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_BATCH wake-up granularity.
66 * (default: 10 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_batch_wakeup_granularity = 10000000UL;
75 * SCHED_OTHER wake-up granularity.
76 * (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 unsigned int sysctl_sched_wakeup_granularity = 10000000UL;
84 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
86 /**************************************************************
87 * CFS operations on generic schedulable entities:
90 #ifdef CONFIG_FAIR_GROUP_SCHED
92 /* cpu runqueue to which this cfs_rq is attached */
93 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
98 /* An entity is a task if it doesn't "own" a runqueue */
99 #define entity_is_task(se) (!se->my_q)
101 #else /* CONFIG_FAIR_GROUP_SCHED */
103 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
105 return container_of(cfs_rq, struct rq, cfs);
108 #define entity_is_task(se) 1
110 #endif /* CONFIG_FAIR_GROUP_SCHED */
112 static inline struct task_struct *task_of(struct sched_entity *se)
114 return container_of(se, struct task_struct, se);
118 /**************************************************************
119 * Scheduling class tree data structure manipulation methods:
122 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
124 s64 delta = (s64)(vruntime - min_vruntime);
126 min_vruntime = vruntime;
131 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
133 s64 delta = (s64)(vruntime - min_vruntime);
135 min_vruntime = vruntime;
140 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
142 return se->vruntime - cfs_rq->min_vruntime;
146 * Enqueue an entity into the rb-tree:
148 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
150 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
151 struct rb_node *parent = NULL;
152 struct sched_entity *entry;
153 s64 key = entity_key(cfs_rq, se);
157 * Find the right place in the rbtree:
161 entry = rb_entry(parent, struct sched_entity, run_node);
163 * We dont care about collisions. Nodes with
164 * the same key stay together.
166 if (key < entity_key(cfs_rq, entry)) {
167 link = &parent->rb_left;
169 link = &parent->rb_right;
175 * Maintain a cache of leftmost tree entries (it is frequently
179 cfs_rq->rb_leftmost = &se->run_node;
181 * maintain cfs_rq->min_vruntime to be a monotonic increasing
182 * value tracking the leftmost vruntime in the tree.
184 cfs_rq->min_vruntime =
185 max_vruntime(cfs_rq->min_vruntime, se->vruntime);
188 rb_link_node(&se->run_node, parent, link);
189 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
192 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
194 if (cfs_rq->rb_leftmost == &se->run_node) {
195 struct rb_node *next_node;
196 struct sched_entity *next;
198 next_node = rb_next(&se->run_node);
199 cfs_rq->rb_leftmost = next_node;
202 next = rb_entry(next_node,
203 struct sched_entity, run_node);
204 cfs_rq->min_vruntime =
205 max_vruntime(cfs_rq->min_vruntime,
210 if (cfs_rq->next == se)
213 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
216 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
218 return cfs_rq->rb_leftmost;
221 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
223 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
226 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
228 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
233 return rb_entry(last, struct sched_entity, run_node);
236 /**************************************************************
237 * Scheduling class statistics methods:
240 #ifdef CONFIG_SCHED_DEBUG
241 int sched_nr_latency_handler(struct ctl_table *table, int write,
242 struct file *filp, void __user *buffer, size_t *lenp,
245 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
250 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
251 sysctl_sched_min_granularity);
258 * The idea is to set a period in which each task runs once.
260 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
261 * this period because otherwise the slices get too small.
263 * p = (nr <= nl) ? l : l*nr/nl
265 static u64 __sched_period(unsigned long nr_running)
267 u64 period = sysctl_sched_latency;
268 unsigned long nr_latency = sched_nr_latency;
270 if (unlikely(nr_running > nr_latency)) {
271 period = sysctl_sched_min_granularity;
272 period *= nr_running;
279 * We calculate the wall-time slice from the period by taking a part
280 * proportional to the weight.
284 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
286 return calc_delta_mine(__sched_period(cfs_rq->nr_running),
287 se->load.weight, &cfs_rq->load);
291 * We calculate the vruntime slice.
