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 u64 slice = __sched_period(cfs_rq->nr_running);
288 slice *= se->load.weight;
289 do_div(slice, cfs_rq->load.weight);
295 * We calculate the vruntime slice.
299 static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running)
301 u64 vslice = __sched_period(nr_running);
303 vslice *= NICE_0_LOAD;
304 do_div(vslice, rq_weight);
309 static u64 sched_vslice(struct cfs_rq *cfs_rq)
311 return __sched_vslice(cfs_rq->load.weight, cfs_rq->nr_running);
314 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
316 return __sched_vslice(cfs_rq->load.weight + se->load.weight,
317 cfs_rq->nr_running + 1);
321 * Update the current task's runtime statistics. Skip current tasks that
322 * are not in our scheduling class.
325 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
326 unsigned long delta_exec)
328 unsigned long delta_exec_weighted;
330 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
332 curr->sum_exec_runtime += delta_exec;
333 schedstat_add(cfs_rq, exec_clock, delta_exec);
334 delta_exec_weighted = delta_exec;
335 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
336 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
339 curr->vruntime += delta_exec_weighted;
342 static void update_curr(struct cfs_rq *cfs_rq)
344 struct sched_entity *curr = cfs_rq->curr;
345 u64 now = rq_of(cfs_rq)->clock;
346 unsigned long delta_exec;
352 * Get the amount of time the current task was running
353 * since the last time we changed load (this cannot
354 * overflow on 32 bits):
356 delta_exec = (unsigned long)(now - curr->exec_start);
358 __update_curr(cfs_rq, curr, delta_exec);
359 curr->exec_start = now;
361 if (entity_is_task(curr)) {
362 struct task_struct *curtask = task_of(curr);
364 cpuacct_charge(curtask, delta_exec);
369 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
371 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
375 * Task is being enqueued - update stats:
377 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
380 * Are we enqueueing a waiting task? (for current tasks
381 * a dequeue/enqueue event is a NOP)
383 if (se != cfs_rq->curr)
384 update_stats_wait_start(cfs_rq, se);
388 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
390 schedstat_set(se->wait_max, max(se->wait_max,
391 rq_of(cfs_rq)->clock - se->wait_start));
392 schedstat_set(se->wait_count, se->wait_count + 1);
393 schedstat_set(se->wait_sum, se->wait_sum +
394 rq_of(cfs_rq)->clock - se->wait_start);
395 schedstat_set(se->wait_start, 0);
399 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
402 * Mark the end of the wait period if dequeueing a
405 if (se != cfs_rq->curr)
406 update_stats_wait_end(cfs_rq, se);
410 * We are picking a new current task - update its stats:
413 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
416 * We are starting a new run period:
418 se->exec_start = rq_of(cfs_rq)->clock;
421 /**************************************************
422 * Scheduling class queueing methods:
426 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
428 update_load_add(&cfs_rq->load, se->load.weight);
429 cfs_rq->nr_running++;
434 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
436 update_load_sub(&cfs_rq->load, se->load.weight);
437 cfs_rq->nr_running--;
441 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
443 #ifdef CONFIG_SCHEDSTATS
444 if (se->sleep_start) {
445 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
446 struct task_struct *tsk = task_of(se);
451 if (unlikely(delta > se->sleep_max))
452 se->sleep_max = delta;
455 se->sum_sleep_runtime += delta;
457 account_scheduler_latency(tsk, delta >> 10, 1);
459 if (se->block_start) {
460 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
461 struct task_struct *tsk = task_of(se);
466 if (unlikely(delta > se->block_max))
467 se->block_max = delta;
470 se->sum_sleep_runtime += delta;
473 * Blocking time is in units of nanosecs, so shift by 20 to
474 * get a milliseconds-range estimation of the amount of
475 * time that the task spent sleeping:
477 if (unlikely(prof_on == SLEEP_PROFILING)) {
479 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
482 account_scheduler_latency(tsk, delta >> 10, 0);
487 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 #ifdef CONFIG_SCHED_DEBUG
490 s64 d = se->vruntime - cfs_rq->min_vruntime;
495 if (d > 3*sysctl_sched_latency)
496 schedstat_inc(cfs_rq, nr_spread_over);
501 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
505 if (first_fair(cfs_rq)) {
506 vruntime = min_vruntime(cfs_rq->min_vruntime,
507 __pick_next_entity(cfs_rq)->vruntime);
509 vruntime = cfs_rq->min_vruntime;
511 if (sched_feat(TREE_AVG)) {
512 struct sched_entity *last = __pick_last_entity(cfs_rq);
514 vruntime += last->vruntime;
517 } else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running)
518 vruntime += sched_vslice(cfs_rq)/2;
521 * The 'current' period is already promised to the current tasks,
522 * however the extra weight of the new task will slow them down a
523 * little, place the new task so that it fits in the slot that
524 * stays open at the end.
