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: 2 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 = 2000000000ULL/HZ;
37 * SCHED_BATCH wake-up granularity.
38 * (default: 10 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 =
48 * SCHED_OTHER wake-up granularity.
49 * (default: 1 msec, units: nanoseconds)
51 * This option delays the preemption effects of decoupled workloads
52 * and reduces their over-scheduling. Synchronous workloads will still
53 * have immediate wakeup/sleep latencies.
55 unsigned int sysctl_sched_wakeup_granularity __read_mostly = 1000000000ULL/HZ;
57 unsigned int sysctl_sched_stat_granularity __read_mostly;
60 * Initialized in sched_init_granularity():
62 unsigned int sysctl_sched_runtime_limit __read_mostly;
65 * Debugging: various feature bits
68 SCHED_FEAT_FAIR_SLEEPERS = 1,
69 SCHED_FEAT_SLEEPER_AVG = 2,
70 SCHED_FEAT_SLEEPER_LOAD_AVG = 4,
71 SCHED_FEAT_PRECISE_CPU_LOAD = 8,
72 SCHED_FEAT_START_DEBIT = 16,
73 SCHED_FEAT_SKIP_INITIAL = 32,
76 unsigned int sysctl_sched_features __read_mostly =
77 SCHED_FEAT_FAIR_SLEEPERS *1 |
78 SCHED_FEAT_SLEEPER_AVG *1 |
79 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
80 SCHED_FEAT_PRECISE_CPU_LOAD *1 |
81 SCHED_FEAT_START_DEBIT *1 |
82 SCHED_FEAT_SKIP_INITIAL *0;
84 extern struct sched_class fair_sched_class;
86 /**************************************************************
87 * CFS operations on generic schedulable entities:
90 #ifdef CONFIG_FAIR_GROUP_SCHED
92 /* cpu runqueue to which this cfs_rq is attached */
93 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
98 /* currently running entity (if any) on this cfs_rq */
99 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
104 /* An entity is a task if it doesn't "own" a runqueue */
105 #define entity_is_task(se) (!se->my_q)
108 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
113 #else /* CONFIG_FAIR_GROUP_SCHED */
115 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
117 return container_of(cfs_rq, struct rq, cfs);
120 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
122 struct rq *rq = rq_of(cfs_rq);
124 if (unlikely(rq->curr->sched_class != &fair_sched_class))
127 return &rq->curr->se;
130 #define entity_is_task(se) 1
133 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
135 #endif /* CONFIG_FAIR_GROUP_SCHED */
137 static inline struct task_struct *task_of(struct sched_entity *se)
139 return container_of(se, struct task_struct, se);
143 /**************************************************************
144 * Scheduling class tree data structure manipulation methods:
148 * Enqueue an entity into the rb-tree:
151 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
153 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
154 struct rb_node *parent = NULL;
155 struct sched_entity *entry;
156 s64 key = se->fair_key;
160 * Find the right place in the rbtree:
164 entry = rb_entry(parent, struct sched_entity, run_node);
166 * We dont care about collisions. Nodes with
167 * the same key stay together.
