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:
206 static u64 __sched_period(unsigned long nr_running)
208 u64 period = sysctl_sched_latency;
209 unsigned long nr_latency =
210 sysctl_sched_latency / sysctl_sched_min_granularity;
212 if (unlikely(nr_running > nr_latency)) {
213 period *= nr_running;
214 do_div(period, nr_latency);
221 * Calculate the preemption granularity needed to schedule every
222 * runnable task once per sysctl_sched_latency amount of time.
223 * (down to a sensible low limit on granularity)
225 * For example, if there are 2 tasks running and latency is 10 msecs,
226 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
227 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
228 * for each task. We do finer and finer scheduling up to until we
229 * reach the minimum granularity value.
231 * To achieve this we use the following dynamic-granularity rule:
233 * gran = lat/nr - lat/nr/nr
235 * This comes out of the following equations:
240 * kB2 = kB1 - d + d/nr
243 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
244 * '1' is start of time, '2' is end of time, 'd' is delay between
245 * 1 and 2 (during which task B was running), 'nr' is number of tasks
246 * running, 'lat' is the the period of each task. ('lat' is the
247 * sched_latency that we aim for.)
250 sched_granularity(struct cfs_rq *cfs_rq)
252 unsigned int gran = sysctl_sched_latency;
253 unsigned int nr = cfs_rq->nr_running;
256 gran = gran/nr - gran/nr/nr;
257 gran = max(gran, sysctl_sched_min_granularity);
264 * We rescale the rescheduling granularity of tasks according to their
265 * nice level, but only linearly, not exponentially:
268 niced_granularity(struct sched_entity *curr, unsigned long granularity)
272 if (likely(curr->load.weight == NICE_0_LOAD))
275 * Positive nice levels get the same granularity as nice-0:
277 if (likely(curr->load.weight < NICE_0_LOAD)) {
278 tmp = curr->load.weight * (u64)granularity;
279 return (long) (tmp >> NICE_0_SHIFT);
282 * Negative nice level tasks get linearly finer
285 tmp = curr->load.inv_weight * (u64)granularity;
288 * It will always fit into 'long':
290 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
294 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
296 long limit = sysctl_sched_runtime_limit;
299 * Niced tasks have the same history dynamic range as
302 if (unlikely(se->wait_runtime > limit)) {
303 se->wait_runtime = limit;
304 schedstat_inc(se, wait_runtime_overruns);
305 schedstat_inc(cfs_rq, wait_runtime_overruns);
307 if (unlikely(se->wait_runtime < -limit)) {
308 se->wait_runtime = -limit;
309 schedstat_inc(se, wait_runtime_underruns);
310 schedstat_inc(cfs_rq, wait_runtime_underruns);
315 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
317 se->wait_runtime += delta;
318 schedstat_add(se, sum_wait_runtime, delta);
319 limit_wait_runtime(cfs_rq, se);
323 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
325 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
326 __add_wait_runtime(cfs_rq, se, delta);
327 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
331 * Update the current task's runtime statistics. Skip current tasks that
332 * are not in our scheduling class.
335 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
336 unsigned long delta_exec)
338 unsigned long delta, delta_fair, delta_mine, delta_exec_weighted;
339 struct load_weight *lw = &cfs_rq->load;
340 unsigned long load = lw->weight;
342 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
344 curr->sum_exec_runtime += delta_exec;
345 cfs_rq->exec_clock += delta_exec;
346 delta_exec_weighted = delta_exec;
347 if (unlikely(curr->load.weight != NICE_0_LOAD)) {
348 delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
351 curr->vruntime += delta_exec_weighted;
353 if (!sched_feat(FAIR_SLEEPERS))
359 delta_fair = calc_delta_fair(delta_exec, lw);
360 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
362 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
363 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
364 delta = min(delta, (unsigned long)(
365 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
366 cfs_rq->sleeper_bonus -= delta;
370 cfs_rq->fair_clock += delta_fair;
372 * We executed delta_exec amount of time on the CPU,
373 * but we were only entitled to delta_mine amount of
374 * time during that period (if nr_running == 1 then
375 * the two values are equal)
376 * [Note: delta_mine - delta_exec is negative]:
378 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
381 static void update_curr(struct cfs_rq *cfs_rq)
383 struct sched_entity *curr = cfs_rq->curr;
384 u64 now = rq_of(cfs_rq)->clock;
385 unsigned long delta_exec;
391 * Get the amount of time the current task was running
392 * since the last time we changed load (this cannot
393 * overflow on 32 bits):
395 delta_exec = (unsigned long)(now - curr->exec_start);
397 __update_curr(cfs_rq, curr, delta_exec);
398 curr->exec_start = now;
402 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
404 se->wait_start_fair = cfs_rq->fair_clock;
405 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
408 static inline unsigned long
409 calc_weighted(unsigned long delta, struct sched_entity *se)
411 unsigned long weight = se->load.weight;
413 if (unlikely(weight != NICE_0_LOAD))
414 return (u64)delta * se->load.weight >> NICE_0_SHIFT;
420 * Task is being enqueued - update stats:
422 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
425 * Are we enqueueing a waiting task? (for current tasks
426 * a dequeue/enqueue event is a NOP)
428 if (se != cfs_rq->curr)
429 update_stats_wait_start(cfs_rq, se);
433 se->fair_key = se->vruntime;
437 * Note: must be called with a freshly updated rq->fair_clock.
