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 static inline struct task_struct *task_of(struct sched_entity *se)
84 return container_of(se, struct task_struct, se);
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 /* Walk up scheduling entities hierarchy */
99 #define for_each_sched_entity(se) \
100 for (; se; se = se->parent)
102 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
107 /* runqueue on which this entity is (to be) queued */
108 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
113 /* runqueue "owned" by this group */
114 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
119 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
120 * another cpu ('this_cpu')
122 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
124 return cfs_rq->tg->cfs_rq[this_cpu];
127 /* Iterate thr' all leaf cfs_rq's on a runqueue */
128 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
129 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
131 /* Do the two (enqueued) entities belong to the same group ? */
133 is_same_group(struct sched_entity *se, struct sched_entity *pse)
135 if (se->cfs_rq == pse->cfs_rq)
141 static inline struct sched_entity *parent_entity(struct sched_entity *se)
146 /* return depth at which a sched entity is present in the hierarchy */
147 static inline int depth_se(struct sched_entity *se)
151 for_each_sched_entity(se)
158 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
160 int se_depth, pse_depth;
163 * preemption test can be made between sibling entities who are in the
164 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
165 * both tasks until we find their ancestors who are siblings of common
169 /* First walk up until both entities are at same depth */
170 se_depth = depth_se(*se);
171 pse_depth = depth_se(*pse);
173 while (se_depth > pse_depth) {
175 *se = parent_entity(*se);
178 while (pse_depth > se_depth) {
180 *pse = parent_entity(*pse);
183 while (!is_same_group(*se, *pse)) {
184 *se = parent_entity(*se);
185 *pse = parent_entity(*pse);
189 #else /* CONFIG_FAIR_GROUP_SCHED */
191 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
193 return container_of(cfs_rq, struct rq, cfs);
196 #define entity_is_task(se) 1
198 #define for_each_sched_entity(se) \
199 for (; se; se = NULL)
201 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
203 return &task_rq(p)->cfs;
206 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
208 struct task_struct *p = task_of(se);
209 struct rq *rq = task_rq(p);
214 /* runqueue "owned" by this group */
215 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
220 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
222 return &cpu_rq(this_cpu)->cfs;
225 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
226 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
229 is_same_group(struct sched_entity *se, struct sched_entity *pse)
234 static inline struct sched_entity *parent_entity(struct sched_entity *se)
240 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
244 #endif /* CONFIG_FAIR_GROUP_SCHED */
247 /**************************************************************
248 * Scheduling class tree data structure manipulation methods:
251 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
253 s64 delta = (s64)(vruntime - min_vruntime);
255 min_vruntime = vruntime;
260 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
262 s64 delta = (s64)(vruntime - min_vruntime);
264 min_vruntime = vruntime;
269 static inline int entity_before(struct sched_entity *a,
270 struct sched_entity *b)
272 return (s64)(a->vruntime - b->vruntime) < 0;
275 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
277 return se->vruntime - cfs_rq->min_vruntime;
280 static void update_min_vruntime(struct cfs_rq *cfs_rq)
282 u64 vruntime = cfs_rq->min_vruntime;
285 vruntime = cfs_rq->curr->vruntime;
287 if (cfs_rq->rb_leftmost) {
288 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
293 vruntime = se->vruntime;
295 vruntime = min_vruntime(vruntime, se->vruntime);
298 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
302 * Enqueue an entity into the rb-tree:
304 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
306 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
307 struct rb_node *parent = NULL;
308 struct sched_entity *entry;
309 s64 key = entity_key(cfs_rq, se);
313 * Find the right place in the rbtree:
317 entry = rb_entry(parent, struct sched_entity, run_node);
319 * We dont care about collisions. Nodes with
320 * the same key stay together.
322 if (key < entity_key(cfs_rq, entry)) {
323 link = &parent->rb_left;
325 link = &parent->rb_right;
331 * Maintain a cache of leftmost tree entries (it is frequently
335 cfs_rq->rb_leftmost = &se->run_node;
337 rb_link_node(&se->run_node, parent, link);
338 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
341 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
343 if (cfs_rq->rb_leftmost == &se->run_node) {
344 struct rb_node *next_node;
346 next_node = rb_next(&se->run_node);
347 cfs_rq->rb_leftmost = next_node;
350 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
353 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
355 struct rb_node *left = cfs_rq->rb_leftmost;
360 return rb_entry(left, struct sched_entity, run_node);
363 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
365 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
370 return rb_entry(last, struct sched_entity, run_node);
373 /**************************************************************
374 * Scheduling class statistics methods:
377 #ifdef CONFIG_SCHED_DEBUG
378 int sched_nr_latency_handler(struct ctl_table *table, int write,
379 struct file *filp, void __user *buffer, size_t *lenp,
382 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
387 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
388 sysctl_sched_min_granularity);
397 static inline unsigned long
398 calc_delta_fair(unsigned long delta, struct sched_entity *se)
400 if (unlikely(se->load.weight != NICE_0_LOAD))
401 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
407 * The idea is to set a period in which each task runs once.
