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 s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
271 return se->vruntime - cfs_rq->min_vruntime;
274 static void update_min_vruntime(struct cfs_rq *cfs_rq)
276 u64 vruntime = cfs_rq->min_vruntime;
279 vruntime = cfs_rq->curr->vruntime;
281 if (cfs_rq->rb_leftmost) {
282 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
286 if (vruntime == cfs_rq->min_vruntime)
287 vruntime = se->vruntime;
289 vruntime = min_vruntime(vruntime, se->vruntime);
292 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
296 * Enqueue an entity into the rb-tree:
298 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
300 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
301 struct rb_node *parent = NULL;
302 struct sched_entity *entry;
303 s64 key = entity_key(cfs_rq, se);
307 * Find the right place in the rbtree:
311 entry = rb_entry(parent, struct sched_entity, run_node);
313 * We dont care about collisions. Nodes with
314 * the same key stay together.
316 if (key < entity_key(cfs_rq, entry)) {
317 link = &parent->rb_left;
319 link = &parent->rb_right;
325 * Maintain a cache of leftmost tree entries (it is frequently
329 cfs_rq->rb_leftmost = &se->run_node;
331 rb_link_node(&se->run_node, parent, link);
332 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
335 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
337 if (cfs_rq->rb_leftmost == &se->run_node) {
338 struct rb_node *next_node;
340 next_node = rb_next(&se->run_node);
341 cfs_rq->rb_leftmost = next_node;
344 if (cfs_rq->next == se)
347 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
350 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
352 struct rb_node *left = cfs_rq->rb_leftmost;
357 return rb_entry(left, struct sched_entity, run_node);
360 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
362 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
367 return rb_entry(last, struct sched_entity, run_node);
370 /**************************************************************
371 * Scheduling class statistics methods:
374 #ifdef CONFIG_SCHED_DEBUG
375 int sched_nr_latency_handler(struct ctl_table *table, int write,
376 struct file *filp, void __user *buffer, size_t *lenp,
379 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
384 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
385 sysctl_sched_min_granularity);
394 static inline unsigned long
395 calc_delta_weight(unsigned long delta, struct sched_entity *se)
397 for_each_sched_entity(se) {
398 delta = calc_delta_mine(delta,
399 se->load.weight, &cfs_rq_of(se)->load);
408 static inline unsigned long
409 calc_delta_fair(unsigned long delta, struct sched_entity *se)
411 if (unlikely(se->load.weight != NICE_0_LOAD))
412 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
418 * The idea is to set a period in which each task runs once.
420 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
421 * this period because otherwise the slices get too small.
423 * p = (nr <= nl) ? l : l*nr/nl
425 static u64 __sched_period(unsigned long nr_running)
427 u64 period = sysctl_sched_latency;
428 unsigned long nr_latency = sched_nr_latency;
430 if (unlikely(nr_running > nr_latency)) {
431 period = sysctl_sched_min_granularity;
432 period *= nr_running;
439 * We calculate the wall-time slice from the period by taking a part
440 * proportional to the weight.
444 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
446 unsigned long nr_running = cfs_rq->nr_running;
448 if (unlikely(!se->on_rq))
451 return calc_delta_weight(__sched_period(nr_running), se);
455 * We calculate the vruntime slice of a to be inserted task
459 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
461 return calc_delta_fair(sched_slice(cfs_rq, se), se);
465 * Update the current task's runtime statistics. Skip current tasks that
466 * are not in our scheduling class.