169 if (key - entry->fair_key < 0) {
170 link = &parent->rb_left;
172 link = &parent->rb_right;
178 * Maintain a cache of leftmost tree entries (it is frequently
182 cfs_rq->rb_leftmost = &se->run_node;
184 rb_link_node(&se->run_node, parent, link);
185 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
186 update_load_add(&cfs_rq->load, se->load.weight);
187 cfs_rq->nr_running++;
192 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
194 if (cfs_rq->rb_leftmost == &se->run_node)
195 cfs_rq->rb_leftmost = rb_next(&se->run_node);
196 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
197 update_load_sub(&cfs_rq->load, se->load.weight);
198 cfs_rq->nr_running--;
202 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
204 return cfs_rq->rb_leftmost;
207 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
209 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
212 /**************************************************************
213 * Scheduling class statistics methods:
217 * We rescale the rescheduling granularity of tasks according to their
218 * nice level, but only linearly, not exponentially:
221 niced_granularity(struct sched_entity *curr, unsigned long granularity)
225 if (likely(curr->load.weight == NICE_0_LOAD))
228 * Positive nice levels get the same granularity as nice-0:
230 if (likely(curr->load.weight < NICE_0_LOAD)) {
231 tmp = curr->load.weight * (u64)granularity;
232 return (long) (tmp >> NICE_0_SHIFT);
235 * Negative nice level tasks get linearly finer
238 tmp = curr->load.inv_weight * (u64)granularity;
241 * It will always fit into 'long':
243 return (long) (tmp >> WMULT_SHIFT);
247 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
249 long limit = sysctl_sched_runtime_limit;
252 * Niced tasks have the same history dynamic range as
255 if (unlikely(se->wait_runtime > limit)) {
256 se->wait_runtime = limit;
257 schedstat_inc(se, wait_runtime_overruns);
258 schedstat_inc(cfs_rq, wait_runtime_overruns);
260 if (unlikely(se->wait_runtime < -limit)) {
261 se->wait_runtime = -limit;
262 schedstat_inc(se, wait_runtime_underruns);
263 schedstat_inc(cfs_rq, wait_runtime_underruns);
268 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
270 se->wait_runtime += delta;
271 schedstat_add(se, sum_wait_runtime, delta);
272 limit_wait_runtime(cfs_rq, se);
276 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
278 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
279 __add_wait_runtime(cfs_rq, se, delta);
280 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
284 * Update the current task's runtime statistics. Skip current tasks that
285 * are not in our scheduling class.
288 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr)
290 unsigned long delta, delta_exec, delta_fair, delta_mine;
291 struct load_weight *lw = &cfs_rq->load;
292 unsigned long load = lw->weight;
294 delta_exec = curr->delta_exec;
295 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
297 curr->sum_exec_runtime += delta_exec;
298 cfs_rq->exec_clock += delta_exec;
303 delta_fair = calc_delta_fair(delta_exec, lw);
304 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
306 if (cfs_rq->sleeper_bonus > sysctl_sched_granularity) {
307 delta = calc_delta_mine(cfs_rq->sleeper_bonus,
308 curr->load.weight, lw);
309 if (unlikely(delta > cfs_rq->sleeper_bonus))
310 delta = cfs_rq->sleeper_bonus;
312 cfs_rq->sleeper_bonus -= delta;
316 cfs_rq->fair_clock += delta_fair;
318 * We executed delta_exec amount of time on the CPU,
319 * but we were only entitled to delta_mine amount of
320 * time during that period (if nr_running == 1 then
321 * the two values are equal)
322 * [Note: delta_mine - delta_exec is negative]:
324 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
327 static void update_curr(struct cfs_rq *cfs_rq)
329 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
330 unsigned long delta_exec;
336 * Get the amount of time the current task was running
337 * since the last time we changed load (this cannot
338 * overflow on 32 bits):
340 delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start);
342 curr->delta_exec += delta_exec;
344 if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
345 __update_curr(cfs_rq, curr);
346 curr->delta_exec = 0;
348 curr->exec_start = rq_of(cfs_rq)->clock;
352 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
354 se->wait_start_fair = cfs_rq->fair_clock;
355 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
359 * We calculate fair deltas here, so protect against the random effects
360 * of a multiplication overflow by capping it to the runtime limit:
362 #if BITS_PER_LONG == 32
363 static inline unsigned long
364 calc_weighted(unsigned long delta, unsigned long weight, int shift)
366 u64 tmp = (u64)delta * weight >> shift;
368 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
369 return sysctl_sched_runtime_limit*2;
373 static inline unsigned long
374 calc_weighted(unsigned long delta, unsigned long weight, int shift)
376 return delta * weight >> shift;
381 * Task is being enqueued - update stats:
383 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
388 * Are we enqueueing a waiting task? (for current tasks
389 * a dequeue/enqueue event is a NOP)
391 if (se != cfs_rq_curr(cfs_rq))
392 update_stats_wait_start(cfs_rq, se);
396 key = cfs_rq->fair_clock;
399 * Optimize the common nice 0 case:
401 if (likely(se->load.weight == NICE_0_LOAD)) {
402 key -= se->wait_runtime;
406 if (se->wait_runtime < 0) {
407 tmp = -se->wait_runtime;
408 key += (tmp * se->load.inv_weight) >>
409 (WMULT_SHIFT - NICE_0_SHIFT);
411 tmp = se->wait_runtime;
412 key -= (tmp * se->load.inv_weight) >>
413 (WMULT_SHIFT - NICE_0_SHIFT);
421 * Note: must be called with a freshly updated rq->fair_clock.