409 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
410 * this period because otherwise the slices get too small.
412 * p = (nr <= nl) ? l : l*nr/nl
414 static u64 __sched_period(unsigned long nr_running)
416 u64 period = sysctl_sched_latency;
417 unsigned long nr_latency = sched_nr_latency;
419 if (unlikely(nr_running > nr_latency)) {
420 period = sysctl_sched_min_granularity;
421 period *= nr_running;
428 * We calculate the wall-time slice from the period by taking a part
429 * proportional to the weight.
433 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
435 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
437 for_each_sched_entity(se) {
438 struct load_weight *load;
439 struct load_weight lw;
441 cfs_rq = cfs_rq_of(se);
442 load = &cfs_rq->load;
444 if (unlikely(!se->on_rq)) {
447 update_load_add(&lw, se->load.weight);
450 slice = calc_delta_mine(slice, se->load.weight, load);
456 * We calculate the vruntime slice of a to be inserted task
460 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
462 return calc_delta_fair(sched_slice(cfs_rq, se), se);
466 * Update the current task's runtime statistics. Skip current tasks that
467 * are not in our scheduling class.
470 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
471 unsigned long delta_exec)
473 unsigned long delta_exec_weighted;
475 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
477 curr->sum_exec_runtime += delta_exec;
478 schedstat_add(cfs_rq, exec_clock, delta_exec);
479 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
480 curr->vruntime += delta_exec_weighted;
481 update_min_vruntime(cfs_rq);
484 static void update_curr(struct cfs_rq *cfs_rq)
486 struct sched_entity *curr = cfs_rq->curr;
487 u64 now = rq_of(cfs_rq)->clock;
488 unsigned long delta_exec;
494 * Get the amount of time the current task was running
495 * since the last time we changed load (this cannot
496 * overflow on 32 bits):
498 delta_exec = (unsigned long)(now - curr->exec_start);
502 __update_curr(cfs_rq, curr, delta_exec);
503 curr->exec_start = now;
505 if (entity_is_task(curr)) {
506 struct task_struct *curtask = task_of(curr);
508 cpuacct_charge(curtask, delta_exec);
509 account_group_exec_runtime(curtask, delta_exec);
514 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
516 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
520 * Task is being enqueued - update stats:
522 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 * Are we enqueueing a waiting task? (for current tasks
526 * a dequeue/enqueue event is a NOP)
528 if (se != cfs_rq->curr)
529 update_stats_wait_start(cfs_rq, se);
533 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
535 schedstat_set(se->wait_max, max(se->wait_max,
536 rq_of(cfs_rq)->clock - se->wait_start));
537 schedstat_set(se->wait_count, se->wait_count + 1);
538 schedstat_set(se->wait_sum, se->wait_sum +
539 rq_of(cfs_rq)->clock - se->wait_start);
540 schedstat_set(se->wait_start, 0);
544 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
547 * Mark the end of the wait period if dequeueing a
550 if (se != cfs_rq->curr)
551 update_stats_wait_end(cfs_rq, se);
555 * We are picking a new current task - update its stats:
558 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
561 * We are starting a new run period:
563 se->exec_start = rq_of(cfs_rq)->clock;
566 /**************************************************
567 * Scheduling class queueing methods:
570 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
572 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
574 cfs_rq->task_weight += weight;
578 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
584 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
586 update_load_add(&cfs_rq->load, se->load.weight);
587 if (!parent_entity(se))
588 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
589 if (entity_is_task(se)) {
590 add_cfs_task_weight(cfs_rq, se->load.weight);
591 list_add(&se->group_node, &cfs_rq->tasks);
593 cfs_rq->nr_running++;
598 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
600 update_load_sub(&cfs_rq->load, se->load.weight);
601 if (!parent_entity(se))
602 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
603 if (entity_is_task(se)) {
604 add_cfs_task_weight(cfs_rq, -se->load.weight);
605 list_del_init(&se->group_node);
607 cfs_rq->nr_running--;
611 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
613 #ifdef CONFIG_SCHEDSTATS
614 struct task_struct *tsk = NULL;
616 if (entity_is_task(se))
619 if (se->sleep_start) {
620 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
625 if (unlikely(delta > se->sleep_max))
626 se->sleep_max = delta;
629 se->sum_sleep_runtime += delta;
632 account_scheduler_latency(tsk, delta >> 10, 1);
634 if (se->block_start) {
635 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
640 if (unlikely(delta > se->block_max))
641 se->block_max = delta;
644 se->sum_sleep_runtime += delta;
648 * Blocking time is in units of nanosecs, so shift by
649 * 20 to get a milliseconds-range estimation of the
650 * amount of time that the task spent sleeping:
652 if (unlikely(prof_on == SLEEP_PROFILING)) {
653 profile_hits(SLEEP_PROFILING,
654 (void *)get_wchan(tsk),
657 account_scheduler_latency(tsk, delta >> 10, 0);
663 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
665 #ifdef CONFIG_SCHED_DEBUG
666 s64 d = se->vruntime - cfs_rq->min_vruntime;
671 if (d > 3*sysctl_sched_latency)
672 schedstat_inc(cfs_rq, nr_spread_over);
677 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
679 u64 vruntime = cfs_rq->min_vruntime;
682 * The 'current' period is already promised to the current tasks,
683 * however the extra weight of the new task will slow them down a
684 * little, place the new task so that it fits in the slot that
685 * stays open at the end.