469 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
470 unsigned long delta_exec)
472 unsigned long delta_exec_weighted;
474 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
476 curr->sum_exec_runtime += delta_exec;
477 schedstat_add(cfs_rq, exec_clock, delta_exec);
478 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
479 curr->vruntime += delta_exec_weighted;
480 update_min_vruntime(cfs_rq);
483 static void update_curr(struct cfs_rq *cfs_rq)
485 struct sched_entity *curr = cfs_rq->curr;
486 u64 now = rq_of(cfs_rq)->clock;
487 unsigned long delta_exec;
493 * Get the amount of time the current task was running
494 * since the last time we changed load (this cannot
495 * overflow on 32 bits):
497 delta_exec = (unsigned long)(now - curr->exec_start);
499 __update_curr(cfs_rq, curr, delta_exec);
500 curr->exec_start = now;
502 if (entity_is_task(curr)) {
503 struct task_struct *curtask = task_of(curr);
505 cpuacct_charge(curtask, delta_exec);
506 account_group_exec_runtime(curtask, delta_exec);
511 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
513 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
517 * Task is being enqueued - update stats:
519 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
522 * Are we enqueueing a waiting task? (for current tasks
523 * a dequeue/enqueue event is a NOP)
525 if (se != cfs_rq->curr)
526 update_stats_wait_start(cfs_rq, se);
530 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
532 schedstat_set(se->wait_max, max(se->wait_max,
533 rq_of(cfs_rq)->clock - se->wait_start));
534 schedstat_set(se->wait_count, se->wait_count + 1);
535 schedstat_set(se->wait_sum, se->wait_sum +
536 rq_of(cfs_rq)->clock - se->wait_start);
537 schedstat_set(se->wait_start, 0);
541 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
544 * Mark the end of the wait period if dequeueing a
547 if (se != cfs_rq->curr)
548 update_stats_wait_end(cfs_rq, se);
552 * We are picking a new current task - update its stats:
555 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
558 * We are starting a new run period:
560 se->exec_start = rq_of(cfs_rq)->clock;
563 /**************************************************
564 * Scheduling class queueing methods:
567 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
569 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
571 cfs_rq->task_weight += weight;
575 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
581 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
583 update_load_add(&cfs_rq->load, se->load.weight);
584 if (!parent_entity(se))
585 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
586 if (entity_is_task(se)) {
587 add_cfs_task_weight(cfs_rq, se->load.weight);
588 list_add(&se->group_node, &cfs_rq->tasks);
590 cfs_rq->nr_running++;
595 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
597 update_load_sub(&cfs_rq->load, se->load.weight);
598 if (!parent_entity(se))
599 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
600 if (entity_is_task(se)) {
601 add_cfs_task_weight(cfs_rq, -se->load.weight);
602 list_del_init(&se->group_node);
604 cfs_rq->nr_running--;
608 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
610 #ifdef CONFIG_SCHEDSTATS
611 if (se->sleep_start) {
612 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
613 struct task_struct *tsk = task_of(se);
618 if (unlikely(delta > se->sleep_max))
619 se->sleep_max = delta;
622 se->sum_sleep_runtime += delta;
624 account_scheduler_latency(tsk, delta >> 10, 1);
626 if (se->block_start) {
627 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
628 struct task_struct *tsk = task_of(se);
633 if (unlikely(delta > se->block_max))
634 se->block_max = delta;
637 se->sum_sleep_runtime += delta;
640 * Blocking time is in units of nanosecs, so shift by 20 to
641 * get a milliseconds-range estimation of the amount of
642 * time that the task spent sleeping:
644 if (unlikely(prof_on == SLEEP_PROFILING)) {
646 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
649 account_scheduler_latency(tsk, delta >> 10, 0);
654 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
656 #ifdef CONFIG_SCHED_DEBUG
657 s64 d = se->vruntime - cfs_rq->min_vruntime;
662 if (d > 3*sysctl_sched_latency)
663 schedstat_inc(cfs_rq, nr_spread_over);
668 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
670 u64 vruntime = cfs_rq->min_vruntime;
673 * The 'current' period is already promised to the current tasks,
674 * however the extra weight of the new task will slow them down a
675 * little, place the new task so that it fits in the slot that
676 * stays open at the end.
678 if (initial && sched_feat(START_DEBIT))
679 vruntime += sched_vslice(cfs_rq, se);
682 /* sleeps upto a single latency don't count. */
683 if (sched_feat(NEW_FAIR_SLEEPERS)) {
684 unsigned long thresh = sysctl_sched_latency;
687 * convert the sleeper threshold into virtual time
689 if (sched_feat(NORMALIZED_SLEEPER))
690 thresh = calc_delta_fair(thresh, se);
695 /* ensure we never gain time by being placed backwards. */
696 vruntime = max_vruntime(se->vruntime, vruntime);
699 se->vruntime = vruntime;
703 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
706 * Update run-time statistics of the 'current'.
709 account_entity_enqueue(cfs_rq, se);
712 place_entity(cfs_rq, se, 0);
713 enqueue_sleeper(cfs_rq, se);
716 update_stats_enqueue(cfs_rq, se);
717 check_spread(cfs_rq, se);
718 if (se != cfs_rq->curr)
719 __enqueue_entity(cfs_rq, se);
723 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
726 * Update run-time statistics of the 'current'.
