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 /**************************************************************
77 * CFS operations on generic schedulable entities:
80 static inline struct task_struct *task_of(struct sched_entity *se)
82 return container_of(se, struct task_struct, se);
85 #ifdef CONFIG_FAIR_GROUP_SCHED
87 /* cpu runqueue to which this cfs_rq is attached */
88 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
93 /* An entity is a task if it doesn't "own" a runqueue */
94 #define entity_is_task(se) (!se->my_q)
96 /* Walk up scheduling entities hierarchy */
97 #define for_each_sched_entity(se) \
98 for (; se; se = se->parent)
100 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
105 /* runqueue on which this entity is (to be) queued */
106 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
111 /* runqueue "owned" by this group */
112 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
117 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
118 * another cpu ('this_cpu')
120 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
122 return cfs_rq->tg->cfs_rq[this_cpu];
125 /* Iterate thr' all leaf cfs_rq's on a runqueue */
126 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
127 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
129 /* Do the two (enqueued) entities belong to the same group ? */
131 is_same_group(struct sched_entity *se, struct sched_entity *pse)
133 if (se->cfs_rq == pse->cfs_rq)
139 static inline struct sched_entity *parent_entity(struct sched_entity *se)
144 #else /* CONFIG_FAIR_GROUP_SCHED */
146 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
148 return container_of(cfs_rq, struct rq, cfs);
151 #define entity_is_task(se) 1
153 #define for_each_sched_entity(se) \
154 for (; se; se = NULL)
156 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
158 return &task_rq(p)->cfs;
161 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
163 struct task_struct *p = task_of(se);
164 struct rq *rq = task_rq(p);
169 /* runqueue "owned" by this group */
170 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
175 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
177 return &cpu_rq(this_cpu)->cfs;
180 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
181 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
184 is_same_group(struct sched_entity *se, struct sched_entity *pse)
189 static inline struct sched_entity *parent_entity(struct sched_entity *se)
194 #endif /* CONFIG_FAIR_GROUP_SCHED */
197 /**************************************************************
198 * Scheduling class tree data structure manipulation methods:
201 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
203 s64 delta = (s64)(vruntime - min_vruntime);
205 min_vruntime = vruntime;
210 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
212 s64 delta = (s64)(vruntime - min_vruntime);
214 min_vruntime = vruntime;
219 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
221 return se->vruntime - cfs_rq->min_vruntime;
225 * Enqueue an entity into the rb-tree:
227 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
229 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
230 struct rb_node *parent = NULL;
231 struct sched_entity *entry;
232 s64 key = entity_key(cfs_rq, se);
236 * Find the right place in the rbtree:
240 entry = rb_entry(parent, struct sched_entity, run_node);
242 * We dont care about collisions. Nodes with
243 * the same key stay together.
245 if (key < entity_key(cfs_rq, entry)) {
246 link = &parent->rb_left;
248 link = &parent->rb_right;
254 * Maintain a cache of leftmost tree entries (it is frequently
258 cfs_rq->rb_leftmost = &se->run_node;
260 * maintain cfs_rq->min_vruntime to be a monotonic increasing
261 * value tracking the leftmost vruntime in the tree.
