2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
26 * Targeted preemption latency for CPU-bound tasks:
27 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
29 * NOTE: this latency value is not the same as the concept of
30 * 'timeslice length' - timeslices in CFS are of variable length
31 * and have no persistent notion like in traditional, time-slice
32 * based scheduling concepts.
34 * (to see the precise effective timeslice length of your workload,
35 * run vmstat and monitor the context-switches (cs) field)
37 unsigned int sysctl_sched_latency = 20000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity = 4000000ULL;
46 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
48 static unsigned int sched_nr_latency = 5;
51 * After fork, child runs first. (default) If set to 0 then
52 * parent will (try to) run first.
54 const_debug unsigned int sysctl_sched_child_runs_first = 1;
57 * sys_sched_yield() compat mode
59 * This option switches the agressive yield implementation of the
60 * old scheduler back on.
62 unsigned int __read_mostly sysctl_sched_compat_yield;
65 * SCHED_OTHER wake-up granularity.
66 * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 * This option delays the preemption effects of decoupled workloads
69 * and reduces their over-scheduling. Synchronous workloads will still
70 * have immediate wakeup/sleep latencies.
72 unsigned int sysctl_sched_wakeup_granularity = 5000000UL;
74 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
76 static const struct sched_class fair_sched_class;
78 /**************************************************************
79 * CFS operations on generic schedulable entities:
82 #ifdef CONFIG_FAIR_GROUP_SCHED
84 /* cpu runqueue to which this cfs_rq is attached */
85 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
90 /* An entity is a task if it doesn't "own" a runqueue */
91 #define entity_is_task(se) (!se->my_q)
93 static inline struct task_struct *task_of(struct sched_entity *se)
95 #ifdef CONFIG_SCHED_DEBUG
96 WARN_ON_ONCE(!entity_is_task(se));
98 return container_of(se, struct task_struct, se);
101 /* Walk up scheduling entities hierarchy */
102 #define for_each_sched_entity(se) \
103 for (; se; se = se->parent)
105 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
110 /* runqueue on which this entity is (to be) queued */
111 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
116 /* runqueue "owned" by this group */
117 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
122 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
123 * another cpu ('this_cpu')
125 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
127 return cfs_rq->tg->cfs_rq[this_cpu];
130 /* Iterate thr' all leaf cfs_rq's on a runqueue */
131 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
132 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
134 /* Do the two (enqueued) entities belong to the same group ? */
136 is_same_group(struct sched_entity *se, struct sched_entity *pse)
138 if (se->cfs_rq == pse->cfs_rq)
144 static inline struct sched_entity *parent_entity(struct sched_entity *se)
149 /* return depth at which a sched entity is present in the hierarchy */
150 static inline int depth_se(struct sched_entity *se)
154 for_each_sched_entity(se)
161 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
163 int se_depth, pse_depth;
166 * preemption test can be made between sibling entities who are in the
167 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
168 * both tasks until we find their ancestors who are siblings of common
172 /* First walk up until both entities are at same depth */
173 se_depth = depth_se(*se);
174 pse_depth = depth_se(*pse);
176 while (se_depth > pse_depth) {
178 *se = parent_entity(*se);
181 while (pse_depth > se_depth) {
183 *pse = parent_entity(*pse);
186 while (!is_same_group(*se, *pse)) {
187 *se = parent_entity(*se);
188 *pse = parent_entity(*pse);
192 #else /* !CONFIG_FAIR_GROUP_SCHED */
194 static inline struct task_struct *task_of(struct sched_entity *se)
196 return container_of(se, struct task_struct, se);
199 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
201 return container_of(cfs_rq, struct rq, cfs);
204 #define entity_is_task(se) 1
206 #define for_each_sched_entity(se) \
207 for (; se; se = NULL)
209 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
211 return &task_rq(p)->cfs;
214 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
216 struct task_struct *p = task_of(se);
217 struct rq *rq = task_rq(p);
222 /* runqueue "owned" by this group */
223 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
228 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
230 return &cpu_rq(this_cpu)->cfs;
233 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
234 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
237 is_same_group(struct sched_entity *se, struct sched_entity *pse)
242 static inline struct sched_entity *parent_entity(struct sched_entity *se)
248 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
252 #endif /* CONFIG_FAIR_GROUP_SCHED */
255 /**************************************************************
256 * Scheduling class tree data structure manipulation methods:
259 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
261 s64 delta = (s64)(vruntime - min_vruntime);
263 min_vruntime = vruntime;
268 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
270 s64 delta = (s64)(vruntime - min_vruntime);
272 min_vruntime = vruntime;
277 static inline int entity_before(struct sched_entity *a,
278 struct sched_entity *b)
280 return (s64)(a->vruntime - b->vruntime) < 0;
283 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
285 return se->vruntime - cfs_rq->min_vruntime;
288 static void update_min_vruntime(struct cfs_rq *cfs_rq)
290 u64 vruntime = cfs_rq->min_vruntime;
293 vruntime = cfs_rq->curr->vruntime;
295 if (cfs_rq->rb_leftmost) {
296 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
301 vruntime = se->vruntime;
303 vruntime = min_vruntime(vruntime, se->vruntime);
306 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
310 * Enqueue an entity into the rb-tree:
312 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
314 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
315 struct rb_node *parent = NULL;
316 struct sched_entity *entry;
317 s64 key = entity_key(cfs_rq, se);
321 * Find the right place in the rbtree:
325 entry = rb_entry(parent, struct sched_entity, run_node);
327 * We dont care about collisions. Nodes with
328 * the same key stay together.
