_start(cfs_rq, se);
385 se->fair_key = se->vruntime;
389 * Note: must be called with a freshly updated rq->fair_clock.
392 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se,
393 unsigned long delta_fair)
395 schedstat_set(se->wait_max, max(se->wait_max,
396 rq_of(cfs_rq)->clock - se->wait_start));
398 delta_fair = calc_weighted(delta_fair, se);
400 add_wait_runtime(cfs_rq, se, delta_fair);
404 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
406 unsigned long delta_fair;
408 if (unlikely(!se->wait_start_fair))
411 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
412 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
414 __update_stats_wait_end(cfs_rq, se, delta_fair);
416 se->wait_start_fair = 0;
417 schedstat_set(se->wait_start, 0);
421 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
425 * Mark the end of the wait period if dequeueing a
428 if (se != cfs_rq->curr)
429 update_stats_wait_end(cfs_rq, se);
433 * We are picking a new current task - update its stats:
436 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
439 * We are starting a new run period:
441 se->exec_start = rq_of(cfs_rq)->clock;
445 * We are descheduling a task - update its stats:
448 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
453 /**************************************************
454 * Scheduling class queueing methods:
457 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
458 unsigned long delta_fair)
460 unsigned long load = cfs_rq->load.weight;
464 * Do not boost sleepers if there's too much bonus 'in flight'
467 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
470 if (sched_feat(SLEEPER_LOAD_AVG))
471 load = rq_of(cfs_rq)->cpu_load[2];
474 * Fix up delta_fair with the effect of us running
475 * during the whole sleep period:
477 if (sched_feat(SLEEPER_AVG))
478 delta_fair = div64_likely32((u64)delta_fair * load,
479 load + se->load.weight);
481 delta_fair = calc_weighted(delta_fair, se);
483 prev_runtime = se->wait_runtime;
484 __add_wait_runtime(cfs_rq, se, delta_fair);
485 delta_fair = se->wait_runtime - prev_runtime;
488 * Track the amount of bonus we've given to sleepers:
490 cfs_rq->sleeper_bonus += delta_fair;
493 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
495 struct task_struct *tsk = task_of(se);
496 unsigned long delta_fair;
498 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
499 !sched_feat(FAIR_SLEEPERS))
502 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
503 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
505 __enqueue_sleeper(cfs_rq, se, delta_fair);
507 se->sleep_start_fair = 0;
509 #ifdef CONFIG_SCHEDSTATS
510 if (se->sleep_start) {
511 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
516 if (unlikely(delta > se->sleep_max))
517 se->sleep_max = delta;
520 se->sum_sleep_runtime += delta;
522 if (se->block_start) {
523 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
528 if (unlikely(delta > se->block_max))
529 se->block_max = delta;
532 se->sum_sleep_runtime += delta;
535 * Blocking time is in units of nanosecs, so shift by 20 to
536 * get a milliseconds-range estimation of the amount of
537 * time that the task spent sleeping:
539 if (unlikely(prof_on == SLEEP_PROFILING)) {
540 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
548 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
550 u64 min_runtime, latency;
552 min_runtime = cfs_rq->min_vruntime;
554 if (sched_feat(USE_TREE_AVG)) {
555 struct sched_entity *last = __pick_last_entity(cfs_rq);
557 min_runtime = __pick_next_entity(cfs_rq)->vruntime;
558 min_runtime += last->vruntime;
561 } else if (sched_feat(APPROX_AVG))
562 min_runtime += sysctl_sched_latency/2;
564 if (initial && sched_feat(START_DEBIT))
565 min_runtime += sched_slice(cfs_rq, se);
567 if (!initial && sched_feat(NEW_FAIR_SLEEPERS)) {
568 latency = sysctl_sched_latency;
569 if (min_runtime > latency)
570 min_runtime -= latency;
575 se->vruntime = max(se->vruntime, min_runtime);
579 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
582 * Update the fair clock.
