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>
21 * Preemption granularity:
22 * (default: 10 msec, units: nanoseconds)
24 * NOTE: this granularity value is not the same as the concept of
25 * 'timeslice length' - timeslices in CFS will typically be somewhat
26 * larger than this value. (to see the precise effective timeslice
27 * length of your workload, run vmstat and monitor the context-switches
30 * On SMP systems the value of this is multiplied by the log2 of the
31 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
32 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
34 unsigned int sysctl_sched_granularity __read_mostly = 10000000UL;
37 * SCHED_BATCH wake-up granularity.
38 * (default: 25 msec, units: nanoseconds)
40 * This option delays the preemption effects of decoupled workloads
41 * and reduces their over-scheduling. Synchronous workloads will still
42 * have immediate wakeup/sleep latencies.
44 unsigned int sysctl_sched_batch_wakeup_granularity __read_mostly = 25000000UL;
47 * SCHED_OTHER wake-up granularity.
48 * (default: 1 msec, units: nanoseconds)
50 * This option delays the preemption effects of decoupled workloads
51 * and reduces their over-scheduling. Synchronous workloads will still
52 * have immediate wakeup/sleep latencies.
54 unsigned int sysctl_sched_wakeup_granularity __read_mostly = 1000000UL;
56 unsigned int sysctl_sched_stat_granularity __read_mostly;
59 * Initialized in sched_init_granularity() [to 5 times the base granularity]:
61 unsigned int sysctl_sched_runtime_limit __read_mostly;
64 * Debugging: various feature bits
67 SCHED_FEAT_FAIR_SLEEPERS = 1,
68 SCHED_FEAT_SLEEPER_AVG = 2,
69 SCHED_FEAT_SLEEPER_LOAD_AVG = 4,
70 SCHED_FEAT_PRECISE_CPU_LOAD = 8,
71 SCHED_FEAT_START_DEBIT = 16,
72 SCHED_FEAT_SKIP_INITIAL = 32,
75 unsigned int sysctl_sched_features __read_mostly =
76 SCHED_FEAT_FAIR_SLEEPERS *1 |
77 SCHED_FEAT_SLEEPER_AVG *0 |
78 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
79 SCHED_FEAT_PRECISE_CPU_LOAD *1 |
80 SCHED_FEAT_START_DEBIT *1 |
81 SCHED_FEAT_SKIP_INITIAL *0;
83 extern struct sched_class fair_sched_class;
85 /**************************************************************
86 * CFS operations on generic schedulable entities:
89 #ifdef CONFIG_FAIR_GROUP_SCHED
91 /* cpu runqueue to which this cfs_rq is attached */
92 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
97 /* currently running entity (if any) on this cfs_rq */
98 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
103 /* An entity is a task if it doesn't "own" a runqueue */
104 #define entity_is_task(se) (!se->my_q)
107 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
112 #else /* CONFIG_FAIR_GROUP_SCHED */
114 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
116 return container_of(cfs_rq, struct rq, cfs);
119 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
121 struct rq *rq = rq_of(cfs_rq);
123 if (unlikely(rq->curr->sched_class != &fair_sched_class))
126 return &rq->curr->se;
129 #define entity_is_task(se) 1
132 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
134 #endif /* CONFIG_FAIR_GROUP_SCHED */
136 static inline struct task_struct *task_of(struct sched_entity *se)
138 return container_of(se, struct task_struct, se);
142 /**************************************************************
143 * Scheduling class tree data structure manipulation methods:
147 * Enqueue an entity into the rb-tree:
150 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
152 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
153 struct rb_node *parent = NULL;
154 struct sched_entity *entry;
155 s64 key = se->fair_key;
159 * Find the right place in the rbtree:
163 entry = rb_entry(parent, struct sched_entity, run_node);
165 * We dont care about collisions. Nodes with
166 * the same key stay together.