730 update_stats_dequeue(cfs_rq, se);
732 #ifdef CONFIG_SCHEDSTATS
733 if (entity_is_task(se)) {
734 struct task_struct *tsk = task_of(se);
736 if (tsk->state & TASK_INTERRUPTIBLE)
737 se->sleep_start = rq_of(cfs_rq)->clock;
738 if (tsk->state & TASK_UNINTERRUPTIBLE)
739 se->block_start = rq_of(cfs_rq)->clock;
744 if (se != cfs_rq->curr)
745 __dequeue_entity(cfs_rq, se);
746 account_entity_dequeue(cfs_rq, se);
747 update_min_vruntime(cfs_rq);
751 * Preempt the current task with a newly woken task if needed:
754 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
756 unsigned long ideal_runtime, delta_exec;
758 ideal_runtime = sched_slice(cfs_rq, curr);
759 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
760 if (delta_exec > ideal_runtime)
761 resched_task(rq_of(cfs_rq)->curr);
765 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
767 /* 'current' is not kept within the tree. */
770 * Any task has to be enqueued before it get to execute on
771 * a CPU. So account for the time it spent waiting on the
774 update_stats_wait_end(cfs_rq, se);
775 __dequeue_entity(cfs_rq, se);
778 update_stats_curr_start(cfs_rq, se);
780 #ifdef CONFIG_SCHEDSTATS
782 * Track our maximum slice length, if the CPU's load is at
783 * least twice that of our own weight (i.e. dont track it
784 * when there are only lesser-weight tasks around):
786 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
787 se->slice_max = max(se->slice_max,
788 se->sum_exec_runtime - se->prev_sum_exec_runtime);
791 se->prev_sum_exec_runtime = se->sum_exec_runtime;
795 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
797 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
799 struct sched_entity *se = __pick_next_entity(cfs_rq);
801 if (!cfs_rq->next || wakeup_preempt_entity(cfs_rq->next, se) == 1)
807 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
810 * If still on the runqueue then deactivate_task()
811 * was not called and update_curr() has to be done:
816 check_spread(cfs_rq, prev);
818 update_stats_wait_start(cfs_rq, prev);
819 /* Put 'current' back into the tree. */
820 __enqueue_entity(cfs_rq, prev);
826 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
829 * Update run-time statistics of the 'current'.
833 #ifdef CONFIG_SCHED_HRTICK
835 * queued ticks are scheduled to match the slice, so don't bother
836 * validating it and just reschedule.
839 resched_task(rq_of(cfs_rq)->curr);
843 * don't let the period tick interfere with the hrtick preemption
845 if (!sched_feat(DOUBLE_TICK) &&
846 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
850 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
851 check_preempt_tick(cfs_rq, curr);
854 /**************************************************
855 * CFS operations on tasks:
858 #ifdef CONFIG_SCHED_HRTICK
859 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
861 struct sched_entity *se = &p->se;
862 struct cfs_rq *cfs_rq = cfs_rq_of(se);
864 WARN_ON(task_rq(p) != rq);
866 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
867 u64 slice = sched_slice(cfs_rq, se);
868 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
869 s64 delta = slice - ran;
878 * Don't schedule slices shorter than 10000ns, that just
879 * doesn't make sense. Rely on vruntime for fairness.
882 delta = max_t(s64, 10000LL, delta);
884 hrtick_start(rq, delta);
889 * called from enqueue/dequeue and updates the hrtick when the
890 * current task is from our class and nr_running is low enough
893 static void hrtick_update(struct rq *rq)
895 struct task_struct *curr = rq->curr;
897 if (curr->sched_class != &fair_sched_class)
900 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
901 hrtick_start_fair(rq, curr);
903 #else /* !CONFIG_SCHED_HRTICK */
905 hrtick_start_fair(struct rq *rq, struct task_struct *p)
909 static inline void hrtick_update(struct rq *rq)
915 * The enqueue_task method is called before nr_running is
916 * increased. Here we update the fair scheduling stats and
917 * then put the task into the rbtree:
919 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
921 struct cfs_rq *cfs_rq;
922 struct sched_entity *se = &p->se;
924 for_each_sched_entity(se) {
927 cfs_rq = cfs_rq_of(se);
928 enqueue_entity(cfs_rq, se, wakeup);
936 * The dequeue_task method is called before nr_running is
937 * decreased. We remove the task from the rbtree and
938 * update the fair scheduling stats:
940 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
942 struct cfs_rq *cfs_rq;
943 struct sched_entity *se = &p->se;
945 for_each_sched_entity(se) {
946 cfs_rq = cfs_rq_of(se);
947 dequeue_entity(cfs_rq, se, sleep);
948 /* Don't dequeue parent if it has other entities besides us */
949 if (cfs_rq->load.weight)
958 * sched_yield() support is very simple - we dequeue and enqueue.