587 place_entity(cfs_rq, se, 0);
588 enqueue_sleeper(cfs_rq, se);
591 update_stats_enqueue(cfs_rq, se);
592 __enqueue_entity(cfs_rq, se);
596 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
598 update_stats_dequeue(cfs_rq, se);
600 se->sleep_start_fair = cfs_rq->fair_clock;
601 #ifdef CONFIG_SCHEDSTATS
602 if (entity_is_task(se)) {
603 struct task_struct *tsk = task_of(se);
605 if (tsk->state & TASK_INTERRUPTIBLE)
606 se->sleep_start = rq_of(cfs_rq)->clock;
607 if (tsk->state & TASK_UNINTERRUPTIBLE)
608 se->block_start = rq_of(cfs_rq)->clock;
612 __dequeue_entity(cfs_rq, se);
616 * Preempt the current task with a newly woken task if needed:
619 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
621 unsigned long ideal_runtime, delta_exec;
623 ideal_runtime = sched_slice(cfs_rq, curr);
624 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
625 if (delta_exec > ideal_runtime)
626 resched_task(rq_of(cfs_rq)->curr);
630 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
633 * Any task has to be enqueued before it get to execute on
634 * a CPU. So account for the time it spent waiting on the
635 * runqueue. (note, here we rely on pick_next_task() having
636 * done a put_prev_task_fair() shortly before this, which
637 * updated rq->fair_clock - used by update_stats_wait_end())
639 update_stats_wait_end(cfs_rq, se);
640 update_stats_curr_start(cfs_rq, se);
642 #ifdef CONFIG_SCHEDSTATS
644 * Track our maximum slice length, if the CPU's load is at
645 * least twice that of our own weight (i.e. dont track it
646 * when there are only lesser-weight tasks around):
648 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
649 se->slice_max = max(se->slice_max,
650 se->sum_exec_runtime - se->prev_sum_exec_runtime);
653 se->prev_sum_exec_runtime = se->sum_exec_runtime;
656 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
658 struct sched_entity *se = __pick_next_entity(cfs_rq);
660 set_next_entity(cfs_rq, se);
665 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
668 * If still on the runqueue then deactivate_task()
669 * was not called and update_curr() has to be done:
674 update_stats_curr_end(cfs_rq, prev);
677 update_stats_wait_start(cfs_rq, prev);
681 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
684 * Dequeue and enqueue the task to update its
685 * position within the tree:
687 dequeue_entity(cfs_rq, curr, 0);
688 enqueue_entity(cfs_rq, curr, 0);
690 if (cfs_rq->nr_running > 1)
691 check_preempt_tick(cfs_rq, curr);
694 /**************************************************
695 * CFS operations on tasks:
698 #ifdef CONFIG_FAIR_GROUP_SCHED
700 /* Walk up scheduling entities hierarchy */
701 #define for_each_sched_entity(se) \
702 for (; se; se = se->parent)
704 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
709 /* runqueue on which this entity is (to be) queued */
710 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
715 /* runqueue "owned" by this group */
716 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
721 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
722 * another cpu ('this_cpu')
724 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
726 /* A later patch will take group into account */
727 return &cpu_rq(this_cpu)->cfs;
730 /* Iterate thr' all leaf cfs_rq's on a runqueue */
731 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
732 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
734 /* Do the two (enqueued) tasks belong to the same group ? */
735 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
737 if (curr->se.cfs_rq == p->se.cfs_rq)
743 #else /* CONFIG_FAIR_GROUP_SCHED */
745 #define for_each_sched_entity(se) \
746 for (; se; se = NULL)
748 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
750 return &task_rq(p)->cfs;
753 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
755 struct task_struct *p = task_of(se);
756 struct rq *rq = task_rq(p);
761 /* runqueue "owned" by this group */
762 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
767 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
769 return &cpu_rq(this_cpu)->cfs;
772 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
773 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
775 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
780 #endif /* CONFIG_FAIR_GROUP_SCHED */
783 * The enqueue_task method is called before nr_running is
784 * increased. Here we update the fair scheduling stats and
785 * then put the task into the rbtree:
787 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
789 struct cfs_rq *cfs_rq;
790 struct sched_entity *se = &p->se;
792 for_each_sched_entity(se) {
795 cfs_rq = cfs_rq_of(se);
796 enqueue_entity(cfs_rq, se, wakeup);
801 * The dequeue_task method is called before nr_running is
802 * decreased. We remove the task from the rbtree and
803 * update the fair scheduling stats:
805 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
807 struct cfs_rq *cfs_rq;
808 struct sched_entity *se = &p->se;
810 for_each_sched_entity(se) {
811 cfs_rq = cfs_rq_of(se);
812 dequeue_entity(cfs_rq, se, sleep);
813 /* Don't dequeue parent if it has other entities besides us */
814 if (cfs_rq->load.weight)
820 * sched_yield() support is very simple - we dequeue and enqueue.