168 if (key - entry->fair_key < 0) {
169 link = &parent->rb_left;
171 link = &parent->rb_right;
177 * Maintain a cache of leftmost tree entries (it is frequently
181 cfs_rq->rb_leftmost = &se->run_node;
183 rb_link_node(&se->run_node, parent, link);
184 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
185 update_load_add(&cfs_rq->load, se->load.weight);
186 cfs_rq->nr_running++;
191 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
193 if (cfs_rq->rb_leftmost == &se->run_node)
194 cfs_rq->rb_leftmost = rb_next(&se->run_node);
195 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
196 update_load_sub(&cfs_rq->load, se->load.weight);
197 cfs_rq->nr_running--;
201 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
203 return cfs_rq->rb_leftmost;
206 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
208 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
211 /**************************************************************
212 * Scheduling class statistics methods:
216 * We rescale the rescheduling granularity of tasks according to their
217 * nice level, but only linearly, not exponentially:
220 niced_granularity(struct sched_entity *curr, unsigned long granularity)
224 if (likely(curr->load.weight == NICE_0_LOAD))
227 * Positive nice levels get the same granularity as nice-0:
229 if (likely(curr->load.weight < NICE_0_LOAD)) {
230 tmp = curr->load.weight * (u64)granularity;
231 return (long) (tmp >> NICE_0_SHIFT);
234 * Negative nice level tasks get linearly finer
237 tmp = curr->load.inv_weight * (u64)granularity;
240 * It will always fit into 'long':
242 return (long) (tmp >> WMULT_SHIFT);
246 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
248 long limit = sysctl_sched_runtime_limit;
251 * Niced tasks have the same history dynamic range as
254 if (unlikely(se->wait_runtime > limit)) {
255 se->wait_runtime = limit;
256 schedstat_inc(se, wait_runtime_overruns);
257 schedstat_inc(cfs_rq, wait_runtime_overruns);
259 if (unlikely(se->wait_runtime < -limit)) {
260 se->wait_runtime = -limit;
261 schedstat_inc(se, wait_runtime_underruns);
262 schedstat_inc(cfs_rq, wait_runtime_underruns);
267 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
269 se->wait_runtime += delta;
270 schedstat_add(se, sum_wait_runtime, delta);
271 limit_wait_runtime(cfs_rq, se);
275 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
277 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
278 __add_wait_runtime(cfs_rq, se, delta);
279 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
283 * Update the current task's runtime statistics. Skip current tasks that
284 * are not in our scheduling class.
287 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr)
289 unsigned long delta, delta_exec, delta_fair, delta_mine;
290 struct load_weight *lw = &cfs_rq->load;
291 unsigned long load = lw->weight;
293 delta_exec = curr->delta_exec;
294 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
296 curr->sum_exec_runtime += delta_exec;
297 cfs_rq->exec_clock += delta_exec;
302 delta_fair = calc_delta_fair(delta_exec, lw);
303 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
305 if (cfs_rq->sleeper_bonus > sysctl_sched_granularity) {
306 delta = min(cfs_rq->sleeper_bonus, (u64)delta_exec);
307 delta = calc_delta_mine(delta, curr->load.weight, lw);
308 delta = min((u64)delta, cfs_rq->sleeper_bonus);
309 delta = min(delta, (unsigned long)(
310 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
311 cfs_rq->sleeper_bonus -= delta;
315 cfs_rq->fair_clock += delta_fair;
317 * We executed delta_exec amount of time on the CPU,
318 * but we were only entitled to delta_mine amount of
319 * time during that period (if nr_running == 1 then
320 * the two values are equal)
321 * [Note: delta_mine - delta_exec is negative]:
323 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
326 static void update_curr(struct cfs_rq *cfs_rq)
328 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
329 unsigned long delta_exec;
335 * Get the amount of time the current task was running
336 * since the last time we changed load (this cannot
337 * overflow on 32 bits):
339 delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start);
341 curr->delta_exec += delta_exec;
343 if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
344 __update_curr(cfs_rq, curr);
345 curr->delta_exec = 0;
347 curr->exec_start = rq_of(cfs_rq)->clock;
351 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
353 se->wait_start_fair = cfs_rq->fair_clock;
354 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
358 * We calculate fair deltas here, so protect against the random effects
359 * of a multiplication overflow by capping it to the runtime limit:
361 #if BITS_PER_LONG == 32
362 static inline unsigned long
363 calc_weighted(unsigned long delta, unsigned long weight, int shift)
365 u64 tmp = (u64)delta * weight >> shift;
367 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
368 return sysctl_sched_runtime_limit*2;
372 static inline unsigned long
373 calc_weighted(unsigned long delta, unsigned long weight, int shift)
375 return delta * weight >> shift;
380 * Task is being enqueued - update stats:
382 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
387 * Are we enqueueing a waiting task? (for current tasks
388 * a dequeue/enqueue event is a NOP)
390 if (se != cfs_rq_curr(cfs_rq))
391 update_stats_wait_start(cfs_rq, se);
395 key = cfs_rq->fair_clock;
398 * Optimize the common nice 0 case:
400 if (likely(se->load.weight == NICE_0_LOAD)) {
401 key -= se->wait_runtime;
405 if (se->wait_runtime < 0) {
406 tmp = -se->wait_runtime;
407 key += (tmp * se->load.inv_weight) >>
408 (WMULT_SHIFT - NICE_0_SHIFT);
410 tmp = se->wait_runtime;
411 key -= (tmp * se->load.inv_weight) >>
412 (WMULT_SHIFT - NICE_0_SHIFT);
420 * Note: must be called with a freshly updated rq->fair_clock.
423 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
425 unsigned long delta_fair = se->delta_fair_run;
427 schedstat_set(se->wait_max, max(se->wait_max,
428 rq_of(cfs_rq)->clock - se->wait_start));
430 if (unlikely(se->load.weight != NICE_0_LOAD))
431 delta_fair = calc_weighted(delta_fair, se->load.weight,
434 add_wait_runtime(cfs_rq, se, delta_fair);
438 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
440 unsigned long delta_fair;
442 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
443 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
445 se->delta_fair_run += delta_fair;
446 if (unlikely(abs(se->delta_fair_run) >=
447 sysctl_sched_stat_granularity)) {
448 __update_stats_wait_end(cfs_rq, se);
449 se->delta_fair_run = 0;
452 se->wait_start_fair = 0;
453 schedstat_set(se->wait_start, 0);
457 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
461 * Mark the end of the wait period if dequeueing a
464 if (se != cfs_rq_curr(cfs_rq))
465 update_stats_wait_end(cfs_rq, se);
469 * We are picking a new current task - update its stats:
472 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
475 * We are starting a new run period:
477 se->exec_start = rq_of(cfs_rq)->clock;
481 * We are descheduling a task - update its stats:
484 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
489 /**************************************************
490 * Scheduling class queueing methods:
493 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
495 unsigned long load = cfs_rq->load.weight, delta_fair;
499 * Do not boost sleepers if there's too much bonus 'in flight'
502 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
505 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
506 load = rq_of(cfs_rq)->cpu_load[2];
508 delta_fair = se->delta_fair_sleep;
511 * Fix up delta_fair with the effect of us running
512 * during the whole sleep period:
514 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
515 delta_fair = div64_likely32((u64)delta_fair * load,
516 load + se->load.weight);
518 if (unlikely(se->load.weight != NICE_0_LOAD))
519 delta_fair = calc_weighted(delta_fair, se->load.weight,
522 prev_runtime = se->wait_runtime;
523 __add_wait_runtime(cfs_rq, se, delta_fair);
524 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
525 delta_fair = se->wait_runtime - prev_runtime;
528 * Track the amount of bonus we've given to sleepers:
530 cfs_rq->sleeper_bonus += delta_fair;
533 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
535 struct task_struct *tsk = task_of(se);
536 unsigned long delta_fair;
538 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
539 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
542 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
543 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
545 se->delta_fair_sleep += delta_fair;
546 if (unlikely(abs(se->delta_fair_sleep) >=
547 sysctl_sched_stat_granularity)) {
548 __enqueue_sleeper(cfs_rq, se);
549 se->delta_fair_sleep = 0;
552 se->sleep_start_fair = 0;
554 #ifdef CONFIG_SCHEDSTATS
555 if (se->sleep_start) {
556 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
561 if (unlikely(delta > se->sleep_max))
562 se->sleep_max = delta;
565 se->sum_sleep_runtime += delta;
567 if (se->block_start) {
568 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
573 if (unlikely(delta > se->block_max))
574 se->block_max = delta;
577 se->sum_sleep_runtime += delta;
583 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
586 * Update the fair clock.
