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 /**************************************************************
203 * Scheduling class statistics methods:
206 static u64 __sched_period(unsigned long nr_running)
208 u64 period = sysctl_sched_latency;
209 unsigned long nr_latency =
210 sysctl_sched_latency / sysctl_sched_min_granularity;
212 if (unlikely(nr_running > nr_latency)) {
213 period *= nr_running;
214 do_div(period, nr_latency);
220 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
222 u64 period = __sched_period(cfs_rq->nr_running);
224 period *= se->load.weight;
225 do_div(period, cfs_rq->load.weight);
231 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
233 long limit = sysctl_sched_runtime_limit;
236 * Niced tasks have the same history dynamic range as
239 if (unlikely(se->wait_runtime > limit)) {
240 se->wait_runtime = limit;
241 schedstat_inc(se, wait_runtime_overruns);
242 schedstat_inc(cfs_rq, wait_runtime_overruns);
244 if (unlikely(se->wait_runtime < -limit)) {
245 se->wait_runtime = -limit;
246 schedstat_inc(se, wait_runtime_underruns);
247 schedstat_inc(cfs_rq, wait_runtime_underruns);
252 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
254 se->wait_runtime += delta;
255 schedstat_add(se, sum_wait_runtime, delta);
256 limit_wait_runtime(cfs_rq, se);
260 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
262 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
263 __add_wait_runtime(cfs_rq, se, delta);
264 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
268 * Update the current task's runtime statistics. Skip current tasks that
269 * are not in our scheduling class.
272 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
273 unsigned long delta_exec)
275 unsigned long delta, delta_fair, delta_mine, delta_exec_weighted;
276 struct load_weight *lw = &cfs_rq->load;
277 unsigned long load = lw->weight;
279 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
281 curr->sum_exec_runtime += delta_exec;
282 cfs_rq->exec_clock += delta_exec;
283 delta_exec_weighted = delta_exec;
284 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
285 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
288 curr->vruntime += delta_exec_weighted;
290 if (!sched_feat(FAIR_SLEEPERS))
296 delta_fair = calc_delta_fair(delta_exec, lw);
297 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
299 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
300 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
301 delta = min(delta, (unsigned long)(
302 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
303 cfs_rq->sleeper_bonus -= delta;
307 cfs_rq->fair_clock += delta_fair;
309 * We executed delta_exec amount of time on the CPU,
310 * but we were only entitled to delta_mine amount of
311 * time during that period (if nr_running == 1 then
312 * the two values are equal)
313 * [Note: delta_mine - delta_exec is negative]:
315 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
318 static void update_curr(struct cfs_rq *cfs_rq)
320 struct sched_entity *curr = cfs_rq->curr;
321 u64 now = rq_of(cfs_rq)->clock;
322 unsigned long delta_exec;
328 * Get the amount of time the current task was running
329 * since the last time we changed load (this cannot
330 * overflow on 32 bits):
332 delta_exec = (unsigned long)(now - curr->exec_start);
334 __update_curr(cfs_rq, curr, delta_exec);
335 curr->exec_start = now;
339 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
341 se->wait_start_fair = cfs_rq->fair_clock;
342 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
345 static inline unsigned long
346 calc_weighted(unsigned long delta, struct sched_entity *se)
348 unsigned long weight = se->load.weight;
350 if (unlikely(weight != NICE_0_LOAD))
351 return (u64)delta * se->load.weight >> NICE_0_SHIFT;
357 * Task is being enqueued - update stats:
359 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
362 * Are we enqueueing a waiting task? (for current tasks
363 * a dequeue/enqueue event is a NOP)
365 if (se != cfs_rq->curr)
366 update_stats_wait_start(cfs_rq, se);
370 se->fair_key = se->vruntime;
374 * Note: must be called with a freshly updated rq->fair_clock.