330 if (key < entity_key(cfs_rq, entry)) {
331 link = &parent->rb_left;
333 link = &parent->rb_right;
339 * Maintain a cache of leftmost tree entries (it is frequently
343 cfs_rq->rb_leftmost = &se->run_node;
345 rb_link_node(&se->run_node, parent, link);
346 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
349 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
351 if (cfs_rq->rb_leftmost == &se->run_node) {
352 struct rb_node *next_node;
354 next_node = rb_next(&se->run_node);
355 cfs_rq->rb_leftmost = next_node;
358 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
361 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
363 struct rb_node *left = cfs_rq->rb_leftmost;
368 return rb_entry(left, struct sched_entity, run_node);
371 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
373 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
378 return rb_entry(last, struct sched_entity, run_node);
381 /**************************************************************
382 * Scheduling class statistics methods:
385 #ifdef CONFIG_SCHED_DEBUG
386 int sched_nr_latency_handler(struct ctl_table *table, int write,
387 struct file *filp, void __user *buffer, size_t *lenp,
390 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
395 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
396 sysctl_sched_min_granularity);
405 static inline unsigned long
406 calc_delta_fair(unsigned long delta, struct sched_entity *se)
408 if (unlikely(se->load.weight != NICE_0_LOAD))
409 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
415 * The idea is to set a period in which each task runs once.
417 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
418 * this period because otherwise the slices get too small.
420 * p = (nr <= nl) ? l : l*nr/nl
422 static u64 __sched_period(unsigned long nr_running)
424 u64 period = sysctl_sched_latency;
425 unsigned long nr_latency = sched_nr_latency;
427 if (unlikely(nr_running > nr_latency)) {
428 period = sysctl_sched_min_granularity;
429 period *= nr_running;
436 * We calculate the wall-time slice from the period by taking a part
437 * proportional to the weight.
441 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
443 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
445 for_each_sched_entity(se) {
446 struct load_weight *load;
447 struct load_weight lw;
449 cfs_rq = cfs_rq_of(se);
450 load = &cfs_rq->load;
452 if (unlikely(!se->on_rq)) {
455 update_load_add(&lw, se->load.weight);
458 slice = calc_delta_mine(slice, se->load.weight, load);
464 * We calculate the vruntime slice of a to be inserted task
468 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
470 return calc_delta_fair(sched_slice(cfs_rq, se), se);
474 * Update the current task's runtime statistics. Skip current tasks that
475 * are not in our scheduling class.
478 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
479 unsigned long delta_exec)
481 unsigned long delta_exec_weighted;
483 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
485 curr->sum_exec_runtime += delta_exec;
486 schedstat_add(cfs_rq, exec_clock, delta_exec);
487 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
488 curr->vruntime += delta_exec_weighted;
489 update_min_vruntime(cfs_rq);
492 static void update_curr(struct cfs_rq *cfs_rq)
494 struct sched_entity *curr = cfs_rq->curr;
495 u64 now = rq_of(cfs_rq)->clock;
496 unsigned long delta_exec;
502 * Get the amount of time the current task was running
503 * since the last time we changed load (this cannot
504 * overflow on 32 bits):
506 delta_exec = (unsigned long)(now - curr->exec_start);
510 __update_curr(cfs_rq, curr, delta_exec);
511 curr->exec_start = now;
513 if (entity_is_task(curr)) {
514 struct task_struct *curtask = task_of(curr);
516 cpuacct_charge(curtask, delta_exec);
517 account_group_exec_runtime(curtask, delta_exec);
522 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
528 * Task is being enqueued - update stats:
530 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
533 * Are we enqueueing a waiting task? (for current tasks
534 * a dequeue/enqueue event is a NOP)
536 if (se != cfs_rq->curr)
537 update_stats_wait_start(cfs_rq, se);
541 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 schedstat_set(se->wait_max, max(se->wait_max,
544 rq_of(cfs_rq)->clock - se->wait_start));
545 schedstat_set(se->wait_count, se->wait_count + 1);
546 schedstat_set(se->wait_sum, se->wait_sum +
