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 unsigned int sysctl_sched_latency __read_mostly = 20000000ULL;
40 * Minimal preemption granularity for CPU-bound tasks:
41 * (default: 2 msec, units: nanoseconds)
43 unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL;
46 * SCHED_BATCH wake-up granularity.
47 * (default: 25 msec, units: nanoseconds)
49 * This option delays the preemption effects of decoupled workloads
50 * and reduces their over-scheduling. Synchronous workloads will still
51 * have immediate wakeup/sleep latencies.
53 unsigned int sysctl_sched_batch_wakeup_granularity __read_mostly = 25000000UL;
56 * SCHED_OTHER wake-up granularity.
57 * (default: 1 msec, units: nanoseconds)
59 * This option delays the preemption effects of decoupled workloads
60 * and reduces their over-scheduling. Synchronous workloads will still
61 * have immediate wakeup/sleep latencies.
63 unsigned int sysctl_sched_wakeup_granularity __read_mostly = 1000000UL;
65 unsigned int sysctl_sched_stat_granularity __read_mostly;
68 * Initialized in sched_init_granularity() [to 5 times the base granularity]:
70 unsigned int sysctl_sched_runtime_limit __read_mostly;
73 * Debugging: various feature bits
76 SCHED_FEAT_FAIR_SLEEPERS = 1,
77 SCHED_FEAT_SLEEPER_AVG = 2,
78 SCHED_FEAT_SLEEPER_LOAD_AVG = 4,
79 SCHED_FEAT_PRECISE_CPU_LOAD = 8,
80 SCHED_FEAT_START_DEBIT = 16,
81 SCHED_FEAT_SKIP_INITIAL = 32,
84 unsigned int sysctl_sched_features __read_mostly =
85 SCHED_FEAT_FAIR_SLEEPERS *1 |
86 SCHED_FEAT_SLEEPER_AVG *0 |
87 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
88 SCHED_FEAT_PRECISE_CPU_LOAD *1 |
89 SCHED_FEAT_START_DEBIT *1 |
90 SCHED_FEAT_SKIP_INITIAL *0;
92 extern struct sched_class fair_sched_class;
94 /**************************************************************
95 * CFS operations on generic schedulable entities:
98 #ifdef CONFIG_FAIR_GROUP_SCHED
100 /* cpu runqueue to which this cfs_rq is attached */
101 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
106 /* currently running entity (if any) on this cfs_rq */
107 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
112 /* An entity is a task if it doesn't "own" a runqueue */
113 #define entity_is_task(se) (!se->my_q)
116 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
121 #else /* CONFIG_FAIR_GROUP_SCHED */
123 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
125 return container_of(cfs_rq, struct rq, cfs);
128 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
130 struct rq *rq = rq_of(cfs_rq);
132 if (unlikely(rq->curr->sched_class != &fair_sched_class))
135 return &rq->curr->se;
138 #define entity_is_task(se) 1
141 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
143 #endif /* CONFIG_FAIR_GROUP_SCHED */
145 static inline struct task_struct *task_of(struct sched_entity *se)
147 return container_of(se, struct task_struct, se);
151 /**************************************************************
152 * Scheduling class tree data structure manipulation methods:
156 * Enqueue an entity into the rb-tree:
159 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
161 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
162 struct rb_node *parent = NULL;
163 struct sched_entity *entry;
164 s64 key = se->fair_key;
168 * Find the right place in the rbtree:
172 entry = rb_entry(parent, struct sched_entity, run_node);
174 * We dont care about collisions. Nodes with
175 * the same key stay together.
177 if (key - entry->fair_key < 0) {
178 link = &parent->rb_left;
180 link = &parent->rb_right;
186 * Maintain a cache of leftmost tree entries (it is frequently
190 cfs_rq->rb_leftmost = &se->run_node;
192 rb_link_node(&se->run_node, parent, link);
193 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
194 update_load_add(&cfs_rq->load, se->load.weight);
195 cfs_rq->nr_running++;
200 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
202 if (cfs_rq->rb_leftmost == &se->run_node)
203 cfs_rq->rb_leftmost = rb_next(&se->run_node);
204 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
205 update_load_sub(&cfs_rq->load, se->load.weight);
206 cfs_rq->nr_running--;
210 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
212 return cfs_rq->rb_leftmost;
215 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
217 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
220 /**************************************************************
221 * Scheduling class statistics methods:
225 * Calculate the preemption granularity needed to schedule every
226 * runnable task once per sysctl_sched_latency amount of time.