377 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se,
378 unsigned long delta_fair)
380 schedstat_set(se->wait_max, max(se->wait_max,
381 rq_of(cfs_rq)->clock - se->wait_start));
383 delta_fair = calc_weighted(delta_fair, se);
385 add_wait_runtime(cfs_rq, se, delta_fair);
389 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
391 unsigned long delta_fair;
393 if (unlikely(!se->wait_start_fair))
396 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
397 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
399 __update_stats_wait_end(cfs_rq, se, delta_fair);
401 se->wait_start_fair = 0;
402 schedstat_set(se->wait_start, 0);
406 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
410 * Mark the end of the wait period if dequeueing a
413 if (se != cfs_rq->curr)
414 update_stats_wait_end(cfs_rq, se);
418 * We are picking a new current task - update its stats:
421 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
424 * We are starting a new run period:
426 se->exec_start = rq_of(cfs_rq)->clock;
430 * We are descheduling a task - update its stats:
433 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
438 /**************************************************
439 * Scheduling class queueing methods:
442 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
443 unsigned long delta_fair)
445 unsigned long load = cfs_rq->load.weight;
449 * Do not boost sleepers if there's too much bonus 'in flight'
452 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
455 if (sched_feat(SLEEPER_LOAD_AVG))
456 load = rq_of(cfs_rq)->cpu_load[2];
459 * Fix up delta_fair with the effect of us running
460 * during the whole sleep period:
462 if (sched_feat(SLEEPER_AVG))
463 delta_fair = div64_likely32((u64)delta_fair * load,
464 load + se->load.weight);
466 delta_fair = calc_weighted(delta_fair, se);
468 prev_runtime = se->wait_runtime;
469 __add_wait_runtime(cfs_rq, se, delta_fair);
470 delta_fair = se->wait_runtime - prev_runtime;
473 * Track the amount of bonus we've given to sleepers:
475 cfs_rq->sleeper_bonus += delta_fair;
478 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
480 struct task_struct *tsk = task_of(se);
481 unsigned long delta_fair;
483 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
484 !sched_feat(FAIR_SLEEPERS))
487 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
488 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
490 __enqueue_sleeper(cfs_rq, se, delta_fair);
492 se->sleep_start_fair = 0;
494 #ifdef CONFIG_SCHEDSTATS
495 if (se->sleep_start) {
496 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
501 if (unlikely(delta > se->sleep_max))
502 se->sleep_max = delta;
505 se->sum_sleep_runtime += delta;
507 if (se->block_start) {
508 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
513 if (unlikely(delta > se->block_max))
514 se->block_max = delta;
517 se->sum_sleep_runtime += delta;
520 * Blocking time is in units of nanosecs, so shift by 20 to
521 * get a milliseconds-range estimation of the amount of
522 * time that the task spent sleeping:
524 if (unlikely(prof_on == SLEEP_PROFILING)) {
525 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
533 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
536 * Update the fair clock.
541 u64 min_runtime, latency;
543 min_runtime = cfs_rq->min_vruntime;
544 min_runtime += sysctl_sched_latency/2;
546 if (sched_feat(NEW_FAIR_SLEEPERS)) {
547 latency = calc_weighted(sysctl_sched_latency, se);
548 if (min_runtime > latency)
549 min_runtime -= latency;
552 se->vruntime = max(se->vruntime, min_runtime);
554 enqueue_sleeper(cfs_rq, se);
557 update_stats_enqueue(cfs_rq, se);
558 __enqueue_entity(cfs_rq, se);
562 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
564 update_stats_dequeue(cfs_rq, se);
566 se->sleep_start_fair = cfs_rq->fair_clock;
567 #ifdef CONFIG_SCHEDSTATS
568 if (entity_is_task(se)) {
569 struct task_struct *tsk = task_of(se);
571 if (tsk->state & TASK_INTERRUPTIBLE)
572 se->sleep_start = rq_of(cfs_rq)->clock;
573 if (tsk->state & TASK_UNINTERRUPTIBLE)
574 se->block_start = rq_of(cfs_rq)->clock;
578 __dequeue_entity(cfs_rq, se);
582 * Preempt the current task with a newly woken task if needed:
585 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
587 unsigned long ideal_runtime, delta_exec;
589 ideal_runtime = sched_slice(cfs_rq, curr);
590 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