591 enqueue_sleeper(cfs_rq, se);
593 update_stats_enqueue(cfs_rq, se);
594 __enqueue_entity(cfs_rq, se);
598 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
600 update_stats_dequeue(cfs_rq, se);
602 se->sleep_start_fair = cfs_rq->fair_clock;
603 #ifdef CONFIG_SCHEDSTATS
604 if (entity_is_task(se)) {
605 struct task_struct *tsk = task_of(se);
607 if (tsk->state & TASK_INTERRUPTIBLE)
608 se->sleep_start = rq_of(cfs_rq)->clock;
609 if (tsk->state & TASK_UNINTERRUPTIBLE)
610 se->block_start = rq_of(cfs_rq)->clock;
612 cfs_rq->wait_runtime -= se->wait_runtime;
615 __dequeue_entity(cfs_rq, se);
619 * Preempt the current task with a newly woken task if needed:
622 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
623 struct sched_entity *curr, unsigned long granularity)
625 s64 __delta = curr->fair_key - se->fair_key;
628 * Take scheduling granularity into account - do not
629 * preempt the current task unless the best task has
630 * a larger than sched_granularity fairness advantage:
632 if (__delta > niced_granularity(curr, granularity))
633 resched_task(rq_of(cfs_rq)->curr);
637 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 * Any task has to be enqueued before it get to execute on
641 * a CPU. So account for the time it spent waiting on the
642 * runqueue. (note, here we rely on pick_next_task() having
643 * done a put_prev_task_fair() shortly before this, which
644 * updated rq->fair_clock - used by update_stats_wait_end())
646 update_stats_wait_end(cfs_rq, se);
647 update_stats_curr_start(cfs_rq, se);
648 set_cfs_rq_curr(cfs_rq, se);
651 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
653 struct sched_entity *se = __pick_next_entity(cfs_rq);
655 set_next_entity(cfs_rq, se);
660 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
663 * If still on the runqueue then deactivate_task()
664 * was not called and update_curr() has to be done:
669 update_stats_curr_end(cfs_rq, prev);
672 update_stats_wait_start(cfs_rq, prev);
673 set_cfs_rq_curr(cfs_rq, NULL);
676 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
678 struct sched_entity *next;
681 * Dequeue and enqueue the task to update its
682 * position within the tree:
684 dequeue_entity(cfs_rq, curr, 0);
685 enqueue_entity(cfs_rq, curr, 0);
688 * Reschedule if another task tops the current one.