960 * If compat_yield is turned on then we requeue to the end of the tree.
962 static void yield_task_fair(struct rq *rq)
964 struct task_struct *curr = rq->curr;
965 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
966 struct sched_entity *rightmost, *se = &curr->se;
969 * Are we the only task in the tree?
971 if (unlikely(cfs_rq->nr_running == 1))
974 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
977 * Update run-time statistics of the 'current'.
984 * Find the rightmost entry in the rbtree:
986 rightmost = __pick_last_entity(cfs_rq);
988 * Already in the rightmost position?
990 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
994 * Minimally necessary key value to be last in the tree:
995 * Upon rescheduling, sched_class::put_prev_task() will place
996 * 'current' within the tree based on its new key value.
998 se->vruntime = rightmost->vruntime + 1;
1002 * wake_idle() will wake a task on an idle cpu if task->cpu is
1003 * not idle and an idle cpu is available. The span of cpus to
1004 * search starts with cpus closest then further out as needed,
1005 * so we always favor a closer, idle cpu.
1006 * Domains may include CPUs that are not usable for migration,
1007 * hence we need to mask them out (cpu_active_map)
1009 * Returns the CPU we should wake onto.
1011 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1012 static int wake_idle(int cpu, struct task_struct *p)
1015 struct sched_domain *sd;
1019 * If it is idle, then it is the best cpu to run this task.
1021 * This cpu is also the best, if it has more than one task already.
1022 * Siblings must be also busy(in most cases) as they didn't already
1023 * pickup the extra load from this cpu and hence we need not check
1024 * sibling runqueue info. This will avoid the checks and cache miss
1025 * penalities associated with that.
1027 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
1030 for_each_domain(cpu, sd) {
1031 if ((sd->flags & SD_WAKE_IDLE)
1032 || ((sd->flags & SD_WAKE_IDLE_FAR)
1033 && !task_hot(p, task_rq(p)->clock, sd))) {
1034 cpus_and(tmp, sd->span, p->cpus_allowed);
1035 cpus_and(tmp, tmp, cpu_active_map);
1036 for_each_cpu_mask_nr(i, tmp) {
1038 if (i != task_cpu(p)) {
1040 se.nr_wakeups_idle);
1051 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1052 static inline int wake_idle(int cpu, struct task_struct *p)
1060 #ifdef CONFIG_FAIR_GROUP_SCHED
1062 * effective_load() calculates the load change as seen from the root_task_group
1064 * Adding load to a group doesn't make a group heavier, but can cause movement
1065 * of group shares between cpus. Assuming the shares were perfectly aligned one
1066 * can calculate the shift in shares.
1068 * The problem is that perfectly aligning the shares is rather expensive, hence
1069 * we try to avoid doing that too often - see update_shares(), which ratelimits
1072 * We compensate this by not only taking the current delta into account, but
1073 * also considering the delta between when the shares were last adjusted and
1076 * We still saw a performance dip, some tracing learned us that between
1077 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1078 * significantly. Therefore try to bias the error in direction of failing
1079 * the affine wakeup.
1082 static long effective_load(struct task_group *tg, int cpu,
1085 struct sched_entity *se = tg->se[cpu];
1091 * By not taking the decrease of shares on the other cpu into
1092 * account our error leans towards reducing the affine wakeups.
1094 if (!wl && sched_feat(ASYM_EFF_LOAD))
1097 for_each_sched_entity(se) {
1098 long S, rw, s, a, b;
1102 * Instead of using this increment, also add the difference
1103 * between when the shares were last updated and now.