547 rq_of(cfs_rq)->clock - se->wait_start);
548 schedstat_set(se->wait_start, 0);
552 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
555 * Mark the end of the wait period if dequeueing a
558 if (se != cfs_rq->curr)
559 update_stats_wait_end(cfs_rq, se);
563 * We are picking a new current task - update its stats:
566 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
569 * We are starting a new run period:
571 se->exec_start = rq_of(cfs_rq)->clock;
574 /**************************************************
575 * Scheduling class queueing methods:
578 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
580 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
582 cfs_rq->task_weight += weight;
586 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
592 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
594 update_load_add(&cfs_rq->load, se->load.weight);
595 if (!parent_entity(se))
596 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
597 if (entity_is_task(se)) {
598 add_cfs_task_weight(cfs_rq, se->load.weight);
599 list_add(&se->group_node, &cfs_rq->tasks);
601 cfs_rq->nr_running++;
606 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
608 update_load_sub(&cfs_rq->load, se->load.weight);
609 if (!parent_entity(se))
610 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
611 if (entity_is_task(se)) {
612 add_cfs_task_weight(cfs_rq, -se->load.weight);
613 list_del_init(&se->group_node);
615 cfs_rq->nr_running--;
619 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
621 #ifdef CONFIG_SCHEDSTATS
622 struct task_struct *tsk = NULL;
624 if (entity_is_task(se))
627 if (se->sleep_start) {
628 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
633 if (unlikely(delta > se->sleep_max))
634 se->sleep_max = delta;
637 se->sum_sleep_runtime += delta;
640 account_scheduler_latency(tsk, delta >> 10, 1);
642 if (se->block_start) {
643 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
648 if (unlikely(delta > se->block_max))
649 se->block_max = delta;
652 se->sum_sleep_runtime += delta;
655 if (tsk->in_iowait) {
656 se->iowait_sum += delta;
661 * Blocking time is in units of nanosecs, so shift by
662 * 20 to get a milliseconds-range estimation of the
663 * amount of time that the task spent sleeping:
665 if (unlikely(prof_on == SLEEP_PROFILING)) {
666 profile_hits(SLEEP_PROFILING,
667 (void *)get_wchan(tsk),
670 account_scheduler_latency(tsk, delta >> 10, 0);
676 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
678 #ifdef CONFIG_SCHED_DEBUG
679 s64 d = se->vruntime - cfs_rq->min_vruntime;
684 if (d > 3*sysctl_sched_latency)
685 schedstat_inc(cfs_rq, nr_spread_over);
690 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
692 u64 vruntime = cfs_rq->min_vruntime;
695 * The 'current' period is already promised to the current tasks,
696 * however the extra weight of the new task will slow them down a
697 * little, place the new task so that it fits in the slot that
698 * stays open at the end.
700 if (initial && sched_feat(START_DEBIT))
701 vruntime += sched_vslice(cfs_rq, se);
704 /* sleeps upto a single latency don't count. */
705 if (sched_feat(NEW_FAIR_SLEEPERS)) {
706 unsigned long thresh = sysctl_sched_latency;
709 * Convert the sleeper threshold into virtual time.
710 * SCHED_IDLE is a special sub-class. We care about
711 * fairness only relative to other SCHED_IDLE tasks,
712 * all of which have the same weight.
714 if (sched_feat(NORMALIZED_SLEEPER) &&
715 (!entity_is_task(se) ||
716 task_of(se)->policy != SCHED_IDLE))
717 thresh = calc_delta_fair(thresh, se);
722 /* ensure we never gain time by being placed backwards. */
723 vruntime = max_vruntime(se->vruntime, vruntime);
726 se->vruntime = vruntime;
730 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
733 * Update run-time statistics of the 'current'.
736 account_entity_enqueue(cfs_rq, se);
739 place_entity(cfs_rq, se, 0);
740 enqueue_sleeper(cfs_rq, se);
743 update_stats_enqueue(cfs_rq, se);
744 check_spread(cfs_rq, se);
745 if (se != cfs_rq->curr)
746 __enqueue_entity(cfs_rq, se);
749 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
751 if (cfs_rq->last == se)
754 if (cfs_rq->next == se)
758 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
760 for_each_sched_entity(se)
761 __clear_buddies(cfs_rq_of(se), se);
765 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
768 * Update run-time statistics of the 'current'.
