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: 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_wakeup_granularity = 10000000UL;
74 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
76 /**************************************************************
77 * CFS operations on generic schedulable entities:
80 #ifdef CONFIG_FAIR_GROUP_SCHED
82 /* cpu runqueue to which this cfs_rq is attached */
83 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
88 /* An entity is a task if it doesn't "own" a runqueue */
89 #define entity_is_task(se) (!se->my_q)
91 #else /* CONFIG_FAIR_GROUP_SCHED */
93 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
95 return container_of(cfs_rq, struct rq, cfs);
98 #define entity_is_task(se) 1
100 #endif /* CONFIG_FAIR_GROUP_SCHED */
102 static inline struct task_struct *task_of(struct sched_entity *se)
104 return container_of(se, struct task_struct, se);
108 /**************************************************************
109 * Scheduling class tree data structure manipulation methods:
112 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
114 s64 delta = (s64)(vruntime - min_vruntime);
116 min_vruntime = vruntime;
121 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
123 s64 delta = (s64)(vruntime - min_vruntime);
125 min_vruntime = vruntime;
130 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
132 return se->vruntime - cfs_rq->min_vruntime;
136 * Enqueue an entity into the rb-tree:
138 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
140 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
141 struct rb_node *parent = NULL;
142 struct sched_entity *entry;
143 s64 key = entity_key(cfs_rq, se);
147 * Find the right place in the rbtree:
151 entry = rb_entry(parent, struct sched_entity, run_node);
153 * We dont care about collisions. Nodes with
154 * the same key stay together.
156 if (key < entity_key(cfs_rq, entry)) {
157 link = &parent->rb_left;
159 link = &parent->rb_right;
165 * Maintain a cache of leftmost tree entries (it is frequently
169 cfs_rq->rb_leftmost = &se->run_node;
171 * maintain cfs_rq->min_vruntime to be a monotonic increasing
172 * value tracking the leftmost vruntime in the tree.
174 cfs_rq->min_vruntime =
175 max_vruntime(cfs_rq->min_vruntime, se->vruntime);
178 rb_link_node(&se->run_node, parent, link);
179 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
182 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
184 if (cfs_rq->rb_leftmost == &se->run_node) {
185 struct rb_node *next_node;
186 struct sched_entity *next;
188 next_node = rb_next(&se->run_node);
189 cfs_rq->rb_leftmost = next_node;
192 next = rb_entry(next_node,
193 struct sched_entity, run_node);
194 cfs_rq->min_vruntime =
195 max_vruntime(cfs_rq->min_vruntime,
200 if (cfs_rq->next == se)
203 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
206 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
208 return cfs_rq->rb_leftmost;
211 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
213 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
216 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
218 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
223 return rb_entry(last, struct sched_entity, run_node);
226 /**************************************************************
227 * Scheduling class statistics methods:
230 #ifdef CONFIG_SCHED_DEBUG
231 int sched_nr_latency_handler(struct ctl_table *table, int write,
232 struct file *filp, void __user *buffer, size_t *lenp,
235 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
240 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
241 sysctl_sched_min_granularity);
248 * The idea is to set a period in which each task runs once.
250 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
251 * this period because otherwise the slices get too small.
253 * p = (nr <= nl) ? l : l*nr/nl
255 static u64 __sched_period(unsigned long nr_running)
257 u64 period = sysctl_sched_latency;
258 unsigned long nr_latency = sched_nr_latency;
260 if (unlikely(nr_running > nr_latency)) {
261 period = sysctl_sched_min_granularity;
262 period *= nr_running;
269 * We calculate the wall-time slice from the period by taking a part
270 * proportional to the weight.
274 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
276 return calc_delta_mine(__sched_period(cfs_rq->nr_running),
277 se->load.weight, &cfs_rq->load);
281 * We calculate the vruntime slice.
285 static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running)
287 u64 vslice = __sched_period(nr_running);
289 vslice *= NICE_0_LOAD;
290 do_div(vslice, rq_weight);
295 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
297 return __sched_vslice(cfs_rq->load.weight + se->load.weight,
298 cfs_rq->nr_running + 1);
302 * Update the current task's runtime statistics. Skip current tasks that
303 * are not in our scheduling class.