1105 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1109 S = se->my_q->tg->shares;
1110 s = se->my_q->shares;
1111 rw = se->my_q->rq_weight;
1122 * Assume the group is already running and will
1123 * thus already be accounted for in the weight.
1125 * That is, moving shares between CPUs, does not
1126 * alter the group weight.
1136 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1137 unsigned long wl, unsigned long wg)
1145 wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
1146 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1147 int idx, unsigned long load, unsigned long this_load,
1148 unsigned int imbalance)
1150 struct task_struct *curr = this_rq->curr;
1151 struct task_group *tg;
1152 unsigned long tl = this_load;
1153 unsigned long tl_per_task;
1154 unsigned long weight;
1157 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1160 if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
1161 p->se.avg_overlap > sysctl_sched_migration_cost))
1165 * If sync wakeup then subtract the (maximum possible)
1166 * effect of the currently running task from the load
1167 * of the current CPU:
1170 tg = task_group(current);
1171 weight = current->se.load.weight;
1173 tl += effective_load(tg, this_cpu, -weight, -weight);
1174 load += effective_load(tg, prev_cpu, 0, -weight);
1178 weight = p->se.load.weight;
1180 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1181 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1184 * If the currently running task will sleep within
1185 * a reasonable amount of time then attract this newly
1188 if (sync && balanced)
1191 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1192 tl_per_task = cpu_avg_load_per_task(this_cpu);
1194 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1197 * This domain has SD_WAKE_AFFINE and
1198 * p is cache cold in this domain, and
1199 * there is no bad imbalance.
1201 schedstat_inc(this_sd, ttwu_move_affine);
1202 schedstat_inc(p, se.nr_wakeups_affine);
1209 static int select_task_rq_fair(struct task_struct *p, int sync)
1211 struct sched_domain *sd, *this_sd = NULL;
1212 int prev_cpu, this_cpu, new_cpu;
1213 unsigned long load, this_load;
1215 unsigned int imbalance;
1218 prev_cpu = task_cpu(p);
1219 this_cpu = smp_processor_id();
1220 this_rq = cpu_rq(this_cpu);
1223 if (prev_cpu == this_cpu)
1226 * 'this_sd' is the first domain that both
1227 * this_cpu and prev_cpu are present in:
1229 for_each_domain(this_cpu, sd) {
1230 if (cpu_isset(prev_cpu, sd->span)) {
1236 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1240 * Check for affine wakeup and passive balancing possibilities.
1245 idx = this_sd->wake_idx;
1247 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1249 load = source_load(prev_cpu, idx);
1250 this_load = target_load(this_cpu, idx);
1252 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1253 load, this_load, imbalance))
1257 * Start passive balancing when half the imbalance_pct
1260 if (this_sd->flags & SD_WAKE_BALANCE) {
1261 if (imbalance*this_load <= 100*load) {
1262 schedstat_inc(this_sd, ttwu_move_balance);
1263 schedstat_inc(p, se.nr_wakeups_passive);
1269 return wake_idle(new_cpu, p);
1271 #endif /* CONFIG_SMP */
1273 static unsigned long wakeup_gran(struct sched_entity *se)
1275 unsigned long gran = sysctl_sched_wakeup_granularity;
1278 * More easily preempt - nice tasks, while not making it harder for
1281 if (!sched_feat(ASYM_GRAN) || se->load.weight > NICE_0_LOAD)
1282 gran = calc_delta_fair(sysctl_sched_wakeup_granularity, se);
1288 * Should 'se' preempt 'curr'.
