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>
24 * Targeted preemption latency for CPU-bound tasks:
25 * (default: 20ms, units: nanoseconds)
27 * NOTE: this latency value is not the same as the concept of
28 * 'timeslice length' - timeslices in CFS are of variable length.
29 * (to see the precise effective timeslice length of your workload,
30 * run vmstat and monitor the context-switches field)
32 * On SMP systems the value of this is multiplied by the log2 of the
33 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
34 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
35 * Targeted preemption latency for CPU-bound tasks:
37 const_debug unsigned int sysctl_sched_latency = 20000000ULL;
40 * After fork, child runs first. (default) If set to 0 then
41 * parent will (try to) run first.
43 const_debug unsigned int sysctl_sched_child_runs_first = 1;
46 * Minimal preemption granularity for CPU-bound tasks:
47 * (default: 2 msec, units: nanoseconds)
49 unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL;
52 * sys_sched_yield() compat mode
54 * This option switches the agressive yield implementation of the
55 * old scheduler back on.
57 unsigned int __read_mostly sysctl_sched_compat_yield;
60 * SCHED_BATCH wake-up granularity.
61 * (default: 25 msec, units: nanoseconds)
63 * This option delays the preemption effects of decoupled workloads
64 * and reduces their over-scheduling. Synchronous workloads will still
65 * have immediate wakeup/sleep latencies.
67 const_debug unsigned int sysctl_sched_batch_wakeup_granularity = 25000000UL;
70 * SCHED_OTHER wake-up granularity.
71 * (default: 1 msec, units: nanoseconds)
73 * This option delays the preemption effects of decoupled workloads
74 * and reduces their over-scheduling. Synchronous workloads will still
75 * have immediate wakeup/sleep latencies.
77 const_debug unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
79 unsigned int sysctl_sched_runtime_limit __read_mostly;
81 extern struct sched_class fair_sched_class;
83 /**************************************************************
84 * CFS operations on generic schedulable entities:
87 #ifdef CONFIG_FAIR_GROUP_SCHED
89 /* cpu runqueue to which this cfs_rq is attached */
90 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
95 /* An entity is a task if it doesn't "own" a runqueue */
96 #define entity_is_task(se) (!se->my_q)
98 #else /* CONFIG_FAIR_GROUP_SCHED */
100 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
102 return container_of(cfs_rq, struct rq, cfs);
105 #define entity_is_task(se) 1
107 #endif /* CONFIG_FAIR_GROUP_SCHED */
109 static inline struct task_struct *task_of(struct sched_entity *se)
111 return container_of(se, struct task_struct, se);
115 /**************************************************************
116 * Scheduling class tree data structure manipulation methods:
120 set_leftmost(struct cfs_rq *cfs_rq, struct rb_node *leftmost)
122 struct sched_entity *se;
124 cfs_rq->rb_leftmost = leftmost;
126 se = rb_entry(leftmost, struct sched_entity, run_node);
127 cfs_rq->min_vruntime = max(se->vruntime,
128 cfs_rq->min_vruntime);
133 * Enqueue an entity into the rb-tree:
136 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
138 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
139 struct rb_node *parent = NULL;
140 struct sched_entity *entry;
141 s64 key = se->fair_key;
145 * Find the right place in the rbtree:
149 entry = rb_entry(parent, struct sched_entity, run_node);
151 * We dont care about collisions. Nodes with
152 * the same key stay together.
154 if (key - entry->fair_key < 0) {
155 link = &parent->rb_left;
157 link = &parent->rb_right;
163 * Maintain a cache of leftmost tree entries (it is frequently
167 set_leftmost(cfs_rq, &se->run_node);
169 rb_link_node(&se->run_node, parent, link);
170 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
171 update_load_add(&cfs_rq->load, se->load.weight);
172 cfs_rq->nr_running++;
175 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
179 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
181 if (cfs_rq->rb_leftmost == &se->run_node)
182 set_leftmost(cfs_rq, rb_next(&se->run_node));
184 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
185 update_load_sub(&cfs_rq->load, se->load.weight);
186 cfs_rq->nr_running--;
189 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
192 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
194 return cfs_rq->rb_leftmost;
197 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
199 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
202 /**************************************************************
203 * Scheduling class statistics methods:
207 * Calculate the preemption granularity needed to schedule every
208 * runnable task once per sysctl_sched_latency amount of time.