263 cfs_rq->min_vruntime =
264 max_vruntime(cfs_rq->min_vruntime, se->vruntime);
267 rb_link_node(&se->run_node, parent, link);
268 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
271 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
273 if (cfs_rq->rb_leftmost == &se->run_node) {
274 struct rb_node *next_node;
275 struct sched_entity *next;
277 next_node = rb_next(&se->run_node);
278 cfs_rq->rb_leftmost = next_node;
281 next = rb_entry(next_node,
282 struct sched_entity, run_node);
283 cfs_rq->min_vruntime =
284 max_vruntime(cfs_rq->min_vruntime,
289 if (cfs_rq->next == se)
292 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
295 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
297 return cfs_rq->rb_leftmost;
300 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
302 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
305 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
307 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
312 return rb_entry(last, struct sched_entity, run_node);
315 /**************************************************************
316 * Scheduling class statistics methods:
319 #ifdef CONFIG_SCHED_DEBUG
320 int sched_nr_latency_handler(struct ctl_table *table, int write,
321 struct file *filp, void __user *buffer, size_t *lenp,
324 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
329 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
330 sysctl_sched_min_granularity);
339 static inline unsigned long
340 calc_delta_weight(unsigned long delta, struct sched_entity *se)
342 for_each_sched_entity(se) {
343 delta = calc_delta_mine(delta,
344 se->load.weight, &cfs_rq_of(se)->load);
353 static inline unsigned long
354 calc_delta_fair(unsigned long delta, struct sched_entity *se)
356 for_each_sched_entity(se) {
357 delta = calc_delta_mine(delta,
358 cfs_rq_of(se)->load.weight, &se->load);
365 * The idea is to set a period in which each task runs once.
367 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
368 * this period because otherwise the slices get too small.
370 * p = (nr <= nl) ? l : l*nr/nl
372 static u64 __sched_period(unsigned long nr_running)
374 u64 period = sysctl_sched_latency;
375 unsigned long nr_latency = sched_nr_latency;
377 if (unlikely(nr_running > nr_latency)) {
378 period = sysctl_sched_min_granularity;
379 period *= nr_running;
386 * We calculate the wall-time slice from the period by taking a part
387 * proportional to the weight.
391 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
393 return calc_delta_weight(__sched_period(cfs_rq->nr_running), se);
397 * We calculate the vruntime slice of a to be inserted task
401 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
403 unsigned long nr_running = cfs_rq->nr_running;
408 return __sched_period(nr_running);
412 * The goal of calc_delta_asym() is to be asymmetrically around NICE_0_LOAD, in
413 * that it favours >=0 over <0.
423 calc_delta_asym(unsigned long delta, struct sched_entity *se)
425 struct load_weight lw = {
426 .weight = NICE_0_LOAD,
427 .inv_weight = 1UL << (WMULT_SHIFT-NICE_0_SHIFT)
430 for_each_sched_entity(se) {
431 struct load_weight *se_lw = &se->load;
432 unsigned long rw = cfs_rq_of(se)->load.weight;
434 #ifdef CONFIG_FAIR_SCHED_GROUP
435 struct cfs_rq *cfs_rq = se->my_q;
436 struct task_group *tg = NULL
441 if (tg && tg->shares < NICE_0_LOAD) {
443 * scale shares to what it would have been had
444 * tg->weight been NICE_0_LOAD:
446 * weight = 1024 * shares / tg->weight
448 lw.weight *= se->load.weight;
449 lw.weight /= tg->shares;
454 rw += lw.weight - se->load.weight;
458 if (se->load.weight < NICE_0_LOAD) {
460 rw += NICE_0_LOAD - se->load.weight;
463 delta = calc_delta_mine(delta, rw, se_lw);
470 * Update the current task's runtime statistics. Skip current tasks that
471 * are not in our scheduling class.