687 if (initial && sched_feat(START_DEBIT))
688 vruntime += sched_vslice(cfs_rq, se);
691 /* sleeps upto a single latency don't count. */
692 if (sched_feat(NEW_FAIR_SLEEPERS)) {
693 unsigned long thresh = sysctl_sched_latency;
696 * Convert the sleeper threshold into virtual time.
697 * SCHED_IDLE is a special sub-class. We care about
698 * fairness only relative to other SCHED_IDLE tasks,
699 * all of which have the same weight.
701 if (sched_feat(NORMALIZED_SLEEPER) &&
702 (!entity_is_task(se) ||
703 task_of(se)->policy != SCHED_IDLE))
704 thresh = calc_delta_fair(thresh, se);
709 /* ensure we never gain time by being placed backwards. */
710 vruntime = max_vruntime(se->vruntime, vruntime);
713 se->vruntime = vruntime;
717 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
720 * Update run-time statistics of the 'current'.
723 account_entity_enqueue(cfs_rq, se);
726 place_entity(cfs_rq, se, 0);
727 enqueue_sleeper(cfs_rq, se);
730 update_stats_enqueue(cfs_rq, se);
731 check_spread(cfs_rq, se);
732 if (se != cfs_rq->curr)
733 __enqueue_entity(cfs_rq, se);
736 static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
738 if (cfs_rq->last == se)
741 if (cfs_rq->next == se)
745 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
747 for_each_sched_entity(se)
748 __clear_buddies(cfs_rq_of(se), se);
752 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
755 * Update run-time statistics of the 'current'.
759 update_stats_dequeue(cfs_rq, se);
761 #ifdef CONFIG_SCHEDSTATS
762 if (entity_is_task(se)) {
763 struct task_struct *tsk = task_of(se);
765 if (tsk->state & TASK_INTERRUPTIBLE)
766 se->sleep_start = rq_of(cfs_rq)->clock;
767 if (tsk->state & TASK_UNINTERRUPTIBLE)
768 se->block_start = rq_of(cfs_rq)->clock;
773 clear_buddies(cfs_rq, se);
775 if (se != cfs_rq->curr)
776 __dequeue_entity(cfs_rq, se);
777 account_entity_dequeue(cfs_rq, se);
778 update_min_vruntime(cfs_rq);
782 * Preempt the current task with a newly woken task if needed:
785 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
787 unsigned long ideal_runtime, delta_exec;
789 ideal_runtime = sched_slice(cfs_rq, curr);
790 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
791 if (delta_exec > ideal_runtime) {
792 resched_task(rq_of(cfs_rq)->curr);
794 * The current task ran long enough, ensure it doesn't get
795 * re-elected due to buddy favours.
797 clear_buddies(cfs_rq, curr);
802 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
804 /* 'current' is not kept within the tree. */
807 * Any task has to be enqueued before it get to execute on
808 * a CPU. So account for the time it spent waiting on the
811 update_stats_wait_end(cfs_rq, se);
812 __dequeue_entity(cfs_rq, se);
815 update_stats_curr_start(cfs_rq, se);
817 #ifdef CONFIG_SCHEDSTATS
819 * Track our maximum slice length, if the CPU's load is at
820 * least twice that of our own weight (i.e. dont track it
821 * when there are only lesser-weight tasks around):
823 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
824 se->slice_max = max(se->slice_max,
825 se->sum_exec_runtime - se->prev_sum_exec_runtime);
828 se->prev_sum_exec_runtime = se->sum_exec_runtime;
832 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
834 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
836 struct sched_entity *se = __pick_next_entity(cfs_rq);
838 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
841 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
847 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
850 * If still on the runqueue then deactivate_task()
851 * was not called and update_curr() has to be done:
856 check_spread(cfs_rq, prev);
858 update_stats_wait_start(cfs_rq, prev);
859 /* Put 'current' back into the tree. */
860 __enqueue_entity(cfs_rq, prev);
866 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
869 * Update run-time statistics of the 'current'.