424 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
426 unsigned long delta_fair = se->delta_fair_run;
428 schedstat_set(se->wait_max, max(se->wait_max,
429 rq_of(cfs_rq)->clock - se->wait_start));
431 if (unlikely(se->load.weight != NICE_0_LOAD))
432 delta_fair = calc_weighted(delta_fair, se->load.weight,
435 add_wait_runtime(cfs_rq, se, delta_fair);
439 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
441 unsigned long delta_fair;
443 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
444 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
446 se->delta_fair_run += delta_fair;
447 if (unlikely(abs(se->delta_fair_run) >=
448 sysctl_sched_stat_granularity)) {
449 __update_stats_wait_end(cfs_rq, se);
450 se->delta_fair_run = 0;
453 se->wait_start_fair = 0;
454 schedstat_set(se->wait_start, 0);
458 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
462 * Mark the end of the wait period if dequeueing a
465 if (se != cfs_rq_curr(cfs_rq))
466 update_stats_wait_end(cfs_rq, se);
470 * We are picking a new current task - update its stats:
473 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
476 * We are starting a new run period:
478 se->exec_start = rq_of(cfs_rq)->clock;
482 * We are descheduling a task - update its stats:
485 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
490 /**************************************************
491 * Scheduling class queueing methods:
494 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
496 unsigned long load = cfs_rq->load.weight, delta_fair;
499 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
500 load = rq_of(cfs_rq)->cpu_load[2];
502 delta_fair = se->delta_fair_sleep;
505 * Fix up delta_fair with the effect of us running
506 * during the whole sleep period:
508 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
509 delta_fair = div64_likely32((u64)delta_fair * load,
510 load + se->load.weight);
512 if (unlikely(se->load.weight != NICE_0_LOAD))
513 delta_fair = calc_weighted(delta_fair, se->load.weight,
516 prev_runtime = se->wait_runtime;
517 __add_wait_runtime(cfs_rq, se, delta_fair);
518 delta_fair = se->wait_runtime - prev_runtime;
521 * Track the amount of bonus we've given to sleepers:
523 cfs_rq->sleeper_bonus += delta_fair;
525 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
528 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
530 struct task_struct *tsk = task_of(se);
531 unsigned long delta_fair;
533 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
534 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
537 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
538 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
540 se->delta_fair_sleep += delta_fair;
541 if (unlikely(abs(se->delta_fair_sleep) >=
542 sysctl_sched_stat_granularity)) {
543 __enqueue_sleeper(cfs_rq, se);
544 se->delta_fair_sleep = 0;
547 se->sleep_start_fair = 0;
549 #ifdef CONFIG_SCHEDSTATS
550 if (se->sleep_start) {
551 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
556 if (unlikely(delta > se->sleep_max))
557 se->sleep_max = delta;
560 se->sum_sleep_runtime += delta;
562 if (se->block_start) {
563 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
568 if (unlikely(delta > se->block_max))
569 se->block_max = delta;
572 se->sum_sleep_runtime += delta;
578 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
581 * Update the fair clock.