822 * If compat_yield is turned on then we requeue to the end of the tree.
824 static void yield_task_fair(struct rq *rq, struct task_struct *p)
826 struct cfs_rq *cfs_rq = task_cfs_rq(p);
827 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
828 struct sched_entity *rightmost, *se = &p->se;
829 struct rb_node *parent;
832 * Are we the only task in the tree?
834 if (unlikely(cfs_rq->nr_running == 1))
837 if (likely(!sysctl_sched_compat_yield)) {
838 __update_rq_clock(rq);
840 * Dequeue and enqueue the task to update its
841 * position within the tree:
843 dequeue_entity(cfs_rq, &p->se, 0);
844 enqueue_entity(cfs_rq, &p->se, 0);
849 * Find the rightmost entry in the rbtree:
853 link = &parent->rb_right;
856 rightmost = rb_entry(parent, struct sched_entity, run_node);
858 * Already in the rightmost position?
860 if (unlikely(rightmost == se))
864 * Minimally necessary key value to be last in the tree:
866 se->fair_key = rightmost->fair_key + 1;
868 if (cfs_rq->rb_leftmost == &se->run_node)
869 cfs_rq->rb_leftmost = rb_next(&se->run_node);
871 * Relink the task to the rightmost position:
873 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
874 rb_link_node(&se->run_node, parent, link);
875 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
879 * Preempt the current task with a newly woken task if needed:
881 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
883 struct task_struct *curr = rq->curr;
884 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
886 if (unlikely(rt_prio(p->prio))) {
892 if (is_same_group(curr, p)) {
893 s64 delta = curr->se.vruntime - p->se.vruntime;
895 if (delta > (s64)sysctl_sched_wakeup_granularity)
900 static struct task_struct *pick_next_task_fair(struct rq *rq)
902 struct cfs_rq *cfs_rq = &rq->cfs;
903 struct sched_entity *se;
905 if (unlikely(!cfs_rq->nr_running))
909 se = pick_next_entity(cfs_rq);
910 cfs_rq = group_cfs_rq(se);
917 * Account for a descheduled task:
919 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
921 struct sched_entity *se = &prev->se;
922 struct cfs_rq *cfs_rq;
924 for_each_sched_entity(se) {
925 cfs_rq = cfs_rq_of(se);
926 put_prev_entity(cfs_rq, se);
930 /**************************************************
931 * Fair scheduling class load-balancing methods:
935 * Load-balancing iterator. Note: while the runqueue stays locked
936 * during the whole iteration, the current task might be
937 * dequeued so the iterator has to be dequeue-safe. Here we
938 * achieve that by always pre-iterating before returning
941 static inline struct task_struct *
942 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
944 struct task_struct *p;
949 p = rb_entry(curr, struct task_struct, se.run_node);
950 cfs_rq->rb_load_balance_curr = rb_next(curr);
955 static struct task_struct *load_balance_start_fair(void *arg)
957 struct cfs_rq *cfs_rq = arg;
959 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
962 static struct task_struct *load_balance_next_fair(void *arg)
964 struct cfs_rq *cfs_rq = arg;
966 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
969 #ifdef CONFIG_FAIR_GROUP_SCHED
970 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
972 struct sched_entity *curr;
973 struct task_struct *p;
975 if (!cfs_rq->nr_running)
978 curr = __pick_next_entity(cfs_rq);
986 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
987 unsigned long max_nr_move, unsigned long max_load_move,
988 struct sched_domain *sd, enum cpu_idle_type idle,
989 int *all_pinned, int *this_best_prio)
991 struct cfs_rq *busy_cfs_rq;
992 unsigned long load_moved, total_nr_moved = 0, nr_moved;
993 long rem_load_move = max_load_move;
994 struct rq_iterator cfs_rq_iterator;