295 static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running)
297 u64 vslice = __sched_period(nr_running);
299 vslice *= NICE_0_LOAD;
300 do_div(vslice, rq_weight);
305 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
307 return __sched_vslice(cfs_rq->load.weight + se->load.weight,
308 cfs_rq->nr_running + 1);
312 * Update the current task's runtime statistics. Skip current tasks that
313 * are not in our scheduling class.
316 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
317 unsigned long delta_exec)
319 unsigned long delta_exec_weighted;
321 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
323 curr->sum_exec_runtime += delta_exec;
324 schedstat_add(cfs_rq, exec_clock, delta_exec);
325 delta_exec_weighted = delta_exec;
326 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
327 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
330 curr->vruntime += delta_exec_weighted;
333 static void update_curr(struct cfs_rq *cfs_rq)
335 struct sched_entity *curr = cfs_rq->curr;
336 u64 now = rq_of(cfs_rq)->clock;
337 unsigned long delta_exec;
343 * Get the amount of time the current task was running
344 * since the last time we changed load (this cannot
345 * overflow on 32 bits):
347 delta_exec = (unsigned long)(now - curr->exec_start);
349 __update_curr(cfs_rq, curr, delta_exec);
350 curr->exec_start = now;
352 if (entity_is_task(curr)) {
353 struct task_struct *curtask = task_of(curr);
355 cpuacct_charge(curtask, delta_exec);
360 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
362 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
366 * Task is being enqueued - update stats:
368 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
371 * Are we enqueueing a waiting task? (for current tasks
372 * a dequeue/enqueue event is a NOP)
374 if (se != cfs_rq->curr)
375 update_stats_wait_start(cfs_rq, se);
379 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
381 schedstat_set(se->wait_max, max(se->wait_max,
382 rq_of(cfs_rq)->clock - se->wait_start));
383 schedstat_set(se->wait_count, se->wait_count + 1);
384 schedstat_set(se->wait_sum, se->wait_sum +
385 rq_of(cfs_rq)->clock - se->wait_start);
386 schedstat_set(se->wait_start, 0);
390 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
393 * Mark the end of the wait period if dequeueing a
396 if (se != cfs_rq->curr)
397 update_stats_wait_end(cfs_rq, se);
401 * We are picking a new current task - update its stats:
404 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
407 * We are starting a new run period:
409 se->exec_start = rq_of(cfs_rq)->clock;
412 /**************************************************
413 * Scheduling class queueing methods:
417 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
419 update_load_add(&cfs_rq->load, se->load.weight);
420 cfs_rq->nr_running++;
425 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
427 update_load_sub(&cfs_rq->load, se->load.weight);
428 cfs_rq->nr_running--;
432 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
434 #ifdef CONFIG_SCHEDSTATS
435 if (se->sleep_start) {
436 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
437 struct task_struct *tsk = task_of(se);
442 if (unlikely(delta > se->sleep_max))
443 se->sleep_max = delta;
446 se->sum_sleep_runtime += delta;
448 account_scheduler_latency(tsk, delta >> 10, 1);
450 if (se->block_start) {
451 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
452 struct task_struct *tsk = task_of(se);
457 if (unlikely(delta > se->block_max))
458 se->block_max = delta;
461 se->sum_sleep_runtime += delta;
464 * Blocking time is in units of nanosecs, so shift by 20 to
465 * get a milliseconds-range estimation of the amount of
466 * time that the task spent sleeping:
468 if (unlikely(prof_on == SLEEP_PROFILING)) {
470 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
473 account_scheduler_latency(tsk, delta >> 10, 0);
478 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
480 #ifdef CONFIG_SCHED_DEBUG
481 s64 d = se->vruntime - cfs_rq->min_vruntime;
486 if (d > 3*sysctl_sched_latency)
487 schedstat_inc(cfs_rq, nr_spread_over);
492 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
496 if (first_fair(cfs_rq)) {
497 vruntime = min_vruntime(cfs_rq->min_vruntime,
498 __pick_next_entity(cfs_rq)->vruntime);
500 vruntime = cfs_rq->min_vruntime;
503 * The 'current' period is already promised to the current tasks,
504 * however the extra weight of the new task will slow them down a
505 * little, place the new task so that it fits in the slot that
506 * stays open at the end.