306 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
307 unsigned long delta_exec)
309 unsigned long delta_exec_weighted;
311 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
313 curr->sum_exec_runtime += delta_exec;
314 schedstat_add(cfs_rq, exec_clock, delta_exec);
315 delta_exec_weighted = delta_exec;
316 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
317 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
320 curr->vruntime += delta_exec_weighted;
323 static void update_curr(struct cfs_rq *cfs_rq)
325 struct sched_entity *curr = cfs_rq->curr;
326 u64 now = rq_of(cfs_rq)->clock;
327 unsigned long delta_exec;
333 * Get the amount of time the current task was running
334 * since the last time we changed load (this cannot
335 * overflow on 32 bits):
337 delta_exec = (unsigned long)(now - curr->exec_start);
339 __update_curr(cfs_rq, curr, delta_exec);
340 curr->exec_start = now;
342 if (entity_is_task(curr)) {
343 struct task_struct *curtask = task_of(curr);
345 cpuacct_charge(curtask, delta_exec);
350 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
352 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
356 * Task is being enqueued - update stats:
358 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
361 * Are we enqueueing a waiting task? (for current tasks
362 * a dequeue/enqueue event is a NOP)
364 if (se != cfs_rq->curr)
365 update_stats_wait_start(cfs_rq, se);
369 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
371 schedstat_set(se->wait_max, max(se->wait_max,
372 rq_of(cfs_rq)->clock - se->wait_start));
373 schedstat_set(se->wait_count, se->wait_count + 1);
374 schedstat_set(se->wait_sum, se->wait_sum +
375 rq_of(cfs_rq)->clock - se->wait_start);
376 schedstat_set(se->wait_start, 0);
380 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
383 * Mark the end of the wait period if dequeueing a
386 if (se != cfs_rq->curr)
387 update_stats_wait_end(cfs_rq, se);
391 * We are picking a new current task - update its stats:
394 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
397 * We are starting a new run period:
399 se->exec_start = rq_of(cfs_rq)->clock;
402 /**************************************************
403 * Scheduling class queueing methods:
407 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
409 update_load_add(&cfs_rq->load, se->load.weight);
410 cfs_rq->nr_running++;
415 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
417 update_load_sub(&cfs_rq->load, se->load.weight);
418 cfs_rq->nr_running--;
422 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
424 #ifdef CONFIG_SCHEDSTATS
425 if (se->sleep_start) {
426 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
427 struct task_struct *tsk = task_of(se);
432 if (unlikely(delta > se->sleep_max))
433 se->sleep_max = delta;
436 se->sum_sleep_runtime += delta;
438 account_scheduler_latency(tsk, delta >> 10, 1);
440 if (se->block_start) {
441 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
442 struct task_struct *tsk = task_of(se);
447 if (unlikely(delta > se->block_max))
448 se->block_max = delta;
451 se->sum_sleep_runtime += delta;
454 * Blocking time is in units of nanosecs, so shift by 20 to
455 * get a milliseconds-range estimation of the amount of
456 * time that the task spent sleeping:
458 if (unlikely(prof_on == SLEEP_PROFILING)) {
460 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
463 account_scheduler_latency(tsk, delta >> 10, 0);
468 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
470 #ifdef CONFIG_SCHED_DEBUG
471 s64 d = se->vruntime - cfs_rq->min_vruntime;
476 if (d > 3*sysctl_sched_latency)
477 schedstat_inc(cfs_rq, nr_spread_over);
482 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
486 if (first_fair(cfs_rq)) {
487 vruntime = min_vruntime(cfs_rq->min_vruntime,
488 __pick_next_entity(cfs_rq)->vruntime);
490 vruntime = cfs_rq->min_vruntime;
493 * The 'current' period is already promised to the current tasks,
494 * however the extra weight of the new task will slow them down a
495 * little, place the new task so that it fits in the slot that
496 * stays open at the end.
498 if (initial && sched_feat(START_DEBIT))
499 vruntime += sched_vslice_add(cfs_rq, se);
502 /* sleeps upto a single latency don't count. */
503 if (sched_feat(NEW_FAIR_SLEEPERS)) {
504 if (sched_feat(NORMALIZED_SLEEPER))
505 vruntime -= calc_delta_fair(sysctl_sched_latency,
508 vruntime -= sysctl_sched_latency;
511 /* ensure we never gain time by being placed backwards. */
512 vruntime = max_vruntime(se->vruntime, vruntime);
515 se->vruntime = vruntime;
519 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
522 * Update run-time statistics of the 'current'.