209 * (down to a sensible low limit on granularity)
211 * For example, if there are 2 tasks running and latency is 10 msecs,
212 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
213 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
214 * for each task. We do finer and finer scheduling up to until we
215 * reach the minimum granularity value.
217 * To achieve this we use the following dynamic-granularity rule:
219 * gran = lat/nr - lat/nr/nr
221 * This comes out of the following equations:
226 * kB2 = kB1 - d + d/nr
229 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
230 * '1' is start of time, '2' is end of time, 'd' is delay between
231 * 1 and 2 (during which task B was running), 'nr' is number of tasks
232 * running, 'lat' is the the period of each task. ('lat' is the
233 * sched_latency that we aim for.)
236 sched_granularity(struct cfs_rq *cfs_rq)
238 unsigned int gran = sysctl_sched_latency;
239 unsigned int nr = cfs_rq->nr_running;
242 gran = gran/nr - gran/nr/nr;
243 gran = max(gran, sysctl_sched_min_granularity);
250 * We rescale the rescheduling granularity of tasks according to their
251 * nice level, but only linearly, not exponentially:
254 niced_granularity(struct sched_entity *curr, unsigned long granularity)
258 if (likely(curr->load.weight == NICE_0_LOAD))
261 * Positive nice levels get the same granularity as nice-0:
263 if (likely(curr->load.weight < NICE_0_LOAD)) {
264 tmp = curr->load.weight * (u64)granularity;
265 return (long) (tmp >> NICE_0_SHIFT);
268 * Negative nice level tasks get linearly finer
271 tmp = curr->load.inv_weight * (u64)granularity;
274 * It will always fit into 'long':
276 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
280 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
282 long limit = sysctl_sched_runtime_limit;
285 * Niced tasks have the same history dynamic range as
288 if (unlikely(se->wait_runtime > limit)) {
289 se->wait_runtime = limit;
290 schedstat_inc(se, wait_runtime_overruns);
291 schedstat_inc(cfs_rq, wait_runtime_overruns);
293 if (unlikely(se->wait_runtime < -limit)) {
294 se->wait_runtime = -limit;
295 schedstat_inc(se, wait_runtime_underruns);
296 schedstat_inc(cfs_rq, wait_runtime_underruns);
301 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
303 se->wait_runtime += delta;
304 schedstat_add(se, sum_wait_runtime, delta);
305 limit_wait_runtime(cfs_rq, se);
309 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
311 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
312 __add_wait_runtime(cfs_rq, se, delta);
313 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
317 * Update the current task's runtime statistics. Skip current tasks that
318 * are not in our scheduling class.
321 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
322 unsigned long delta_exec)
324 unsigned long delta, delta_fair, delta_mine, delta_exec_weighted;
325 struct load_weight *lw = &cfs_rq->load;
326 unsigned long load = lw->weight;
328 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
330 curr->sum_exec_runtime += delta_exec;
331 cfs_rq->exec_clock += delta_exec;
332 delta_exec_weighted = delta_exec;
333 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
334 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
337 curr->vruntime += delta_exec_weighted;
342 delta_fair = calc_delta_fair(delta_exec, lw);
343 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
345 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
346 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
347 delta = min(delta, (unsigned long)(
348 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
349 cfs_rq->sleeper_bonus -= delta;
353 cfs_rq->fair_clock += delta_fair;
355 * We executed delta_exec amount of time on the CPU,
356 * but we were only entitled to delta_mine amount of
357 * time during that period (if nr_running == 1 then
358 * the two values are equal)
359 * [Note: delta_mine - delta_exec is negative]:
361 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
364 static void update_curr(struct cfs_rq *cfs_rq)
366 struct sched_entity *curr = cfs_rq->curr;
367 u64 now = rq_of(cfs_rq)->clock;
368 unsigned long delta_exec;
374 * Get the amount of time the current task was running
375 * since the last time we changed load (this cannot
376 * overflow on 32 bits):
378 delta_exec = (unsigned long)(now - curr->exec_start);
380 __update_curr(cfs_rq, curr, delta_exec);
381 curr->exec_start = now;
385 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
387 se->wait_start_fair = cfs_rq->fair_clock;
388 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
391 static inline unsigned long
392 calc_weighted(unsigned long delta, struct sched_entity *se)
394 unsigned long weight = se->load.weight;
396 if (unlikely(weight != NICE_0_LOAD))
397 return (u64)delta * se->load.weight >> NICE_0_SHIFT;