772 update_stats_dequeue(cfs_rq, se);
774 #ifdef CONFIG_SCHEDSTATS
775 if (entity_is_task(se)) {
776 struct task_struct *tsk = task_of(se);
778 if (tsk->state & TASK_INTERRUPTIBLE)
779 se->sleep_start = rq_of(cfs_rq)->clock;
780 if (tsk->state & TASK_UNINTERRUPTIBLE)
781 se->block_start = rq_of(cfs_rq)->clock;
786 clear_buddies(cfs_rq, se);
788 if (se != cfs_rq->curr)
789 __dequeue_entity(cfs_rq, se);
790 account_entity_dequeue(cfs_rq, se);
791 update_min_vruntime(cfs_rq);
795 * Preempt the current task with a newly woken task if needed:
798 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
800 unsigned long ideal_runtime, delta_exec;
802 ideal_runtime = sched_slice(cfs_rq, curr);
803 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
804 if (delta_exec > ideal_runtime) {
805 resched_task(rq_of(cfs_rq)->curr);
807 * The current task ran long enough, ensure it doesn't get
808 * re-elected due to buddy favours.
810 clear_buddies(cfs_rq, curr);
815 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
817 /* 'current' is not kept within the tree. */
820 * Any task has to be enqueued before it get to execute on
821 * a CPU. So account for the time it spent waiting on the
824 update_stats_wait_end(cfs_rq, se);
825 __dequeue_entity(cfs_rq, se);
828 update_stats_curr_start(cfs_rq, se);
830 #ifdef CONFIG_SCHEDSTATS
832 * Track our maximum slice length, if the CPU's load is at
833 * least twice that of our own weight (i.e. dont track it
834 * when there are only lesser-weight tasks around):
836 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
837 se->slice_max = max(se->slice_max,
838 se->sum_exec_runtime - se->prev_sum_exec_runtime);
841 se->prev_sum_exec_runtime = se->sum_exec_runtime;
845 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
847 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
849 struct sched_entity *se = __pick_next_entity(cfs_rq);
851 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
854 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
860 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
863 * If still on the runqueue then deactivate_task()
864 * was not called and update_curr() has to be done:
869 check_spread(cfs_rq, prev);
871 update_stats_wait_start(cfs_rq, prev);
872 /* Put 'current' back into the tree. */
873 __enqueue_entity(cfs_rq, prev);
879 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
882 * Update run-time statistics of the 'current'.
886 #ifdef CONFIG_SCHED_HRTICK
888 * queued ticks are scheduled to match the slice, so don't bother
889 * validating it and just reschedule.
892 resched_task(rq_of(cfs_rq)->curr);
896 * don't let the period tick interfere with the hrtick preemption
898 if (!sched_feat(DOUBLE_TICK) &&
899 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
903 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
904 check_preempt_tick(cfs_rq, curr);
907 /**************************************************
908 * CFS operations on tasks:
911 #ifdef CONFIG_SCHED_HRTICK
912 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
914 struct sched_entity *se = &p->se;
915 struct cfs_rq *cfs_rq = cfs_rq_of(se);
917 WARN_ON(task_rq(p) != rq);
919 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
920 u64 slice = sched_slice(cfs_rq, se);
921 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
922 s64 delta = slice - ran;
931 * Don't schedule slices shorter than 10000ns, that just
932 * doesn't make sense. Rely on vruntime for fairness.
935 delta = max_t(s64, 10000LL, delta);
937 hrtick_start(rq, delta);
942 * called from enqueue/dequeue and updates the hrtick when the
943 * current task is from our class and nr_running is low enough
946 static void hrtick_update(struct rq *rq)
948 struct task_struct *curr = rq->curr;
950 if (curr->sched_class != &fair_sched_class)
953 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
954 hrtick_start_fair(rq, curr);
956 #else /* !CONFIG_SCHED_HRTICK */
958 hrtick_start_fair(struct rq *rq, struct task_struct *p)
962 static inline void hrtick_update(struct rq *rq)
968 * The enqueue_task method is called before nr_running is
969 * increased. Here we update the fair scheduling stats and
970 * then put the task into the rbtree:
972 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
974 struct cfs_rq *cfs_rq;
975 struct sched_entity *se = &p->se;
977 for_each_sched_entity(se) {
980 cfs_rq = cfs_rq_of(se);
981 enqueue_entity(cfs_rq, se, wakeup);
989 * The dequeue_task method is called before nr_running is
990 * decreased. We remove the task from the rbtree and
991 * update the fair scheduling stats:
993 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
995 struct cfs_rq *cfs_rq;
996 struct sched_entity *se = &p->se;
998 for_each_sched_entity(se) {
999 cfs_rq = cfs_rq_of(se);
1000 dequeue_entity(cfs_rq, se, sleep);
1001 /* Don't dequeue parent if it has other entities besides us */
1002 if (cfs_rq->load.weight)
1011 * sched_yield() support is very simple - we dequeue and enqueue.
1013 * If compat_yield is turned on then we requeue to the end of the tree.
1015 static void yield_task_fair(struct rq *rq)
1017 struct task_struct *curr = rq->curr;
1018 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1019 struct sched_entity *rightmost, *se = &curr->se;
1022 * Are we the only task in the tree?
