PCI: update pci_create_slot() to take a 'hotplug' param
[linux-2.6/mini2440.git] / kernel / sched_fair.c
blobf604dae71316264445e63b4d09f26a483d61113e
1 /*
2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
23 #include <linux/latencytop.h>
26 * Targeted preemption latency for CPU-bound tasks:
27 * (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
29 * NOTE: this latency value is not the same as the concept of
30 * 'timeslice length' - timeslices in CFS are of variable length
31 * and have no persistent notion like in traditional, time-slice
32 * based scheduling concepts.
34 * (to see the precise effective timeslice length of your workload,
35 * run vmstat and monitor the context-switches (cs) field)
37 unsigned int sysctl_sched_latency = 20000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity = 4000000ULL;
46 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
48 static unsigned int sched_nr_latency = 5;
51 * After fork, child runs first. (default) If set to 0 then
52 * parent will (try to) run first.
54 const_debug unsigned int sysctl_sched_child_runs_first = 1;
57 * sys_sched_yield() compat mode
59 * This option switches the agressive yield implementation of the
60 * old scheduler back on.
62 unsigned int __read_mostly sysctl_sched_compat_yield;
65 * SCHED_OTHER wake-up granularity.
66 * (default: 5 msec * (1 + ilog(ncpus)), units: nanoseconds)
68 * This option delays the preemption effects of decoupled workloads
69 * and reduces their over-scheduling. Synchronous workloads will still
70 * have immediate wakeup/sleep latencies.
72 unsigned int sysctl_sched_wakeup_granularity = 5000000UL;
74 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
76 /**************************************************************
77 * CFS operations on generic schedulable entities:
80 static inline struct task_struct *task_of(struct sched_entity *se)
82 return container_of(se, struct task_struct, se);
85 #ifdef CONFIG_FAIR_GROUP_SCHED
87 /* cpu runqueue to which this cfs_rq is attached */
88 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
90 return cfs_rq->rq;
93 /* An entity is a task if it doesn't "own" a runqueue */
94 #define entity_is_task(se) (!se->my_q)
96 /* Walk up scheduling entities hierarchy */
97 #define for_each_sched_entity(se) \
98 for (; se; se = se->parent)
100 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
102 return p->se.cfs_rq;
105 /* runqueue on which this entity is (to be) queued */
106 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
108 return se->cfs_rq;
111 /* runqueue "owned" by this group */
112 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
114 return grp->my_q;
117 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
118 * another cpu ('this_cpu')
120 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
122 return cfs_rq->tg->cfs_rq[this_cpu];
125 /* Iterate thr' all leaf cfs_rq's on a runqueue */
126 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
127 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
129 /* Do the two (enqueued) entities belong to the same group ? */
130 static inline int
131 is_same_group(struct sched_entity *se, struct sched_entity *pse)
133 if (se->cfs_rq == pse->cfs_rq)
134 return 1;
136 return 0;
139 static inline struct sched_entity *parent_entity(struct sched_entity *se)
141 return se->parent;
144 #else /* CONFIG_FAIR_GROUP_SCHED */
146 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
148 return container_of(cfs_rq, struct rq, cfs);
151 #define entity_is_task(se) 1
153 #define for_each_sched_entity(se) \
154 for (; se; se = NULL)
156 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
158 return &task_rq(p)->cfs;
161 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
163 struct task_struct *p = task_of(se);
164 struct rq *rq = task_rq(p);
166 return &rq->cfs;
169 /* runqueue "owned" by this group */
170 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
172 return NULL;
175 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
177 return &cpu_rq(this_cpu)->cfs;
180 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
181 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
183 static inline int
184 is_same_group(struct sched_entity *se, struct sched_entity *pse)
186 return 1;
189 static inline struct sched_entity *parent_entity(struct sched_entity *se)
191 return NULL;
194 #endif /* CONFIG_FAIR_GROUP_SCHED */
197 /**************************************************************
198 * Scheduling class tree data structure manipulation methods:
201 static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
203 s64 delta = (s64)(vruntime - min_vruntime);
204 if (delta > 0)
205 min_vruntime = vruntime;
207 return min_vruntime;
210 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
212 s64 delta = (s64)(vruntime - min_vruntime);
213 if (delta < 0)
214 min_vruntime = vruntime;
216 return min_vruntime;
219 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
221 return se->vruntime - cfs_rq->min_vruntime;
225 * Enqueue an entity into the rb-tree:
227 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
229 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
230 struct rb_node *parent = NULL;
231 struct sched_entity *entry;
232 s64 key = entity_key(cfs_rq, se);
233 int leftmost = 1;
236 * Find the right place in the rbtree:
238 while (*link) {
239 parent = *link;
240 entry = rb_entry(parent, struct sched_entity, run_node);
242 * We dont care about collisions. Nodes with
243 * the same key stay together.
245 if (key < entity_key(cfs_rq, entry)) {
246 link = &parent->rb_left;
247 } else {
248 link = &parent->rb_right;
249 leftmost = 0;
254 * Maintain a cache of leftmost tree entries (it is frequently
255 * used):
257 if (leftmost) {
258 cfs_rq->rb_leftmost = &se->run_node;
260 * maintain cfs_rq->min_vruntime to be a monotonic increasing
261 * value tracking the leftmost vruntime in the tree.