440 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se,
441 unsigned long delta_fair)
443 schedstat_set(se->wait_max, max(se->wait_max,
444 rq_of(cfs_rq)->clock - se->wait_start));
446 delta_fair = calc_weighted(delta_fair, se);
448 add_wait_runtime(cfs_rq, se, delta_fair);
452 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
454 unsigned long delta_fair;
456 if (unlikely(!se->wait_start_fair))
459 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
460 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
462 __update_stats_wait_end(cfs_rq, se, delta_fair);
464 se->wait_start_fair = 0;
465 schedstat_set(se->wait_start, 0);
469 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
473 * Mark the end of the wait period if dequeueing a
476 if (se != cfs_rq->curr)
477 update_stats_wait_end(cfs_rq, se);
481 * We are picking a new current task - update its stats:
484 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
487 * We are starting a new run period:
489 se->exec_start = rq_of(cfs_rq)->clock;
493 * We are descheduling a task - update its stats:
496 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
501 /**************************************************
502 * Scheduling class queueing methods:
505 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
506 unsigned long delta_fair)
508 unsigned long load = cfs_rq->load.weight;
512 * Do not boost sleepers if there's too much bonus 'in flight'
515 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
518 if (sched_feat(SLEEPER_LOAD_AVG))
519 load = rq_of(cfs_rq)->cpu_load[2];
522 * Fix up delta_fair with the effect of us running
523 * during the whole sleep period:
525 if (sched_feat(SLEEPER_AVG))
526 delta_fair = div64_likely32((u64)delta_fair * load,
527 load + se->load.weight);
529 delta_fair = calc_weighted(delta_fair, se);
531 prev_runtime = se->wait_runtime;
532 __add_wait_runtime(cfs_rq, se, delta_fair);
533 delta_fair = se->wait_runtime - prev_runtime;
536 * Track the amount of bonus we've given to sleepers:
538 cfs_rq->sleeper_bonus += delta_fair;
541 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
543 struct task_struct *tsk = task_of(se);
544 unsigned long delta_fair;
546 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
547 !sched_feat(FAIR_SLEEPERS))
550 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
551 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
553 __enqueue_sleeper(cfs_rq, se, delta_fair);
555 se->sleep_start_fair = 0;
557 #ifdef CONFIG_SCHEDSTATS
558 if (se->sleep_start) {
559 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
564 if (unlikely(delta > se->sleep_max))
565 se->sleep_max = delta;
568 se->sum_sleep_runtime += delta;
570 if (se->block_start) {
571 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
576 if (unlikely(delta > se->block_max))
577 se->block_max = delta;
580 se->sum_sleep_runtime += delta;
583 * Blocking time is in units of nanosecs, so shift by 20 to
584 * get a milliseconds-range estimation of the amount of
585 * time that the task spent sleeping:
587 if (unlikely(prof_on == SLEEP_PROFILING)) {
588 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
596 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
599 * Update the fair clock.