508 if (initial && sched_feat(START_DEBIT))
509 vruntime += sched_vslice_add(cfs_rq, se);
512 /* sleeps upto a single latency don't count. */
513 if (sched_feat(NEW_FAIR_SLEEPERS)) {
514 vruntime -= calc_delta_fair(sysctl_sched_latency,
518 /* ensure we never gain time by being placed backwards. */
519 vruntime = max_vruntime(se->vruntime, vruntime);
522 se->vruntime = vruntime;
526 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
529 * Update run-time statistics of the 'current'.
534 place_entity(cfs_rq, se, 0);
535 enqueue_sleeper(cfs_rq, se);
538 update_stats_enqueue(cfs_rq, se);
539 check_spread(cfs_rq, se);
540 if (se != cfs_rq->curr)
541 __enqueue_entity(cfs_rq, se);
542 account_entity_enqueue(cfs_rq, se);
545 static void update_avg(u64 *avg, u64 sample)
547 s64 diff = sample - *avg;
551 static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se)
553 if (!se->last_wakeup)
556 update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup);
561 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
564 * Update run-time statistics of the 'current'.
568 update_stats_dequeue(cfs_rq, se);
570 update_avg_stats(cfs_rq, se);
571 #ifdef CONFIG_SCHEDSTATS
572 if (entity_is_task(se)) {
573 struct task_struct *tsk = task_of(se);
575 if (tsk->state & TASK_INTERRUPTIBLE)
576 se->sleep_start = rq_of(cfs_rq)->clock;
577 if (tsk->state & TASK_UNINTERRUPTIBLE)
578 se->block_start = rq_of(cfs_rq)->clock;
583 if (se != cfs_rq->curr)
584 __dequeue_entity(cfs_rq, se);
585 account_entity_dequeue(cfs_rq, se);
589 * Preempt the current task with a newly woken task if needed:
592 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
594 unsigned long ideal_runtime, delta_exec;
596 ideal_runtime = sched_slice(cfs_rq, curr);
597 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
598 if (delta_exec > ideal_runtime)
599 resched_task(rq_of(cfs_rq)->curr);
603 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
605 /* 'current' is not kept within the tree. */
608 * Any task has to be enqueued before it get to execute on
609 * a CPU. So account for the time it spent waiting on the
612 update_stats_wait_end(cfs_rq, se);
613 __dequeue_entity(cfs_rq, se);
616 update_stats_curr_start(cfs_rq, se);
618 #ifdef CONFIG_SCHEDSTATS
620 * Track our maximum slice length, if the CPU's load is at
621 * least twice that of our own weight (i.e. dont track it
622 * when there are only lesser-weight tasks around):
624 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
625 se->slice_max = max(se->slice_max,
626 se->sum_exec_runtime - se->prev_sum_exec_runtime);
629 se->prev_sum_exec_runtime = se->sum_exec_runtime;
633 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
635 static struct sched_entity *
636 pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se)
641 if (wakeup_preempt_entity(cfs_rq->next, se) != 0)
647 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
649 struct sched_entity *se = NULL;
651 if (first_fair(cfs_rq)) {
652 se = __pick_next_entity(cfs_rq);
653 se = pick_next(cfs_rq, se);
654 set_next_entity(cfs_rq, se);
660 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
663 * If still on the runqueue then deactivate_task()
664 * was not called and update_curr() has to be done:
669 check_spread(cfs_rq, prev);
671 update_stats_wait_start(cfs_rq, prev);
672 /* Put 'current' back into the tree. */
673 __enqueue_entity(cfs_rq, prev);
679 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
682 * Update run-time statistics of the 'current'.
686 #ifdef CONFIG_SCHED_HRTICK
688 * queued ticks are scheduled to match the slice, so don't bother
689 * validating it and just reschedule.