1024 if (unlikely(cfs_rq->nr_running == 1))
1027 clear_buddies(cfs_rq, se);
1029 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1030 update_rq_clock(rq);
1032 * Update run-time statistics of the 'current'.
1034 update_curr(cfs_rq);
1039 * Find the rightmost entry in the rbtree:
1041 rightmost = __pick_last_entity(cfs_rq);
1043 * Already in the rightmost position?
1045 if (unlikely(!rightmost || entity_before(rightmost, se)))
1049 * Minimally necessary key value to be last in the tree:
1050 * Upon rescheduling, sched_class::put_prev_task() will place
1051 * 'current' within the tree based on its new key value.
1053 se->vruntime = rightmost->vruntime + 1;
1057 * wake_idle() will wake a task on an idle cpu if task->cpu is
1058 * not idle and an idle cpu is available. The span of cpus to
1059 * search starts with cpus closest then further out as needed,
1060 * so we always favor a closer, idle cpu.
1061 * Domains may include CPUs that are not usable for migration,
1062 * hence we need to mask them out (rq->rd->online)
1064 * Returns the CPU we should wake onto.
1066 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1068 #define cpu_rd_active(cpu, rq) cpumask_test_cpu(cpu, rq->rd->online)
1070 static int wake_idle(int cpu, struct task_struct *p)
1072 struct sched_domain *sd;
1074 unsigned int chosen_wakeup_cpu;
1076 struct rq *task_rq = task_rq(p);
1079 * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu
1080 * are idle and this is not a kernel thread and this task's affinity
1081 * allows it to be moved to preferred cpu, then just move!
1084 this_cpu = smp_processor_id();
1086 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu;
1088 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP &&
1089 idle_cpu(cpu) && idle_cpu(this_cpu) &&
1090 p->mm && !(p->flags & PF_KTHREAD) &&
1091 cpu_isset(chosen_wakeup_cpu, p->cpus_allowed))
1092 return chosen_wakeup_cpu;
1095 * If it is idle, then it is the best cpu to run this task.
1097 * This cpu is also the best, if it has more than one task already.
1098 * Siblings must be also busy(in most cases) as they didn't already
1099 * pickup the extra load from this cpu and hence we need not check
1100 * sibling runqueue info. This will avoid the checks and cache miss
1101 * penalities associated with that.
1103 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
1106 for_each_domain(cpu, sd) {
1107 if ((sd->flags & SD_WAKE_IDLE)
1108 || ((sd->flags & SD_WAKE_IDLE_FAR)
1109 && !task_hot(p, task_rq->clock, sd))) {
1110 for_each_cpu_and(i, sched_domain_span(sd),
1112 if (cpu_rd_active(i, task_rq) && idle_cpu(i)) {
1113 if (i != task_cpu(p)) {
1115 se.nr_wakeups_idle);
1126 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1127 static inline int wake_idle(int cpu, struct task_struct *p)
1135 #ifdef CONFIG_FAIR_GROUP_SCHED
1137 * effective_load() calculates the load change as seen from the root_task_group
1139 * Adding load to a group doesn't make a group heavier, but can cause movement
1140 * of group shares between cpus. Assuming the shares were perfectly aligned one
1141 * can calculate the shift in shares.
1143 * The problem is that perfectly aligning the shares is rather expensive, hence
1144 * we try to avoid doing that too often - see update_shares(), which ratelimits
1147 * We compensate this by not only taking the current delta into account, but
1148 * also considering the delta between when the shares were last adjusted and
1151 * We still saw a performance dip, some tracing learned us that between
1152 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1153 * significantly. Therefore try to bias the error in direction of failing
1154 * the affine wakeup.
1157 static long effective_load(struct task_group *tg, int cpu,
1160 struct sched_entity *se = tg->se[cpu];
1166 * By not taking the decrease of shares on the other cpu into
1167 * account our error leans towards reducing the affine wakeups.
1169 if (!wl && sched_feat(ASYM_EFF_LOAD))
1172 for_each_sched_entity(se) {
1173 long S, rw, s, a, b;
1177 * Instead of using this increment, also add the difference
1178 * between when the shares were last updated and now.
1180 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1184 S = se->my_q->tg->shares;
1185 s = se->my_q->shares;
1186 rw = se->my_q->rq_weight;
1197 * Assume the group is already running and will
1198 * thus already be accounted for in the weight.
1200 * That is, moving shares between CPUs, does not
1201 * alter the group weight.