1302 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1304 s64 gran, vdiff = curr->vruntime - se->vruntime;
1309 gran = wakeup_gran(curr);
1317 * Preempt the current task with a newly woken task if needed:
1319 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
1321 struct task_struct *curr = rq->curr;
1322 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1323 struct sched_entity *se = &curr->se, *pse = &p->se;
1325 if (unlikely(rt_prio(p->prio))) {
1326 update_rq_clock(rq);
1327 update_curr(cfs_rq);
1332 if (unlikely(se == pse))
1335 cfs_rq_of(pse)->next = pse;
1338 * We can come here with TIF_NEED_RESCHED already set from new task
1341 if (test_tsk_need_resched(curr))
1345 * Batch tasks do not preempt (their preemption is driven by
1348 if (unlikely(p->policy == SCHED_BATCH))
1351 if (!sched_feat(WAKEUP_PREEMPT))
1354 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1355 (se->avg_overlap < sysctl_sched_migration_cost &&
1356 pse->avg_overlap < sysctl_sched_migration_cost))) {
1361 find_matching_se(&se, &pse);
1366 if (wakeup_preempt_entity(se, pse) == 1) {
1371 se = parent_entity(se);
1372 pse = parent_entity(pse);
1376 static struct task_struct *pick_next_task_fair(struct rq *rq)
1378 struct task_struct *p;
1379 struct cfs_rq *cfs_rq = &rq->cfs;
1380 struct sched_entity *se;
1382 if (unlikely(!cfs_rq->nr_running))
1386 se = pick_next_entity(cfs_rq);
1387 set_next_entity(cfs_rq, se);
1388 cfs_rq = group_cfs_rq(se);
1392 hrtick_start_fair(rq, p);
1398 * Account for a descheduled task:
1400 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1402 struct sched_entity *se = &prev->se;
1403 struct cfs_rq *cfs_rq;
1405 for_each_sched_entity(se) {
1406 cfs_rq = cfs_rq_of(se);
1407 put_prev_entity(cfs_rq, se);
1412 /**************************************************
1413 * Fair scheduling class load-balancing methods:
1417 * Load-balancing iterator. Note: while the runqueue stays locked
1418 * during the whole iteration, the current task might be
1419 * dequeued so the iterator has to be dequeue-safe. Here we
1420 * achieve that by always pre-iterating before returning
1423 static struct task_struct *
1424 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1426 struct task_struct *p = NULL;
1427 struct sched_entity *se;
1429 if (next == &cfs_rq->tasks)
1432 se = list_entry(next, struct sched_entity, group_node);
1434 cfs_rq->balance_iterator = next->next;
1439 static struct task_struct *load_balance_start_fair(void *arg)
1441 struct cfs_rq *cfs_rq = arg;
1443 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1446 static struct task_struct *load_balance_next_fair(void *arg)
1448 struct cfs_rq *cfs_rq = arg;
1450 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1453 static unsigned long
1454 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1455 unsigned long max_load_move, struct sched_domain *sd,
1456 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1457 struct cfs_rq *cfs_rq)
1459 struct rq_iterator cfs_rq_iterator;
1461 cfs_rq_iterator.start = load_balance_start_fair;
1462 cfs_rq_iterator.next = load_balance_next_fair;
1463 cfs_rq_iterator.arg = cfs_rq;
1465 return balance_tasks(this_rq, this_cpu, busiest,
1466 max_load_move, sd, idle, all_pinned,
1467 this_best_prio, &cfs_rq_iterator);
1470 #ifdef CONFIG_FAIR_GROUP_SCHED
1471 static unsigned long
1472 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1473 unsigned long max_load_move,
1474 struct sched_domain *sd, enum cpu_idle_type idle,
1475 int *all_pinned, int *this_best_prio)
1477 long rem_load_move = max_load_move;
1478 int busiest_cpu = cpu_of(busiest);
1479 struct task_group *tg;
1482 update_h_load(busiest_cpu);
1484 list_for_each_entry_rcu(tg, &task_groups, list) {
1485 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1486 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1487 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1488 u64 rem_load, moved_load;
1493 if (!busiest_cfs_rq->task_weight)
1496 rem_load = (u64)rem_load_move * busiest_weight;
1497 rem_load = div_u64(rem_load, busiest_h_load + 1);
1499 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1500 rem_load, sd, idle, all_pinned, this_best_prio,
1501 tg->cfs_rq[busiest_cpu]);
1506 moved_load *= busiest_h_load;
1507 moved_load = div_u64(moved_load, busiest_weight + 1);
1509 rem_load_move -= moved_load;
1510 if (rem_load_move < 0)
1515 return max_load_move - rem_load_move;
1518 static unsigned long
1519 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1520 unsigned long max_load_move,
1521 struct sched_domain *sd, enum cpu_idle_type idle,
1522 int *all_pinned, int *this_best_prio)
1524 return __load_balance_fair(this_rq, this_cpu, busiest,
1525 max_load_move, sd, idle, all_pinned,
1526 this_best_prio, &busiest->cfs);
1531 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1532 struct sched_domain *sd, enum cpu_idle_type idle)
1534 struct cfs_rq *busy_cfs_rq;
1535 struct rq_iterator cfs_rq_iterator;
1537 cfs_rq_iterator.start = load_balance_start_fair;
1538 cfs_rq_iterator.next = load_balance_next_fair;
1540 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1542 * pass busy_cfs_rq argument into
1543 * load_balance_[start|next]_fair iterators
1545 cfs_rq_iterator.arg = busy_cfs_rq;
1546 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1553 #endif /* CONFIG_SMP */
1556 * scheduler tick hitting a task of our scheduling class:
1558 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1560 struct cfs_rq *cfs_rq;
1561 struct sched_entity *se = &curr->se;
1563 for_each_sched_entity(se) {
1564 cfs_rq = cfs_rq_of(se);
1565 entity_tick(cfs_rq, se, queued);
1569 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1572 * Share the fairness runtime between parent and child, thus the
1573 * total amount of pressure for CPU stays equal - new tasks
1574 * get a chance to run but frequent forkers are not allowed to
1575 * monopolize the CPU. Note: the parent runqueue is locked,
1576 * the child is not running yet.