591 if (delta_exec > ideal_runtime)
592 resched_task(rq_of(cfs_rq)->curr);
596 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
599 * Any task has to be enqueued before it get to execute on
600 * a CPU. So account for the time it spent waiting on the
601 * runqueue. (note, here we rely on pick_next_task() having
602 * done a put_prev_task_fair() shortly before this, which
603 * updated rq->fair_clock - used by update_stats_wait_end())
605 update_stats_wait_end(cfs_rq, se);
606 update_stats_curr_start(cfs_rq, se);
608 #ifdef CONFIG_SCHEDSTATS
610 * Track our maximum slice length, if the CPU's load is at
611 * least twice that of our own weight (i.e. dont track it
612 * when there are only lesser-weight tasks around):
614 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
615 se->slice_max = max(se->slice_max,
616 se->sum_exec_runtime - se->prev_sum_exec_runtime);
619 se->prev_sum_exec_runtime = se->sum_exec_runtime;
622 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
624 struct sched_entity *se = __pick_next_entity(cfs_rq);
626 set_next_entity(cfs_rq, se);
631 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
634 * If still on the runqueue then deactivate_task()
635 * was not called and update_curr() has to be done:
640 update_stats_curr_end(cfs_rq, prev);
643 update_stats_wait_start(cfs_rq, prev);
647 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
650 * Dequeue and enqueue the task to update its
651 * position within the tree:
653 dequeue_entity(cfs_rq, curr, 0);
654 enqueue_entity(cfs_rq, curr, 0);
656 if (cfs_rq->nr_running > 1)
657 check_preempt_tick(cfs_rq, curr);
660 /**************************************************
661 * CFS operations on tasks:
664 #ifdef CONFIG_FAIR_GROUP_SCHED
666 /* Walk up scheduling entities hierarchy */
667 #define for_each_sched_entity(se) \
668 for (; se; se = se->parent)
670 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
675 /* runqueue on which this entity is (to be) queued */
676 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
681 /* runqueue "owned" by this group */
682 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
687 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
688 * another cpu ('this_cpu')
690 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
692 /* A later patch will take group into account */
693 return &cpu_rq(this_cpu)->cfs;
696 /* Iterate thr' all leaf cfs_rq's on a runqueue */
697 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
698 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
700 /* Do the two (enqueued) tasks belong to the same group ? */
701 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
703 if (curr->se.cfs_rq == p->se.cfs_rq)
709 #else /* CONFIG_FAIR_GROUP_SCHED */
711 #define for_each_sched_entity(se) \
712 for (; se; se = NULL)
714 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
716 return &task_rq(p)->cfs;
719 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
721 struct task_struct *p = task_of(se);
722 struct rq *rq = task_rq(p);
727 /* runqueue "owned" by this group */
728 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
733 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
735 return &cpu_rq(this_cpu)->cfs;
738 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
739 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
741 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
746 #endif /* CONFIG_FAIR_GROUP_SCHED */
749 * The enqueue_task method is called before nr_running is
750 * increased. Here we update the fair scheduling stats and
751 * then put the task into the rbtree:
753 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
755 struct cfs_rq *cfs_rq;
756 struct sched_entity *se = &p->se;
758 for_each_sched_entity(se) {
761 cfs_rq = cfs_rq_of(se);
762 enqueue_entity(cfs_rq, se, wakeup);
767 * The dequeue_task method is called before nr_running is
768 * decreased. We remove the task from the rbtree and
769 * update the fair scheduling stats:
771 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
773 struct cfs_rq *cfs_rq;
774 struct sched_entity *se = &p->se;
776 for_each_sched_entity(se) {
777 cfs_rq = cfs_rq_of(se);
778 dequeue_entity(cfs_rq, se, sleep);
779 /* Don't dequeue parent if it has other entities besides us */
780 if (cfs_rq->load.weight)
786 * sched_yield() support is very simple - we dequeue and enqueue.