586 enqueue_sleeper(cfs_rq, se);
588 update_stats_enqueue(cfs_rq, se);
589 __enqueue_entity(cfs_rq, se);
593 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
595 update_stats_dequeue(cfs_rq, se);
597 se->sleep_start_fair = cfs_rq->fair_clock;
598 #ifdef CONFIG_SCHEDSTATS
599 if (entity_is_task(se)) {
600 struct task_struct *tsk = task_of(se);
602 if (tsk->state & TASK_INTERRUPTIBLE)
603 se->sleep_start = rq_of(cfs_rq)->clock;
604 if (tsk->state & TASK_UNINTERRUPTIBLE)
605 se->block_start = rq_of(cfs_rq)->clock;
607 cfs_rq->wait_runtime -= se->wait_runtime;
610 __dequeue_entity(cfs_rq, se);
614 * Preempt the current task with a newly woken task if needed:
617 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
618 struct sched_entity *curr, unsigned long granularity)
620 s64 __delta = curr->fair_key - se->fair_key;
623 * Take scheduling granularity into account - do not
624 * preempt the current task unless the best task has
625 * a larger than sched_granularity fairness advantage:
627 if (__delta > niced_granularity(curr, granularity))
628 resched_task(rq_of(cfs_rq)->curr);
632 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
635 * Any task has to be enqueued before it get to execute on
636 * a CPU. So account for the time it spent waiting on the
637 * runqueue. (note, here we rely on pick_next_task() having
638 * done a put_prev_task_fair() shortly before this, which
639 * updated rq->fair_clock - used by update_stats_wait_end())
641 update_stats_wait_end(cfs_rq, se);
642 update_stats_curr_start(cfs_rq, se);
643 set_cfs_rq_curr(cfs_rq, se);
646 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
648 struct sched_entity *se = __pick_next_entity(cfs_rq);
650 set_next_entity(cfs_rq, se);
655 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
658 * If still on the runqueue then deactivate_task()
659 * was not called and update_curr() has to be done:
664 update_stats_curr_end(cfs_rq, prev);
667 update_stats_wait_start(cfs_rq, prev);
668 set_cfs_rq_curr(cfs_rq, NULL);
671 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
673 struct sched_entity *next;
676 * Dequeue and enqueue the task to update its
677 * position within the tree:
679 dequeue_entity(cfs_rq, curr, 0);
680 enqueue_entity(cfs_rq, curr, 0);
683 * Reschedule if another task tops the current one.
685 next = __pick_next_entity(cfs_rq);
689 __check_preempt_curr_fair(cfs_rq, next, curr, sysctl_sched_granularity);
692 /**************************************************
693 * CFS operations on tasks:
696 #ifdef CONFIG_FAIR_GROUP_SCHED
698 /* Walk up scheduling entities hierarchy */
699 #define for_each_sched_entity(se) \
700 for (; se; se = se->parent)
702 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
707 /* runqueue on which this entity is (to be) queued */
708 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
713 /* runqueue "owned" by this group */
714 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
719 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
720 * another cpu ('this_cpu')
722 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
724 /* A later patch will take group into account */
725 return &cpu_rq(this_cpu)->cfs;
728 /* Iterate thr' all leaf cfs_rq's on a runqueue */
729 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
730 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
732 /* Do the two (enqueued) tasks belong to the same group ? */
733 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
735 if (curr->se.cfs_rq == p->se.cfs_rq)
741 #else /* CONFIG_FAIR_GROUP_SCHED */
743 #define for_each_sched_entity(se) \
744 for (; se; se = NULL)
746 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
748 return &task_rq(p)->cfs;
751 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
753 struct task_struct *p = task_of(se);
754 struct rq *rq = task_rq(p);
759 /* runqueue "owned" by this group */
760 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