1211 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1212 unsigned long wl, unsigned long wg)
1220 wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
1221 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1222 int idx, unsigned long load, unsigned long this_load,
1223 unsigned int imbalance)
1225 struct task_struct *curr = this_rq->curr;
1226 struct task_group *tg;
1227 unsigned long tl = this_load;
1228 unsigned long tl_per_task;
1229 unsigned long weight;
1232 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1235 if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1236 p->se.avg_overlap > sysctl_sched_migration_cost))
1240 * If sync wakeup then subtract the (maximum possible)
1241 * effect of the currently running task from the load
1242 * of the current CPU:
1245 tg = task_group(current);
1246 weight = current->se.load.weight;
1248 tl += effective_load(tg, this_cpu, -weight, -weight);
1249 load += effective_load(tg, prev_cpu, 0, -weight);
1253 weight = p->se.load.weight;
1255 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1256 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1259 * If the currently running task will sleep within
1260 * a reasonable amount of time then attract this newly
1263 if (sync && balanced)
1266 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1267 tl_per_task = cpu_avg_load_per_task(this_cpu);
1269 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1272 * This domain has SD_WAKE_AFFINE and
1273 * p is cache cold in this domain, and
1274 * there is no bad imbalance.
1276 schedstat_inc(this_sd, ttwu_move_affine);
1277 schedstat_inc(p, se.nr_wakeups_affine);
1284 static int select_task_rq_fair(struct task_struct *p, int sync)
1286 struct sched_domain *sd, *this_sd = NULL;
1287 int prev_cpu, this_cpu, new_cpu;
1288 unsigned long load, this_load;
1290 unsigned int imbalance;
1293 prev_cpu = task_cpu(p);
1294 this_cpu = smp_processor_id();
1295 this_rq = cpu_rq(this_cpu);
1298 if (prev_cpu == this_cpu)
1301 * 'this_sd' is the first domain that both
1302 * this_cpu and prev_cpu are present in:
1304 for_each_domain(this_cpu, sd) {
1305 if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) {
1311 if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed)))
1315 * Check for affine wakeup and passive balancing possibilities.
1320 idx = this_sd->wake_idx;
1322 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1324 load = source_load(prev_cpu, idx);
1325 this_load = target_load(this_cpu, idx);
1327 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1328 load, this_load, imbalance))
1332 * Start passive balancing when half the imbalance_pct
1335 if (this_sd->flags & SD_WAKE_BALANCE) {
1336 if (imbalance*this_load <= 100*load) {
1337 schedstat_inc(this_sd, ttwu_move_balance);
1338 schedstat_inc(p, se.nr_wakeups_passive);
1344 return wake_idle(new_cpu, p);
1346 #endif /* CONFIG_SMP */
1349 * Adaptive granularity
1351 * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1352 * with the limit of wakeup_gran -- when it never does a wakeup.
1354 * So the smaller avg_wakeup is the faster we want this task to preempt,
1355 * but we don't want to treat the preemptee unfairly and therefore allow it
1356 * to run for at least the amount of time we'd like to run.
1358 * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1360 * NOTE: we use *nr_running to scale with load, this nicely matches the
1361 * degrading latency on load.
1363 static unsigned long
1364 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1366 u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1367 u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1370 if (this_run < expected_wakeup)
1371 gran = expected_wakeup - this_run;
1373 return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1376 static unsigned long
1377 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1379 unsigned long gran = sysctl_sched_wakeup_granularity;
1381 if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1382 gran = adaptive_gran(curr, se);
1385 * Since its curr running now, convert the gran from real-time
1386 * to virtual-time in his units.
1388 if (sched_feat(ASYM_GRAN)) {
1390 * By using 'se' instead of 'curr' we penalize light tasks, so
1391 * they get preempted easier. That is, if 'se' < 'curr' then
1392 * the resulting gran will be larger, therefore penalizing the
1393 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1394 * be smaller, again penalizing the lighter task.
1396 * This is especially important for buddies when the leftmost
1397 * task is higher priority than the buddy.
1399 if (unlikely(se->load.weight != NICE_0_LOAD))
1400 gran = calc_delta_fair(gran, se);
1402 if (unlikely(curr->load.weight != NICE_0_LOAD))
1403 gran = calc_delta_fair(gran, curr);
1410 * Should 'se' preempt 'curr'.
1424 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1426 s64 gran, vdiff = curr->vruntime - se->vruntime;
1431 gran = wakeup_gran(curr, se);
1438 static void set_last_buddy(struct sched_entity *se)
1440 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1441 for_each_sched_entity(se)
1442 cfs_rq_of(se)->last = se;
1446 static void set_next_buddy(struct sched_entity *se)
1448 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1449 for_each_sched_entity(se)
1450 cfs_rq_of(se)->next = se;
1455 * Preempt the current task with a newly woken task if needed:
1457 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
1459 struct task_struct *curr = rq->curr;
1460 struct sched_entity *se = &curr->se, *pse = &p->se;
1461 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1463 update_curr(cfs_rq);
1465 if (unlikely(rt_prio(p->prio))) {
1470 if (unlikely(p->sched_class != &fair_sched_class))
1473 if (unlikely(se == pse))
1477 * Only set the backward buddy when the current task is still on the
1478 * rq. This can happen when a wakeup gets interleaved with schedule on
1479 * the ->pre_schedule() or idle_balance() point, either of which can
1482 * Also, during early boot the idle thread is in the fair class, for
1483 * obvious reasons its a bad idea to schedule back to the idle thread.