604 u64 min_runtime, latency;
606 min_runtime = cfs_rq->min_vruntime;
607 min_runtime += sysctl_sched_latency/2;
609 if (sched_feat(NEW_FAIR_SLEEPERS)) {
610 latency = calc_weighted(sysctl_sched_latency, se);
611 if (min_runtime > latency)
612 min_runtime -= latency;
615 se->vruntime = max(se->vruntime, min_runtime);
617 enqueue_sleeper(cfs_rq, se);
620 update_stats_enqueue(cfs_rq, se);
621 __enqueue_entity(cfs_rq, se);
625 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
627 update_stats_dequeue(cfs_rq, se);
629 se->sleep_start_fair = cfs_rq->fair_clock;
630 #ifdef CONFIG_SCHEDSTATS
631 if (entity_is_task(se)) {
632 struct task_struct *tsk = task_of(se);
634 if (tsk->state & TASK_INTERRUPTIBLE)
635 se->sleep_start = rq_of(cfs_rq)->clock;
636 if (tsk->state & TASK_UNINTERRUPTIBLE)
637 se->block_start = rq_of(cfs_rq)->clock;
641 __dequeue_entity(cfs_rq, se);
645 * Preempt the current task with a newly woken task if needed:
648 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
649 struct sched_entity *curr, unsigned long granularity)
651 s64 __delta = curr->fair_key - se->fair_key;
652 unsigned long ideal_runtime, delta_exec;
655 * ideal_runtime is compared against sum_exec_runtime, which is
656 * walltime, hence do not scale.
658 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
659 (unsigned long)sysctl_sched_min_granularity);
662 * If we executed more than what the latency constraint suggests,
663 * reduce the rescheduling granularity. This way the total latency
664 * of how much a task is not scheduled converges to
665 * sysctl_sched_latency:
667 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
668 if (delta_exec > ideal_runtime)
672 * Take scheduling granularity into account - do not
673 * preempt the current task unless the best task has
674 * a larger than sched_granularity fairness advantage:
676 * scale granularity as key space is in fair_clock.
678 if (__delta > niced_granularity(curr, granularity))
679 resched_task(rq_of(cfs_rq)->curr);
683 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 * Any task has to be enqueued before it get to execute on
687 * a CPU. So account for the time it spent waiting on the
688 * runqueue. (note, here we rely on pick_next_task() having
689 * done a put_prev_task_fair() shortly before this, which
690 * updated rq->fair_clock - used by update_stats_wait_end())
692 update_stats_wait_end(cfs_rq, se);
693 update_stats_curr_start(cfs_rq, se);
695 #ifdef CONFIG_SCHEDSTATS
697 * Track our maximum slice length, if the CPU's load is at
698 * least twice that of our own weight (i.e. dont track it
699 * when there are only lesser-weight tasks around):
701 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
702 se->slice_max = max(se->slice_max,
703 se->sum_exec_runtime - se->prev_sum_exec_runtime);
706 se->prev_sum_exec_runtime = se->sum_exec_runtime;
709 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
711 struct sched_entity *se = __pick_next_entity(cfs_rq);
713 set_next_entity(cfs_rq, se);
718 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
721 * If still on the runqueue then deactivate_task()
722 * was not called and update_curr() has to be done:
727 update_stats_curr_end(cfs_rq, prev);
730 update_stats_wait_start(cfs_rq, prev);
734 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
736 struct sched_entity *next;
739 * Dequeue and enqueue the task to update its
740 * position within the tree:
742 dequeue_entity(cfs_rq, curr, 0);
743 enqueue_entity(cfs_rq, curr, 0);
746 * Reschedule if another task tops the current one.
748 next = __pick_next_entity(cfs_rq);
752 __check_preempt_curr_fair(cfs_rq, next, curr,
753 sched_granularity(cfs_rq));
756 /**************************************************
757 * CFS operations on tasks:
760 #ifdef CONFIG_FAIR_GROUP_SCHED
762 /* Walk up scheduling entities hierarchy */
763 #define for_each_sched_entity(se) \
764 for (; se; se = se->parent)
766 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
771 /* runqueue on which this entity is (to be) queued */
772 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
777 /* runqueue "owned" by this group */
778 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
783 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
784 * another cpu ('this_cpu')
786 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
788 /* A later patch will take group into account */
789 return &cpu_rq(this_cpu)->cfs;
792 /* Iterate thr' all leaf cfs_rq's on a runqueue */
793 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
794 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
796 /* Do the two (enqueued) tasks belong to the same group ? */
797 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
799 if (curr->se.cfs_rq == p->se.cfs_rq)
805 #else /* CONFIG_FAIR_GROUP_SCHED */
807 #define for_each_sched_entity(se) \
808 for (; se; se = NULL)
810 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
812 return &task_rq(p)->cfs;
815 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
817 struct task_struct *p = task_of(se);
818 struct rq *rq = task_rq(p);
823 /* runqueue "owned" by this group */
824 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
829 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
831 return &cpu_rq(this_cpu)->cfs;
834 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
835 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
837 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
842 #endif /* CONFIG_FAIR_GROUP_SCHED */
845 * The enqueue_task method is called before nr_running is
846 * increased. Here we update the fair scheduling stats and
847 * then put the task into the rbtree:
849 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
851 struct cfs_rq *cfs_rq;
852 struct sched_entity *se = &p->se;
854 for_each_sched_entity(se) {
857 cfs_rq = cfs_rq_of(se);
858 enqueue_entity(cfs_rq, se, wakeup);
863 * The dequeue_task method is called before nr_running is
864 * decreased. We remove the task from the rbtree and
865 * update the fair scheduling stats:
867 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
869 struct cfs_rq *cfs_rq;
870 struct sched_entity *se = &p->se;
872 for_each_sched_entity(se) {
873 cfs_rq = cfs_rq_of(se);
874 dequeue_entity(cfs_rq, se, sleep);
875 /* Don't dequeue parent if it has other entities besides us */
876 if (cfs_rq->load.weight)
882 * sched_yield() support is very simple - we dequeue and enqueue.