788 * If compat_yield is turned on then we requeue to the end of the tree.
790 static void yield_task_fair(struct rq *rq, struct task_struct *p)
792 struct cfs_rq *cfs_rq = task_cfs_rq(p);
793 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
794 struct sched_entity *rightmost, *se = &p->se;
795 struct rb_node *parent;
798 * Are we the only task in the tree?
800 if (unlikely(cfs_rq->nr_running == 1))
803 if (likely(!sysctl_sched_compat_yield)) {
804 __update_rq_clock(rq);
806 * Dequeue and enqueue the task to update its
807 * position within the tree:
809 dequeue_entity(cfs_rq, &p->se, 0);
810 enqueue_entity(cfs_rq, &p->se, 0);
815 * Find the rightmost entry in the rbtree:
819 link = &parent->rb_right;
822 rightmost = rb_entry(parent, struct sched_entity, run_node);
824 * Already in the rightmost position?
826 if (unlikely(rightmost == se))
830 * Minimally necessary key value to be last in the tree:
832 se->fair_key = rightmost->fair_key + 1;
834 if (cfs_rq->rb_leftmost == &se->run_node)
835 cfs_rq->rb_leftmost = rb_next(&se->run_node);
837 * Relink the task to the rightmost position:
839 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
840 rb_link_node(&se->run_node, parent, link);
841 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
845 * Preempt the current task with a newly woken task if needed:
847 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
849 struct task_struct *curr = rq->curr;
850 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
852 if (unlikely(rt_prio(p->prio))) {
858 if (is_same_group(curr, p)) {
859 s64 delta = curr->se.vruntime - p->se.vruntime;
861 if (delta > (s64)sysctl_sched_wakeup_granularity)
866 static struct task_struct *pick_next_task_fair(struct rq *rq)
868 struct cfs_rq *cfs_rq = &rq->cfs;
869 struct sched_entity *se;
871 if (unlikely(!cfs_rq->nr_running))
875 se = pick_next_entity(cfs_rq);
876 cfs_rq = group_cfs_rq(se);
883 * Account for a descheduled task:
885 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
887 struct sched_entity *se = &prev->se;
888 struct cfs_rq *cfs_rq;
890 for_each_sched_entity(se) {
891 cfs_rq = cfs_rq_of(se);
892 put_prev_entity(cfs_rq, se);
896 /**************************************************
897 * Fair scheduling class load-balancing methods:
901 * Load-balancing iterator. Note: while the runqueue stays locked
902 * during the whole iteration, the current task might be
903 * dequeued so the iterator has to be dequeue-safe. Here we
904 * achieve that by always pre-iterating before returning
907 static inline struct task_struct *
908 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
910 struct task_struct *p;
915 p = rb_entry(curr, struct task_struct, se.run_node);
916 cfs_rq->rb_load_balance_curr = rb_next(curr);
921 static struct task_struct *load_balance_start_fair(void *arg)
923 struct cfs_rq *cfs_rq = arg;
925 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
928 static struct task_struct *load_balance_next_fair(void *arg)
930 struct cfs_rq *cfs_rq = arg;
932 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
935 #ifdef CONFIG_FAIR_GROUP_SCHED
936 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
938 struct sched_entity *curr;
939 struct task_struct *p;
941 if (!cfs_rq->nr_running)
944 curr = __pick_next_entity(cfs_rq);
952 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
953 unsigned long max_nr_move, unsigned long max_load_move,
954 struct sched_domain *sd, enum cpu_idle_type idle,
955 int *all_pinned, int *this_best_prio)
957 struct cfs_rq *busy_cfs_rq;
958 unsigned long load_moved, total_nr_moved = 0, nr_moved;
959 long rem_load_move = max_load_move;
960 struct rq_iterator cfs_rq_iterator;
962 cfs_rq_iterator.start = load_balance_start_fair;
963 cfs_rq_iterator.next = load_balance_next_fair;
