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;
339 if (!sched_feat(FAIR_SLEEPERS))
345 delta_fair = calc_delta_fair(delta_exec, lw);
346 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
348 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
349 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
350 delta = min(delta, (unsigned long)(
351 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
352 cfs_rq->sleeper_bonus -= delta;
356 cfs_rq->fair_clock += delta_fair;
358 * We executed delta_exec amount of time on the CPU,
359 * but we were only entitled to delta_mine amount of
360 * time during that period (if nr_running == 1 then
361 * the two values are equal)
362 * [Note: delta_mine - delta_exec is negative]:
364 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
367 static void update_curr(struct cfs_rq *cfs_rq)
369 struct sched_entity *curr = cfs_rq->curr;
370 u64 now = rq_of(cfs_rq)->clock;
371 unsigned long delta_exec;
377 * Get the amount of time the current task was running
378 * since the last time we changed load (this cannot
379 * overflow on 32 bits):
381 delta_exec = (unsigned long)(now - curr->exec_start);
383 __update_curr(cfs_rq, curr, delta_exec);
384 curr->exec_start = now;
388 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
390 se->wait_start_fair = cfs_rq->fair_clock;
391 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
394 static inline unsigned long
395 calc_weighted(unsigned long delta, struct sched_entity *se)
397 unsigned long weight = se->load.weight;
399 if (unlikely(weight != NICE_0_LOAD))
400 return (u64)delta * se->load.weight >> NICE_0_SHIFT;
406 * Task is being enqueued - update stats:
408 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
411 * Are we enqueueing a waiting task? (for current tasks
412 * a dequeue/enqueue event is a NOP)
414 if (se != cfs_rq->curr)
415 update_stats_wait_start(cfs_rq, se);
419 se->fair_key = se->vruntime;
423 * Note: must be called with a freshly updated rq->fair_clock.
426 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se,
427 unsigned long delta_fair)
429 schedstat_set(se->wait_max, max(se->wait_max,
430 rq_of(cfs_rq)->clock - se->wait_start));
432 delta_fair = calc_weighted(delta_fair, se);
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 if (unlikely(!se->wait_start_fair))
445 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
446 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
448 __update_stats_wait_end(cfs_rq, se, delta_fair);
450 se->wait_start_fair = 0;
451 schedstat_set(se->wait_start, 0);
455 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
459 * Mark the end of the wait period if dequeueing a
462 if (se != cfs_rq->curr)
463 update_stats_wait_end(cfs_rq, se);
467 * We are picking a new current task - update its stats:
470 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
473 * We are starting a new run period:
475 se->exec_start = rq_of(cfs_rq)->clock;
479 * We are descheduling a task - update its stats:
482 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
487 /**************************************************
488 * Scheduling class queueing methods:
491 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
492 unsigned long delta_fair)
494 unsigned long load = cfs_rq->load.weight;
498 * Do not boost sleepers if there's too much bonus 'in flight'
501 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
504 if (sched_feat(SLEEPER_LOAD_AVG))
505 load = rq_of(cfs_rq)->cpu_load[2];
508 * Fix up delta_fair with the effect of us running
509 * during the whole sleep period:
511 if (sched_feat(SLEEPER_AVG))
512 delta_fair = div64_likely32((u64)delta_fair * load,
513 load + se->load.weight);
515 delta_fair = calc_weighted(delta_fair, se);
517 prev_runtime = se->wait_runtime;
518 __add_wait_runtime(cfs_rq, se, delta_fair);
519 delta_fair = se->wait_runtime - prev_runtime;
522 * Track the amount of bonus we've given to sleepers:
524 cfs_rq->sleeper_bonus += delta_fair;
527 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
529 struct task_struct *tsk = task_of(se);
530 unsigned long delta_fair;
532 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
533 !sched_feat(FAIR_SLEEPERS))
536 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
537 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
539 __enqueue_sleeper(cfs_rq, se, delta_fair);
541 se->sleep_start_fair = 0;
543 #ifdef CONFIG_SCHEDSTATS
544 if (se->sleep_start) {
545 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
550 if (unlikely(delta > se->sleep_max))
551 se->sleep_max = delta;
554 se->sum_sleep_runtime += delta;
556 if (se->block_start) {
557 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
562 if (unlikely(delta > se->block_max))
563 se->block_max = delta;
566 se->sum_sleep_runtime += delta;
569 * Blocking time is in units of nanosecs, so shift by 20 to
570 * get a milliseconds-range estimation of the amount of
571 * time that the task spent sleeping:
573 if (unlikely(prof_on == SLEEP_PROFILING)) {
574 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
582 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
585 * Update the fair clock.