692 return resched_task(rq_of(cfs_rq)->curr);
694 * don't let the period tick interfere with the hrtick preemption
696 if (!sched_feat(DOUBLE_TICK) &&
697 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
701 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
702 check_preempt_tick(cfs_rq, curr);
705 /**************************************************
706 * CFS operations on tasks:
709 #ifdef CONFIG_FAIR_GROUP_SCHED
711 /* Walk up scheduling entities hierarchy */
712 #define for_each_sched_entity(se) \
713 for (; se; se = se->parent)
715 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
720 /* runqueue on which this entity is (to be) queued */
721 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
726 /* runqueue "owned" by this group */
727 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
732 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
733 * another cpu ('this_cpu')
735 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
737 return cfs_rq->tg->cfs_rq[this_cpu];
740 /* Iterate thr' all leaf cfs_rq's on a runqueue */
741 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
742 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
744 /* Do the two (enqueued) entities belong to the same group ? */
746 is_same_group(struct sched_entity *se, struct sched_entity *pse)
748 if (se->cfs_rq == pse->cfs_rq)
754 static inline struct sched_entity *parent_entity(struct sched_entity *se)
759 #else /* CONFIG_FAIR_GROUP_SCHED */
761 #define for_each_sched_entity(se) \
762 for (; se; se = NULL)
764 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
766 return &task_rq(p)->cfs;
769 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
771 struct task_struct *p = task_of(se);
772 struct rq *rq = task_rq(p);
777 /* runqueue "owned" by this group */
778 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
783 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
785 return &cpu_rq(this_cpu)->cfs;
788 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
789 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
792 is_same_group(struct sched_entity *se, struct sched_entity *pse)
797 static inline struct sched_entity *parent_entity(struct sched_entity *se)
802 #endif /* CONFIG_FAIR_GROUP_SCHED */
804 #ifdef CONFIG_SCHED_HRTICK
805 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
807 int requeue = rq->curr == p;
808 struct sched_entity *se = &p->se;
809 struct cfs_rq *cfs_rq = cfs_rq_of(se);
811 WARN_ON(task_rq(p) != rq);
813 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
814 u64 slice = sched_slice(cfs_rq, se);
815 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
816 s64 delta = slice - ran;
825 * Don't schedule slices shorter than 10000ns, that just
826 * doesn't make sense. Rely on vruntime for fairness.
829 delta = max(10000LL, delta);
831 hrtick_start(rq, delta, requeue);
836 hrtick_start_fair(struct rq *rq, struct task_struct *p)
842 * The enqueue_task method is called before nr_running is
843 * increased. Here we update the fair scheduling stats and
844 * then put the task into the rbtree:
846 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
848 struct cfs_rq *cfs_rq;
849 struct sched_entity *se = &p->se;
851 for_each_sched_entity(se) {
854 cfs_rq = cfs_rq_of(se);
855 enqueue_entity(cfs_rq, se, wakeup);
859 hrtick_start_fair(rq, rq->curr);
863 * The dequeue_task method is called before nr_running is
864 * decreased. We remove the task from the rbtree and
865 * update the fair scheduling stats:
867 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
869 struct cfs_rq *cfs_rq;
870 struct sched_entity *se = &p->se;
872 for_each_sched_entity(se) {
873 cfs_rq = cfs_rq_of(se);
874 dequeue_entity(cfs_rq, se, sleep);
875 /* Don't dequeue parent if it has other entities besides us */
876 if (cfs_rq->load.weight)
881 hrtick_start_fair(rq, rq->curr);
885 * sched_yield() support is very simple - we dequeue and enqueue.
887 * If compat_yield is turned on then we requeue to the end of the tree.
889 static void yield_task_fair(struct rq *rq)
891 struct task_struct *curr = rq->curr;
892 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
893 struct sched_entity *rightmost, *se = &curr->se;
896 * Are we the only task in the tree?
898 if (unlikely(cfs_rq->nr_running == 1))
901 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
902 __update_rq_clock(rq);
904 * Update run-time statistics of the 'current'.
911 * Find the rightmost entry in the rbtree:
913 rightmost = __pick_last_entity(cfs_rq);
915 * Already in the rightmost position?
917 if (unlikely(rightmost->vruntime < se->vruntime))
921 * Minimally necessary key value to be last in the tree:
922 * Upon rescheduling, sched_class::put_prev_task() will place
923 * 'current' within the tree based on its new key value.
925 se->vruntime = rightmost->vruntime + 1;
929 * wake_idle() will wake a task on an idle cpu if task->cpu is
930 * not idle and an idle cpu is available. The span of cpus to
931 * search starts with cpus closest then further out as needed,
932 * so we always favor a closer, idle cpu.