526 if (initial && sched_feat(START_DEBIT))
527 vruntime += sched_vslice_add(cfs_rq, se);
530 /* sleeps upto a single latency don't count. */
531 if (sched_feat(NEW_FAIR_SLEEPERS))
532 vruntime -= sysctl_sched_latency;
534 /* ensure we never gain time by being placed backwards. */
535 vruntime = max_vruntime(se->vruntime, vruntime);
538 se->vruntime = vruntime;
542 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
545 * Update run-time statistics of the 'current'.
550 place_entity(cfs_rq, se, 0);
551 enqueue_sleeper(cfs_rq, se);
554 update_stats_enqueue(cfs_rq, se);
555 check_spread(cfs_rq, se);
556 if (se != cfs_rq->curr)
557 __enqueue_entity(cfs_rq, se);
558 account_entity_enqueue(cfs_rq, se);
562 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
565 * Update run-time statistics of the 'current'.
569 update_stats_dequeue(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;
632 static struct sched_entity *
633 pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 diff = cfs_rq->next->vruntime - se->vruntime;
644 gran = calc_delta_fair(sysctl_sched_wakeup_granularity, &cfs_rq->load);
651 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
653 struct sched_entity *se = NULL;
655 if (first_fair(cfs_rq)) {
656 se = __pick_next_entity(cfs_rq);
657 se = pick_next(cfs_rq, se);
658 set_next_entity(cfs_rq, se);
664 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
667 * If still on the runqueue then deactivate_task()
668 * was not called and update_curr() has to be done:
673 check_spread(cfs_rq, prev);
675 update_stats_wait_start(cfs_rq, prev);
676 /* Put 'current' back into the tree. */
677 __enqueue_entity(cfs_rq, prev);
683 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
686 * Update run-time statistics of the 'current'.
690 #ifdef CONFIG_SCHED_HRTICK
692 * queued ticks are scheduled to match the slice, so don't bother
693 * validating it and just reschedule.
696 return resched_task(rq_of(cfs_rq)->curr);
698 * don't let the period tick interfere with the hrtick preemption
700 if (!sched_feat(DOUBLE_TICK) &&
701 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
705 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
706 check_preempt_tick(cfs_rq, curr);
709 /**************************************************
710 * CFS operations on tasks:
713 #ifdef CONFIG_FAIR_GROUP_SCHED
715 /* Walk up scheduling entities hierarchy */
716 #define for_each_sched_entity(se) \
717 for (; se; se = se->parent)
719 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
724 /* runqueue on which this entity is (to be) queued */
725 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
730 /* runqueue "owned" by this group */
731 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
736 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
737 * another cpu ('this_cpu')
739 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
741 return cfs_rq->tg->cfs_rq[this_cpu];
744 /* Iterate thr' all leaf cfs_rq's on a runqueue */
745 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
746 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
748 /* Do the two (enqueued) entities belong to the same group ? */
750 is_same_group(struct sched_entity *se, struct sched_entity *pse)
752 if (se->cfs_rq == pse->cfs_rq)
758 static inline struct sched_entity *parent_entity(struct sched_entity *se)
763 #else /* CONFIG_FAIR_GROUP_SCHED */
765 #define for_each_sched_entity(se) \
766 for (; se; se = NULL)
768 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
770 return &task_rq(p)->cfs;
773 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
775 struct task_struct *p = task_of(se);
776 struct rq *rq = task_rq(p);
781 /* runqueue "owned" by this group */
782 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
787 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
789 return &cpu_rq(this_cpu)->cfs;
792 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
793 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
796 is_same_group(struct sched_entity *se, struct sched_entity *pse)
801 static inline struct sched_entity *parent_entity(struct sched_entity *se)
806 #endif /* CONFIG_FAIR_GROUP_SCHED */
808 #ifdef CONFIG_SCHED_HRTICK
809 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
811 int requeue = rq->curr == p;
812 struct sched_entity *se = &p->se;
813 struct cfs_rq *cfs_rq = cfs_rq_of(se);
815 WARN_ON(task_rq(p) != rq);
817 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
818 u64 slice = sched_slice(cfs_rq, se);
819 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
820 s64 delta = slice - ran;
829 * Don't schedule slices shorter than 10000ns, that just
830 * doesn't make sense. Rely on vruntime for fairness.