403 * Task is being enqueued - update stats:
405 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
408 * Are we enqueueing a waiting task? (for current tasks
409 * a dequeue/enqueue event is a NOP)
411 if (se != cfs_rq->curr)
412 update_stats_wait_start(cfs_rq, se);
416 se->fair_key = se->vruntime;
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,
424 unsigned long delta_fair)
426 schedstat_set(se->wait_max, max(se->wait_max,
427 rq_of(cfs_rq)->clock - se->wait_start));
429 delta_fair = calc_weighted(delta_fair, se);
431 add_wait_runtime(cfs_rq, se, delta_fair);
435 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
437 unsigned long delta_fair;
439 if (unlikely(!se->wait_start_fair))
442 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
443 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
445 __update_stats_wait_end(cfs_rq, se, delta_fair);
447 se->wait_start_fair = 0;
448 schedstat_set(se->wait_start, 0);
452 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
456 * Mark the end of the wait period if dequeueing a
459 if (se != cfs_rq->curr)
460 update_stats_wait_end(cfs_rq, se);
464 * We are picking a new current task - update its stats:
467 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
470 * We are starting a new run period:
472 se->exec_start = rq_of(cfs_rq)->clock;
476 * We are descheduling a task - update its stats:
479 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
484 /**************************************************
485 * Scheduling class queueing methods:
488 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
489 unsigned long delta_fair)
491 unsigned long load = cfs_rq->load.weight;
495 * Do not boost sleepers if there's too much bonus 'in flight'
498 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
501 if (sched_feat(SLEEPER_LOAD_AVG))
502 load = rq_of(cfs_rq)->cpu_load[2];
505 * Fix up delta_fair with the effect of us running
506 * during the whole sleep period:
508 if (sched_feat(SLEEPER_AVG))
509 delta_fair = div64_likely32((u64)delta_fair * load,
510 load + se->load.weight);
512 delta_fair = calc_weighted(delta_fair, se);
514 prev_runtime = se->wait_runtime;
515 __add_wait_runtime(cfs_rq, se, delta_fair);
516 delta_fair = se->wait_runtime - prev_runtime;
519 * Track the amount of bonus we've given to sleepers:
521 cfs_rq->sleeper_bonus += delta_fair;
524 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
526 struct task_struct *tsk = task_of(se);
527 unsigned long delta_fair;
529 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
530 !sched_feat(FAIR_SLEEPERS))
533 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
534 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
536 __enqueue_sleeper(cfs_rq, se, delta_fair);
538 se->sleep_start_fair = 0;
540 #ifdef CONFIG_SCHEDSTATS
541 if (se->sleep_start) {
542 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
547 if (unlikely(delta > se->sleep_max))
548 se->sleep_max = delta;
551 se->sum_sleep_runtime += delta;
553 if (se->block_start) {
554 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
559 if (unlikely(delta > se->block_max))
560 se->block_max = delta;
563 se->sum_sleep_runtime += delta;
566 * Blocking time is in units of nanosecs, so shift by 20 to
567 * get a milliseconds-range estimation of the amount of
568 * time that the task spent sleeping:
570 if (unlikely(prof_on == SLEEP_PROFILING)) {
571 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
579 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
582 * Update the fair clock.
587 u64 min_runtime, latency;
589 min_runtime = cfs_rq->min_vruntime;
590 min_runtime += sysctl_sched_latency/2;
592 if (sched_feat(NEW_FAIR_SLEEPERS)) {
593 latency = calc_weighted(sysctl_sched_latency, se);
594 if (min_runtime > latency)
595 min_runtime -= latency;
598 se->vruntime = max(se->vruntime, min_runtime);
600 enqueue_sleeper(cfs_rq, se);
603 update_stats_enqueue(cfs_rq, se);
604 __enqueue_entity(cfs_rq, se);
608 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
610 update_stats_dequeue(cfs_rq, se);
612 se->sleep_start_fair = cfs_rq->fair_clock;
613 #ifdef CONFIG_SCHEDSTATS
614 if (entity_is_task(se)) {
615 struct task_struct *tsk = task_of(se);
617 if (tsk->state & TASK_INTERRUPTIBLE)
618 se->sleep_start = rq_of(cfs_rq)->clock;
619 if (tsk->state & TASK_UNINTERRUPTIBLE)
620 se->block_start = rq_of(cfs_rq)->clock;
624 __dequeue_entity(cfs_rq, se);
628 * Preempt the current task with a newly woken task if needed:
631 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
632 struct sched_entity *curr, unsigned long granularity)
634 s64 __delta = curr->fair_key - se->fair_key;
635 unsigned long ideal_runtime, delta_exec;
638 * ideal_runtime is compared against sum_exec_runtime, which is
639 * walltime, hence do not scale.