474 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
475 unsigned long delta_exec)
477 unsigned long delta_exec_weighted;
479 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
481 curr->sum_exec_runtime += delta_exec;
482 schedstat_add(cfs_rq, exec_clock, delta_exec);
483 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
484 curr->vruntime += delta_exec_weighted;
487 static void update_curr(struct cfs_rq *cfs_rq)
489 struct sched_entity *curr = cfs_rq->curr;
490 u64 now = rq_of(cfs_rq)->clock;
491 unsigned long delta_exec;
497 * Get the amount of time the current task was running
498 * since the last time we changed load (this cannot
499 * overflow on 32 bits):
501 delta_exec = (unsigned long)(now - curr->exec_start);
503 __update_curr(cfs_rq, curr, delta_exec);
504 curr->exec_start = now;
506 if (entity_is_task(curr)) {
507 struct task_struct *curtask = task_of(curr);
509 cpuacct_charge(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 cfs_rq->nr_running++;
593 list_add(&se->group_node, &cfs_rq->tasks);
597 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
599 update_load_sub(&cfs_rq->load, se->load.weight);
600 if (!parent_entity(se))
601 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
602 if (entity_is_task(se))
603 add_cfs_task_weight(cfs_rq, -se->load.weight);
604 cfs_rq->nr_running--;
606 list_del_init(&se->group_node);
609 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
611 #ifdef CONFIG_SCHEDSTATS
612 if (se->sleep_start) {
613 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
614 struct task_struct *tsk = task_of(se);
619 if (unlikely(delta > se->sleep_max))
620 se->sleep_max = delta;
623 se->sum_sleep_runtime += delta;
625 account_scheduler_latency(tsk, delta >> 10, 1);
627 if (se->block_start) {
628 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
629 struct task_struct *tsk = task_of(se);
634 if (unlikely(delta > se->block_max))
635 se->block_max = delta;
638 se->sum_sleep_runtime += delta;
641 * Blocking time is in units of nanosecs, so shift by 20 to
642 * get a milliseconds-range estimation of the amount of
643 * time that the task spent sleeping:
645 if (unlikely(prof_on == SLEEP_PROFILING)) {
647 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
650 account_scheduler_latency(tsk, delta >> 10, 0);
655 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
657 #ifdef CONFIG_SCHED_DEBUG
658 s64 d = se->vruntime - cfs_rq->min_vruntime;
663 if (d > 3*sysctl_sched_latency)
664 schedstat_inc(cfs_rq, nr_spread_over);
669 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
673 if (first_fair(cfs_rq)) {
674 vruntime = min_vruntime(cfs_rq->min_vruntime,
675 __pick_next_entity(cfs_rq)->vruntime);
677 vruntime = cfs_rq->min_vruntime;
680 * The 'current' period is already promised to the current tasks,
681 * however the extra weight of the new task will slow them down a
682 * little, place the new task so that it fits in the slot that
683 * stays open at the end.
685 if (initial && sched_feat(START_DEBIT))
686 vruntime += sched_vslice_add(cfs_rq, se);
689 /* sleeps upto a single latency don't count. */
690 if (sched_feat(NEW_FAIR_SLEEPERS)) {
691 unsigned long thresh = sysctl_sched_latency;
694 * convert the sleeper threshold into virtual time
696 if (sched_feat(NORMALIZED_SLEEPER))
697 thresh = calc_delta_fair(thresh, se);
702 /* ensure we never gain time by being placed backwards. */
703 vruntime = max_vruntime(se->vruntime, vruntime);
706 se->vruntime = vruntime;
710 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
713 * Update run-time statistics of the 'current'.
716 account_entity_enqueue(cfs_rq, se);
719 place_entity(cfs_rq, se, 0);
720 enqueue_sleeper(cfs_rq, se);
723 update_stats_enqueue(cfs_rq, se);
724 check_spread(cfs_rq, se);
725 if (se != cfs_rq->curr)
726 __enqueue_entity(cfs_rq, se);
730 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
733 * Update run-time statistics of the 'current'.