227 * (down to a sensible low limit on granularity)
229 * For example, if there are 2 tasks running and latency is 10 msecs,
230 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
231 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
232 * for each task. We do finer and finer scheduling up to until we
233 * reach the minimum granularity value.
235 * To achieve this we use the following dynamic-granularity rule:
237 * gran = lat/nr - lat/nr/nr
239 * This comes out of the following equations:
244 * kB2 = kB1 - d + d/nr
247 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
248 * '1' is start of time, '2' is end of time, 'd' is delay between
249 * 1 and 2 (during which task B was running), 'nr' is number of tasks
250 * running, 'lat' is the the period of each task. ('lat' is the
251 * sched_latency that we aim for.)
254 sched_granularity(struct cfs_rq *cfs_rq)
256 unsigned int gran = sysctl_sched_latency;
257 unsigned int nr = cfs_rq->nr_running;
260 gran = gran/nr - gran/nr/nr;
261 gran = max(gran, sysctl_sched_min_granularity);
268 * We rescale the rescheduling granularity of tasks according to their
269 * nice level, but only linearly, not exponentially:
272 niced_granularity(struct sched_entity *curr, unsigned long granularity)
276 if (likely(curr->load.weight == NICE_0_LOAD))
279 * Positive nice levels get the same granularity as nice-0:
281 if (likely(curr->load.weight < NICE_0_LOAD)) {
282 tmp = curr->load.weight * (u64)granularity;
283 return (long) (tmp >> NICE_0_SHIFT);
286 * Negative nice level tasks get linearly finer
289 tmp = curr->load.inv_weight * (u64)granularity;
292 * It will always fit into 'long':
294 return (long) (tmp >> WMULT_SHIFT);
298 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
300 long limit = sysctl_sched_runtime_limit;
303 * Niced tasks have the same history dynamic range as
306 if (unlikely(se->wait_runtime > limit)) {
307 se->wait_runtime = limit;
308 schedstat_inc(se, wait_runtime_overruns);
309 schedstat_inc(cfs_rq, wait_runtime_overruns);
311 if (unlikely(se->wait_runtime < -limit)) {
312 se->wait_runtime = -limit;
313 schedstat_inc(se, wait_runtime_underruns);
314 schedstat_inc(cfs_rq, wait_runtime_underruns);
319 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
321 se->wait_runtime += delta;
322 schedstat_add(se, sum_wait_runtime, delta);
323 limit_wait_runtime(cfs_rq, se);
327 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
329 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
330 __add_wait_runtime(cfs_rq, se, delta);
331 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
335 * Update the current task's runtime statistics. Skip current tasks that
336 * are not in our scheduling class.
339 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr)
341 unsigned long delta, delta_exec, delta_fair, delta_mine;
342 struct load_weight *lw = &cfs_rq->load;
343 unsigned long load = lw->weight;
345 delta_exec = curr->delta_exec;
346 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
348 curr->sum_exec_runtime += delta_exec;
349 cfs_rq->exec_clock += delta_exec;
354 delta_fair = calc_delta_fair(delta_exec, lw);
355 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
357 if (cfs_rq->sleeper_bonus > sysctl_sched_latency) {
358 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
359 delta = min(delta, (unsigned long)(
360 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
361 cfs_rq->sleeper_bonus -= delta;
365 cfs_rq->fair_clock += delta_fair;
367 * We executed delta_exec amount of time on the CPU,
368 * but we were only entitled to delta_mine amount of
369 * time during that period (if nr_running == 1 then
370 * the two values are equal)
371 * [Note: delta_mine - delta_exec is negative]:
373 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
376 static void update_curr(struct cfs_rq *cfs_rq)
378 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
379 unsigned long delta_exec;
385 * Get the amount of time the current task was running
386 * since the last time we changed load (this cannot
387 * overflow on 32 bits):
389 delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start);
391 curr->delta_exec += delta_exec;
393 if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
394 __update_curr(cfs_rq, curr);
395 curr->delta_exec = 0;
397 curr->exec_start = rq_of(cfs_rq)->clock;
401 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
403 se->wait_start_fair = cfs_rq->fair_clock;
404 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
408 * We calculate fair deltas here, so protect against the random effects
409 * of a multiplication overflow by capping it to the runtime limit:
411 #if BITS_PER_LONG == 32
412 static inline unsigned long
413 calc_weighted(unsigned long delta, unsigned long weight, int shift)
415 u64 tmp = (u64)delta * weight >> shift;
417 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
418 return sysctl_sched_runtime_limit*2;
422 static inline unsigned long
423 calc_weighted(unsigned long delta, unsigned long weight, int shift)
425 return delta * weight >> shift;
430 * Task is being enqueued - update stats:
432 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
437 * Are we enqueueing a waiting task? (for current tasks
438 * a dequeue/enqueue event is a NOP)
440 if (se != cfs_rq_curr(cfs_rq))
441 update_stats_wait_start(cfs_rq, se);
445 key = cfs_rq->fair_clock;
448 * Optimize the common nice 0 case:
450 if (likely(se->load.weight == NICE_0_LOAD)) {
451 key -= se->wait_runtime;
455 if (se->wait_runtime < 0) {
456 tmp = -se->wait_runtime;
457 key += (tmp * se->load.inv_weight) >>
458 (WMULT_SHIFT - NICE_0_SHIFT);
460 tmp = se->wait_runtime;
461 key -= (tmp * se->load.inv_weight) >>
462 (WMULT_SHIFT - NICE_0_SHIFT);
470 * Note: must be called with a freshly updated rq->fair_clock.