765 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
767 return &cpu_rq(this_cpu)->cfs;
770 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
771 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
773 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
778 #endif /* CONFIG_FAIR_GROUP_SCHED */
781 * The enqueue_task method is called before nr_running is
782 * increased. Here we update the fair scheduling stats and
783 * then put the task into the rbtree:
785 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
787 struct cfs_rq *cfs_rq;
788 struct sched_entity *se = &p->se;
790 for_each_sched_entity(se) {
793 cfs_rq = cfs_rq_of(se);
794 enqueue_entity(cfs_rq, se, wakeup);
799 * The dequeue_task method is called before nr_running is
800 * decreased. We remove the task from the rbtree and
801 * update the fair scheduling stats:
803 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
805 struct cfs_rq *cfs_rq;
806 struct sched_entity *se = &p->se;
808 for_each_sched_entity(se) {
809 cfs_rq = cfs_rq_of(se);
810 dequeue_entity(cfs_rq, se, sleep);
811 /* Don't dequeue parent if it has other entities besides us */
812 if (cfs_rq->load.weight)
818 * sched_yield() support is very simple - we dequeue and enqueue
820 static void yield_task_fair(struct rq *rq, struct task_struct *p)
822 struct cfs_rq *cfs_rq = task_cfs_rq(p);
824 __update_rq_clock(rq);
826 * Dequeue and enqueue the task to update its
827 * position within the tree:
829 dequeue_entity(cfs_rq, &p->se, 0);
830 enqueue_entity(cfs_rq, &p->se, 0);
834 * Preempt the current task with a newly woken task if needed:
836 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
838 struct task_struct *curr = rq->curr;
839 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
842 if (unlikely(rt_prio(p->prio))) {
849 gran = sysctl_sched_wakeup_granularity;
851 * Batch tasks prefer throughput over latency:
853 if (unlikely(p->policy == SCHED_BATCH))
854 gran = sysctl_sched_batch_wakeup_granularity;
856 if (is_same_group(curr, p))
857 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
860 static struct task_struct *pick_next_task_fair(struct rq *rq)
862 struct cfs_rq *cfs_rq = &rq->cfs;
863 struct sched_entity *se;
865 if (unlikely(!cfs_rq->nr_running))
869 se = pick_next_entity(cfs_rq);
870 cfs_rq = group_cfs_rq(se);
877 * Account for a descheduled task:
879 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
881 struct sched_entity *se = &prev->se;
882 struct cfs_rq *cfs_rq;
884 for_each_sched_entity(se) {
885 cfs_rq = cfs_rq_of(se);
886 put_prev_entity(cfs_rq, se);
890 /**************************************************
891 * Fair scheduling class load-balancing methods:
895 * Load-balancing iterator. Note: while the runqueue stays locked
896 * during the whole iteration, the current task might be
897 * dequeued so the iterator has to be dequeue-safe. Here we
898 * achieve that by always pre-iterating before returning
901 static inline struct task_struct *
902 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
904 struct task_struct *p;
909 p = rb_entry(curr, struct task_struct, se.run_node);
910 cfs_rq->rb_load_balance_curr = rb_next(curr);
915 static struct task_struct *load_balance_start_fair(void *arg)
917 struct cfs_rq *cfs_rq = arg;
919 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
922 static struct task_struct *load_balance_next_fair(void *arg)
924 struct cfs_rq *cfs_rq = arg;
926 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
929 #ifdef CONFIG_FAIR_GROUP_SCHED
930 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
932 struct sched_entity *curr;
933 struct task_struct *p;
935 if (!cfs_rq->nr_running)
938 curr = __pick_next_entity(cfs_rq);
946 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
947 unsigned long max_nr_move, unsigned long max_load_move,
948 struct sched_domain *sd, enum cpu_idle_type idle,
949 int *all_pinned, int *this_best_prio)
951 struct cfs_rq *busy_cfs_rq;