527 place_entity(cfs_rq, se, 0);
528 enqueue_sleeper(cfs_rq, se);
531 update_stats_enqueue(cfs_rq, se);
532 check_spread(cfs_rq, se);
533 if (se != cfs_rq->curr)
534 __enqueue_entity(cfs_rq, se);
535 account_entity_enqueue(cfs_rq, se);
538 static void update_avg(u64 *avg, u64 sample)
540 s64 diff = sample - *avg;
544 static void update_avg_stats(struct cfs_rq *cfs_rq, struct sched_entity *se)
546 if (!se->last_wakeup)
549 update_avg(&se->avg_overlap, se->sum_exec_runtime - se->last_wakeup);
554 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
557 * Update run-time statistics of the 'current'.
561 update_stats_dequeue(cfs_rq, se);
563 update_avg_stats(cfs_rq, se);
564 #ifdef CONFIG_SCHEDSTATS
565 if (entity_is_task(se)) {
566 struct task_struct *tsk = task_of(se);
568 if (tsk->state & TASK_INTERRUPTIBLE)
569 se->sleep_start = rq_of(cfs_rq)->clock;
570 if (tsk->state & TASK_UNINTERRUPTIBLE)
571 se->block_start = rq_of(cfs_rq)->clock;
576 if (se != cfs_rq->curr)
577 __dequeue_entity(cfs_rq, se);
578 account_entity_dequeue(cfs_rq, se);
582 * Preempt the current task with a newly woken task if needed:
585 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
587 unsigned long ideal_runtime, delta_exec;
589 ideal_runtime = sched_slice(cfs_rq, curr);
590 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
591 if (delta_exec > ideal_runtime)
592 resched_task(rq_of(cfs_rq)->curr);
596 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
598 /* 'current' is not kept within the tree. */
601 * Any task has to be enqueued before it get to execute on
602 * a CPU. So account for the time it spent waiting on the
605 update_stats_wait_end(cfs_rq, se);
606 __dequeue_entity(cfs_rq, se);
609 update_stats_curr_start(cfs_rq, se);
611 #ifdef CONFIG_SCHEDSTATS
613 * Track our maximum slice length, if the CPU's load is at
614 * least twice that of our own weight (i.e. dont track it
615 * when there are only lesser-weight tasks around):
617 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
618 se->slice_max = max(se->slice_max,
619 se->sum_exec_runtime - se->prev_sum_exec_runtime);
622 se->prev_sum_exec_runtime = se->sum_exec_runtime;
626 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
628 static struct sched_entity *
629 pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se)
634 if (wakeup_preempt_entity(cfs_rq->next, se) != 0)
640 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
642 struct sched_entity *se = NULL;
644 if (first_fair(cfs_rq)) {
645 se = __pick_next_entity(cfs_rq);
646 se = pick_next(cfs_rq, se);
647 set_next_entity(cfs_rq, se);
653 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
656 * If still on the runqueue then deactivate_task()
657 * was not called and update_curr() has to be done:
662 check_spread(cfs_rq, prev);
664 update_stats_wait_start(cfs_rq, prev);
665 /* Put 'current' back into the tree. */
666 __enqueue_entity(cfs_rq, prev);
672 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
675 * Update run-time statistics of the 'current'.
679 #ifdef CONFIG_SCHED_HRTICK
681 * queued ticks are scheduled to match the slice, so don't bother
682 * validating it and just reschedule.