873 #ifdef CONFIG_SCHED_HRTICK
875 * queued ticks are scheduled to match the slice, so don't bother
876 * validating it and just reschedule.
879 resched_task(rq_of(cfs_rq)->curr);
883 * don't let the period tick interfere with the hrtick preemption
885 if (!sched_feat(DOUBLE_TICK) &&
886 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
890 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
891 check_preempt_tick(cfs_rq, curr);
894 /**************************************************
895 * CFS operations on tasks:
898 #ifdef CONFIG_SCHED_HRTICK
899 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
901 struct sched_entity *se = &p->se;
902 struct cfs_rq *cfs_rq = cfs_rq_of(se);
904 WARN_ON(task_rq(p) != rq);
906 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
907 u64 slice = sched_slice(cfs_rq, se);
908 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
909 s64 delta = slice - ran;
918 * Don't schedule slices shorter than 10000ns, that just
919 * doesn't make sense. Rely on vruntime for fairness.
922 delta = max_t(s64, 10000LL, delta);
924 hrtick_start(rq, delta);
929 * called from enqueue/dequeue and updates the hrtick when the
930 * current task is from our class and nr_running is low enough
933 static void hrtick_update(struct rq *rq)
935 struct task_struct *curr = rq->curr;
937 if (curr->sched_class != &fair_sched_class)
940 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
941 hrtick_start_fair(rq, curr);
943 #else /* !CONFIG_SCHED_HRTICK */
945 hrtick_start_fair(struct rq *rq, struct task_struct *p)
949 static inline void hrtick_update(struct rq *rq)
955 * The enqueue_task method is called before nr_running is
956 * increased. Here we update the fair scheduling stats and
957 * then put the task into the rbtree:
959 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
961 struct cfs_rq *cfs_rq;
962 struct sched_entity *se = &p->se;
964 for_each_sched_entity(se) {
967 cfs_rq = cfs_rq_of(se);
968 enqueue_entity(cfs_rq, se, wakeup);
976 * The dequeue_task method is called before nr_running is
977 * decreased. We remove the task from the rbtree and
978 * update the fair scheduling stats:
980 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
982 struct cfs_rq *cfs_rq;
983 struct sched_entity *se = &p->se;
985 for_each_sched_entity(se) {
986 cfs_rq = cfs_rq_of(se);
987 dequeue_entity(cfs_rq, se, sleep);
988 /* Don't dequeue parent if it has other entities besides us */
989 if (cfs_rq->load.weight)
998 * sched_yield() support is very simple - we dequeue and enqueue.
1000 * If compat_yield is turned on then we requeue to the end of the tree.
1002 static void yield_task_fair(struct rq *rq)
1004 struct task_struct *curr = rq->curr;
1005 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1006 struct sched_entity *rightmost, *se = &curr->se;
1009 * Are we the only task in the tree?
1011 if (unlikely(cfs_rq->nr_running == 1))
1014 clear_buddies(cfs_rq, se);
1016 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
1017 update_rq_clock(rq);
1019 * Update run-time statistics of the 'current'.
1021 update_curr(cfs_rq);
1026 * Find the rightmost entry in the rbtree:
1028 rightmost = __pick_last_entity(cfs_rq);
1030 * Already in the rightmost position?
1032 if (unlikely(!rightmost || entity_before(rightmost, se)))
1036 * Minimally necessary key value to be last in the tree:
1037 * Upon rescheduling, sched_class::put_prev_task() will place
1038 * 'current' within the tree based on its new key value.
1040 se->vruntime = rightmost->vruntime + 1;
1044 * wake_idle() will wake a task on an idle cpu if task->cpu is
1045 * not idle and an idle cpu is available. The span of cpus to
1046 * search starts with cpus closest then further out as needed,
1047 * so we always favor a closer, idle cpu.
1048 * Domains may include CPUs that are not usable for migration,
1049 * hence we need to mask them out (rq->rd->online)
1051 * Returns the CPU we should wake onto.