952 unsigned long load_moved, total_nr_moved = 0, nr_moved;
953 long rem_load_move = max_load_move;
954 struct rq_iterator cfs_rq_iterator;
956 cfs_rq_iterator.start = load_balance_start_fair;
957 cfs_rq_iterator.next = load_balance_next_fair;
959 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
960 #ifdef CONFIG_FAIR_GROUP_SCHED
961 struct cfs_rq *this_cfs_rq;
963 unsigned long maxload;
965 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
967 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
968 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
972 /* Don't pull more than imbalance/2 */
974 maxload = min(rem_load_move, imbalance);
976 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
978 # define maxload rem_load_move
980 /* pass busy_cfs_rq argument into
981 * load_balance_[start|next]_fair iterators
983 cfs_rq_iterator.arg = busy_cfs_rq;
984 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
985 max_nr_move, maxload, sd, idle, all_pinned,
986 &load_moved, this_best_prio, &cfs_rq_iterator);
988 total_nr_moved += nr_moved;
989 max_nr_move -= nr_moved;
990 rem_load_move -= load_moved;
992 if (max_nr_move <= 0 || rem_load_move <= 0)
996 return max_load_move - rem_load_move;
1000 * scheduler tick hitting a task of our scheduling class:
1002 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1004 struct cfs_rq *cfs_rq;
1005 struct sched_entity *se = &curr->se;
1007 for_each_sched_entity(se) {
1008 cfs_rq = cfs_rq_of(se);
1009 entity_tick(cfs_rq, se);
1014 * Share the fairness runtime between parent and child, thus the
1015 * total amount of pressure for CPU stays equal - new tasks
1016 * get a chance to run but frequent forkers are not allowed to
1017 * monopolize the CPU. Note: the parent runqueue is locked,
1018 * the child is not running yet.
1020 static void task_new_fair(struct rq *rq, struct task_struct *p)
1022 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1023 struct sched_entity *se = &p->se;
1025 sched_info_queued(p);
1027 update_stats_enqueue(cfs_rq, se);
1029 * Child runs first: we let it run before the parent
1030 * until it reschedules once. We set up the key so that
1031 * it will preempt the parent:
1033 p->se.fair_key = current->se.fair_key -
1034 niced_granularity(&rq->curr->se, sysctl_sched_granularity) - 1;
1036 * The first wait is dominated by the child-runs-first logic,
1037 * so do not credit it with that waiting time yet:
1039 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1040 p->se.wait_start_fair = 0;
1043 * The statistical average of wait_runtime is about
1044 * -granularity/2, so initialize the task with that:
1046 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1047 p->se.wait_runtime = -(sysctl_sched_granularity / 2);
1049 __enqueue_entity(cfs_rq, se);
1052 #ifdef CONFIG_FAIR_GROUP_SCHED
1053 /* Account for a task changing its policy or group.
1055 * This routine is mostly called to set cfs_rq->curr field when a task
1056 * migrates between groups/classes.
1058 static void set_curr_task_fair(struct rq *rq)
1060 struct sched_entity *se = &rq->curr.se;
1062 for_each_sched_entity(se)
1063 set_next_entity(cfs_rq_of(se), se);
1066 static void set_curr_task_fair(struct rq *rq)
1072 * All the scheduling class methods:
1074 struct sched_class fair_sched_class __read_mostly = {
1075 .enqueue_task = enqueue_task_fair,
1076 .dequeue_task = dequeue_task_fair,
1077 .yield_task = yield_task_fair,
1079 .check_preempt_curr = check_preempt_curr_fair,
1081 .pick_next_task = pick_next_task_fair,
1082 .put_prev_task = put_prev_task_fair,
1084 .load_balance = load_balance_fair,
1086 .set_curr_task = set_curr_task_fair,
1087 .task_tick = task_tick_fair,
1088 .task_new = task_new_fair,
1091 #ifdef CONFIG_SCHED_DEBUG
1092 static void print_cfs_stats(struct seq_file *m, int cpu)
1094 struct cfs_rq *cfs_rq;
1096 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1097 print_cfs_rq(m, cpu, cfs_rq);