833 delta = max(10000LL, delta);
835 hrtick_start(rq, delta, requeue);
840 hrtick_start_fair(struct rq *rq, struct task_struct *p)
846 * The enqueue_task method is called before nr_running is
847 * increased. Here we update the fair scheduling stats and
848 * then put the task into the rbtree:
850 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
852 struct cfs_rq *cfs_rq;
853 struct sched_entity *se = &p->se;
855 for_each_sched_entity(se) {
858 cfs_rq = cfs_rq_of(se);
859 enqueue_entity(cfs_rq, se, wakeup);
863 hrtick_start_fair(rq, rq->curr);
867 * The dequeue_task method is called before nr_running is
868 * decreased. We remove the task from the rbtree and
869 * update the fair scheduling stats:
871 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
873 struct cfs_rq *cfs_rq;
874 struct sched_entity *se = &p->se;
876 for_each_sched_entity(se) {
877 cfs_rq = cfs_rq_of(se);
878 dequeue_entity(cfs_rq, se, sleep);
879 /* Don't dequeue parent if it has other entities besides us */
880 if (cfs_rq->load.weight)
885 hrtick_start_fair(rq, rq->curr);
889 * sched_yield() support is very simple - we dequeue and enqueue.
891 * If compat_yield is turned on then we requeue to the end of the tree.
893 static void yield_task_fair(struct rq *rq)
895 struct task_struct *curr = rq->curr;
896 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
897 struct sched_entity *rightmost, *se = &curr->se;
900 * Are we the only task in the tree?
902 if (unlikely(cfs_rq->nr_running == 1))
905 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
906 __update_rq_clock(rq);
908 * Update run-time statistics of the 'current'.
915 * Find the rightmost entry in the rbtree:
917 rightmost = __pick_last_entity(cfs_rq);
919 * Already in the rightmost position?
921 if (unlikely(rightmost->vruntime < se->vruntime))
925 * Minimally necessary key value to be last in the tree:
926 * Upon rescheduling, sched_class::put_prev_task() will place
927 * 'current' within the tree based on its new key value.
929 se->vruntime = rightmost->vruntime + 1;
933 * wake_idle() will wake a task on an idle cpu if task->cpu is
934 * not idle and an idle cpu is available. The span of cpus to
935 * search starts with cpus closest then further out as needed,
936 * so we always favor a closer, idle cpu.
938 * Returns the CPU we should wake onto.
940 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
941 static int wake_idle(int cpu, struct task_struct *p)
944 struct sched_domain *sd;
948 * If it is idle, then it is the best cpu to run this task.
950 * This cpu is also the best, if it has more than one task already.
951 * Siblings must be also busy(in most cases) as they didn't already
952 * pickup the extra load from this cpu and hence we need not check
953 * sibling runqueue info. This will avoid the checks and cache miss
954 * penalities associated with that.
956 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
959 for_each_domain(cpu, sd) {
960 if (sd->flags & SD_WAKE_IDLE) {
961 cpus_and(tmp, sd->span, p->cpus_allowed);
962 for_each_cpu_mask(i, tmp) {
964 if (i != task_cpu(p)) {
978 static inline int wake_idle(int cpu, struct task_struct *p)
985 static int select_task_rq_fair(struct task_struct *p, int sync)
989 struct sched_domain *sd, *this_sd = NULL;
994 this_cpu = smp_processor_id();
1000 for_each_domain(this_cpu, sd) {
1001 if (cpu_isset(cpu, sd->span)) {
1007 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1011 * Check for affine wakeup and passive balancing possibilities.
1014 int idx = this_sd->wake_idx;
1015 unsigned int imbalance;
1016 unsigned long load, this_load;
1018 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1020 load = source_load(cpu, idx);
1021 this_load = target_load(this_cpu, idx);
1023 new_cpu = this_cpu; /* Wake to this CPU if we can */
1025 if (this_sd->flags & SD_WAKE_AFFINE) {
1026 unsigned long tl = this_load;
1027 unsigned long tl_per_task;
1030 * Attract cache-cold tasks on sync wakeups:
1032 if (sync && !task_hot(p, rq->clock, this_sd))
1035 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1036 tl_per_task = cpu_avg_load_per_task(this_cpu);
1039 * If sync wakeup then subtract the (maximum possible)
1040 * effect of the currently running task from the load
1041 * of the current CPU:
1044 tl -= current->se.load.weight;
1047 tl + target_load(cpu, idx) <= tl_per_task) ||
1048 100*(tl + p->se.load.weight) <= imbalance*load) {
1050 * This domain has SD_WAKE_AFFINE and
1051 * p is cache cold in this domain, and
1052 * there is no bad imbalance.