1053 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1055 #define cpu_rd_active(cpu, rq) cpumask_test_cpu(cpu, rq->rd->online)
1057 static int wake_idle(int cpu, struct task_struct *p)
1059 struct sched_domain *sd;
1061 unsigned int chosen_wakeup_cpu;
1063 struct rq *task_rq = task_rq(p);
1066 * At POWERSAVINGS_BALANCE_WAKEUP level, if both this_cpu and prev_cpu
1067 * are idle and this is not a kernel thread and this task's affinity
1068 * allows it to be moved to preferred cpu, then just move!
1071 this_cpu = smp_processor_id();
1073 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu;
1075 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP &&
1076 idle_cpu(cpu) && idle_cpu(this_cpu) &&
1077 p->mm && !(p->flags & PF_KTHREAD) &&
1078 cpu_isset(chosen_wakeup_cpu, p->cpus_allowed))
1079 return chosen_wakeup_cpu;
1082 * If it is idle, then it is the best cpu to run this task.
1084 * This cpu is also the best, if it has more than one task already.
1085 * Siblings must be also busy(in most cases) as they didn't already
1086 * pickup the extra load from this cpu and hence we need not check
1087 * sibling runqueue info. This will avoid the checks and cache miss
1088 * penalities associated with that.
1090 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
1093 for_each_domain(cpu, sd) {
1094 if ((sd->flags & SD_WAKE_IDLE)
1095 || ((sd->flags & SD_WAKE_IDLE_FAR)
1096 && !task_hot(p, task_rq->clock, sd))) {
1097 for_each_cpu_and(i, sched_domain_span(sd),
1099 if (cpu_rd_active(i, task_rq) && idle_cpu(i)) {
1100 if (i != task_cpu(p)) {
1102 se.nr_wakeups_idle);
1113 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1114 static inline int wake_idle(int cpu, struct task_struct *p)
1122 #ifdef CONFIG_FAIR_GROUP_SCHED
1124 * effective_load() calculates the load change as seen from the root_task_group
1126 * Adding load to a group doesn't make a group heavier, but can cause movement
1127 * of group shares between cpus. Assuming the shares were perfectly aligned one
1128 * can calculate the shift in shares.
1130 * The problem is that perfectly aligning the shares is rather expensive, hence
1131 * we try to avoid doing that too often - see update_shares(), which ratelimits
1134 * We compensate this by not only taking the current delta into account, but
1135 * also considering the delta between when the shares were last adjusted and
1138 * We still saw a performance dip, some tracing learned us that between
1139 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1140 * significantly. Therefore try to bias the error in direction of failing
1141 * the affine wakeup.
1144 static long effective_load(struct task_group *tg, int cpu,
1147 struct sched_entity *se = tg->se[cpu];
1153 * By not taking the decrease of shares on the other cpu into
1154 * account our error leans towards reducing the affine wakeups.
1156 if (!wl && sched_feat(ASYM_EFF_LOAD))
1159 for_each_sched_entity(se) {
1160 long S, rw, s, a, b;
1164 * Instead of using this increment, also add the difference
1165 * between when the shares were last updated and now.
1167 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1171 S = se->my_q->tg->shares;
1172 s = se->my_q->shares;
1173 rw = se->my_q->rq_weight;
1184 * Assume the group is already running and will
1185 * thus already be accounted for in the weight.
1187 * That is, moving shares between CPUs, does not
1188 * alter the group weight.
1198 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1199 unsigned long wl, unsigned long wg)
1207 wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
1208 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1209 int idx, unsigned long load, unsigned long this_load,
1210 unsigned int imbalance)
1212 struct task_struct *curr = this_rq->curr;
1213 struct task_group *tg;
1214 unsigned long tl = this_load;
1215 unsigned long tl_per_task;
1216 unsigned long weight;
1219 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1222 if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1223 p->se.avg_overlap > sysctl_sched_migration_cost))
1227 * If sync wakeup then subtract the (maximum possible)
1228 * effect of the currently running task from the load
1229 * of the current CPU:
1232 tg = task_group(current);
1233 weight = current->se.load.weight;
1235 tl += effective_load(tg, this_cpu, -weight, -weight);
1236 load += effective_load(tg, prev_cpu, 0, -weight);
1240 weight = p->se.load.weight;
1242 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1243 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1246 * If the currently running task will sleep within
1247 * a reasonable amount of time then attract this newly
1250 if (sync && balanced)
1253 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1254 tl_per_task = cpu_avg_load_per_task(this_cpu);
1256 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1259 * This domain has SD_WAKE_AFFINE and
1260 * p is cache cold in this domain, and
1261 * there is no bad imbalance.