737 update_stats_dequeue(cfs_rq, se);
739 #ifdef CONFIG_SCHEDSTATS
740 if (entity_is_task(se)) {
741 struct task_struct *tsk = task_of(se);
743 if (tsk->state & TASK_INTERRUPTIBLE)
744 se->sleep_start = rq_of(cfs_rq)->clock;
745 if (tsk->state & TASK_UNINTERRUPTIBLE)
746 se->block_start = rq_of(cfs_rq)->clock;
751 if (se != cfs_rq->curr)
752 __dequeue_entity(cfs_rq, se);
753 account_entity_dequeue(cfs_rq, se);
757 * Preempt the current task with a newly woken task if needed:
760 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
762 unsigned long ideal_runtime, delta_exec;
764 ideal_runtime = sched_slice(cfs_rq, curr);
765 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
766 if (delta_exec > ideal_runtime)
767 resched_task(rq_of(cfs_rq)->curr);
771 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 /* 'current' is not kept within the tree. */
776 * Any task has to be enqueued before it get to execute on
777 * a CPU. So account for the time it spent waiting on the
780 update_stats_wait_end(cfs_rq, se);
781 __dequeue_entity(cfs_rq, se);
784 update_stats_curr_start(cfs_rq, se);
786 #ifdef CONFIG_SCHEDSTATS
788 * Track our maximum slice length, if the CPU's load is at
789 * least twice that of our own weight (i.e. dont track it
790 * when there are only lesser-weight tasks around):
792 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
793 se->slice_max = max(se->slice_max,
794 se->sum_exec_runtime - se->prev_sum_exec_runtime);
797 se->prev_sum_exec_runtime = se->sum_exec_runtime;
800 static struct sched_entity *
801 pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se)
803 struct rq *rq = rq_of(cfs_rq);
804 u64 pair_slice = rq->clock - cfs_rq->pair_start;
806 if (!cfs_rq->next || pair_slice > sched_slice(cfs_rq, cfs_rq->next)) {
807 cfs_rq->pair_start = rq->clock;
814 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
816 struct sched_entity *se = NULL;
818 if (first_fair(cfs_rq)) {
819 se = __pick_next_entity(cfs_rq);
820 se = pick_next(cfs_rq, se);
821 set_next_entity(cfs_rq, se);
827 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
830 * If still on the runqueue then deactivate_task()
831 * was not called and update_curr() has to be done:
836 check_spread(cfs_rq, prev);
838 update_stats_wait_start(cfs_rq, prev);
839 /* Put 'current' back into the tree. */
840 __enqueue_entity(cfs_rq, prev);
846 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
849 * Update run-time statistics of the 'current'.
853 #ifdef CONFIG_SCHED_HRTICK
855 * queued ticks are scheduled to match the slice, so don't bother
856 * validating it and just reschedule.
859 resched_task(rq_of(cfs_rq)->curr);
863 * don't let the period tick interfere with the hrtick preemption
865 if (!sched_feat(DOUBLE_TICK) &&
866 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
870 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
871 check_preempt_tick(cfs_rq, curr);
874 /**************************************************
875 * CFS operations on tasks:
878 #ifdef CONFIG_SCHED_HRTICK
879 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
881 int requeue = rq->curr == p;
882 struct sched_entity *se = &p->se;
883 struct cfs_rq *cfs_rq = cfs_rq_of(se);
885 WARN_ON(task_rq(p) != rq);
887 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
888 u64 slice = sched_slice(cfs_rq, se);
889 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
890 s64 delta = slice - ran;
899 * Don't schedule slices shorter than 10000ns, that just
900 * doesn't make sense. Rely on vruntime for fairness.
903 delta = max(10000LL, delta);
905 hrtick_start(rq, delta, requeue);
908 #else /* !CONFIG_SCHED_HRTICK */
910 hrtick_start_fair(struct rq *rq, struct task_struct *p)
916 * The enqueue_task method is called before nr_running is
917 * increased. Here we update the fair scheduling stats and
918 * then put the task into the rbtree:
920 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
922 struct cfs_rq *cfs_rq;
923 struct sched_entity *se = &p->se;
925 for_each_sched_entity(se) {
928 cfs_rq = cfs_rq_of(se);
929 enqueue_entity(cfs_rq, se, wakeup);
933 hrtick_start_fair(rq, rq->curr);
937 * The dequeue_task method is called before nr_running is
938 * decreased. We remove the task from the rbtree and
939 * update the fair scheduling stats:
941 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
943 struct cfs_rq *cfs_rq;
944 struct sched_entity *se = &p->se;
946 for_each_sched_entity(se) {
947 cfs_rq = cfs_rq_of(se);
948 dequeue_entity(cfs_rq, se, sleep);
949 /* Don't dequeue parent if it has other entities besides us */
950 if (cfs_rq->load.weight)
955 hrtick_start_fair(rq, rq->curr);
959 * sched_yield() support is very simple - we dequeue and enqueue.