934 * Returns the CPU we should wake onto.
936 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
937 static int wake_idle(int cpu, struct task_struct *p)
940 struct sched_domain *sd;
944 * If it is idle, then it is the best cpu to run this task.
946 * This cpu is also the best, if it has more than one task already.
947 * Siblings must be also busy(in most cases) as they didn't already
948 * pickup the extra load from this cpu and hence we need not check
949 * sibling runqueue info. This will avoid the checks and cache miss
950 * penalities associated with that.
952 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
955 for_each_domain(cpu, sd) {
956 if (sd->flags & SD_WAKE_IDLE) {
957 cpus_and(tmp, sd->span, p->cpus_allowed);
958 for_each_cpu_mask(i, tmp) {
960 if (i != task_cpu(p)) {
974 static inline int wake_idle(int cpu, struct task_struct *p)
982 static const struct sched_class fair_sched_class;
985 wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq,
986 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
987 int idx, unsigned long load, unsigned long this_load,
988 unsigned int imbalance)
990 struct task_struct *curr = this_rq->curr;
991 unsigned long tl = this_load;
992 unsigned long tl_per_task;
994 if (!(this_sd->flags & SD_WAKE_AFFINE))
998 * If the currently running task will sleep within
999 * a reasonable amount of time then attract this newly
1002 if (sync && curr->sched_class == &fair_sched_class) {
1003 if (curr->se.avg_overlap < sysctl_sched_migration_cost &&
1004 p->se.avg_overlap < sysctl_sched_migration_cost)
1008 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1009 tl_per_task = cpu_avg_load_per_task(this_cpu);
1012 * If sync wakeup then subtract the (maximum possible)
1013 * effect of the currently running task from the load
1014 * of the current CPU:
1017 tl -= current->se.load.weight;
1019 if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) ||
1020 100*(tl + p->se.load.weight) <= imbalance*load) {
1022 * This domain has SD_WAKE_AFFINE and
1023 * p is cache cold in this domain, and
1024 * there is no bad imbalance.
1026 schedstat_inc(this_sd, ttwu_move_affine);
1027 schedstat_inc(p, se.nr_wakeups_affine);
1034 static int select_task_rq_fair(struct task_struct *p, int sync)
1036 struct sched_domain *sd, *this_sd = NULL;
1037 int prev_cpu, this_cpu, new_cpu;
1038 unsigned long load, this_load;
1039 struct rq *rq, *this_rq;
1040 unsigned int imbalance;
1043 prev_cpu = task_cpu(p);
1045 this_cpu = smp_processor_id();
1046 this_rq = cpu_rq(this_cpu);
1050 * 'this_sd' is the first domain that both
1051 * this_cpu and prev_cpu are present in:
1053 for_each_domain(this_cpu, sd) {
1054 if (cpu_isset(prev_cpu, sd->span)) {
1060 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1064 * Check for affine wakeup and passive balancing possibilities.
1069 idx = this_sd->wake_idx;
1071 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1073 load = source_load(prev_cpu, idx);
1074 this_load = target_load(this_cpu, idx);
1076 if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1077 load, this_load, imbalance))
1080 if (prev_cpu == this_cpu)
1084 * Start passive balancing when half the imbalance_pct
1087 if (this_sd->flags & SD_WAKE_BALANCE) {
1088 if (imbalance*this_load <= 100*load) {
1089 schedstat_inc(this_sd, ttwu_move_balance);
1090 schedstat_inc(p, se.nr_wakeups_passive);
1096 return wake_idle(new_cpu, p);
1098 #endif /* CONFIG_SMP */
1100 static unsigned long wakeup_gran(struct sched_entity *se)
1102 unsigned long gran = sysctl_sched_wakeup_granularity;
1105 * More easily preempt - nice tasks, while not making
1106 * it harder for + nice tasks.
1108 if (unlikely(se->load.weight > NICE_0_LOAD))
1109 gran = calc_delta_fair(gran, &se->load);
1115 * Should 'se' preempt 'curr'.