263 cfs_rq->min_vruntime =
264 max_vruntime(cfs_rq->min_vruntime, se->vruntime);
267 rb_link_node(&se->run_node, parent, link);
268 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
271 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
273 if (cfs_rq->rb_leftmost == &se->run_node) {
274 struct rb_node *next_node;
275 struct sched_entity *next;
277 next_node = rb_next(&se->run_node);
278 cfs_rq->rb_leftmost = next_node;
280 if (next_node) {
281 next = rb_entry(next_node,
282 struct sched_entity, run_node);
283 cfs_rq->min_vruntime =
284 max_vruntime(cfs_rq->min_vruntime,
285 next->vruntime);
289 if (cfs_rq->next == se)
290 cfs_rq->next = NULL;
292 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
295 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
297 return cfs_rq->rb_leftmost;
300 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
302 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
305 static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
307 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
309 if (!last)
310 return NULL;
312 return rb_entry(last, struct sched_entity, run_node);
315 /**************************************************************
316 * Scheduling class statistics methods:
319 #ifdef CONFIG_SCHED_DEBUG
320 int sched_nr_latency_handler(struct ctl_table *table, int write,
321 struct file *filp, void __user *buffer, size_t *lenp,
322 loff_t *ppos)
324 int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
326 if (ret || !write)
327 return ret;
329 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
330 sysctl_sched_min_granularity);
332 return 0;
334 #endif
337 * delta *= w / rw
339 static inline unsigned long
340 calc_delta_weight(unsigned long delta, struct sched_entity *se)
342 for_each_sched_entity(se) {
343 delta = calc_delta_mine(delta,
344 se->load.weight, &cfs_rq_of(se)->load);
347 return delta;
351 * delta *= rw / w
353 static inline unsigned long
354 calc_delta_fair(unsigned long delta, struct sched_entity *se)
356 for_each_sched_entity(se) {
357 delta = calc_delta_mine(delta,
358 cfs_rq_of(se)->load.weight, &se->load);
361 return delta;
365 * The idea is to set a period in which each task runs once.
367 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
368 * this period because otherwise the slices get too small.
370 * p = (nr <= nl) ? l : l*nr/nl
372 static u64 __sched_period(unsigned long nr_running)
374 u64 period = sysctl_sched_latency;
375 unsigned long nr_latency = sched_nr_latency;
377 if (unlikely(nr_running > nr_latency)) {
378 period = sysctl_sched_min_granularity;
379 period *= nr_running;
382 return period;
386 * We calculate the wall-time slice from the period by taking a part
387 * proportional to the weight.
389 * s = p*w/rw
391 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
393 return calc_delta_weight(__sched_period(cfs_rq->nr_running), se);
397 * We calculate the vruntime slice of a to be inserted task
399 * vs = s*rw/w = p
401 static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
403 unsigned long nr_running = cfs_rq->nr_running;
405 if (!se->on_rq)
406 nr_running++;
408 return __sched_period(nr_running);
412 * Update the current task's runtime statistics. Skip current tasks that
413 * are not in our scheduling class.
415 static inline void
416 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
417 unsigned long delta_exec)
419 unsigned long delta_exec_weighted;
421 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
423 curr->sum_exec_runtime += delta_exec;
424 schedstat_add(cfs_rq, exec_clock, delta_exec);
425 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
426 curr->vruntime += delta_exec_weighted;
429 static void update_curr(struct cfs_rq *cfs_rq)
431 struct sched_entity *curr = cfs_rq->curr;
432 u64 now = rq_of(cfs_rq)->clock;
433 unsigned long delta_exec;
435 if (unlikely(!curr))
436 return;
439 * Get the amount of time the current task was running
440 * since the last time we changed load (this cannot
441 * overflow on 32 bits):
443 delta_exec = (unsigned long)(now - curr->exec_start);
445 __update_curr(cfs_rq, curr, delta_exec);
446 curr->exec_start = now;
448 if (entity_is_task(curr)) {
449 struct task_struct *curtask = task_of(curr);
451 cpuacct_charge(curtask, delta_exec);
452 account_group_exec_runtime(curtask, delta_exec);
456 static inline void
457 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
459 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
463 * Task is being enqueued - update stats:
465 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
468 * Are we enqueueing a waiting task? (for current tasks
469 * a dequeue/enqueue event is a NOP)
471 if (se != cfs_rq->curr)
472 update_stats_wait_start(cfs_rq, se);
475 static void
476 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