590 u64 min_runtime, latency;
592 min_runtime = cfs_rq->min_vruntime;
593 min_runtime += sysctl_sched_latency/2;
595 if (sched_feat(NEW_FAIR_SLEEPERS)) {
596 latency = calc_weighted(sysctl_sched_latency, se);
597 if (min_runtime > latency)
598 min_runtime -= latency;
601 se->vruntime = max(se->vruntime, min_runtime);
603 enqueue_sleeper(cfs_rq, se);
606 update_stats_enqueue(cfs_rq, se);
607 __enqueue_entity(cfs_rq, se);
611 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
613 update_stats_dequeue(cfs_rq, se);
615 se->sleep_start_fair = cfs_rq->fair_clock;
616 #ifdef CONFIG_SCHEDSTATS
617 if (entity_is_task(se)) {
618 struct task_struct *tsk = task_of(se);
620 if (tsk->state & TASK_INTERRUPTIBLE)
621 se->sleep_start = rq_of(cfs_rq)->clock;
622 if (tsk->state & TASK_UNINTERRUPTIBLE)
623 se->block_start = rq_of(cfs_rq)->clock;
627 __dequeue_entity(cfs_rq, se);
631 * Preempt the current task with a newly woken task if needed:
634 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
635 struct sched_entity *curr, unsigned long granularity)
637 s64 __delta = curr->fair_key - se->fair_key;
638 unsigned long ideal_runtime, delta_exec;
641 * ideal_runtime is compared against sum_exec_runtime, which is
642 * walltime, hence do not scale.
644 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
645 (unsigned long)sysctl_sched_min_granularity);
648 * If we executed more than what the latency constraint suggests,
649 * reduce the rescheduling granularity. This way the total latency
650 * of how much a task is not scheduled converges to
651 * sysctl_sched_latency:
653 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
654 if (delta_exec > ideal_runtime)
658 * Take scheduling granularity into account - do not
659 * preempt the current task unless the best task has
660 * a larger than sched_granularity fairness advantage:
662 * scale granularity as key space is in fair_clock.
664 if (__delta > niced_granularity(curr, granularity))
665 resched_task(rq_of(cfs_rq)->curr);
669 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
672 * Any task has to be enqueued before it get to execute on
673 * a CPU. So account for the time it spent waiting on the
674 * runqueue. (note, here we rely on pick_next_task() having
675 * done a put_prev_task_fair() shortly before this, which
676 * updated rq->fair_clock - used by update_stats_wait_end())
678 update_stats_wait_end(cfs_rq, se);
679 update_stats_curr_start(cfs_rq, se);
681 #ifdef CONFIG_SCHEDSTATS
683 * Track our maximum slice length, if the CPU's load is at
684 * least twice that of our own weight (i.e. dont track it
685 * when there are only lesser-weight tasks around):
687 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
688 se->slice_max = max(se->slice_max,
689 se->sum_exec_runtime - se->prev_sum_exec_runtime);
692 se->prev_sum_exec_runtime = se->sum_exec_runtime;
695 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
697 struct sched_entity *se = __pick_next_entity(cfs_rq);
699 set_next_entity(cfs_rq, se);
704 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
707 * If still on the runqueue then deactivate_task()
708 * was not called and update_curr() has to be done:
713 update_stats_curr_end(cfs_rq, prev);
716 update_stats_wait_start(cfs_rq, prev);
720 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
722 struct sched_entity *next;
725 * Dequeue and enqueue the task to update its
726 * position within the tree:
728 dequeue_entity(cfs_rq, curr, 0);
729 enqueue_entity(cfs_rq, curr, 0);
732 * Reschedule if another task tops the current one.