641 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
642 (unsigned long)sysctl_sched_min_granularity);
645 * If we executed more than what the latency constraint suggests,
646 * reduce the rescheduling granularity. This way the total latency
647 * of how much a task is not scheduled converges to
648 * sysctl_sched_latency:
650 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
651 if (delta_exec > ideal_runtime)
655 * Take scheduling granularity into account - do not
656 * preempt the current task unless the best task has
657 * a larger than sched_granularity fairness advantage:
659 * scale granularity as key space is in fair_clock.
661 if (__delta > niced_granularity(curr, granularity))
662 resched_task(rq_of(cfs_rq)->curr);
666 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
669 * Any task has to be enqueued before it get to execute on
670 * a CPU. So account for the time it spent waiting on the
671 * runqueue. (note, here we rely on pick_next_task() having
672 * done a put_prev_task_fair() shortly before this, which
673 * updated rq->fair_clock - used by update_stats_wait_end())
675 update_stats_wait_end(cfs_rq, se);
676 update_stats_curr_start(cfs_rq, se);
678 #ifdef CONFIG_SCHEDSTATS
680 * Track our maximum slice length, if the CPU's load is at
681 * least twice that of our own weight (i.e. dont track it
682 * when there are only lesser-weight tasks around):
684 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
685 se->slice_max = max(se->slice_max,
686 se->sum_exec_runtime - se->prev_sum_exec_runtime);
689 se->prev_sum_exec_runtime = se->sum_exec_runtime;
692 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
694 struct sched_entity *se = __pick_next_entity(cfs_rq);
696 set_next_entity(cfs_rq, se);
701 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
704 * If still on the runqueue then deactivate_task()
705 * was not called and update_curr() has to be done:
710 update_stats_curr_end(cfs_rq, prev);
713 update_stats_wait_start(cfs_rq, prev);
717 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
719 struct sched_entity *next;
722 * Dequeue and enqueue the task to update its
723 * position within the tree:
725 dequeue_entity(cfs_rq, curr, 0);
726 enqueue_entity(cfs_rq, curr, 0);
729 * Reschedule if another task tops the current one.
731 next = __pick_next_entity(cfs_rq);
735 __check_preempt_curr_fair(cfs_rq, next, curr,
736 sched_granularity(cfs_rq));
739 /**************************************************
740 * CFS operations on tasks:
743 #ifdef CONFIG_FAIR_GROUP_SCHED
745 /* Walk up scheduling entities hierarchy */
746 #define for_each_sched_entity(se) \
747 for (; se; se = se->parent)
749 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
754 /* runqueue on which this entity is (to be) queued */
755 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
760 /* runqueue "owned" by this group */
761 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
766 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
767 * another cpu ('this_cpu')
769 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
771 /* A later patch will take group into account */
772 return &cpu_rq(this_cpu)->cfs;
775 /* Iterate thr' all leaf cfs_rq's on a runqueue */
776 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
777 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
779 /* Do the two (enqueued) tasks belong to the same group ? */
780 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
782 if (curr->se.cfs_rq == p->se.cfs_rq)
788 #else /* CONFIG_FAIR_GROUP_SCHED */
790 #define for_each_sched_entity(se) \
791 for (; se; se = NULL)
793 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
795 return &task_rq(p)->cfs;
798 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
800 struct task_struct *p = task_of(se);
801 struct rq *rq = task_rq(p);
806 /* runqueue "owned" by this group */
807 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
812 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
814 return &cpu_rq(this_cpu)->cfs;
817 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
818 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
820 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
825 #endif /* CONFIG_FAIR_GROUP_SCHED */
828 * The enqueue_task method is called before nr_running is
829 * increased. Here we update the fair scheduling stats and
830 * then put the task into the rbtree:
832 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
834 struct cfs_rq *cfs_rq;
835 struct sched_entity *se = &p->se;
837 for_each_sched_entity(se) {
840 cfs_rq = cfs_rq_of(se);
841 enqueue_entity(cfs_rq, se, wakeup);
846 * The dequeue_task method is called before nr_running is
847 * decreased. We remove the task from the rbtree and
848 * update the fair scheduling stats:
850 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
852 struct cfs_rq *cfs_rq;
853 struct sched_entity *se = &p->se;
855 for_each_sched_entity(se) {
856 cfs_rq = cfs_rq_of(se);
857 dequeue_entity(cfs_rq, se, sleep);
858 /* Don't dequeue parent if it has other entities besides us */
859 if (cfs_rq->load.weight)
865 * sched_yield() support is very simple - we dequeue and enqueue.