685 return resched_task(rq_of(cfs_rq)->curr);
687 * don't let the period tick interfere with the hrtick preemption
689 if (!sched_feat(DOUBLE_TICK) &&
690 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
694 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
695 check_preempt_tick(cfs_rq, curr);
698 /**************************************************
699 * CFS operations on tasks:
702 #ifdef CONFIG_FAIR_GROUP_SCHED
704 /* Walk up scheduling entities hierarchy */
705 #define for_each_sched_entity(se) \
706 for (; se; se = se->parent)
708 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
713 /* runqueue on which this entity is (to be) queued */
714 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
719 /* runqueue "owned" by this group */
720 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
725 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
726 * another cpu ('this_cpu')
728 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
730 return cfs_rq->tg->cfs_rq[this_cpu];
733 /* Iterate thr' all leaf cfs_rq's on a runqueue */
734 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
735 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
737 /* Do the two (enqueued) entities belong to the same group ? */
739 is_same_group(struct sched_entity *se, struct sched_entity *pse)
741 if (se->cfs_rq == pse->cfs_rq)
747 static inline struct sched_entity *parent_entity(struct sched_entity *se)
752 #else /* CONFIG_FAIR_GROUP_SCHED */
754 #define for_each_sched_entity(se) \
755 for (; se; se = NULL)
757 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
759 return &task_rq(p)->cfs;
762 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
764 struct task_struct *p = task_of(se);
765 struct rq *rq = task_rq(p);
770 /* runqueue "owned" by this group */
771 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
776 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
778 return &cpu_rq(this_cpu)->cfs;
781 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
782 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
785 is_same_group(struct sched_entity *se, struct sched_entity *pse)
790 static inline struct sched_entity *parent_entity(struct sched_entity *se)
795 #endif /* CONFIG_FAIR_GROUP_SCHED */
797 #ifdef CONFIG_SCHED_HRTICK
798 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
800 int requeue = rq->curr == p;
801 struct sched_entity *se = &p->se;
802 struct cfs_rq *cfs_rq = cfs_rq_of(se);
804 WARN_ON(task_rq(p) != rq);
806 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
807 u64 slice = sched_slice(cfs_rq, se);
808 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
809 s64 delta = slice - ran;
818 * Don't schedule slices shorter than 10000ns, that just
819 * doesn't make sense. Rely on vruntime for fairness.
822 delta = max(10000LL, delta);
824 hrtick_start(rq, delta, requeue);
829 hrtick_start_fair(struct rq *rq, struct task_struct *p)
835 * The enqueue_task method is called before nr_running is
836 * increased. Here we update the fair scheduling stats and
837 * then put the task into the rbtree:
839 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
841 struct cfs_rq *cfs_rq;
842 struct sched_entity *se = &p->se;
844 for_each_sched_entity(se) {
847 cfs_rq = cfs_rq_of(se);
848 enqueue_entity(cfs_rq, se, wakeup);
852 hrtick_start_fair(rq, rq->curr);
856 * The dequeue_task method is called before nr_running is
857 * decreased. We remove the task from the rbtree and
858 * update the fair scheduling stats:
860 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
862 struct cfs_rq *cfs_rq;
863 struct sched_entity *se = &p->se;
865 for_each_sched_entity(se) {
866 cfs_rq = cfs_rq_of(se);
867 dequeue_entity(cfs_rq, se, sleep);
868 /* Don't dequeue parent if it has other entities besides us */
869 if (cfs_rq->load.weight)
874 hrtick_start_fair(rq, rq->curr);
878 * sched_yield() support is very simple - we dequeue and enqueue.
880 * If compat_yield is turned on then we requeue to the end of the tree.
882 static void yield_task_fair(struct rq *rq)
884 struct task_struct *curr = rq->curr;
885 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
886 struct sched_entity *rightmost, *se = &curr->se;
889 * Are we the only task in the tree?
891 if (unlikely(cfs_rq->nr_running == 1))
894 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
895 __update_rq_clock(rq);
897 * Update run-time statistics of the 'current'.
904 * Find the rightmost entry in the rbtree:
906 rightmost = __pick_last_entity(cfs_rq);
908 * Already in the rightmost position?
910 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
914 * Minimally necessary key value to be last in the tree:
915 * Upon rescheduling, sched_class::put_prev_task() will place
916 * 'current' within the tree based on its new key value.
918 se->vruntime = rightmost->vruntime + 1;
922 * wake_idle() will wake a task on an idle cpu if task->cpu is
923 * not idle and an idle cpu is available. The span of cpus to
924 * search starts with cpus closest then further out as needed,
925 * so we always favor a closer, idle cpu.
927 * Returns the CPU we should wake onto.