478 schedstat_set(se->wait_max, max(se->wait_max,
479 rq_of(cfs_rq)->clock - se->wait_start));
480 schedstat_set(se->wait_count, se->wait_count + 1);
481 schedstat_set(se->wait_sum, se->wait_sum +
482 rq_of(cfs_rq)->clock - se->wait_start);
483 schedstat_set(se->wait_start, 0);
486 static inline void
487 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
490 * Mark the end of the wait period if dequeueing a
491 * waiting task:
493 if (se != cfs_rq->curr)
494 update_stats_wait_end(cfs_rq, se);
498 * We are picking a new current task - update its stats:
500 static inline void
501 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
504 * We are starting a new run period:
506 se->exec_start = rq_of(cfs_rq)->clock;
509 /**************************************************
510 * Scheduling class queueing methods:
513 #if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
514 static void
515 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
517 cfs_rq->task_weight += weight;
519 #else
520 static inline void
521 add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
524 #endif
526 static void
527 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
529 update_load_add(&cfs_rq->load, se->load.weight);
530 if (!parent_entity(se))
531 inc_cpu_load(rq_of(cfs_rq), se->load.weight);
532 if (entity_is_task(se)) {
533 add_cfs_task_weight(cfs_rq, se->load.weight);
534 list_add(&se->group_node, &cfs_rq->tasks);
536 cfs_rq->nr_running++;
537 se->on_rq = 1;
540 static void
541 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 update_load_sub(&cfs_rq->load, se->load.weight);
544 if (!parent_entity(se))
545 dec_cpu_load(rq_of(cfs_rq), se->load.weight);
546 if (entity_is_task(se)) {
547 add_cfs_task_weight(cfs_rq, -se->load.weight);
548 list_del_init(&se->group_node);
550 cfs_rq->nr_running--;
551 se->on_rq = 0;
554 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
556 #ifdef CONFIG_SCHEDSTATS
557 if (se->sleep_start) {
558 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
559 struct task_struct *tsk = task_of(se);
561 if ((s64)delta < 0)
562 delta = 0;
564 if (unlikely(delta > se->sleep_max))
565 se->sleep_max = delta;
567 se->sleep_start = 0;
568 se->sum_sleep_runtime += delta;
570 account_scheduler_latency(tsk, delta >> 10, 1);
572 if (se->block_start) {
573 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
574 struct task_struct *tsk = task_of(se);
576 if ((s64)delta < 0)
577 delta = 0;
579 if (unlikely(delta > se->block_max))
580 se->block_max = delta;
582 se->block_start = 0;
583 se->sum_sleep_runtime += delta;
586 * Blocking time is in units of nanosecs, so shift by 20 to
587 * get a milliseconds-range estimation of the amount of
588 * time that the task spent sleeping:
590 if (unlikely(prof_on == SLEEP_PROFILING)) {
592 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
593 delta >> 20);
595 account_scheduler_latency(tsk, delta >> 10, 0);
597 #endif
600 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
602 #ifdef CONFIG_SCHED_DEBUG
603 s64 d = se->vruntime - cfs_rq->min_vruntime;
605 if (d < 0)
606 d = -d;
608 if (d > 3*sysctl_sched_latency)
609 schedstat_inc(cfs_rq, nr_spread_over);
610 #endif
613 static void
614 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
616 u64 vruntime;
618 if (first_fair(cfs_rq)) {
619 vruntime = min_vruntime(cfs_rq->min_vruntime,
620 __pick_next_entity(cfs_rq)->vruntime);
621 } else
622 vruntime = cfs_rq->min_vruntime;
625 * The 'current' period is already promised to the current tasks,
626 * however the extra weight of the new task will slow them down a
627 * little, place the new task so that it fits in the slot that
628 * stays open at the end.
630 if (initial && sched_feat(START_DEBIT))
631 vruntime += sched_vslice_add(cfs_rq, se);
633 if (!initial) {
634 /* sleeps upto a single latency don't count. */
635 if (sched_feat(NEW_FAIR_SLEEPERS)) {
636 unsigned long thresh = sysctl_sched_latency;
639 * convert the sleeper threshold into virtual time
641 if (sched_feat(NORMALIZED_SLEEPER))
642 thresh = calc_delta_fair(thresh, se);
644 vruntime -= thresh;
647 /* ensure we never gain time by being placed backwards. */
648 vruntime = max_vruntime(se->vruntime, vruntime);
651 se->vruntime = vruntime;
654 static void
655 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
658 * Update run-time statistics of the 'current'.
660 update_curr(cfs_rq);
661 account_entity_enqueue(cfs_rq, se);
663 if (wakeup) {
664 place_entity(cfs_rq, se, 0);
665 enqueue_sleeper(cfs_rq, se);
668 update_stats_enqueue(cfs_rq, se);
669 check_spread(cfs_rq, se);
670 if (se != cfs_rq->curr)
671 __enqueue_entity(cfs_rq, se);
674 static void
675 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
678 * Update run-time statistics of the 'current'.