1129 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1131 s64 gran, vdiff = curr->vruntime - se->vruntime;
1136 gran = wakeup_gran(curr);
1144 * Preempt the current task with a newly woken task if needed:
1146 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
1148 struct task_struct *curr = rq->curr;
1149 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1150 struct sched_entity *se = &curr->se, *pse = &p->se;
1152 if (unlikely(rt_prio(p->prio))) {
1153 update_rq_clock(rq);
1154 update_curr(cfs_rq);
1159 se->last_wakeup = se->sum_exec_runtime;
1160 if (unlikely(se == pse))
1163 cfs_rq_of(pse)->next = pse;
1166 * Batch tasks do not preempt (their preemption is driven by
1169 if (unlikely(p->policy == SCHED_BATCH))
1172 if (!sched_feat(WAKEUP_PREEMPT))
1175 while (!is_same_group(se, pse)) {
1176 se = parent_entity(se);
1177 pse = parent_entity(pse);
1180 if (wakeup_preempt_entity(se, pse) == 1)
1184 static struct task_struct *pick_next_task_fair(struct rq *rq)
1186 struct task_struct *p;
1187 struct cfs_rq *cfs_rq = &rq->cfs;
1188 struct sched_entity *se;
1190 if (unlikely(!cfs_rq->nr_running))
1194 se = pick_next_entity(cfs_rq);
1195 cfs_rq = group_cfs_rq(se);
1199 hrtick_start_fair(rq, p);
1205 * Account for a descheduled task:
1207 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1209 struct sched_entity *se = &prev->se;
1210 struct cfs_rq *cfs_rq;
1212 for_each_sched_entity(se) {
1213 cfs_rq = cfs_rq_of(se);
1214 put_prev_entity(cfs_rq, se);
1219 /**************************************************
1220 * Fair scheduling class load-balancing methods:
1224 * Load-balancing iterator. Note: while the runqueue stays locked
1225 * during the whole iteration, the current task might be
1226 * dequeued so the iterator has to be dequeue-safe. Here we
1227 * achieve that by always pre-iterating before returning
1230 static struct task_struct *
1231 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1233 struct task_struct *p;
1238 p = rb_entry(curr, struct task_struct, se.run_node);
1239 cfs_rq->rb_load_balance_curr = rb_next(curr);
1244 static struct task_struct *load_balance_start_fair(void *arg)
1246 struct cfs_rq *cfs_rq = arg;
1248 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1251 static struct task_struct *load_balance_next_fair(void *arg)
1253 struct cfs_rq *cfs_rq = arg;
1255 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1258 #ifdef CONFIG_FAIR_GROUP_SCHED
1259 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1261 struct sched_entity *curr;
1262 struct task_struct *p;
1264 if (!cfs_rq->nr_running || !first_fair(cfs_rq))
1267 curr = cfs_rq->curr;
1269 curr = __pick_next_entity(cfs_rq);
1277 static unsigned long
1278 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1279 unsigned long max_load_move,
1280 struct sched_domain *sd, enum cpu_idle_type idle,
1281 int *all_pinned, int *this_best_prio)
1283 struct cfs_rq *busy_cfs_rq;
1284 long rem_load_move = max_load_move;
1285 struct rq_iterator cfs_rq_iterator;
1287 cfs_rq_iterator.start = load_balance_start_fair;
1288 cfs_rq_iterator.next = load_balance_next_fair;
1290 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1291 #ifdef CONFIG_FAIR_GROUP_SCHED
1292 struct cfs_rq *this_cfs_rq;
1294 unsigned long maxload;
1296 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1298 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1299 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1303 /* Don't pull more than imbalance/2 */
1305 maxload = min(rem_load_move, imbalance);
1307 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1309 # define maxload rem_load_move
1312 * pass busy_cfs_rq argument into
1313 * load_balance_[start|next]_fair iterators
1315 cfs_rq_iterator.arg = busy_cfs_rq;
1316 rem_load_move -= balance_tasks(this_rq, this_cpu, busiest,
1317 maxload, sd, idle, all_pinned,
1321 if (rem_load_move <= 0)
1325 return max_load_move - rem_load_move;
1329 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1330 struct sched_domain *sd, enum cpu_idle_type idle)
1332 struct cfs_rq *busy_cfs_rq;
1333 struct rq_iterator cfs_rq_iterator;
1335 cfs_rq_iterator.start = load_balance_start_fair;
1336 cfs_rq_iterator.next = load_balance_next_fair;
1338 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1340 * pass busy_cfs_rq argument into
1341 * load_balance_[start|next]_fair iterators
1343 cfs_rq_iterator.arg = busy_cfs_rq;
1344 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1354 * scheduler tick hitting a task of our scheduling class:
1356 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1358 struct cfs_rq *cfs_rq;
1359 struct sched_entity *se = &curr->se;
1361 for_each_sched_entity(se) {
1362 cfs_rq = cfs_rq_of(se);
1363 entity_tick(cfs_rq, se, queued);
1367 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1370 * Share the fairness runtime between parent and child, thus the
1371 * total amount of pressure for CPU stays equal - new tasks
1372 * get a chance to run but frequent forkers are not allowed to
1373 * monopolize the CPU. Note: the parent runqueue is locked,
1374 * the child is not running yet.