1263 schedstat_inc(this_sd, ttwu_move_affine);
1264 schedstat_inc(p, se.nr_wakeups_affine);
1271 static int select_task_rq_fair(struct task_struct *p, int sync)
1273 struct sched_domain *sd, *this_sd = NULL;
1274 int prev_cpu, this_cpu, new_cpu;
1275 unsigned long load, this_load;
1277 unsigned int imbalance;
1280 prev_cpu = task_cpu(p);
1281 this_cpu = smp_processor_id();
1282 this_rq = cpu_rq(this_cpu);
1285 if (prev_cpu == this_cpu)
1288 * 'this_sd' is the first domain that both
1289 * this_cpu and prev_cpu are present in:
1291 for_each_domain(this_cpu, sd) {
1292 if (cpumask_test_cpu(prev_cpu, sched_domain_span(sd))) {
1298 if (unlikely(!cpumask_test_cpu(this_cpu, &p->cpus_allowed)))
1302 * Check for affine wakeup and passive balancing possibilities.
1307 idx = this_sd->wake_idx;
1309 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1311 load = source_load(prev_cpu, idx);
1312 this_load = target_load(this_cpu, idx);
1314 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1315 load, this_load, imbalance))
1319 * Start passive balancing when half the imbalance_pct
1322 if (this_sd->flags & SD_WAKE_BALANCE) {
1323 if (imbalance*this_load <= 100*load) {
1324 schedstat_inc(this_sd, ttwu_move_balance);
1325 schedstat_inc(p, se.nr_wakeups_passive);
1331 return wake_idle(new_cpu, p);
1333 #endif /* CONFIG_SMP */
1336 * Adaptive granularity
1338 * se->avg_wakeup gives the average time a task runs until it does a wakeup,
1339 * with the limit of wakeup_gran -- when it never does a wakeup.
1341 * So the smaller avg_wakeup is the faster we want this task to preempt,
1342 * but we don't want to treat the preemptee unfairly and therefore allow it
1343 * to run for at least the amount of time we'd like to run.
1345 * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
1347 * NOTE: we use *nr_running to scale with load, this nicely matches the
1348 * degrading latency on load.
1350 static unsigned long
1351 adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
1353 u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1354 u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
1357 if (this_run < expected_wakeup)
1358 gran = expected_wakeup - this_run;
1360 return min_t(s64, gran, sysctl_sched_wakeup_granularity);
1363 static unsigned long
1364 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
1366 unsigned long gran = sysctl_sched_wakeup_granularity;
1368 if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
1369 gran = adaptive_gran(curr, se);
1372 * Since its curr running now, convert the gran from real-time
1373 * to virtual-time in his units.
1375 if (sched_feat(ASYM_GRAN)) {
1377 * By using 'se' instead of 'curr' we penalize light tasks, so
1378 * they get preempted easier. That is, if 'se' < 'curr' then
1379 * the resulting gran will be larger, therefore penalizing the
1380 * lighter, if otoh 'se' > 'curr' then the resulting gran will
1381 * be smaller, again penalizing the lighter task.
1383 * This is especially important for buddies when the leftmost
1384 * task is higher priority than the buddy.
1386 if (unlikely(se->load.weight != NICE_0_LOAD))
1387 gran = calc_delta_fair(gran, se);
1389 if (unlikely(curr->load.weight != NICE_0_LOAD))
1390 gran = calc_delta_fair(gran, curr);
1397 * Should 'se' preempt 'curr'.
1411 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1413 s64 gran, vdiff = curr->vruntime - se->vruntime;
1418 gran = wakeup_gran(curr, se);
1425 static void set_last_buddy(struct sched_entity *se)
1427 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1428 for_each_sched_entity(se)
1429 cfs_rq_of(se)->last = se;
1433 static void set_next_buddy(struct sched_entity *se)
1435 if (likely(task_of(se)->policy != SCHED_IDLE)) {
1436 for_each_sched_entity(se)
1437 cfs_rq_of(se)->next = se;
1442 * Preempt the current task with a newly woken task if needed:
1444 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
1446 struct task_struct *curr = rq->curr;
1447 struct sched_entity *se = &curr->se, *pse = &p->se;
1448 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1450 update_curr(cfs_rq);
1452 if (unlikely(rt_prio(p->prio))) {
1457 if (unlikely(p->sched_class != &fair_sched_class))
1460 if (unlikely(se == pse))
1464 * Only set the backward buddy when the current task is still on the
1465 * rq. This can happen when a wakeup gets interleaved with schedule on
1466 * the ->pre_schedule() or idle_balance() point, either of which can
1469 * Also, during early boot the idle thread is in the fair class, for
1470 * obvious reasons its a bad idea to schedule back to the idle thread.