867 * If compat_yield is turned on then we requeue to the end of the tree.
869 static void yield_task_fair(struct rq *rq, struct task_struct *p)
871 struct cfs_rq *cfs_rq = task_cfs_rq(p);
872 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
873 struct sched_entity *rightmost, *se = &p->se;
874 struct rb_node *parent;
877 * Are we the only task in the tree?
879 if (unlikely(cfs_rq->nr_running == 1))
882 if (likely(!sysctl_sched_compat_yield)) {
883 __update_rq_clock(rq);
885 * Dequeue and enqueue the task to update its
886 * position within the tree:
888 dequeue_entity(cfs_rq, &p->se, 0);
889 enqueue_entity(cfs_rq, &p->se, 0);
894 * Find the rightmost entry in the rbtree:
898 link = &parent->rb_right;
901 rightmost = rb_entry(parent, struct sched_entity, run_node);
903 * Already in the rightmost position?
905 if (unlikely(rightmost == se))
909 * Minimally necessary key value to be last in the tree:
911 se->fair_key = rightmost->fair_key + 1;
913 if (cfs_rq->rb_leftmost == &se->run_node)
914 cfs_rq->rb_leftmost = rb_next(&se->run_node);
916 * Relink the task to the rightmost position:
918 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
919 rb_link_node(&se->run_node, parent, link);
920 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
924 * Preempt the current task with a newly woken task if needed:
926 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
928 struct task_struct *curr = rq->curr;
929 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
932 if (unlikely(rt_prio(p->prio))) {
939 gran = sysctl_sched_wakeup_granularity;
941 * Batch tasks prefer throughput over latency:
943 if (unlikely(p->policy == SCHED_BATCH))
944 gran = sysctl_sched_batch_wakeup_granularity;
946 if (is_same_group(curr, p))
947 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
950 static struct task_struct *pick_next_task_fair(struct rq *rq)
952 struct cfs_rq *cfs_rq = &rq->cfs;
953 struct sched_entity *se;
955 if (unlikely(!cfs_rq->nr_running))
959 se = pick_next_entity(cfs_rq);
960 cfs_rq = group_cfs_rq(se);
967 * Account for a descheduled task:
969 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
971 struct sched_entity *se = &prev->se;
972 struct cfs_rq *cfs_rq;
974 for_each_sched_entity(se) {
975 cfs_rq = cfs_rq_of(se);
976 put_prev_entity(cfs_rq, se);
980 /**************************************************
981 * Fair scheduling class load-balancing methods:
985 * Load-balancing iterator. Note: while the runqueue stays locked
986 * during the whole iteration, the current task might be
987 * dequeued so the iterator has to be dequeue-safe. Here we
988 * achieve that by always pre-iterating before returning
991 static inline struct task_struct *
992 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
994 struct task_struct *p;
999 p = rb_entry(curr, struct task_struct, se.run_node);
1000 cfs_rq->rb_load_balance_curr = rb_next(curr);
1005 static struct task_struct *load_balance_start_fair(void *arg)
1007 struct cfs_rq *cfs_rq = arg;
1009 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1012 static struct task_struct *load_balance_next_fair(void *arg)
1014 struct cfs_rq *cfs_rq = arg;
1016 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1019 #ifdef CONFIG_FAIR_GROUP_SCHED
1020 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1022 struct sched_entity *curr;
1023 struct task_struct *p;
1025 if (!cfs_rq->nr_running)
1028 curr = __pick_next_entity(cfs_rq);
1035 static unsigned long
1036 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1037 unsigned long max_nr_move, unsigned long max_load_move,
1038 struct sched_domain *sd, enum cpu_idle_type idle,
1039 int *all_pinned, int *this_best_prio)
1041 struct cfs_rq *busy_cfs_rq;
1042 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1043 long rem_load_move = max_load_move;
1044 struct rq_iterator cfs_rq_iterator;