884 * If compat_yield is turned on then we requeue to the end of the tree.
886 static void yield_task_fair(struct rq *rq, struct task_struct *p)
888 struct cfs_rq *cfs_rq = task_cfs_rq(p);
889 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
890 struct sched_entity *rightmost, *se = &p->se;
891 struct rb_node *parent;
894 * Are we the only task in the tree?
896 if (unlikely(cfs_rq->nr_running == 1))
899 if (likely(!sysctl_sched_compat_yield)) {
900 __update_rq_clock(rq);
902 * Dequeue and enqueue the task to update its
903 * position within the tree:
905 dequeue_entity(cfs_rq, &p->se, 0);
906 enqueue_entity(cfs_rq, &p->se, 0);
911 * Find the rightmost entry in the rbtree:
915 link = &parent->rb_right;
918 rightmost = rb_entry(parent, struct sched_entity, run_node);
920 * Already in the rightmost position?
922 if (unlikely(rightmost == se))
926 * Minimally necessary key value to be last in the tree:
928 se->fair_key = rightmost->fair_key + 1;
930 if (cfs_rq->rb_leftmost == &se->run_node)
931 cfs_rq->rb_leftmost = rb_next(&se->run_node);
933 * Relink the task to the rightmost position:
935 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
936 rb_link_node(&se->run_node, parent, link);
937 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
941 * Preempt the current task with a newly woken task if needed:
943 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
945 struct task_struct *curr = rq->curr;
946 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
949 if (unlikely(rt_prio(p->prio))) {
956 gran = sysctl_sched_wakeup_granularity;
958 * Batch tasks prefer throughput over latency:
960 if (unlikely(p->policy == SCHED_BATCH))
961 gran = sysctl_sched_batch_wakeup_granularity;
963 if (is_same_group(curr, p))
964 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
967 static struct task_struct *pick_next_task_fair(struct rq *rq)
969 struct cfs_rq *cfs_rq = &rq->cfs;
970 struct sched_entity *se;
972 if (unlikely(!cfs_rq->nr_running))
976 se = pick_next_entity(cfs_rq);
977 cfs_rq = group_cfs_rq(se);
984 * Account for a descheduled task:
986 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
988 struct sched_entity *se = &prev->se;
989 struct cfs_rq *cfs_rq;
991 for_each_sched_entity(se) {
992 cfs_rq = cfs_rq_of(se);
993 put_prev_entity(cfs_rq, se);
997 /**************************************************
998 * Fair scheduling class load-balancing methods:
1002 * Load-balancing iterator. Note: while the runqueue stays locked
1003 * during the whole iteration, the current task might be
1004 * dequeued so the iterator has to be dequeue-safe. Here we
1005 * achieve that by always pre-iterating before returning
1008 static inline struct task_struct *
1009 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1011 struct task_struct *p;
1016 p = rb_entry(curr, struct task_struct, se.run_node);
1017 cfs_rq->rb_load_balance_curr = rb_next(curr);
1022 static struct task_struct *load_balance_start_fair(void *arg)
1024 struct cfs_rq *cfs_rq = arg;
1026 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1029 static struct task_struct *load_balance_next_fair(void *arg)
1031 struct cfs_rq *cfs_rq = arg;
1033 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1036 #ifdef CONFIG_FAIR_GROUP_SCHED
1037 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1039 struct sched_entity *curr;
1040 struct task_struct *p;
1042 if (!cfs_rq->nr_running)
1045 curr = __pick_next_entity(cfs_rq);
1052 static unsigned long
1053 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1054 unsigned long max_nr_move, unsigned long max_load_move,
1055 struct sched_domain *sd, enum cpu_idle_type idle,
1056 int *all_pinned, int *this_best_prio)
1058 struct cfs_rq *busy_cfs_rq;
1059 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1060 long rem_load_move = max_load_move;
1061 struct rq_iterator cfs_rq_iterator;
1063 cfs_rq_iterator.start = load_balance_start_fair;