1472 if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
1474 set_next_buddy(pse);
1477 * We can come here with TIF_NEED_RESCHED already set from new task
1480 if (test_tsk_need_resched(curr))
1484 * Batch and idle tasks do not preempt (their preemption is driven by
1487 if (unlikely(p->policy != SCHED_NORMAL))
1490 /* Idle tasks are by definition preempted by everybody. */
1491 if (unlikely(curr->policy == SCHED_IDLE)) {
1496 if (!sched_feat(WAKEUP_PREEMPT))
1499 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1500 (se->avg_overlap < sysctl_sched_migration_cost &&
1501 pse->avg_overlap < sysctl_sched_migration_cost))) {
1506 find_matching_se(&se, &pse);
1510 if (wakeup_preempt_entity(se, pse) == 1)
1514 static struct task_struct *pick_next_task_fair(struct rq *rq)
1516 struct task_struct *p;
1517 struct cfs_rq *cfs_rq = &rq->cfs;
1518 struct sched_entity *se;
1520 if (unlikely(!cfs_rq->nr_running))
1524 se = pick_next_entity(cfs_rq);
1526 * If se was a buddy, clear it so that it will have to earn
1529 __clear_buddies(cfs_rq, se);
1530 set_next_entity(cfs_rq, se);
1531 cfs_rq = group_cfs_rq(se);
1535 hrtick_start_fair(rq, p);
1541 * Account for a descheduled task:
1543 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1545 struct sched_entity *se = &prev->se;
1546 struct cfs_rq *cfs_rq;
1548 for_each_sched_entity(se) {
1549 cfs_rq = cfs_rq_of(se);
1550 put_prev_entity(cfs_rq, se);
1555 /**************************************************
1556 * Fair scheduling class load-balancing methods:
1560 * Load-balancing iterator. Note: while the runqueue stays locked
1561 * during the whole iteration, the current task might be
1562 * dequeued so the iterator has to be dequeue-safe. Here we
1563 * achieve that by always pre-iterating before returning
1566 static struct task_struct *
1567 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1569 struct task_struct *p = NULL;
1570 struct sched_entity *se;
1572 if (next == &cfs_rq->tasks)
1575 se = list_entry(next, struct sched_entity, group_node);
1577 cfs_rq->balance_iterator = next->next;
1582 static struct task_struct *load_balance_start_fair(void *arg)
1584 struct cfs_rq *cfs_rq = arg;
1586 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1589 static struct task_struct *load_balance_next_fair(void *arg)
1591 struct cfs_rq *cfs_rq = arg;
1593 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1596 static unsigned long
1597 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1598 unsigned long max_load_move, struct sched_domain *sd,
1599 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1600 struct cfs_rq *cfs_rq)
1602 struct rq_iterator cfs_rq_iterator;
1604 cfs_rq_iterator.start = load_balance_start_fair;
1605 cfs_rq_iterator.next = load_balance_next_fair;
1606 cfs_rq_iterator.arg = cfs_rq;
1608 return balance_tasks(this_rq, this_cpu, busiest,
1609 max_load_move, sd, idle, all_pinned,
1610 this_best_prio, &cfs_rq_iterator);
1613 #ifdef CONFIG_FAIR_GROUP_SCHED
1614 static unsigned long
1615 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1616 unsigned long max_load_move,
1617 struct sched_domain *sd, enum cpu_idle_type idle,
1618 int *all_pinned, int *this_best_prio)
1620 long rem_load_move = max_load_move;
1621 int busiest_cpu = cpu_of(busiest);
1622 struct task_group *tg;
1625 update_h_load(busiest_cpu);
1627 list_for_each_entry_rcu(tg, &task_groups, list) {
1628 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1629 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1630 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1631 u64 rem_load, moved_load;
1636 if (!busiest_cfs_rq->task_weight)
1639 rem_load = (u64)rem_load_move * busiest_weight;
1640 rem_load = div_u64(rem_load, busiest_h_load + 1);
1642 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1643 rem_load, sd, idle, all_pinned, this_best_prio,
1644 tg->cfs_rq[busiest_cpu]);
1649 moved_load *= busiest_h_load;
1650 moved_load = div_u64(moved_load, busiest_weight + 1);
1652 rem_load_move -= moved_load;
1653 if (rem_load_move < 0)
1658 return max_load_move - rem_load_move;
1661 static unsigned long
1662 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1663 unsigned long max_load_move,
1664 struct sched_domain *sd, enum cpu_idle_type idle,
1665 int *all_pinned, int *this_best_prio)