1485 if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1487 set_next_buddy(pse);
1490 * We can come here with TIF_NEED_RESCHED already set from new task
1493 if (test_tsk_need_resched(curr))
1497 * Batch and idle tasks do not preempt (their preemption is driven by
1500 if (unlikely(p->policy != SCHED_NORMAL))
1503 /* Idle tasks are by definition preempted by everybody. */
1504 if (unlikely(curr->policy == SCHED_IDLE)) {
1509 if (!sched_feat(WAKEUP_PREEMPT))
1512 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1513 (se->avg_overlap < sysctl_sched_migration_cost &&
1514 pse->avg_overlap < sysctl_sched_migration_cost))) {
1519 find_matching_se(&se, &pse);
1523 if (wakeup_preempt_entity(se, pse) == 1)
1527 static struct task_struct *pick_next_task_fair(struct rq *rq)
1529 struct task_struct *p;
1530 struct cfs_rq *cfs_rq = &rq->cfs;
1531 struct sched_entity *se;
1533 if (unlikely(!cfs_rq->nr_running))
1537 se = pick_next_entity(cfs_rq);
1539 * If se was a buddy, clear it so that it will have to earn
1542 __clear_buddies(cfs_rq, se);
1543 set_next_entity(cfs_rq, se);
1544 cfs_rq = group_cfs_rq(se);
1548 hrtick_start_fair(rq, p);
1554 * Account for a descheduled task:
1556 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1558 struct sched_entity *se = &prev->se;
1559 struct cfs_rq *cfs_rq;
1561 for_each_sched_entity(se) {
1562 cfs_rq = cfs_rq_of(se);
1563 put_prev_entity(cfs_rq, se);
1568 /**************************************************
1569 * Fair scheduling class load-balancing methods:
1573 * Load-balancing iterator. Note: while the runqueue stays locked
1574 * during the whole iteration, the current task might be
1575 * dequeued so the iterator has to be dequeue-safe. Here we
1576 * achieve that by always pre-iterating before returning
1579 static struct task_struct *
1580 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1582 struct task_struct *p = NULL;
1583 struct sched_entity *se;
1585 if (next == &cfs_rq->tasks)
1588 se = list_entry(next, struct sched_entity, group_node);
1590 cfs_rq->balance_iterator = next->next;
1595 static struct task_struct *load_balance_start_fair(void *arg)
1597 struct cfs_rq *cfs_rq = arg;
1599 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1602 static struct task_struct *load_balance_next_fair(void *arg)
1604 struct cfs_rq *cfs_rq = arg;
1606 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1609 static unsigned long
1610 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1611 unsigned long max_load_move, struct sched_domain *sd,
1612 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1613 struct cfs_rq *cfs_rq)
1615 struct rq_iterator cfs_rq_iterator;
1617 cfs_rq_iterator.start = load_balance_start_fair;
1618 cfs_rq_iterator.next = load_balance_next_fair;
1619 cfs_rq_iterator.arg = cfs_rq;
1621 return balance_tasks(this_rq, this_cpu, busiest,
1622 max_load_move, sd, idle, all_pinned,
1623 this_best_prio, &cfs_rq_iterator);
1626 #ifdef CONFIG_FAIR_GROUP_SCHED
1627 static unsigned long
1628 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1629 unsigned long max_load_move,
1630 struct sched_domain *sd, enum cpu_idle_type idle,
1631 int *all_pinned, int *this_best_prio)
1633 long rem_load_move = max_load_move;
1634 int busiest_cpu = cpu_of(busiest);
1635 struct task_group *tg;
1638 update_h_load(busiest_cpu);
1640 list_for_each_entry_rcu(tg, &task_groups, list) {
1641 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1642 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1643 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1644 u64 rem_load, moved_load;
1649 if (!busiest_cfs_rq->task_weight)
1652 rem_load = (u64)rem_load_move * busiest_weight;
1653 rem_load = div_u64(rem_load, busiest_h_load + 1);
1655 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1656 rem_load, sd, idle, all_pinned, this_best_prio,
1657 tg->cfs_rq[busiest_cpu]);
1662 moved_load *= busiest_h_load;
1663 moved_load = div_u64(moved_load, busiest_weight + 1);
1665 rem_load_move -= moved_load;
1666 if (rem_load_move < 0)
1671 return max_load_move - rem_load_move;
1674 static unsigned long
1675 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1676 unsigned long max_load_move,
1677 struct sched_domain *sd, enum cpu_idle_type idle,
1678 int *all_pinned, int *this_best_prio)