961 * If compat_yield is turned on then we requeue to the end of the tree.
963 static void yield_task_fair(struct rq *rq)
965 struct task_struct *curr = rq->curr;
966 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
967 struct sched_entity *rightmost, *se = &curr->se;
970 * Are we the only task in the tree?
972 if (unlikely(cfs_rq->nr_running == 1))
975 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
978 * Update run-time statistics of the 'current'.
985 * Find the rightmost entry in the rbtree:
987 rightmost = __pick_last_entity(cfs_rq);
989 * Already in the rightmost position?
991 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
995 * Minimally necessary key value to be last in the tree:
996 * Upon rescheduling, sched_class::put_prev_task() will place
997 * 'current' within the tree based on its new key value.
999 se->vruntime = rightmost->vruntime + 1;
1003 * wake_idle() will wake a task on an idle cpu if task->cpu is
1004 * not idle and an idle cpu is available. The span of cpus to
1005 * search starts with cpus closest then further out as needed,
1006 * so we always favor a closer, idle cpu.
1008 * Returns the CPU we should wake onto.
1010 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1011 static int wake_idle(int cpu, struct task_struct *p)
1014 struct sched_domain *sd;
1018 * If it is idle, then it is the best cpu to run this task.
1020 * This cpu is also the best, if it has more than one task already.
1021 * Siblings must be also busy(in most cases) as they didn't already
1022 * pickup the extra load from this cpu and hence we need not check
1023 * sibling runqueue info. This will avoid the checks and cache miss
1024 * penalities associated with that.
1026 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
1029 for_each_domain(cpu, sd) {
1030 if ((sd->flags & SD_WAKE_IDLE)
1031 || ((sd->flags & SD_WAKE_IDLE_FAR)
1032 && !task_hot(p, task_rq(p)->clock, sd))) {
1033 cpus_and(tmp, sd->span, p->cpus_allowed);
1034 for_each_cpu_mask(i, tmp) {
1036 if (i != task_cpu(p)) {
1038 se.nr_wakeups_idle);
1049 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
1050 static inline int wake_idle(int cpu, struct task_struct *p)
1058 static const struct sched_class fair_sched_class;
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];
1092 * By not taking the decrease of shares on the other cpu into
1093 * account our error leans towards reducing the affine wakeups.
1095 if (!wl && sched_feat(ASYM_EFF_LOAD))
1099 * Instead of using this increment, also add the difference
1100 * between when the shares were last updated and now.
1102 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1106 for_each_sched_entity(se) {
1107 #define D(n) (likely(n) ? (n) : 1)
1109 long S, rw, s, a, b;
1111 S = se->my_q->tg->shares;
1112 s = se->my_q->shares;
1113 rw = se->my_q->rq_weight;
1120 * Assume the group is already running and will
1121 * thus already be accounted for in the weight.
1123 * That is, moving shares between CPUs, does not
1124 * alter the group weight.
1135 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1136 unsigned long wl, unsigned long wg)
1144 wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq,
1145 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1146 int idx, unsigned long load, unsigned long this_load,
1147 unsigned int imbalance)
1149 struct task_struct *curr = this_rq->curr;
1150 struct task_group *tg;
1151 unsigned long tl = this_load;
1152 unsigned long tl_per_task;
1153 unsigned long weight;
1156 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1160 * If sync wakeup then subtract the (maximum possible)
1161 * effect of the currently running task from the load
1162 * of the current CPU:
1165 tg = task_group(current);
1166 weight = current->se.load.weight;
1168 tl += effective_load(tg, this_cpu, -weight, -weight);
1169 load += effective_load(tg, prev_cpu, 0, -weight);
1173 weight = p->se.load.weight;
1175 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1176 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1179 * If the currently running task will sleep within
1180 * a reasonable amount of time then attract this newly
1183 if (sync && balanced) {
1184 if (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1185 p->se.avg_overlap < sysctl_sched_migration_cost)
1189 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1190 tl_per_task = cpu_avg_load_per_task(this_cpu);
1192 if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) ||
1195 * This domain has SD_WAKE_AFFINE and
1196 * p is cache cold in this domain, and
1197 * there is no bad imbalance.