1578 static void task_new_fair(struct rq *rq, struct task_struct *p)
1580 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1581 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1582 int this_cpu = smp_processor_id();
1584 sched_info_queued(p);
1586 update_curr(cfs_rq);
1587 place_entity(cfs_rq, se, 1);
1589 /* 'curr' will be NULL if the child belongs to a different group */
1590 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1591 curr && curr->vruntime < se->vruntime) {
1593 * Upon rescheduling, sched_class::put_prev_task() will place
1594 * 'current' within the tree based on its new key value.
1596 swap(curr->vruntime, se->vruntime);
1597 resched_task(rq->curr);
1600 enqueue_task_fair(rq, p, 0);
1604 * Priority of the task has changed. Check to see if we preempt
1607 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1608 int oldprio, int running)
1611 * Reschedule if we are currently running on this runqueue and
1612 * our priority decreased, or if we are not currently running on
1613 * this runqueue and our priority is higher than the current's
1616 if (p->prio > oldprio)
1617 resched_task(rq->curr);
1619 check_preempt_curr(rq, p, 0);
1623 * We switched to the sched_fair class.
1625 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1629 * We were most likely switched from sched_rt, so
1630 * kick off the schedule if running, otherwise just see
1631 * if we can still preempt the current task.
1634 resched_task(rq->curr);
1636 check_preempt_curr(rq, p, 0);
1639 /* Account for a task changing its policy or group.
1641 * This routine is mostly called to set cfs_rq->curr field when a task
1642 * migrates between groups/classes.
1644 static void set_curr_task_fair(struct rq *rq)
1646 struct sched_entity *se = &rq->curr->se;
1648 for_each_sched_entity(se)
1649 set_next_entity(cfs_rq_of(se), se);
1652 #ifdef CONFIG_FAIR_GROUP_SCHED
1653 static void moved_group_fair(struct task_struct *p)
1655 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1657 update_curr(cfs_rq);
1658 place_entity(cfs_rq, &p->se, 1);
1663 * All the scheduling class methods:
1665 static const struct sched_class fair_sched_class = {
1666 .next = &idle_sched_class,
1667 .enqueue_task = enqueue_task_fair,
1668 .dequeue_task = dequeue_task_fair,
1669 .yield_task = yield_task_fair,
1671 .check_preempt_curr = check_preempt_wakeup,
1673 .pick_next_task = pick_next_task_fair,
1674 .put_prev_task = put_prev_task_fair,
1677 .select_task_rq = select_task_rq_fair,
1679 .load_balance = load_balance_fair,
1680 .move_one_task = move_one_task_fair,
1683 .set_curr_task = set_curr_task_fair,
1684 .task_tick = task_tick_fair,
1685 .task_new = task_new_fair,
1687 .prio_changed = prio_changed_fair,
1688 .switched_to = switched_to_fair,
1690 #ifdef CONFIG_FAIR_GROUP_SCHED
1691 .moved_group = moved_group_fair,
1695 #ifdef CONFIG_SCHED_DEBUG
1696 static void print_cfs_stats(struct seq_file *m, int cpu)
1698 struct cfs_rq *cfs_rq;
1701 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1702 print_cfs_rq(m, cpu, cfs_rq);