965 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
966 #ifdef CONFIG_FAIR_GROUP_SCHED
967 struct cfs_rq *this_cfs_rq;
969 unsigned long maxload;
971 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
973 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
974 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
978 /* Don't pull more than imbalance/2 */
980 maxload = min(rem_load_move, imbalance);
982 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
984 # define maxload rem_load_move
986 /* pass busy_cfs_rq argument into
987 * load_balance_[start|next]_fair iterators
989 cfs_rq_iterator.arg = busy_cfs_rq;
990 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
991 max_nr_move, maxload, sd, idle, all_pinned,
992 &load_moved, this_best_prio, &cfs_rq_iterator);
994 total_nr_moved += nr_moved;
995 max_nr_move -= nr_moved;
996 rem_load_move -= load_moved;
998 if (max_nr_move <= 0 || rem_load_move <= 0)
1002 return max_load_move - rem_load_move;
1006 * scheduler tick hitting a task of our scheduling class:
1008 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1010 struct cfs_rq *cfs_rq;
1011 struct sched_entity *se = &curr->se;
1013 for_each_sched_entity(se) {
1014 cfs_rq = cfs_rq_of(se);
1015 entity_tick(cfs_rq, se);
1019 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1022 * Share the fairness runtime between parent and child, thus the
1023 * total amount of pressure for CPU stays equal - new tasks
1024 * get a chance to run but frequent forkers are not allowed to
1025 * monopolize the CPU. Note: the parent runqueue is locked,
1026 * the child is not running yet.
1028 static void task_new_fair(struct rq *rq, struct task_struct *p)
1030 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1031 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1033 sched_info_queued(p);
1035 update_curr(cfs_rq);
1036 se->vruntime = cfs_rq->min_vruntime;
1037 update_stats_enqueue(cfs_rq, se);
1040 * The first wait is dominated by the child-runs-first logic,
1041 * so do not credit it with that waiting time yet:
1043 if (sched_feat(SKIP_INITIAL))
1044 se->wait_start_fair = 0;
1047 * The statistical average of wait_runtime is about
1048 * -granularity/2, so initialize the task with that:
1050 if (sched_feat(START_DEBIT))
1051 se->wait_runtime = -(__sched_period(cfs_rq->nr_running+1) / 2);
1053 if (sysctl_sched_child_runs_first &&
1054 curr->vruntime < se->vruntime) {
1056 dequeue_entity(cfs_rq, curr, 0);
1057 swap(curr->vruntime, se->vruntime);
1058 enqueue_entity(cfs_rq, curr, 0);
1061 update_stats_enqueue(cfs_rq, se);
1062 __enqueue_entity(cfs_rq, se);
1063 resched_task(rq->curr);
1066 #ifdef CONFIG_FAIR_GROUP_SCHED
1067 /* Account for a task changing its policy or group.
1069 * This routine is mostly called to set cfs_rq->curr field when a task
1070 * migrates between groups/classes.
1072 static void set_curr_task_fair(struct rq *rq)
1074 struct sched_entity *se = &rq->curr->se;
1076 for_each_sched_entity(se)
1077 set_next_entity(cfs_rq_of(se), se);
1080 static void set_curr_task_fair(struct rq *rq)
1086 * All the scheduling class methods:
1088 struct sched_class fair_sched_class __read_mostly = {
1089 .enqueue_task = enqueue_task_fair,
1090 .dequeue_task = dequeue_task_fair,
1091 .yield_task = yield_task_fair,
1093 .check_preempt_curr = check_preempt_wakeup,
1095 .pick_next_task = pick_next_task_fair,
1096 .put_prev_task = put_prev_task_fair,
1098 .load_balance = load_balance_fair,
1100 .set_curr_task = set_curr_task_fair,
1101 .task_tick = task_tick_fair,
1102 .task_new = task_new_fair,
1105 #ifdef CONFIG_SCHED_DEBUG
1106 static void print_cfs_stats(struct seq_file *m, int cpu)
1108 struct cfs_rq *cfs_rq;
1110 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1111 print_cfs_rq(m, cpu, cfs_rq);