1064 cfs_rq_iterator.next = load_balance_next_fair;
1066 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1067 #ifdef CONFIG_FAIR_GROUP_SCHED
1068 struct cfs_rq *this_cfs_rq;
1070 unsigned long maxload;
1072 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1074 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1075 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1079 /* Don't pull more than imbalance/2 */
1081 maxload = min(rem_load_move, imbalance);
1083 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1085 # define maxload rem_load_move
1087 /* pass busy_cfs_rq argument into
1088 * load_balance_[start|next]_fair iterators
1090 cfs_rq_iterator.arg = busy_cfs_rq;
1091 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1092 max_nr_move, maxload, sd, idle, all_pinned,
1093 &load_moved, this_best_prio, &cfs_rq_iterator);
1095 total_nr_moved += nr_moved;
1096 max_nr_move -= nr_moved;
1097 rem_load_move -= load_moved;
1099 if (max_nr_move <= 0 || rem_load_move <= 0)
1103 return max_load_move - rem_load_move;
1107 * scheduler tick hitting a task of our scheduling class:
1109 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1111 struct cfs_rq *cfs_rq;
1112 struct sched_entity *se = &curr->se;
1114 for_each_sched_entity(se) {
1115 cfs_rq = cfs_rq_of(se);
1116 entity_tick(cfs_rq, se);
1120 #define swap(a,b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
1123 * Share the fairness runtime between parent and child, thus the
1124 * total amount of pressure for CPU stays equal - new tasks
1125 * get a chance to run but frequent forkers are not allowed to
1126 * monopolize the CPU. Note: the parent runqueue is locked,
1127 * the child is not running yet.
1129 static void task_new_fair(struct rq *rq, struct task_struct *p)
1131 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1132 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1134 sched_info_queued(p);
1136 update_curr(cfs_rq);
1137 se->vruntime = cfs_rq->min_vruntime;
1138 update_stats_enqueue(cfs_rq, se);
1141 * The first wait is dominated by the child-runs-first logic,
1142 * so do not credit it with that waiting time yet:
1144 if (sched_feat(SKIP_INITIAL))
1145 se->wait_start_fair = 0;
1148 * The statistical average of wait_runtime is about
1149 * -granularity/2, so initialize the task with that:
1151 if (sched_feat(START_DEBIT))
1152 se->wait_runtime = -(__sched_period(cfs_rq->nr_running+1) / 2);
1154 if (sysctl_sched_child_runs_first &&
1155 curr->vruntime < se->vruntime) {
1157 dequeue_entity(cfs_rq, curr, 0);
1158 swap(curr->vruntime, se->vruntime);
1159 enqueue_entity(cfs_rq, curr, 0);
1162 update_stats_enqueue(cfs_rq, se);
1163 __enqueue_entity(cfs_rq, se);
1164 resched_task(rq->curr);
1167 #ifdef CONFIG_FAIR_GROUP_SCHED
1168 /* Account for a task changing its policy or group.
1170 * This routine is mostly called to set cfs_rq->curr field when a task
1171 * migrates between groups/classes.
1173 static void set_curr_task_fair(struct rq *rq)
1175 struct sched_entity *se = &rq->curr->se;
1177 for_each_sched_entity(se)
1178 set_next_entity(cfs_rq_of(se), se);
1181 static void set_curr_task_fair(struct rq *rq)
1187 * All the scheduling class methods:
1189 struct sched_class fair_sched_class __read_mostly = {
1190 .enqueue_task = enqueue_task_fair,
1191 .dequeue_task = dequeue_task_fair,
1192 .yield_task = yield_task_fair,
1194 .check_preempt_curr = check_preempt_curr_fair,
1196 .pick_next_task = pick_next_task_fair,
1197 .put_prev_task = put_prev_task_fair,
1199 .load_balance = load_balance_fair,
1201 .set_curr_task = set_curr_task_fair,
1202 .task_tick = task_tick_fair,
1203 .task_new = task_new_fair,
1206 #ifdef CONFIG_SCHED_DEBUG
1207 static void print_cfs_stats(struct seq_file *m, int cpu)
1209 struct cfs_rq *cfs_rq;
1211 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1212 print_cfs_rq(m, cpu, cfs_rq);