929 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
930 static int wake_idle(int cpu, struct task_struct *p)
933 struct sched_domain *sd;
937 * If it is idle, then it is the best cpu to run this task.
939 * This cpu is also the best, if it has more than one task already.
940 * Siblings must be also busy(in most cases) as they didn't already
941 * pickup the extra load from this cpu and hence we need not check
942 * sibling runqueue info. This will avoid the checks and cache miss
943 * penalities associated with that.
945 if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
948 for_each_domain(cpu, sd) {
949 if (sd->flags & SD_WAKE_IDLE) {
950 cpus_and(tmp, sd->span, p->cpus_allowed);
951 for_each_cpu_mask(i, tmp) {
953 if (i != task_cpu(p)) {
967 static inline int wake_idle(int cpu, struct task_struct *p)
975 static const struct sched_class fair_sched_class;
978 wake_affine(struct rq *rq, struct sched_domain *this_sd, struct rq *this_rq,
979 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
980 int idx, unsigned long load, unsigned long this_load,
981 unsigned int imbalance)
983 struct task_struct *curr = this_rq->curr;
984 unsigned long tl = this_load;
985 unsigned long tl_per_task;
987 if (!(this_sd->flags & SD_WAKE_AFFINE))
991 * If the currently running task will sleep within
992 * a reasonable amount of time then attract this newly
995 if (sync && curr->sched_class == &fair_sched_class) {
996 if (curr->se.avg_overlap < sysctl_sched_migration_cost &&
997 p->se.avg_overlap < sysctl_sched_migration_cost)
1001 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1002 tl_per_task = cpu_avg_load_per_task(this_cpu);
1005 * If sync wakeup then subtract the (maximum possible)
1006 * effect of the currently running task from the load
1007 * of the current CPU:
1010 tl -= current->se.load.weight;
1012 if ((tl <= load && tl + target_load(prev_cpu, idx) <= tl_per_task) ||
1013 100*(tl + p->se.load.weight) <= imbalance*load) {
1015 * This domain has SD_WAKE_AFFINE and
1016 * p is cache cold in this domain, and
1017 * there is no bad imbalance.
1019 schedstat_inc(this_sd, ttwu_move_affine);
1020 schedstat_inc(p, se.nr_wakeups_affine);
1027 static int select_task_rq_fair(struct task_struct *p, int sync)
1029 struct sched_domain *sd, *this_sd = NULL;
1030 int prev_cpu, this_cpu, new_cpu;
1031 unsigned long load, this_load;
1032 struct rq *rq, *this_rq;
1033 unsigned int imbalance;
1036 prev_cpu = task_cpu(p);
1038 this_cpu = smp_processor_id();
1039 this_rq = cpu_rq(this_cpu);
1043 * 'this_sd' is the first domain that both
1044 * this_cpu and prev_cpu are present in:
1046 for_each_domain(this_cpu, sd) {
1047 if (cpu_isset(prev_cpu, sd->span)) {
1053 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1057 * Check for affine wakeup and passive balancing possibilities.
1062 idx = this_sd->wake_idx;
1064 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1066 load = source_load(prev_cpu, idx);
1067 this_load = target_load(this_cpu, idx);
1069 if (wake_affine(rq, this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1070 load, this_load, imbalance))
1073 if (prev_cpu == this_cpu)
1077 * Start passive balancing when half the imbalance_pct
1080 if (this_sd->flags & SD_WAKE_BALANCE) {
1081 if (imbalance*this_load <= 100*load) {
1082 schedstat_inc(this_sd, ttwu_move_balance);
1083 schedstat_inc(p, se.nr_wakeups_passive);
1089 return wake_idle(new_cpu, p);
1091 #endif /* CONFIG_SMP */
1093 static unsigned long wakeup_gran(struct sched_entity *se)
1095 unsigned long gran = sysctl_sched_wakeup_granularity;
1098 * More easily preempt - nice tasks, while not making
1099 * it harder for + nice tasks.
1101 if (unlikely(se->load.weight > NICE_0_LOAD))
1102 gran = calc_delta_fair(gran, &se->load);
1108 * Should 'se' preempt 'curr'.