1046 cfs_rq_iterator.start = load_balance_start_fair;
1047 cfs_rq_iterator.next = load_balance_next_fair;
1049 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1050 #ifdef CONFIG_FAIR_GROUP_SCHED
1051 struct cfs_rq *this_cfs_rq;
1053 unsigned long maxload;
1055 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1057 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1058 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1062 /* Don't pull more than imbalance/2 */
1064 maxload = min(rem_load_move, imbalance);
1066 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1068 # define maxload rem_load_move
1070 /* pass busy_cfs_rq argument into
1071 * load_balance_[start|next]_fair iterators
1073 cfs_rq_iterator.arg = busy_cfs_rq;
1074 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1075 max_nr_move, maxload, sd, idle, all_pinned,
1076 &load_moved, this_best_prio, &cfs_rq_iterator);
1078 total_nr_moved += nr_moved;
1079 max_nr_move -= nr_moved;
1080 rem_load_move -= load_moved;
1082 if (max_nr_move <= 0 || rem_load_move <= 0)
1086 return max_load_move - rem_load_move;
1090 * scheduler tick hitting a task of our scheduling class:
1092 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1094 struct cfs_rq *cfs_rq;
1095 struct sched_entity *se = &curr->se;
1097 for_each_sched_entity(se) {
1098 cfs_rq = cfs_rq_of(se);
1099 entity_tick(cfs_rq, se);
1104 * Share the fairness runtime between parent and child, thus the
1105 * total amount of pressure for CPU stays equal - new tasks
1106 * get a chance to run but frequent forkers are not allowed to
1107 * monopolize the CPU. Note: the parent runqueue is locked,
1108 * the child is not running yet.
1110 static void task_new_fair(struct rq *rq, struct task_struct *p)
1112 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1113 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1115 sched_info_queued(p);
1117 update_curr(cfs_rq);
1118 update_stats_enqueue(cfs_rq, se);
1120 * Child runs first: we let it run before the parent
1121 * until it reschedules once. We set up the key so that
1122 * it will preempt the parent:
1124 se->fair_key = curr->fair_key -
1125 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1127 * The first wait is dominated by the child-runs-first logic,
1128 * so do not credit it with that waiting time yet:
1130 if (sched_feat(SKIP_INITIAL))
1131 se->wait_start_fair = 0;
1134 * The statistical average of wait_runtime is about
1135 * -granularity/2, so initialize the task with that:
1137 if (sched_feat(START_DEBIT))
1138 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1140 se->vruntime = cfs_rq->min_vruntime;
1141 update_stats_enqueue(cfs_rq, se);
1142 __enqueue_entity(cfs_rq, se);
1143 resched_task(rq->curr);
1146 #ifdef CONFIG_FAIR_GROUP_SCHED
1147 /* Account for a task changing its policy or group.
1149 * This routine is mostly called to set cfs_rq->curr field when a task
1150 * migrates between groups/classes.
1152 static void set_curr_task_fair(struct rq *rq)
1154 struct sched_entity *se = &rq->curr->se;
1156 for_each_sched_entity(se)
1157 set_next_entity(cfs_rq_of(se), se);
1160 static void set_curr_task_fair(struct rq *rq)
1166 * All the scheduling class methods:
1168 struct sched_class fair_sched_class __read_mostly = {
1169 .enqueue_task = enqueue_task_fair,
1170 .dequeue_task = dequeue_task_fair,
1171 .yield_task = yield_task_fair,
1173 .check_preempt_curr = check_preempt_curr_fair,
1175 .pick_next_task = pick_next_task_fair,
1176 .put_prev_task = put_prev_task_fair,
1178 .load_balance = load_balance_fair,
1180 .set_curr_task = set_curr_task_fair,
1181 .task_tick = task_tick_fair,
1182 .task_new = task_new_fair,
1185 #ifdef CONFIG_SCHED_DEBUG
1186 static void print_cfs_stats(struct seq_file *m, int cpu)
1188 struct cfs_rq *cfs_rq;
1190 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1191 print_cfs_rq(m, cpu, cfs_rq);