680 update_curr(cfs_rq);
682 update_stats_dequeue(cfs_rq, se);
683 if (sleep) {
684 #ifdef CONFIG_SCHEDSTATS
685 if (entity_is_task(se)) {
686 struct task_struct *tsk = task_of(se);
688 if (tsk->state & TASK_INTERRUPTIBLE)
689 se->sleep_start = rq_of(cfs_rq)->clock;
690 if (tsk->state & TASK_UNINTERRUPTIBLE)
691 se->block_start = rq_of(cfs_rq)->clock;
693 #endif
696 if (se != cfs_rq->curr)
697 __dequeue_entity(cfs_rq, se);
698 account_entity_dequeue(cfs_rq, se);
702 * Preempt the current task with a newly woken task if needed:
704 static void
705 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
707 unsigned long ideal_runtime, delta_exec;
709 ideal_runtime = sched_slice(cfs_rq, curr);
710 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
711 if (delta_exec > ideal_runtime)
712 resched_task(rq_of(cfs_rq)->curr);
715 static void
716 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
718 /* 'current' is not kept within the tree. */
719 if (se->on_rq) {
721 * Any task has to be enqueued before it get to execute on
722 * a CPU. So account for the time it spent waiting on the
723 * runqueue.
725 update_stats_wait_end(cfs_rq, se);
726 __dequeue_entity(cfs_rq, se);
729 update_stats_curr_start(cfs_rq, se);
730 cfs_rq->curr = se;
731 #ifdef CONFIG_SCHEDSTATS
733 * Track our maximum slice length, if the CPU's load is at
734 * least twice that of our own weight (i.e. dont track it
735 * when there are only lesser-weight tasks around):
737 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
738 se->slice_max = max(se->slice_max,
739 se->sum_exec_runtime - se->prev_sum_exec_runtime);
741 #endif
742 se->prev_sum_exec_runtime = se->sum_exec_runtime;
745 static struct sched_entity *
746 pick_next(struct cfs_rq *cfs_rq, struct sched_entity *se)
748 struct rq *rq = rq_of(cfs_rq);
749 u64 pair_slice = rq->clock - cfs_rq->pair_start;
751 if (!cfs_rq->next || pair_slice > sched_slice(cfs_rq, cfs_rq->next)) {
752 cfs_rq->pair_start = rq->clock;
753 return se;
756 return cfs_rq->next;
759 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
761 struct sched_entity *se = NULL;
763 if (first_fair(cfs_rq)) {
764 se = __pick_next_entity(cfs_rq);
765 se = pick_next(cfs_rq, se);
766 set_next_entity(cfs_rq, se);
769 return se;
772 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
775 * If still on the runqueue then deactivate_task()
776 * was not called and update_curr() has to be done:
778 if (prev->on_rq)
779 update_curr(cfs_rq);
781 check_spread(cfs_rq, prev);
782 if (prev->on_rq) {
783 update_stats_wait_start(cfs_rq, prev);
784 /* Put 'current' back into the tree. */
785 __enqueue_entity(cfs_rq, prev);
787 cfs_rq->curr = NULL;
790 static void
791 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
794 * Update run-time statistics of the 'current'.
796 update_curr(cfs_rq);
798 #ifdef CONFIG_SCHED_HRTICK
800 * queued ticks are scheduled to match the slice, so don't bother
801 * validating it and just reschedule.
803 if (queued) {
804 resched_task(rq_of(cfs_rq)->curr);
805 return;
808 * don't let the period tick interfere with the hrtick preemption
810 if (!sched_feat(DOUBLE_TICK) &&
811 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
812 return;
813 #endif
815 if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
816 check_preempt_tick(cfs_rq, curr);
819 /**************************************************
820 * CFS operations on tasks:
823 #ifdef CONFIG_SCHED_HRTICK
824 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
826 struct sched_entity *se = &p->se;
827 struct cfs_rq *cfs_rq = cfs_rq_of(se);
829 WARN_ON(task_rq(p) != rq);
831 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
832 u64 slice = sched_slice(cfs_rq, se);
833 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
834 s64 delta = slice - ran;
836 if (delta < 0) {
837 if (rq->curr == p)
838 resched_task(p);
839 return;
843 * Don't schedule slices shorter than 10000ns, that just
844 * doesn't make sense. Rely on vruntime for fairness.
846 if (rq->curr != p)
847 delta = max_t(s64, 10000LL, delta);
849 hrtick_start(rq, delta);
852 #else /* !CONFIG_SCHED_HRTICK */
853 static inline void
854 hrtick_start_fair(struct rq *rq, struct task_struct *p)
857 #endif
860 * The enqueue_task method is called before nr_running is
861 * increased. Here we update the fair scheduling stats and
862 * then put the task into the rbtree:
864 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
866 struct cfs_rq *cfs_rq;
867 struct sched_entity *se = &p->se;
869 for_each_sched_entity(se) {
870 if (se->on_rq)
871 break;
872 cfs_rq = cfs_rq_of(se);
873 enqueue_entity(cfs_rq, se, wakeup);
874 wakeup = 1;
877 hrtick_start_fair(rq, rq->curr);
881 * The dequeue_task method is called before nr_running is
882 * decreased. We remove the task from the rbtree and
883 * update the fair scheduling stats:
885 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
887 struct cfs_rq *cfs_rq;
888 struct sched_entity *se = &p->se;
890 for_each_sched_entity(se) {
891 cfs_rq = cfs_rq_of(se);
892 dequeue_entity(cfs_rq, se, sleep);
893 /* Don't dequeue parent if it has other entities besides us */
894 if (cfs_rq->load.weight)
895 break;
896 sleep = 1;
899 hrtick_start_fair(rq, rq->curr);
903 * sched_yield() support is very simple - we dequeue and enqueue.