1199 schedstat_inc(this_sd, ttwu_move_affine);
1200 schedstat_inc(p, se.nr_wakeups_affine);
1207 static int select_task_rq_fair(struct task_struct *p, int sync)
1209 struct sched_domain *sd, *this_sd = NULL;
1210 int prev_cpu, this_cpu, new_cpu;
1211 unsigned long load, this_load;
1212 struct rq *rq, *this_rq;
1213 unsigned int imbalance;
1216 prev_cpu = task_cpu(p);
1218 this_cpu = smp_processor_id();
1219 this_rq = cpu_rq(this_cpu);
1223 * 'this_sd' is the first domain that both
1224 * this_cpu and prev_cpu are present in:
1226 for_each_domain(this_cpu, sd) {
1227 if (cpu_isset(prev_cpu, sd->span)) {
1233 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1237 * Check for affine wakeup and passive balancing possibilities.
1242 idx = this_sd->wake_idx;
1244 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1246 load = source_load(prev_cpu, idx);
1247 this_load = target_load(this_cpu, idx);
1249 if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1250 load, this_load, imbalance))
1253 if (prev_cpu == this_cpu)
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))
1282 gran = calc_delta_asym(sysctl_sched_wakeup_granularity, se);
1284 gran = calc_delta_fair(sysctl_sched_wakeup_granularity, se);
1290 * Should 'se' preempt 'curr'.
1304 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1306 s64 gran, vdiff = curr->vruntime - se->vruntime;
1311 gran = wakeup_gran(curr);
1318 /* return depth at which a sched entity is present in the hierarchy */
1319 static inline int depth_se(struct sched_entity *se)
1323 for_each_sched_entity(se)
1330 * Preempt the current task with a newly woken task if needed:
1332 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
1334 struct task_struct *curr = rq->curr;
1335 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1336 struct sched_entity *se = &curr->se, *pse = &p->se;
1337 int se_depth, pse_depth;
1339 if (unlikely(rt_prio(p->prio))) {
1340 update_rq_clock(rq);
1341 update_curr(cfs_rq);
1346 if (unlikely(se == pse))
1349 cfs_rq_of(pse)->next = pse;
1352 * Batch tasks do not preempt (their preemption is driven by
1355 if (unlikely(p->policy == SCHED_BATCH))
1358 if (!sched_feat(WAKEUP_PREEMPT))
1362 * preemption test can be made between sibling entities who are in the
1363 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
1364 * both tasks until we find their ancestors who are siblings of common
1368 /* First walk up until both entities are at same depth */
1369 se_depth = depth_se(se);
1370 pse_depth = depth_se(pse);
1372 while (se_depth > pse_depth) {
1374 se = parent_entity(se);
1377 while (pse_depth > se_depth) {
1379 pse = parent_entity(pse);
1382 while (!is_same_group(se, pse)) {
1383 se = parent_entity(se);
1384 pse = parent_entity(pse);
1387 if (wakeup_preempt_entity(se, pse) == 1)
1391 static struct task_struct *pick_next_task_fair(struct rq *rq)
1393 struct task_struct *p;
1394 struct cfs_rq *cfs_rq = &rq->cfs;
1395 struct sched_entity *se;
1397 if (unlikely(!cfs_rq->nr_running))
1401 se = pick_next_entity(cfs_rq);
1402 cfs_rq = group_cfs_rq(se);
1406 hrtick_start_fair(rq, p);
1412 * Account for a descheduled task:
1414 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1416 struct sched_entity *se = &prev->se;
1417 struct cfs_rq *cfs_rq;
1419 for_each_sched_entity(se) {
1420 cfs_rq = cfs_rq_of(se);
1421 put_prev_entity(cfs_rq, se);
1426 /**************************************************
1427 * Fair scheduling class load-balancing methods:
1431 * Load-balancing iterator. Note: while the runqueue stays locked
1432 * during the whole iteration, the current task might be
1433 * dequeued so the iterator has to be dequeue-safe. Here we
1434 * achieve that by always pre-iterating before returning
1437 static struct task_struct *
1438 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1440 struct task_struct *p = NULL;