1054 schedstat_inc(this_sd, ttwu_move_affine);
1055 schedstat_inc(p, se.nr_wakeups_affine);
1061 * Start passive balancing when half the imbalance_pct
1064 if (this_sd->flags & SD_WAKE_BALANCE) {
1065 if (imbalance*this_load <= 100*load) {
1066 schedstat_inc(this_sd, ttwu_move_balance);
1067 schedstat_inc(p, se.nr_wakeups_passive);
1073 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
1075 return wake_idle(new_cpu, p);
1077 #endif /* CONFIG_SMP */
1081 * Preempt the current task with a newly woken task if needed:
1083 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
1085 struct task_struct *curr = rq->curr;
1086 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1087 struct sched_entity *se = &curr->se, *pse = &p->se;
1090 if (unlikely(rt_prio(p->prio))) {
1091 update_rq_clock(rq);
1092 update_curr(cfs_rq);
1097 cfs_rq_of(pse)->next = pse;
1100 * Batch tasks do not preempt (their preemption is driven by
1103 if (unlikely(p->policy == SCHED_BATCH))
1106 if (!sched_feat(WAKEUP_PREEMPT))
1109 while (!is_same_group(se, pse)) {
1110 se = parent_entity(se);
1111 pse = parent_entity(pse);
1114 gran = sysctl_sched_wakeup_granularity;
1116 * More easily preempt - nice tasks, while not making
1117 * it harder for + nice tasks.
1119 if (unlikely(se->load.weight > NICE_0_LOAD))
1120 gran = calc_delta_fair(gran, &se->load);
1122 if (pse->vruntime + gran < se->vruntime)
1126 static struct task_struct *pick_next_task_fair(struct rq *rq)
1128 struct task_struct *p;
1129 struct cfs_rq *cfs_rq = &rq->cfs;
1130 struct sched_entity *se;
1132 if (unlikely(!cfs_rq->nr_running))
1136 se = pick_next_entity(cfs_rq);
1137 cfs_rq = group_cfs_rq(se);
1141 hrtick_start_fair(rq, p);
1147 * Account for a descheduled task:
1149 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1151 struct sched_entity *se = &prev->se;
1152 struct cfs_rq *cfs_rq;
1154 for_each_sched_entity(se) {
1155 cfs_rq = cfs_rq_of(se);
1156 put_prev_entity(cfs_rq, se);
1161 /**************************************************
1162 * Fair scheduling class load-balancing methods:
1166 * Load-balancing iterator. Note: while the runqueue stays locked
1167 * during the whole iteration, the current task might be
1168 * dequeued so the iterator has to be dequeue-safe. Here we
1169 * achieve that by always pre-iterating before returning
1172 static struct task_struct *
1173 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1175 struct task_struct *p;
1180 p = rb_entry(curr, struct task_struct, se.run_node);
1181 cfs_rq->rb_load_balance_curr = rb_next(curr);
1186 static struct task_struct *load_balance_start_fair(void *arg)
1188 struct cfs_rq *cfs_rq = arg;
1190 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1193 static struct task_struct *load_balance_next_fair(void *arg)
1195 struct cfs_rq *cfs_rq = arg;
1197 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1200 #ifdef CONFIG_FAIR_GROUP_SCHED
1201 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1203 struct sched_entity *curr;
1204 struct task_struct *p;
1206 if (!cfs_rq->nr_running || !first_fair(cfs_rq))
1209 curr = cfs_rq->curr;
1211 curr = __pick_next_entity(cfs_rq);
1219 static unsigned long
1220 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1221 unsigned long max_load_move,
1222 struct sched_domain *sd, enum cpu_idle_type idle,
1223 int *all_pinned, int *this_best_prio)
1225 struct cfs_rq *busy_cfs_rq;
1226 long rem_load_move = max_load_move;
1227 struct rq_iterator cfs_rq_iterator;
1229 cfs_rq_iterator.start = load_balance_start_fair;
1230 cfs_rq_iterator.next = load_balance_next_fair;
1232 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1233 #ifdef CONFIG_FAIR_GROUP_SCHED
1234 struct cfs_rq *this_cfs_rq;
1236 unsigned long maxload;
1238 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1240 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1241 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1245 /* Don't pull more than imbalance/2 */
1247 maxload = min(rem_load_move, imbalance);
1249 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1251 # define maxload rem_load_move
1254 * pass busy_cfs_rq argument into
1255 * load_balance_[start|next]_fair iterators
1257 cfs_rq_iterator.arg = busy_cfs_rq;
1258 rem_load_move -= balance_tasks(this_rq, this_cpu, busiest,
1259 maxload, sd, idle, all_pinned,
1263 if (rem_load_move <= 0)
1267 return max_load_move - rem_load_move;
1271 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1272 struct sched_domain *sd, enum cpu_idle_type idle)
1274 struct cfs_rq *busy_cfs_rq;
1275 struct rq_iterator cfs_rq_iterator;
1277 cfs_rq_iterator.start = load_balance_start_fair;
1278 cfs_rq_iterator.next = load_balance_next_fair;
1280 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1282 * pass busy_cfs_rq argument into
1283 * load_balance_[start|next]_fair iterators
1285 cfs_rq_iterator.arg = busy_cfs_rq;
1286 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1296 * scheduler tick hitting a task of our scheduling class:
1298 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1300 struct cfs_rq *cfs_rq;
1301 struct sched_entity *se = &curr->se;
1303 for_each_sched_entity(se) {
1304 cfs_rq = cfs_rq_of(se);
1305 entity_tick(cfs_rq, se, queued);
1309 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1312 * Share the fairness runtime between parent and child, thus the
1313 * total amount of pressure for CPU stays equal - new tasks
1314 * get a chance to run but frequent forkers are not allowed to
1315 * monopolize the CPU. Note: the parent runqueue is locked,
1316 * the child is not running yet.