905 * If compat_yield is turned on then we requeue to the end of the tree.
907 static void yield_task_fair(struct rq *rq)
909 struct task_struct *curr = rq->curr;
910 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
911 struct sched_entity *rightmost, *se = &curr->se;
914 * Are we the only task in the tree?
916 if (unlikely(cfs_rq->nr_running == 1))
917 return;
919 if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
920 update_rq_clock(rq);
922 * Update run-time statistics of the 'current'.
924 update_curr(cfs_rq);
926 return;
929 * Find the rightmost entry in the rbtree:
931 rightmost = __pick_last_entity(cfs_rq);
933 * Already in the rightmost position?
935 if (unlikely(!rightmost || rightmost->vruntime < se->vruntime))
936 return;
939 * Minimally necessary key value to be last in the tree:
940 * Upon rescheduling, sched_class::put_prev_task() will place
941 * 'current' within the tree based on its new key value.
943 se->vruntime = rightmost->vruntime + 1;
947 * wake_idle() will wake a task on an idle cpu if task->cpu is
948 * not idle and an idle cpu is available. The span of cpus to
949 * search starts with cpus closest then further out as needed,
950 * so we always favor a closer, idle cpu.
951 * Domains may include CPUs that are not usable for migration,
952 * hence we need to mask them out (cpu_active_map)
954 * Returns the CPU we should wake onto.
956 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
957 static int wake_idle(int cpu, struct task_struct *p)
959 cpumask_t tmp;
960 struct sched_domain *sd;
961 int i;
964 * If it is idle, then it is the best cpu to run this task.
966 * This cpu is also the best, if it has more than one task already.
967 * Siblings must be also busy(in most cases) as they didn't already
968 * pickup the extra load from this cpu and hence we need not check
969 * sibling runqueue info. This will avoid the checks and cache miss
970 * penalities associated with that.
972 if (idle_cpu(cpu) || cpu_rq(cpu)->cfs.nr_running > 1)
973 return cpu;
975 for_each_domain(cpu, sd) {
976 if ((sd->flags & SD_WAKE_IDLE)
977 || ((sd->flags & SD_WAKE_IDLE_FAR)
978 && !task_hot(p, task_rq(p)->clock, sd))) {
979 cpus_and(tmp, sd->span, p->cpus_allowed);
980 cpus_and(tmp, tmp, cpu_active_map);
981 for_each_cpu_mask_nr(i, tmp) {
982 if (idle_cpu(i)) {
983 if (i != task_cpu(p)) {
984 schedstat_inc(p,
985 se.nr_wakeups_idle);
987 return i;
990 } else {
991 break;
994 return cpu;
996 #else /* !ARCH_HAS_SCHED_WAKE_IDLE*/
997 static inline int wake_idle(int cpu, struct task_struct *p)
999 return cpu;
1001 #endif
1003 #ifdef CONFIG_SMP
1005 static const struct sched_class fair_sched_class;
1007 #ifdef CONFIG_FAIR_GROUP_SCHED
1009 * effective_load() calculates the load change as seen from the root_task_group
1011 * Adding load to a group doesn't make a group heavier, but can cause movement
1012 * of group shares between cpus. Assuming the shares were perfectly aligned one
1013 * can calculate the shift in shares.
1015 * The problem is that perfectly aligning the shares is rather expensive, hence
1016 * we try to avoid doing that too often - see update_shares(), which ratelimits
1017 * this change.
1019 * We compensate this by not only taking the current delta into account, but
1020 * also considering the delta between when the shares were last adjusted and
1021 * now.
1023 * We still saw a performance dip, some tracing learned us that between
1024 * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
1025 * significantly. Therefore try to bias the error in direction of failing
1026 * the affine wakeup.
1029 static long effective_load(struct task_group *tg, int cpu,
1030 long wl, long wg)
1032 struct sched_entity *se = tg->se[cpu];
1034 if (!tg->parent)
1035 return wl;
1038 * By not taking the decrease of shares on the other cpu into
1039 * account our error leans towards reducing the affine wakeups.
1041 if (!wl && sched_feat(ASYM_EFF_LOAD))
1042 return wl;
1044 for_each_sched_entity(se) {
1045 long S, rw, s, a, b;
1046 long more_w;
1049 * Instead of using this increment, also add the difference
1050 * between when the shares were last updated and now.
1052 more_w = se->my_q->load.weight - se->my_q->rq_weight;
1053 wl += more_w;
1054 wg += more_w;
1056 S = se->my_q->tg->shares;
1057 s = se->my_q->shares;
1058 rw = se->my_q->rq_weight;
1060 a = S*(rw + wl);
1061 b = S*rw + s*wg;
1063 wl = s*(a-b);
1065 if (likely(b))
1066 wl /= b;
1069 * Assume the group is already running and will
1070 * thus already be accounted for in the weight.
1072 * That is, moving shares between CPUs, does not
1073 * alter the group weight.