1441 struct sched_entity *se;
1443 while (next != &cfs_rq->tasks) {
1444 se = list_entry(next, struct sched_entity, group_node);
1447 /* Skip over entities that are not tasks */
1448 if (entity_is_task(se)) {
1454 cfs_rq->balance_iterator = next;
1458 static struct task_struct *load_balance_start_fair(void *arg)
1460 struct cfs_rq *cfs_rq = arg;
1462 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1465 static struct task_struct *load_balance_next_fair(void *arg)
1467 struct cfs_rq *cfs_rq = arg;
1469 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1472 static unsigned long
1473 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1474 unsigned long max_load_move, struct sched_domain *sd,
1475 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1476 struct cfs_rq *cfs_rq)
1478 struct rq_iterator cfs_rq_iterator;
1480 cfs_rq_iterator.start = load_balance_start_fair;
1481 cfs_rq_iterator.next = load_balance_next_fair;
1482 cfs_rq_iterator.arg = cfs_rq;
1484 return balance_tasks(this_rq, this_cpu, busiest,
1485 max_load_move, sd, idle, all_pinned,
1486 this_best_prio, &cfs_rq_iterator);
1489 #ifdef CONFIG_FAIR_GROUP_SCHED
1490 static unsigned long
1491 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1492 unsigned long max_load_move,
1493 struct sched_domain *sd, enum cpu_idle_type idle,
1494 int *all_pinned, int *this_best_prio)
1496 long rem_load_move = max_load_move;
1497 int busiest_cpu = cpu_of(busiest);
1498 struct task_group *tg;
1501 update_h_load(busiest_cpu);
1503 list_for_each_entry(tg, &task_groups, list) {
1504 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1505 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1506 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1507 u64 rem_load, moved_load;
1512 if (!busiest_cfs_rq->task_weight)
1515 rem_load = (u64)rem_load_move * busiest_weight;
1516 rem_load = div_u64(rem_load, busiest_h_load + 1);
1518 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1519 rem_load, sd, idle, all_pinned, this_best_prio,
1520 tg->cfs_rq[busiest_cpu]);
1525 moved_load *= busiest_h_load;
1526 moved_load = div_u64(moved_load, busiest_weight + 1);
1528 rem_load_move -= moved_load;
1529 if (rem_load_move < 0)
1534 return max_load_move - rem_load_move;
1537 static unsigned long
1538 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1539 unsigned long max_load_move,
1540 struct sched_domain *sd, enum cpu_idle_type idle,
1541 int *all_pinned, int *this_best_prio)
1543 return __load_balance_fair(this_rq, this_cpu, busiest,
1544 max_load_move, sd, idle, all_pinned,
1545 this_best_prio, &busiest->cfs);
1550 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1551 struct sched_domain *sd, enum cpu_idle_type idle)
1553 struct cfs_rq *busy_cfs_rq;
1554 struct rq_iterator cfs_rq_iterator;
1556 cfs_rq_iterator.start = load_balance_start_fair;
1557 cfs_rq_iterator.next = load_balance_next_fair;
1559 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1561 * pass busy_cfs_rq argument into
1562 * load_balance_[start|next]_fair iterators
1564 cfs_rq_iterator.arg = busy_cfs_rq;
1565 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1572 #endif /* CONFIG_SMP */
1575 * scheduler tick hitting a task of our scheduling class:
1577 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1579 struct cfs_rq *cfs_rq;
1580 struct sched_entity *se = &curr->se;
1582 for_each_sched_entity(se) {
1583 cfs_rq = cfs_rq_of(se);
1584 entity_tick(cfs_rq, se, queued);
1588 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1591 * Share the fairness runtime between parent and child, thus the
1592 * total amount of pressure for CPU stays equal - new tasks
1593 * get a chance to run but frequent forkers are not allowed to
1594 * monopolize the CPU. Note: the parent runqueue is locked,
1595 * the child is not running yet.