1667 return __load_balance_fair(this_rq, this_cpu, busiest,
1668 max_load_move, sd, idle, all_pinned,
1669 this_best_prio, &busiest->cfs);
1674 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1675 struct sched_domain *sd, enum cpu_idle_type idle)
1677 struct cfs_rq *busy_cfs_rq;
1678 struct rq_iterator cfs_rq_iterator;
1680 cfs_rq_iterator.start = load_balance_start_fair;
1681 cfs_rq_iterator.next = load_balance_next_fair;
1683 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1685 * pass busy_cfs_rq argument into
1686 * load_balance_[start|next]_fair iterators
1688 cfs_rq_iterator.arg = busy_cfs_rq;
1689 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1696 #endif /* CONFIG_SMP */
1699 * scheduler tick hitting a task of our scheduling class:
1701 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1703 struct cfs_rq *cfs_rq;
1704 struct sched_entity *se = &curr->se;
1706 for_each_sched_entity(se) {
1707 cfs_rq = cfs_rq_of(se);
1708 entity_tick(cfs_rq, se, queued);
1713 * Share the fairness runtime between parent and child, thus the
1714 * total amount of pressure for CPU stays equal - new tasks
1715 * get a chance to run but frequent forkers are not allowed to
1716 * monopolize the CPU. Note: the parent runqueue is locked,
1717 * the child is not running yet.
1719 static void task_new_fair(struct rq *rq, struct task_struct *p)
1721 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1722 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1723 int this_cpu = smp_processor_id();
1725 sched_info_queued(p);
1727 update_curr(cfs_rq);
1728 place_entity(cfs_rq, se, 1);
1730 /* 'curr' will be NULL if the child belongs to a different group */
1731 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1732 curr && entity_before(curr, se)) {
1734 * Upon rescheduling, sched_class::put_prev_task() will place
1735 * 'current' within the tree based on its new key value.
1737 swap(curr->vruntime, se->vruntime);
1738 resched_task(rq->curr);
1741 enqueue_task_fair(rq, p, 0);
1745 * Priority of the task has changed. Check to see if we preempt
1748 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1749 int oldprio, int running)
1752 * Reschedule if we are currently running on this runqueue and
1753 * our priority decreased, or if we are not currently running on
1754 * this runqueue and our priority is higher than the current's
1757 if (p->prio > oldprio)
1758 resched_task(rq->curr);
1760 check_preempt_curr(rq, p, 0);
1764 * We switched to the sched_fair class.
1766 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1770 * We were most likely switched from sched_rt, so
1771 * kick off the schedule if running, otherwise just see
1772 * if we can still preempt the current task.
1775 resched_task(rq->curr);
1777 check_preempt_curr(rq, p, 0);
1780 /* Account for a task changing its policy or group.
1782 * This routine is mostly called to set cfs_rq->curr field when a task
1783 * migrates between groups/classes.
1785 static void set_curr_task_fair(struct rq *rq)
1787 struct sched_entity *se = &rq->curr->se;
1789 for_each_sched_entity(se)
1790 set_next_entity(cfs_rq_of(se), se);
1793 #ifdef CONFIG_FAIR_GROUP_SCHED
1794 static void moved_group_fair(struct task_struct *p)
1796 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1798 update_curr(cfs_rq);
1799 place_entity(cfs_rq, &p->se, 1);
1804 * All the scheduling class methods:
1806 static const struct sched_class fair_sched_class = {
1807 .next = &idle_sched_class,
1808 .enqueue_task = enqueue_task_fair,
1809 .dequeue_task = dequeue_task_fair,
1810 .yield_task = yield_task_fair,
1812 .check_preempt_curr = check_preempt_wakeup,
1814 .pick_next_task = pick_next_task_fair,
1815 .put_prev_task = put_prev_task_fair,
1818 .select_task_rq = select_task_rq_fair,
1820 .load_balance = load_balance_fair,
1821 .move_one_task = move_one_task_fair,
1824 .set_curr_task = set_curr_task_fair,
1825 .task_tick = task_tick_fair,
1826 .task_new = task_new_fair,
1828 .prio_changed = prio_changed_fair,
1829 .switched_to = switched_to_fair,
1831 #ifdef CONFIG_FAIR_GROUP_SCHED
1832 .moved_group = moved_group_fair,
1836 #ifdef CONFIG_SCHED_DEBUG
1837 static void print_cfs_stats(struct seq_file *m, int cpu)
1839 struct cfs_rq *cfs_rq;
1842 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1843 print_cfs_rq(m, cpu, cfs_rq);