1376 static void task_new_fair(struct rq *rq, struct task_struct *p)
1378 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1379 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1380 int this_cpu = smp_processor_id();
1382 sched_info_queued(p);
1384 update_curr(cfs_rq);
1385 place_entity(cfs_rq, se, 1);
1387 /* 'curr' will be NULL if the child belongs to a different group */
1388 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1389 curr && curr->vruntime < se->vruntime) {
1391 * Upon rescheduling, sched_class::put_prev_task() will place
1392 * 'current' within the tree based on its new key value.
1394 swap(curr->vruntime, se->vruntime);
1397 enqueue_task_fair(rq, p, 0);
1398 resched_task(rq->curr);
1402 * Priority of the task has changed. Check to see if we preempt
1405 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1406 int oldprio, int running)
1409 * Reschedule if we are currently running on this runqueue and
1410 * our priority decreased, or if we are not currently running on
1411 * this runqueue and our priority is higher than the current's
1414 if (p->prio > oldprio)
1415 resched_task(rq->curr);
1417 check_preempt_curr(rq, p);
1421 * We switched to the sched_fair class.
1423 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1427 * We were most likely switched from sched_rt, so
1428 * kick off the schedule if running, otherwise just see
1429 * if we can still preempt the current task.
1432 resched_task(rq->curr);
1434 check_preempt_curr(rq, p);
1437 /* Account for a task changing its policy or group.
1439 * This routine is mostly called to set cfs_rq->curr field when a task
1440 * migrates between groups/classes.
1442 static void set_curr_task_fair(struct rq *rq)
1444 struct sched_entity *se = &rq->curr->se;
1446 for_each_sched_entity(se)
1447 set_next_entity(cfs_rq_of(se), se);
1450 #ifdef CONFIG_FAIR_GROUP_SCHED
1451 static void moved_group_fair(struct task_struct *p)
1453 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1455 update_curr(cfs_rq);
1456 place_entity(cfs_rq, &p->se, 1);
1461 * All the scheduling class methods:
1463 static const struct sched_class fair_sched_class = {
1464 .next = &idle_sched_class,
1465 .enqueue_task = enqueue_task_fair,
1466 .dequeue_task = dequeue_task_fair,
1467 .yield_task = yield_task_fair,
1469 .select_task_rq = select_task_rq_fair,
1470 #endif /* CONFIG_SMP */
1472 .check_preempt_curr = check_preempt_wakeup,
1474 .pick_next_task = pick_next_task_fair,
1475 .put_prev_task = put_prev_task_fair,
1478 .load_balance = load_balance_fair,
1479 .move_one_task = move_one_task_fair,
1482 .set_curr_task = set_curr_task_fair,
1483 .task_tick = task_tick_fair,
1484 .task_new = task_new_fair,
1486 .prio_changed = prio_changed_fair,
1487 .switched_to = switched_to_fair,
1489 #ifdef CONFIG_FAIR_GROUP_SCHED
1490 .moved_group = moved_group_fair,
1494 #ifdef CONFIG_SCHED_DEBUG
1495 static void print_cfs_stats(struct seq_file *m, int cpu)
1497 struct cfs_rq *cfs_rq;
1499 #ifdef CONFIG_FAIR_GROUP_SCHED
1500 print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs);
1503 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1504 print_cfs_rq(m, cpu, cfs_rq);