1680 return __load_balance_fair(this_rq, this_cpu, busiest,
1681 max_load_move, sd, idle, all_pinned,
1682 this_best_prio, &busiest->cfs);
1687 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1688 struct sched_domain *sd, enum cpu_idle_type idle)
1690 struct cfs_rq *busy_cfs_rq;
1691 struct rq_iterator cfs_rq_iterator;
1693 cfs_rq_iterator.start = load_balance_start_fair;
1694 cfs_rq_iterator.next = load_balance_next_fair;
1696 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1698 * pass busy_cfs_rq argument into
1699 * load_balance_[start|next]_fair iterators
1701 cfs_rq_iterator.arg = busy_cfs_rq;
1702 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1709 #endif /* CONFIG_SMP */
1712 * scheduler tick hitting a task of our scheduling class:
1714 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1716 struct cfs_rq *cfs_rq;
1717 struct sched_entity *se = &curr->se;
1719 for_each_sched_entity(se) {
1720 cfs_rq = cfs_rq_of(se);
1721 entity_tick(cfs_rq, se, queued);
1726 * Share the fairness runtime between parent and child, thus the
1727 * total amount of pressure for CPU stays equal - new tasks
1728 * get a chance to run but frequent forkers are not allowed to
1729 * monopolize the CPU. Note: the parent runqueue is locked,
1730 * the child is not running yet.
1732 static void task_new_fair(struct rq *rq, struct task_struct *p)
1734 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1735 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1736 int this_cpu = smp_processor_id();
1738 sched_info_queued(p);
1740 update_curr(cfs_rq);
1741 place_entity(cfs_rq, se, 1);
1743 /* 'curr' will be NULL if the child belongs to a different group */
1744 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1745 curr && entity_before(curr, se)) {
1747 * Upon rescheduling, sched_class::put_prev_task() will place
1748 * 'current' within the tree based on its new key value.
1750 swap(curr->vruntime, se->vruntime);
1751 resched_task(rq->curr);
1754 enqueue_task_fair(rq, p, 0);
1758 * Priority of the task has changed. Check to see if we preempt
1761 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1762 int oldprio, int running)
1765 * Reschedule if we are currently running on this runqueue and
1766 * our priority decreased, or if we are not currently running on
1767 * this runqueue and our priority is higher than the current's
1770 if (p->prio > oldprio)
1771 resched_task(rq->curr);
1773 check_preempt_curr(rq, p, 0);
1777 * We switched to the sched_fair class.
1779 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1783 * We were most likely switched from sched_rt, so
1784 * kick off the schedule if running, otherwise just see
1785 * if we can still preempt the current task.
1788 resched_task(rq->curr);
1790 check_preempt_curr(rq, p, 0);
1793 /* Account for a task changing its policy or group.
1795 * This routine is mostly called to set cfs_rq->curr field when a task
1796 * migrates between groups/classes.
1798 static void set_curr_task_fair(struct rq *rq)
1800 struct sched_entity *se = &rq->curr->se;
1802 for_each_sched_entity(se)
1803 set_next_entity(cfs_rq_of(se), se);
1806 #ifdef CONFIG_FAIR_GROUP_SCHED
1807 static void moved_group_fair(struct task_struct *p)
1809 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1811 update_curr(cfs_rq);
1812 place_entity(cfs_rq, &p->se, 1);
1817 * All the scheduling class methods:
1819 static const struct sched_class fair_sched_class = {
1820 .next = &idle_sched_class,
1821 .enqueue_task = enqueue_task_fair,
1822 .dequeue_task = dequeue_task_fair,
1823 .yield_task = yield_task_fair,
1825 .check_preempt_curr = check_preempt_wakeup,
1827 .pick_next_task = pick_next_task_fair,
1828 .put_prev_task = put_prev_task_fair,
1831 .select_task_rq = select_task_rq_fair,
1833 .load_balance = load_balance_fair,
1834 .move_one_task = move_one_task_fair,
1837 .set_curr_task = set_curr_task_fair,
1838 .task_tick = task_tick_fair,
1839 .task_new = task_new_fair,
1841 .prio_changed = prio_changed_fair,
1842 .switched_to = switched_to_fair,
1844 #ifdef CONFIG_FAIR_GROUP_SCHED
1845 .moved_group = moved_group_fair,
1849 #ifdef CONFIG_SCHED_DEBUG
1850 static void print_cfs_stats(struct seq_file *m, int cpu)
1852 struct cfs_rq *cfs_rq;
1855 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1856 print_cfs_rq(m, cpu, cfs_rq);