1122 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
1124 s64 gran, vdiff = curr->vruntime - se->vruntime;
1129 gran = wakeup_gran(curr);
1137 * Preempt the current task with a newly woken task if needed:
1139 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
1141 struct task_struct *curr = rq->curr;
1142 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1143 struct sched_entity *se = &curr->se, *pse = &p->se;
1145 if (unlikely(rt_prio(p->prio))) {
1146 update_rq_clock(rq);
1147 update_curr(cfs_rq);
1152 se->last_wakeup = se->sum_exec_runtime;
1153 if (unlikely(se == pse))
1156 cfs_rq_of(pse)->next = pse;
1159 * Batch tasks do not preempt (their preemption is driven by
1162 if (unlikely(p->policy == SCHED_BATCH))
1165 if (!sched_feat(WAKEUP_PREEMPT))
1168 while (!is_same_group(se, pse)) {
1169 se = parent_entity(se);
1170 pse = parent_entity(pse);
1173 if (wakeup_preempt_entity(se, pse) == 1)
1177 static struct task_struct *pick_next_task_fair(struct rq *rq)
1179 struct task_struct *p;
1180 struct cfs_rq *cfs_rq = &rq->cfs;
1181 struct sched_entity *se;
1183 if (unlikely(!cfs_rq->nr_running))
1187 se = pick_next_entity(cfs_rq);
1188 cfs_rq = group_cfs_rq(se);
1192 hrtick_start_fair(rq, p);
1198 * Account for a descheduled task:
1200 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1202 struct sched_entity *se = &prev->se;
1203 struct cfs_rq *cfs_rq;
1205 for_each_sched_entity(se) {
1206 cfs_rq = cfs_rq_of(se);
1207 put_prev_entity(cfs_rq, se);
1212 /**************************************************
1213 * Fair scheduling class load-balancing methods:
1217 * Load-balancing iterator. Note: while the runqueue stays locked
1218 * during the whole iteration, the current task might be
1219 * dequeued so the iterator has to be dequeue-safe. Here we
1220 * achieve that by always pre-iterating before returning
1223 static struct task_struct *
1224 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1226 struct task_struct *p;
1231 p = rb_entry(curr, struct task_struct, se.run_node);
1232 cfs_rq->rb_load_balance_curr = rb_next(curr);
1237 static struct task_struct *load_balance_start_fair(void *arg)
1239 struct cfs_rq *cfs_rq = arg;
1241 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1244 static struct task_struct *load_balance_next_fair(void *arg)
1246 struct cfs_rq *cfs_rq = arg;
1248 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1251 #ifdef CONFIG_FAIR_GROUP_SCHED
1252 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1254 struct sched_entity *curr;
1255 struct task_struct *p;
1257 if (!cfs_rq->nr_running || !first_fair(cfs_rq))
1260 curr = cfs_rq->curr;
1262 curr = __pick_next_entity(cfs_rq);
1270 static unsigned long
1271 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1272 unsigned long max_load_move,
1273 struct sched_domain *sd, enum cpu_idle_type idle,
1274 int *all_pinned, int *this_best_prio)
1276 struct cfs_rq *busy_cfs_rq;
1277 long rem_load_move = max_load_move;
1278 struct rq_iterator cfs_rq_iterator;
1280 cfs_rq_iterator.start = load_balance_start_fair;
1281 cfs_rq_iterator.next = load_balance_next_fair;
1283 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1284 #ifdef CONFIG_FAIR_GROUP_SCHED
1285 struct cfs_rq *this_cfs_rq;
1287 unsigned long maxload;
1289 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1291 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1292 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1296 /* Don't pull more than imbalance/2 */
1298 maxload = min(rem_load_move, imbalance);
1300 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1302 # define maxload rem_load_move
1305 * pass busy_cfs_rq argument into
1306 * load_balance_[start|next]_fair iterators
1308 cfs_rq_iterator.arg = busy_cfs_rq;
1309 rem_load_move -= balance_tasks(this_rq, this_cpu, busiest,
1310 maxload, sd, idle, all_pinned,
1314 if (rem_load_move <= 0)
1318 return max_load_move - rem_load_move;
1322 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1323 struct sched_domain *sd, enum cpu_idle_type idle)
1325 struct cfs_rq *busy_cfs_rq;
1326 struct rq_iterator cfs_rq_iterator;
1328 cfs_rq_iterator.start = load_balance_start_fair;
1329 cfs_rq_iterator.next = load_balance_next_fair;
1331 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1333 * pass busy_cfs_rq argument into
1334 * load_balance_[start|next]_fair iterators
1336 cfs_rq_iterator.arg = busy_cfs_rq;
1337 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1347 * scheduler tick hitting a task of our scheduling class:
1349 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1351 struct cfs_rq *cfs_rq;
1352 struct sched_entity *se = &curr->se;
1354 for_each_sched_entity(se) {
1355 cfs_rq = cfs_rq_of(se);
1356 entity_tick(cfs_rq, se, queued);
1360 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1363 * Share the fairness runtime between parent and child, thus the
1364 * total amount of pressure for CPU stays equal - new tasks
1365 * get a chance to run but frequent forkers are not allowed to
1366 * monopolize the CPU. Note: the parent runqueue is locked,
1367 * the child is not running yet.