1075 wg = 0;
1078 return wl;
1081 #else
1083 static inline unsigned long effective_load(struct task_group *tg, int cpu,
1084 unsigned long wl, unsigned long wg)
1086 return wl;
1089 #endif
1091 static int
1092 wake_affine(struct sched_domain *this_sd, struct rq *this_rq,
1093 struct task_struct *p, int prev_cpu, int this_cpu, int sync,
1094 int idx, unsigned long load, unsigned long this_load,
1095 unsigned int imbalance)
1097 struct task_struct *curr = this_rq->curr;
1098 struct task_group *tg;
1099 unsigned long tl = this_load;
1100 unsigned long tl_per_task;
1101 unsigned long weight;
1102 int balanced;
1104 if (!(this_sd->flags & SD_WAKE_AFFINE) || !sched_feat(AFFINE_WAKEUPS))
1105 return 0;
1107 if (!sync && sched_feat(SYNC_WAKEUPS) &&
1108 curr->se.avg_overlap < sysctl_sched_migration_cost &&
1109 p->se.avg_overlap < sysctl_sched_migration_cost)
1110 sync = 1;
1113 * If sync wakeup then subtract the (maximum possible)
1114 * effect of the currently running task from the load
1115 * of the current CPU:
1117 if (sync) {
1118 tg = task_group(current);
1119 weight = current->se.load.weight;
1121 tl += effective_load(tg, this_cpu, -weight, -weight);
1122 load += effective_load(tg, prev_cpu, 0, -weight);
1125 tg = task_group(p);
1126 weight = p->se.load.weight;
1128 balanced = 100*(tl + effective_load(tg, this_cpu, weight, weight)) <=
1129 imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
1132 * If the currently running task will sleep within
1133 * a reasonable amount of time then attract this newly
1134 * woken task:
1136 if (sync && balanced)
1137 return 1;
1139 schedstat_inc(p, se.nr_wakeups_affine_attempts);
1140 tl_per_task = cpu_avg_load_per_task(this_cpu);
1142 if (balanced || (tl <= load && tl + target_load(prev_cpu, idx) <=
1143 tl_per_task)) {
1145 * This domain has SD_WAKE_AFFINE and
1146 * p is cache cold in this domain, and
1147 * there is no bad imbalance.
1149 schedstat_inc(this_sd, ttwu_move_affine);
1150 schedstat_inc(p, se.nr_wakeups_affine);
1152 return 1;
1154 return 0;
1157 static int select_task_rq_fair(struct task_struct *p, int sync)
1159 struct sched_domain *sd, *this_sd = NULL;
1160 int prev_cpu, this_cpu, new_cpu;
1161 unsigned long load, this_load;
1162 struct rq *this_rq;
1163 unsigned int imbalance;
1164 int idx;
1166 prev_cpu = task_cpu(p);
1167 this_cpu = smp_processor_id();
1168 this_rq = cpu_rq(this_cpu);
1169 new_cpu = prev_cpu;
1171 if (prev_cpu == this_cpu)
1172 goto out;
1174 * 'this_sd' is the first domain that both
1175 * this_cpu and prev_cpu are present in:
1177 for_each_domain(this_cpu, sd) {
1178 if (cpu_isset(prev_cpu, sd->span)) {
1179 this_sd = sd;
1180 break;
1184 if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
1185 goto out;
1188 * Check for affine wakeup and passive balancing possibilities.
1190 if (!this_sd)
1191 goto out;
1193 idx = this_sd->wake_idx;
1195 imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
1197 load = source_load(prev_cpu, idx);
1198 this_load = target_load(this_cpu, idx);
1200 if (wake_affine(this_sd, this_rq, p, prev_cpu, this_cpu, sync, idx,
1201 load, this_load, imbalance))
1202 return this_cpu;
1205 * Start passive balancing when half the imbalance_pct
1206 * limit is reached.
1208 if (this_sd->flags & SD_WAKE_BALANCE) {
1209 if (imbalance*this_load <= 100*load) {
1210 schedstat_inc(this_sd, ttwu_move_balance);
1211 schedstat_inc(p, se.nr_wakeups_passive);
1212 return this_cpu;
1216 out:
1217 return wake_idle(new_cpu, p);
1219 #endif /* CONFIG_SMP */
1221 static unsigned long wakeup_gran(struct sched_entity *se)
1223 unsigned long gran = sysctl_sched_wakeup_granularity;
1226 * More easily preempt - nice tasks, while not making it harder for
1227 * + nice tasks.
1229 if (sched_feat(ASYM_GRAN))
1230 gran = calc_delta_mine(gran, NICE_0_LOAD, &se->load);
1232 return gran;
1236 * Preempt the current task with a newly woken task if needed:
1238 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
1240 struct task_struct *curr = rq->curr;
1241 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1242 struct sched_entity *se = &curr->se, *pse = &p->se;
1243 s64 delta_exec;
1245 if (unlikely(rt_prio(p->prio))) {
1246 update_rq_clock(rq);
1247 update_curr(cfs_rq);
1248 resched_task(curr);
1249 return;
1252 if (unlikely(se == pse))
1253 return;
1255 cfs_rq_of(pse)->next = pse;
1258 * We can come here with TIF_NEED_RESCHED already set from new task
1259 * wake up path.