1597 static void task_new_fair(struct rq *rq, struct task_struct *p)
1599 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1600 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1601 int this_cpu = smp_processor_id();
1603 sched_info_queued(p);
1605 update_curr(cfs_rq);
1606 place_entity(cfs_rq, se, 1);
1608 /* 'curr' will be NULL if the child belongs to a different group */
1609 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1610 curr && curr->vruntime < se->vruntime) {
1612 * Upon rescheduling, sched_class::put_prev_task() will place
1613 * 'current' within the tree based on its new key value.
1615 swap(curr->vruntime, se->vruntime);
1618 enqueue_task_fair(rq, p, 0);
1619 resched_task(rq->curr);
1623 * Priority of the task has changed. Check to see if we preempt
1626 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1627 int oldprio, int running)
1630 * Reschedule if we are currently running on this runqueue and
1631 * our priority decreased, or if we are not currently running on
1632 * this runqueue and our priority is higher than the current's
1635 if (p->prio > oldprio)
1636 resched_task(rq->curr);
1638 check_preempt_curr(rq, p);
1642 * We switched to the sched_fair class.
1644 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1648 * We were most likely switched from sched_rt, so
1649 * kick off the schedule if running, otherwise just see
1650 * if we can still preempt the current task.
1653 resched_task(rq->curr);
1655 check_preempt_curr(rq, p);
1658 /* Account for a task changing its policy or group.
1660 * This routine is mostly called to set cfs_rq->curr field when a task
1661 * migrates between groups/classes.
1663 static void set_curr_task_fair(struct rq *rq)
1665 struct sched_entity *se = &rq->curr->se;
1667 for_each_sched_entity(se)
1668 set_next_entity(cfs_rq_of(se), se);
1671 #ifdef CONFIG_FAIR_GROUP_SCHED
1672 static void moved_group_fair(struct task_struct *p)
1674 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1676 update_curr(cfs_rq);
1677 place_entity(cfs_rq, &p->se, 1);
1682 * All the scheduling class methods:
1684 static const struct sched_class fair_sched_class = {
1685 .next = &idle_sched_class,
1686 .enqueue_task = enqueue_task_fair,
1687 .dequeue_task = dequeue_task_fair,
1688 .yield_task = yield_task_fair,
1690 .select_task_rq = select_task_rq_fair,
1691 #endif /* CONFIG_SMP */
1693 .check_preempt_curr = check_preempt_wakeup,
1695 .pick_next_task = pick_next_task_fair,
1696 .put_prev_task = put_prev_task_fair,
1699 .load_balance = load_balance_fair,
1700 .move_one_task = move_one_task_fair,
1703 .set_curr_task = set_curr_task_fair,
1704 .task_tick = task_tick_fair,
1705 .task_new = task_new_fair,
1707 .prio_changed = prio_changed_fair,
1708 .switched_to = switched_to_fair,
1710 #ifdef CONFIG_FAIR_GROUP_SCHED
1711 .moved_group = moved_group_fair,
1715 #ifdef CONFIG_SCHED_DEBUG
1716 static void print_cfs_stats(struct seq_file *m, int cpu)
1718 struct cfs_rq *cfs_rq;
1721 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1722 print_cfs_rq(m, cpu, cfs_rq);