1318 static void task_new_fair(struct rq *rq, struct task_struct *p)
1320 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1321 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1322 int this_cpu = smp_processor_id();
1324 sched_info_queued(p);
1326 update_curr(cfs_rq);
1327 place_entity(cfs_rq, se, 1);
1329 /* 'curr' will be NULL if the child belongs to a different group */
1330 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1331 curr && curr->vruntime < se->vruntime) {
1333 * Upon rescheduling, sched_class::put_prev_task() will place
1334 * 'current' within the tree based on its new key value.
1336 swap(curr->vruntime, se->vruntime);
1339 enqueue_task_fair(rq, p, 0);
1340 resched_task(rq->curr);
1344 * Priority of the task has changed. Check to see if we preempt
1347 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1348 int oldprio, int running)
1351 * Reschedule if we are currently running on this runqueue and
1352 * our priority decreased, or if we are not currently running on
1353 * this runqueue and our priority is higher than the current's
1356 if (p->prio > oldprio)
1357 resched_task(rq->curr);
1359 check_preempt_curr(rq, p);
1363 * We switched to the sched_fair class.
1365 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1369 * We were most likely switched from sched_rt, so
1370 * kick off the schedule if running, otherwise just see
1371 * if we can still preempt the current task.
1374 resched_task(rq->curr);
1376 check_preempt_curr(rq, p);
1379 /* Account for a task changing its policy or group.
1381 * This routine is mostly called to set cfs_rq->curr field when a task
1382 * migrates between groups/classes.
1384 static void set_curr_task_fair(struct rq *rq)
1386 struct sched_entity *se = &rq->curr->se;
1388 for_each_sched_entity(se)
1389 set_next_entity(cfs_rq_of(se), se);
1392 #ifdef CONFIG_FAIR_GROUP_SCHED
1393 static void moved_group_fair(struct task_struct *p)
1395 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1397 update_curr(cfs_rq);
1398 place_entity(cfs_rq, &p->se, 1);
1403 * All the scheduling class methods:
1405 static const struct sched_class fair_sched_class = {
1406 .next = &idle_sched_class,
1407 .enqueue_task = enqueue_task_fair,
1408 .dequeue_task = dequeue_task_fair,
1409 .yield_task = yield_task_fair,
1411 .select_task_rq = select_task_rq_fair,
1412 #endif /* CONFIG_SMP */
1414 .check_preempt_curr = check_preempt_wakeup,
1416 .pick_next_task = pick_next_task_fair,
1417 .put_prev_task = put_prev_task_fair,
1420 .load_balance = load_balance_fair,
1421 .move_one_task = move_one_task_fair,
1424 .set_curr_task = set_curr_task_fair,
1425 .task_tick = task_tick_fair,
1426 .task_new = task_new_fair,
1428 .prio_changed = prio_changed_fair,
1429 .switched_to = switched_to_fair,
1431 #ifdef CONFIG_FAIR_GROUP_SCHED
1432 .moved_group = moved_group_fair,
1436 #ifdef CONFIG_SCHED_DEBUG
1437 static void print_cfs_stats(struct seq_file *m, int cpu)
1439 struct cfs_rq *cfs_rq;
1441 #ifdef CONFIG_FAIR_GROUP_SCHED
1442 print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs);
1445 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1446 print_cfs_rq(m, cpu, cfs_rq);