1261 if (test_tsk_need_resched(curr))
1262 return;
1265 * Batch tasks do not preempt (their preemption is driven by
1266 * the tick):
1268 if (unlikely(p->policy == SCHED_BATCH))
1269 return;
1271 if (!sched_feat(WAKEUP_PREEMPT))
1272 return;
1274 if (sched_feat(WAKEUP_OVERLAP) && (sync ||
1275 (se->avg_overlap < sysctl_sched_migration_cost &&
1276 pse->avg_overlap < sysctl_sched_migration_cost))) {
1277 resched_task(curr);
1278 return;
1281 delta_exec = se->sum_exec_runtime - se->prev_sum_exec_runtime;
1282 if (delta_exec > wakeup_gran(pse))
1283 resched_task(curr);
1286 static struct task_struct *pick_next_task_fair(struct rq *rq)
1288 struct task_struct *p;
1289 struct cfs_rq *cfs_rq = &rq->cfs;
1290 struct sched_entity *se;
1292 if (unlikely(!cfs_rq->nr_running))
1293 return NULL;
1295 do {
1296 se = pick_next_entity(cfs_rq);
1297 cfs_rq = group_cfs_rq(se);
1298 } while (cfs_rq);
1300 p = task_of(se);
1301 hrtick_start_fair(rq, p);
1303 return p;
1307 * Account for a descheduled task:
1309 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1311 struct sched_entity *se = &prev->se;
1312 struct cfs_rq *cfs_rq;
1314 for_each_sched_entity(se) {
1315 cfs_rq = cfs_rq_of(se);
1316 put_prev_entity(cfs_rq, se);
1320 #ifdef CONFIG_SMP
1321 /**************************************************
1322 * Fair scheduling class load-balancing methods:
1326 * Load-balancing iterator. Note: while the runqueue stays locked
1327 * during the whole iteration, the current task might be
1328 * dequeued so the iterator has to be dequeue-safe. Here we
1329 * achieve that by always pre-iterating before returning
1330 * the current task:
1332 static struct task_struct *
1333 __load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
1335 struct task_struct *p = NULL;
1336 struct sched_entity *se;
1338 if (next == &cfs_rq->tasks)
1339 return NULL;
1341 se = list_entry(next, struct sched_entity, group_node);
1342 p = task_of(se);
1343 cfs_rq->balance_iterator = next->next;
1345 return p;
1348 static struct task_struct *load_balance_start_fair(void *arg)
1350 struct cfs_rq *cfs_rq = arg;
1352 return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
1355 static struct task_struct *load_balance_next_fair(void *arg)
1357 struct cfs_rq *cfs_rq = arg;
1359 return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
1362 static unsigned long
1363 __load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1364 unsigned long max_load_move, struct sched_domain *sd,
1365 enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
1366 struct cfs_rq *cfs_rq)
1368 struct rq_iterator cfs_rq_iterator;
1370 cfs_rq_iterator.start = load_balance_start_fair;
1371 cfs_rq_iterator.next = load_balance_next_fair;
1372 cfs_rq_iterator.arg = cfs_rq;
1374 return balance_tasks(this_rq, this_cpu, busiest,
1375 max_load_move, sd, idle, all_pinned,
1376 this_best_prio, &cfs_rq_iterator);
1379 #ifdef CONFIG_FAIR_GROUP_SCHED
1380 static unsigned long
1381 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1382 unsigned long max_load_move,
1383 struct sched_domain *sd, enum cpu_idle_type idle,
1384 int *all_pinned, int *this_best_prio)
1386 long rem_load_move = max_load_move;
1387 int busiest_cpu = cpu_of(busiest);
1388 struct task_group *tg;
1390 rcu_read_lock();
1391 update_h_load(busiest_cpu);
1393 list_for_each_entry_rcu(tg, &task_groups, list) {
1394 struct cfs_rq *busiest_cfs_rq = tg->cfs_rq[busiest_cpu];
1395 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
1396 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
1397 u64 rem_load, moved_load;
1400 * empty group
1402 if (!busiest_cfs_rq->task_weight)
1403 continue;
1405 rem_load = (u64)rem_load_move * busiest_weight;
1406 rem_load = div_u64(rem_load, busiest_h_load + 1);
1408 moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
1409 rem_load, sd, idle, all_pinned, this_best_prio,
1410 tg->cfs_rq[busiest_cpu]);
1412 if (!moved_load)
1413 continue;
1415 moved_load *= busiest_h_load;
1416 moved_load = div_u64(moved_load, busiest_weight + 1);
1418 rem_load_move -= moved_load;
1419 if (rem_load_move < 0)
1420 break;
1422 rcu_read_unlock();
1424 return max_load_move - rem_load_move;
1426 #else
1427 static unsigned long
1428 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1429 unsigned long max_load_move,
1430 struct sched_domain *sd, enum cpu_idle_type idle,
1431 int *all_pinned, int *this_best_prio)
1433 return __load_balance_fair(this_rq, this_cpu, busiest,
1434 max_load_move, sd, idle, all_pinned,
1435 this_best_prio, &busiest->cfs);
1437 #endif
1439 static int
1440 move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1441 struct sched_domain *sd, enum cpu_idle_type idle)
1443 struct cfs_rq *busy_cfs_rq;
1444 struct rq_iterator cfs_rq_iterator;
1446 cfs_rq_iterator.start = load_balance_start_fair;
1447 cfs_rq_iterator.next = load_balance_next_fair;
1449 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1451 * pass busy_cfs_rq argument into
1452 * load_balance_[start|next]_fair iterators
1454 cfs_rq_iterator.arg = busy_cfs_rq;
1455 if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
1456 &cfs_rq_iterator))
1457 return 1;
1460 return 0;
1462 #endif /* CONFIG_SMP */
1465 * scheduler tick hitting a task of our scheduling class:
1467 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
1469 struct cfs_rq *cfs_rq;
1470 struct sched_entity *se = &curr->se;
1472 for_each_sched_entity(se) {
1473 cfs_rq = cfs_rq_of(se);
1474 entity_tick(cfs_rq, se, queued);
1478 #define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1481 * Share the fairness runtime between parent and child, thus the
1482 * total amount of pressure for CPU stays equal - new tasks
1483 * get a chance to run but frequent forkers are not allowed to
1484 * monopolize the CPU. Note: the parent runqueue is locked,
1485 * the child is not running yet.