996 cfs_rq_iterator.start = load_balance_start_fair;
997 cfs_rq_iterator.next = load_balance_next_fair;
999 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1000 #ifdef CONFIG_FAIR_GROUP_SCHED
1001 struct cfs_rq *this_cfs_rq;
1003 unsigned long maxload;
1005 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1007 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1008 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1012 /* Don't pull more than imbalance/2 */
1014 maxload = min(rem_load_move, imbalance);
1016 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1018 # define maxload rem_load_move
1020 /* pass busy_cfs_rq argument into
1021 * load_balance_[start|next]_fair iterators
1023 cfs_rq_iterator.arg = busy_cfs_rq;
1024 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1025 max_nr_move, maxload, sd, idle, all_pinned,
1026 &load_moved, this_best_prio, &cfs_rq_iterator);
1028 total_nr_moved += nr_moved;
1029 max_nr_move -= nr_moved;
1030 rem_load_move -= load_moved;
1032 if (max_nr_move <= 0 || rem_load_move <= 0)
1036 return max_load_move - rem_load_move;
1040 * scheduler tick hitting a task of our scheduling class:
1042 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1044 struct cfs_rq *cfs_rq;
1045 struct sched_entity *se = &curr->se;
1047 for_each_sched_entity(se) {
1048 cfs_rq = cfs_rq_of(se);
1049 entity_tick(cfs_rq, se);
1053 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1056 * Share the fairness runtime between parent and child, thus the
1057 * total amount of pressure for CPU stays equal - new tasks
1058 * get a chance to run but frequent forkers are not allowed to
1059 * monopolize the CPU. Note: the parent runqueue is locked,
1060 * the child is not running yet.
1062 static void task_new_fair(struct rq *rq, struct task_struct *p)
1064 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1065 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1067 sched_info_queued(p);
1069 update_curr(cfs_rq);
1070 place_entity(cfs_rq, se, 1);
1073 * The statistical average of wait_runtime is about
1074 * -granularity/2, so initialize the task with that:
1076 if (sched_feat(START_DEBIT))
1077 se->wait_runtime = -(__sched_period(cfs_rq->nr_running+1) / 2);
1079 if (sysctl_sched_child_runs_first &&
1080 curr->vruntime < se->vruntime) {
1082 dequeue_entity(cfs_rq, curr, 0);
1083 swap(curr->vruntime, se->vruntime);
1084 enqueue_entity(cfs_rq, curr, 0);
1087 update_stats_enqueue(cfs_rq, se);
1088 __enqueue_entity(cfs_rq, se);
1089 resched_task(rq->curr);
1092 #ifdef CONFIG_FAIR_GROUP_SCHED
1093 /* Account for a task changing its policy or group.
1095 * This routine is mostly called to set cfs_rq->curr field when a task
1096 * migrates between groups/classes.
1098 static void set_curr_task_fair(struct rq *rq)
1100 struct sched_entity *se = &rq->curr->se;
1102 for_each_sched_entity(se)
1103 set_next_entity(cfs_rq_of(se), se);
1106 static void set_curr_task_fair(struct rq *rq)
1112 * All the scheduling class methods:
1114 struct sched_class fair_sched_class __read_mostly = {
1115 .enqueue_task = enqueue_task_fair,
1116 .dequeue_task = dequeue_task_fair,
1117 .yield_task = yield_task_fair,
1119 .check_preempt_curr = check_preempt_wakeup,
1121 .pick_next_task = pick_next_task_fair,
1122 .put_prev_task = put_prev_task_fair,
1124 .load_balance = load_balance_fair,
1126 .set_curr_task = set_curr_task_fair,
1127 .task_tick = task_tick_fair,
1128 .task_new = task_new_fair,
1131 #ifdef CONFIG_SCHED_DEBUG
1132 static void print_cfs_stats(struct seq_file *m, int cpu)
1134 struct cfs_rq *cfs_rq;
1136 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1137 print_cfs_rq(m, cpu, cfs_rq);