1369 static void task_new_fair(struct rq *rq, struct task_struct *p)
1371 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1372 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1373 int this_cpu = smp_processor_id();
1375 sched_info_queued(p);
1377 update_curr(cfs_rq);
1378 place_entity(cfs_rq, se, 1);
1380 /* 'curr' will be NULL if the child belongs to a different group */
1381 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1382 curr && curr->vruntime < se->vruntime) {
1384 * Upon rescheduling, sched_class::put_prev_task() will place
1385 * 'current' within the tree based on its new key value.
1387 swap(curr->vruntime, se->vruntime);
1390 enqueue_task_fair(rq, p, 0);
1391 resched_task(rq->curr);
1395 * Priority of the task has changed. Check to see if we preempt
1398 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1399 int oldprio, int running)
1402 * Reschedule if we are currently running on this runqueue and
1403 * our priority decreased, or if we are not currently running on
1404 * this runqueue and our priority is higher than the current's
1407 if (p->prio > oldprio)
1408 resched_task(rq->curr);
1410 check_preempt_curr(rq, p);
1414 * We switched to the sched_fair class.
1416 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1420 * We were most likely switched from sched_rt, so
1421 * kick off the schedule if running, otherwise just see
1422 * if we can still preempt the current task.
1425 resched_task(rq->curr);
1427 check_preempt_curr(rq, p);
1430 /* Account for a task changing its policy or group.
1432 * This routine is mostly called to set cfs_rq->curr field when a task
1433 * migrates between groups/classes.
1435 static void set_curr_task_fair(struct rq *rq)
1437 struct sched_entity *se = &rq->curr->se;
1439 for_each_sched_entity(se)
1440 set_next_entity(cfs_rq_of(se), se);
1443 #ifdef CONFIG_FAIR_GROUP_SCHED
1444 static void moved_group_fair(struct task_struct *p)
1446 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1448 update_curr(cfs_rq);
1449 place_entity(cfs_rq, &p->se, 1);
1454 * All the scheduling class methods:
1456 static const struct sched_class fair_sched_class = {
1457 .next = &idle_sched_class,
1458 .enqueue_task = enqueue_task_fair,
1459 .dequeue_task = dequeue_task_fair,
1460 .yield_task = yield_task_fair,
1462 .select_task_rq = select_task_rq_fair,
1463 #endif /* CONFIG_SMP */
1465 .check_preempt_curr = check_preempt_wakeup,
1467 .pick_next_task = pick_next_task_fair,
1468 .put_prev_task = put_prev_task_fair,
1471 .load_balance = load_balance_fair,
1472 .move_one_task = move_one_task_fair,
1475 .set_curr_task = set_curr_task_fair,
1476 .task_tick = task_tick_fair,
1477 .task_new = task_new_fair,
1479 .prio_changed = prio_changed_fair,
1480 .switched_to = switched_to_fair,
1482 #ifdef CONFIG_FAIR_GROUP_SCHED
1483 .moved_group = moved_group_fair,
1487 #ifdef CONFIG_SCHED_DEBUG
1488 static void print_cfs_stats(struct seq_file *m, int cpu)
1490 struct cfs_rq *cfs_rq;
1492 #ifdef CONFIG_FAIR_GROUP_SCHED
1493 print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs);
1496 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1497 print_cfs_rq(m, cpu, cfs_rq);