690 next = __pick_next_entity(cfs_rq);
694 __check_preempt_curr_fair(cfs_rq, next, curr, sysctl_sched_granularity);
697 /**************************************************
698 * CFS operations on tasks:
701 #ifdef CONFIG_FAIR_GROUP_SCHED
703 /* Walk up scheduling entities hierarchy */
704 #define for_each_sched_entity(se) \
705 for (; se; se = se->parent)
707 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
712 /* runqueue on which this entity is (to be) queued */
713 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
718 /* runqueue "owned" by this group */
719 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
724 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
725 * another cpu ('this_cpu')
727 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
729 /* A later patch will take group into account */
730 return &cpu_rq(this_cpu)->cfs;
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(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
737 /* Do the two (enqueued) tasks belong to the same group ? */
738 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
740 if (curr->se.cfs_rq == p->se.cfs_rq)
746 #else /* CONFIG_FAIR_GROUP_SCHED */
748 #define for_each_sched_entity(se) \
749 for (; se; se = NULL)
751 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
753 return &task_rq(p)->cfs;
756 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
758 struct task_struct *p = task_of(se);
759 struct rq *rq = task_rq(p);
764 /* runqueue "owned" by this group */
765 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
770 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
772 return &cpu_rq(this_cpu)->cfs;
775 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
776 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
778 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
783 #endif /* CONFIG_FAIR_GROUP_SCHED */
786 * The enqueue_task method is called before nr_running is
787 * increased. Here we update the fair scheduling stats and
788 * then put the task into the rbtree:
790 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
792 struct cfs_rq *cfs_rq;
793 struct sched_entity *se = &p->se;
795 for_each_sched_entity(se) {
798 cfs_rq = cfs_rq_of(se);
799 enqueue_entity(cfs_rq, se, wakeup);
804 * The dequeue_task method is called before nr_running is
805 * decreased. We remove the task from the rbtree and
806 * update the fair scheduling stats:
808 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
810 struct cfs_rq *cfs_rq;
811 struct sched_entity *se = &p->se;
813 for_each_sched_entity(se) {
814 cfs_rq = cfs_rq_of(se);
815 dequeue_entity(cfs_rq, se, sleep);
816 /* Don't dequeue parent if it has other entities besides us */
817 if (cfs_rq->load.weight)
823 * sched_yield() support is very simple - we dequeue and enqueue
825 static void yield_task_fair(struct rq *rq, struct task_struct *p)
827 struct cfs_rq *cfs_rq = task_cfs_rq(p);
829 __update_rq_clock(rq);
831 * Dequeue and enqueue the task to update its
832 * position within the tree:
834 dequeue_entity(cfs_rq, &p->se, 0);
835 enqueue_entity(cfs_rq, &p->se, 0);
839 * Preempt the current task with a newly woken task if needed:
841 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
843 struct task_struct *curr = rq->curr;
844 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
847 if (unlikely(rt_prio(p->prio))) {
854 gran = sysctl_sched_wakeup_granularity;
856 * Batch tasks prefer throughput over latency:
858 if (unlikely(p->policy == SCHED_BATCH))
859 gran = sysctl_sched_batch_wakeup_granularity;
861 if (is_same_group(curr, p))
862 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
865 static struct task_struct *pick_next_task_fair(struct rq *rq)
867 struct cfs_rq *cfs_rq = &rq->cfs;
868 struct sched_entity *se;
870 if (unlikely(!cfs_rq->nr_running))
874 se = pick_next_entity(cfs_rq);
875 cfs_rq = group_cfs_rq(se);
882 * Account for a descheduled task:
884 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
886 struct sched_entity *se = &prev->se;
887 struct cfs_rq *cfs_rq;
889 for_each_sched_entity(se) {
890 cfs_rq = cfs_rq_of(se);
891 put_prev_entity(cfs_rq, se);
895 /**************************************************
896 * Fair scheduling class load-balancing methods:
900 * Load-balancing iterator. Note: while the runqueue stays locked
901 * during the whole iteration, the current task might be
902 * dequeued so the iterator has to be dequeue-safe. Here we
903 * achieve that by always pre-iterating before returning
906 static inline struct task_struct *
907 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
909 struct task_struct *p;
914 p = rb_entry(curr, struct task_struct, se.run_node);
915 cfs_rq->rb_load_balance_curr = rb_next(curr);