734 next = __pick_next_entity(cfs_rq);
738 __check_preempt_curr_fair(cfs_rq, next, curr,
739 sched_granularity(cfs_rq));
742 /**************************************************
743 * CFS operations on tasks:
746 #ifdef CONFIG_FAIR_GROUP_SCHED
748 /* Walk up scheduling entities hierarchy */
749 #define for_each_sched_entity(se) \
750 for (; se; se = se->parent)
752 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
757 /* runqueue on which this entity is (to be) queued */
758 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
763 /* runqueue "owned" by this group */
764 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
769 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
770 * another cpu ('this_cpu')
772 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
774 /* A later patch will take group into account */
775 return &cpu_rq(this_cpu)->cfs;
778 /* Iterate thr' all leaf cfs_rq's on a runqueue */
779 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
780 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
782 /* Do the two (enqueued) tasks belong to the same group ? */
783 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
785 if (curr->se.cfs_rq == p->se.cfs_rq)
791 #else /* CONFIG_FAIR_GROUP_SCHED */
793 #define for_each_sched_entity(se) \
794 for (; se; se = NULL)
796 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
798 return &task_rq(p)->cfs;
801 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
803 struct task_struct *p = task_of(se);
804 struct rq *rq = task_rq(p);
809 /* runqueue "owned" by this group */
810 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
815 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
817 return &cpu_rq(this_cpu)->cfs;
820 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
821 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
823 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
828 #endif /* CONFIG_FAIR_GROUP_SCHED */
831 * The enqueue_task method is called before nr_running is
832 * increased. Here we update the fair scheduling stats and
833 * then put the task into the rbtree:
835 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
837 struct cfs_rq *cfs_rq;
838 struct sched_entity *se = &p->se;
840 for_each_sched_entity(se) {
843 cfs_rq = cfs_rq_of(se);
844 enqueue_entity(cfs_rq, se, wakeup);
849 * The dequeue_task method is called before nr_running is
850 * decreased. We remove the task from the rbtree and
851 * update the fair scheduling stats:
853 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
855 struct cfs_rq *cfs_rq;
856 struct sched_entity *se = &p->se;
858 for_each_sched_entity(se) {
859 cfs_rq = cfs_rq_of(se);
860 dequeue_entity(cfs_rq, se, sleep);
861 /* Don't dequeue parent if it has other entities besides us */
862 if (cfs_rq->load.weight)
868 * sched_yield() support is very simple - we dequeue and enqueue.
870 * If compat_yield is turned on then we requeue to the end of the tree.
872 static void yield_task_fair(struct rq *rq, struct task_struct *p)
874 struct cfs_rq *cfs_rq = task_cfs_rq(p);
875 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
876 struct sched_entity *rightmost, *se = &p->se;
877 struct rb_node *parent;
880 * Are we the only task in the tree?
882 if (unlikely(cfs_rq->nr_running == 1))
885 if (likely(!sysctl_sched_compat_yield)) {
886 __update_rq_clock(rq);
888 * Dequeue and enqueue the task to update its
889 * position within the tree:
891 dequeue_entity(cfs_rq, &p->se, 0);
892 enqueue_entity(cfs_rq, &p->se, 0);
897 * Find the rightmost entry in the rbtree:
901 link = &parent->rb_right;
904 rightmost = rb_entry(parent, struct sched_entity, run_node);
906 * Already in the rightmost position?
908 if (unlikely(rightmost == se))
912 * Minimally necessary key value to be last in the tree:
914 se->fair_key = rightmost->fair_key + 1;
916 if (cfs_rq->rb_leftmost == &se->run_node)
917 cfs_rq->rb_leftmost = rb_next(&se->run_node);
919 * Relink the task to the rightmost position:
921 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
922 rb_link_node(&se->run_node, parent, link);
923 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
927 * Preempt the current task with a newly woken task if needed:
929 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
931 struct task_struct *curr = rq->curr;
932 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
935 if (unlikely(rt_prio(p->prio))) {
942 gran = sysctl_sched_wakeup_granularity;
944 * Batch tasks prefer throughput over latency:
946 if (unlikely(p->policy == SCHED_BATCH))
947 gran = sysctl_sched_batch_wakeup_granularity;
949 if (is_same_group(curr, p))
950 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
953 static struct task_struct *pick_next_task_fair(struct rq *rq)
955 struct cfs_rq *cfs_rq = &rq->cfs;
956 struct sched_entity *se;
958 if (unlikely(!cfs_rq->nr_running))
962 se = pick_next_entity(cfs_rq);
963 cfs_rq = group_cfs_rq(se);
970 * Account for a descheduled task:
972 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
974 struct sched_entity *se = &prev->se;
975 struct cfs_rq *cfs_rq;
977 for_each_sched_entity(se) {
978 cfs_rq = cfs_rq_of(se);
979 put_prev_entity(cfs_rq, se);
983 /**************************************************
984 * Fair scheduling class load-balancing methods:
988 * Load-balancing iterator. Note: while the runqueue stays locked
989 * during the whole iteration, the current task might be
990 * dequeued so the iterator has to be dequeue-safe. Here we
991 * achieve that by always pre-iterating before returning
994 static inline struct task_struct *
995 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
997 struct task_struct *p;
1002 p = rb_entry(curr, struct task_struct, se.run_node);
1003 cfs_rq->rb_load_balance_curr = rb_next(curr);
1008 static struct task_struct *load_balance_start_fair(void *arg)
1010 struct cfs_rq *cfs_rq = arg;
1012 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1015 static struct task_struct *load_balance_next_fair(void *arg)
1017 struct cfs_rq *cfs_rq = arg;
1019 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1022 #ifdef CONFIG_FAIR_GROUP_SCHED
1023 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1025 struct sched_entity *curr;
1026 struct task_struct *p;
1028 if (!cfs_rq->nr_running)
1031 curr = __pick_next_entity(cfs_rq);
1038 static unsigned long
1039 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1040 unsigned long max_nr_move, unsigned long max_load_move,
1041 struct sched_domain *sd, enum cpu_idle_type idle,
1042 int *all_pinned, int *this_best_prio)
1044 struct cfs_rq *busy_cfs_rq;
1045 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1046 long rem_load_move = max_load_move;
1047 struct rq_iterator cfs_rq_iterator;
1049 cfs_rq_iterator.start = load_balance_start_fair;
1050 cfs_rq_iterator.next = load_balance_next_fair;
1052 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1053 #ifdef CONFIG_FAIR_GROUP_SCHED
1054 struct cfs_rq *this_cfs_rq;
1056 unsigned long maxload;
1058 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1060 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1061 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1065 /* Don't pull more than imbalance/2 */
1067 maxload = min(rem_load_move, imbalance);
1069 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1071 # define maxload rem_load_move
1073 /* pass busy_cfs_rq argument into
1074 * load_balance_[start|next]_fair iterators
1076 cfs_rq_iterator.arg = busy_cfs_rq;
1077 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1078 max_nr_move, maxload, sd, idle, all_pinned,
1079 &load_moved, this_best_prio, &cfs_rq_iterator);
1081 total_nr_moved += nr_moved;
1082 max_nr_move -= nr_moved;
1083 rem_load_move -= load_moved;
1085 if (max_nr_move <= 0 || rem_load_move <= 0)
1089 return max_load_move - rem_load_move;
1093 * scheduler tick hitting a task of our scheduling class:
1095 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1097 struct cfs_rq *cfs_rq;
1098 struct sched_entity *se = &curr->se;
1100 for_each_sched_entity(se) {
1101 cfs_rq = cfs_rq_of(se);
1102 entity_tick(cfs_rq, se);
1107 * Share the fairness runtime between parent and child, thus the
1108 * total amount of pressure for CPU stays equal - new tasks
1109 * get a chance to run but frequent forkers are not allowed to
1110 * monopolize the CPU. Note: the parent runqueue is locked,
1111 * the child is not running yet.
1113 static void task_new_fair(struct rq *rq, struct task_struct *p)
1115 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1116 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1118 sched_info_queued(p);
1120 update_curr(cfs_rq);
1121 update_stats_enqueue(cfs_rq, se);
1123 * Child runs first: we let it run before the parent
1124 * until it reschedules once. We set up the key so that
1125 * it will preempt the parent:
1127 se->fair_key = curr->fair_key -
1128 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1130 * The first wait is dominated by the child-runs-first logic,
1131 * so do not credit it with that waiting time yet:
1133 if (sched_feat(SKIP_INITIAL))
1134 se->wait_start_fair = 0;
1137 * The statistical average of wait_runtime is about
1138 * -granularity/2, so initialize the task with that:
1140 if (sched_feat(START_DEBIT))
1141 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1143 se->vruntime = cfs_rq->min_vruntime;
1144 update_stats_enqueue(cfs_rq, se);
1145 __enqueue_entity(cfs_rq, se);
1146 resched_task(rq->curr);
1149 #ifdef CONFIG_FAIR_GROUP_SCHED
1150 /* Account for a task changing its policy or group.
1152 * This routine is mostly called to set cfs_rq->curr field when a task
1153 * migrates between groups/classes.
1155 static void set_curr_task_fair(struct rq *rq)
1157 struct sched_entity *se = &rq->curr->se;
1159 for_each_sched_entity(se)
1160 set_next_entity(cfs_rq_of(se), se);
1163 static void set_curr_task_fair(struct rq *rq)
1169 * All the scheduling class methods:
1171 struct sched_class fair_sched_class __read_mostly = {
1172 .enqueue_task = enqueue_task_fair,
1173 .dequeue_task = dequeue_task_fair,
1174 .yield_task = yield_task_fair,
1176 .check_preempt_curr = check_preempt_curr_fair,
1178 .pick_next_task = pick_next_task_fair,
1179 .put_prev_task = put_prev_task_fair,
1181 .load_balance = load_balance_fair,
1183 .set_curr_task = set_curr_task_fair,
1184 .task_tick = task_tick_fair,
1185 .task_new = task_new_fair,
1188 #ifdef CONFIG_SCHED_DEBUG
1189 static void print_cfs_stats(struct seq_file *m, int cpu)
1191 struct cfs_rq *cfs_rq;
1193 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1194 print_cfs_rq(m, cpu, cfs_rq);