1487 static void task_new_fair(struct rq *rq, struct task_struct *p)
1489 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1490 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1491 int this_cpu = smp_processor_id();
1493 sched_info_queued(p);
1495 update_curr(cfs_rq);
1496 place_entity(cfs_rq, se, 1);
1498 /* 'curr' will be NULL if the child belongs to a different group */
1499 if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
1500 curr && curr->vruntime < se->vruntime) {
1502 * Upon rescheduling, sched_class::put_prev_task() will place
1503 * 'current' within the tree based on its new key value.
1505 swap(curr->vruntime, se->vruntime);
1506 resched_task(rq->curr);
1509 enqueue_task_fair(rq, p, 0);
1513 * Priority of the task has changed. Check to see if we preempt
1514 * the current task.
1516 static void prio_changed_fair(struct rq *rq, struct task_struct *p,
1517 int oldprio, int running)
1520 * Reschedule if we are currently running on this runqueue and
1521 * our priority decreased, or if we are not currently running on
1522 * this runqueue and our priority is higher than the current's
1524 if (running) {
1525 if (p->prio > oldprio)
1526 resched_task(rq->curr);
1527 } else
1528 check_preempt_curr(rq, p, 0);
1532 * We switched to the sched_fair class.
1534 static void switched_to_fair(struct rq *rq, struct task_struct *p,
1535 int running)
1538 * We were most likely switched from sched_rt, so
1539 * kick off the schedule if running, otherwise just see
1540 * if we can still preempt the current task.
1542 if (running)
1543 resched_task(rq->curr);
1544 else
1545 check_preempt_curr(rq, p, 0);
1548 /* Account for a task changing its policy or group.
1550 * This routine is mostly called to set cfs_rq->curr field when a task
1551 * migrates between groups/classes.
1553 static void set_curr_task_fair(struct rq *rq)
1555 struct sched_entity *se = &rq->curr->se;
1557 for_each_sched_entity(se)
1558 set_next_entity(cfs_rq_of(se), se);
1561 #ifdef CONFIG_FAIR_GROUP_SCHED
1562 static void moved_group_fair(struct task_struct *p)
1564 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1566 update_curr(cfs_rq);
1567 place_entity(cfs_rq, &p->se, 1);
1569 #endif
1572 * All the scheduling class methods:
1574 static const struct sched_class fair_sched_class = {
1575 .next = &idle_sched_class,
1576 .enqueue_task = enqueue_task_fair,
1577 .dequeue_task = dequeue_task_fair,
1578 .yield_task = yield_task_fair,
1579 #ifdef CONFIG_SMP
1580 .select_task_rq = select_task_rq_fair,
1581 #endif /* CONFIG_SMP */
1583 .check_preempt_curr = check_preempt_wakeup,
1585 .pick_next_task = pick_next_task_fair,
1586 .put_prev_task = put_prev_task_fair,
1588 #ifdef CONFIG_SMP
1589 .load_balance = load_balance_fair,
1590 .move_one_task = move_one_task_fair,
1591 #endif
1593 .set_curr_task = set_curr_task_fair,
1594 .task_tick = task_tick_fair,
1595 .task_new = task_new_fair,
1597 .prio_changed = prio_changed_fair,
1598 .switched_to = switched_to_fair,
1600 #ifdef CONFIG_FAIR_GROUP_SCHED
1601 .moved_group = moved_group_fair,
1602 #endif
1605 #ifdef CONFIG_SCHED_DEBUG
1606 static void print_cfs_stats(struct seq_file *m, int cpu)
1608 struct cfs_rq *cfs_rq;
1610 rcu_read_lock();
1611 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1612 print_cfs_rq(m, cpu, cfs_rq);
1613 rcu_read_unlock();
1615 #endif