920 static struct task_struct *load_balance_start_fair(void *arg)
922 struct cfs_rq *cfs_rq = arg;
924 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
927 static struct task_struct *load_balance_next_fair(void *arg)
929 struct cfs_rq *cfs_rq = arg;
931 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
934 #ifdef CONFIG_FAIR_GROUP_SCHED
935 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
937 struct sched_entity *curr;
938 struct task_struct *p;
940 if (!cfs_rq->nr_running)
943 curr = __pick_next_entity(cfs_rq);
951 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
952 unsigned long max_nr_move, unsigned long max_load_move,
953 struct sched_domain *sd, enum cpu_idle_type idle,
954 int *all_pinned, int *this_best_prio)
956 struct cfs_rq *busy_cfs_rq;
957 unsigned long load_moved, total_nr_moved = 0, nr_moved;
958 long rem_load_move = max_load_move;
959 struct rq_iterator cfs_rq_iterator;
961 cfs_rq_iterator.start = load_balance_start_fair;
962 cfs_rq_iterator.next = load_balance_next_fair;
964 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
965 #ifdef CONFIG_FAIR_GROUP_SCHED
966 struct cfs_rq *this_cfs_rq;
968 unsigned long maxload;
970 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
972 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
973 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
977 /* Don't pull more than imbalance/2 */
979 maxload = min(rem_load_move, imbalance);
981 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
983 # define maxload rem_load_move
985 /* pass busy_cfs_rq argument into
986 * load_balance_[start|next]_fair iterators
988 cfs_rq_iterator.arg = busy_cfs_rq;
989 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
990 max_nr_move, maxload, sd, idle, all_pinned,
991 &load_moved, this_best_prio, &cfs_rq_iterator);
993 total_nr_moved += nr_moved;
994 max_nr_move -= nr_moved;
995 rem_load_move -= load_moved;
997 if (max_nr_move <= 0 || rem_load_move <= 0)
1001 return max_load_move - rem_load_move;
1005 * scheduler tick hitting a task of our scheduling class:
1007 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1009 struct cfs_rq *cfs_rq;
1010 struct sched_entity *se = &curr->se;
1012 for_each_sched_entity(se) {
1013 cfs_rq = cfs_rq_of(se);
1014 entity_tick(cfs_rq, se);
1019 * Share the fairness runtime between parent and child, thus the
1020 * total amount of pressure for CPU stays equal - new tasks
1021 * get a chance to run but frequent forkers are not allowed to
1022 * monopolize the CPU. Note: the parent runqueue is locked,
1023 * the child is not running yet.
1025 static void task_new_fair(struct rq *rq, struct task_struct *p)
1027 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1028 struct sched_entity *se = &p->se;
1030 sched_info_queued(p);
1032 update_stats_enqueue(cfs_rq, se);
1034 * Child runs first: we let it run before the parent
1035 * until it reschedules once. We set up the key so that
1036 * it will preempt the parent:
1038 p->se.fair_key = current->se.fair_key -
1039 niced_granularity(&rq->curr->se, sysctl_sched_granularity) - 1;
1041 * The first wait is dominated by the child-runs-first logic,
1042 * so do not credit it with that waiting time yet:
1044 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1045 p->se.wait_start_fair = 0;
1048 * The statistical average of wait_runtime is about
1049 * -granularity/2, so initialize the task with that:
1051 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1052 p->se.wait_runtime = -(sysctl_sched_granularity / 2);
1054 __enqueue_entity(cfs_rq, se);
1057 #ifdef CONFIG_FAIR_GROUP_SCHED
1058 /* Account for a task changing its policy or group.
1060 * This routine is mostly called to set cfs_rq->curr field when a task
1061 * migrates between groups/classes.
1063 static void set_curr_task_fair(struct rq *rq)
1065 struct sched_entity *se = &rq->curr->se;
1067 for_each_sched_entity(se)
1068 set_next_entity(cfs_rq_of(se), se);
1071 static void set_curr_task_fair(struct rq *rq)
1077 * All the scheduling class methods:
1079 struct sched_class fair_sched_class __read_mostly = {
1080 .enqueue_task = enqueue_task_fair,
1081 .dequeue_task = dequeue_task_fair,
1082 .yield_task = yield_task_fair,
1084 .check_preempt_curr = check_preempt_curr_fair,
1086 .pick_next_task = pick_next_task_fair,
1087 .put_prev_task = put_prev_task_fair,
1089 .load_balance = load_balance_fair,
1091 .set_curr_task = set_curr_task_fair,
1092 .task_tick = task_tick_fair,
1093 .task_new = task_new_fair,
1096 #ifdef CONFIG_SCHED_DEBUG
1097 static void print_cfs_stats(struct seq_file *m, int cpu)
1099 struct cfs_rq *cfs_rq;
1101 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1102 print_cfs_rq(m, cpu, cfs_rq);