473 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
475 unsigned long delta_fair = se->delta_fair_run;
477 schedstat_set(se->wait_max, max(se->wait_max,
478 rq_of(cfs_rq)->clock - se->wait_start));
480 if (unlikely(se->load.weight != NICE_0_LOAD))
481 delta_fair = calc_weighted(delta_fair, se->load.weight,
484 add_wait_runtime(cfs_rq, se, delta_fair);
488 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
490 unsigned long delta_fair;
492 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
493 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
495 se->delta_fair_run += delta_fair;
496 if (unlikely(abs(se->delta_fair_run) >=
497 sysctl_sched_stat_granularity)) {
498 __update_stats_wait_end(cfs_rq, se);
499 se->delta_fair_run = 0;
502 se->wait_start_fair = 0;
503 schedstat_set(se->wait_start, 0);
507 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
511 * Mark the end of the wait period if dequeueing a
514 if (se != cfs_rq_curr(cfs_rq))
515 update_stats_wait_end(cfs_rq, se);
519 * We are picking a new current task - update its stats:
522 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
525 * We are starting a new run period:
527 se->exec_start = rq_of(cfs_rq)->clock;
531 * We are descheduling a task - update its stats:
534 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
539 /**************************************************
540 * Scheduling class queueing methods:
543 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
545 unsigned long load = cfs_rq->load.weight, delta_fair;
549 * Do not boost sleepers if there's too much bonus 'in flight'
552 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
555 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
556 load = rq_of(cfs_rq)->cpu_load[2];
558 delta_fair = se->delta_fair_sleep;
561 * Fix up delta_fair with the effect of us running
562 * during the whole sleep period:
564 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
565 delta_fair = div64_likely32((u64)delta_fair * load,
566 load + se->load.weight);
568 if (unlikely(se->load.weight != NICE_0_LOAD))
569 delta_fair = calc_weighted(delta_fair, se->load.weight,
572 prev_runtime = se->wait_runtime;
573 __add_wait_runtime(cfs_rq, se, delta_fair);
574 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
575 delta_fair = se->wait_runtime - prev_runtime;
578 * Track the amount of bonus we've given to sleepers:
580 cfs_rq->sleeper_bonus += delta_fair;
583 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
585 struct task_struct *tsk = task_of(se);
586 unsigned long delta_fair;
588 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
589 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
592 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
593 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
595 se->delta_fair_sleep += delta_fair;
596 if (unlikely(abs(se->delta_fair_sleep) >=
597 sysctl_sched_stat_granularity)) {
598 __enqueue_sleeper(cfs_rq, se);
599 se->delta_fair_sleep = 0;
602 se->sleep_start_fair = 0;
604 #ifdef CONFIG_SCHEDSTATS
605 if (se->sleep_start) {
606 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
611 if (unlikely(delta > se->sleep_max))
612 se->sleep_max = delta;
615 se->sum_sleep_runtime += delta;
617 if (se->block_start) {
618 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
623 if (unlikely(delta > se->block_max))
624 se->block_max = delta;
627 se->sum_sleep_runtime += delta;
633 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
636 * Update the fair clock.
641 enqueue_sleeper(cfs_rq, se);
643 update_stats_enqueue(cfs_rq, se);
644 __enqueue_entity(cfs_rq, se);
648 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
650 update_stats_dequeue(cfs_rq, se);
652 se->sleep_start_fair = cfs_rq->fair_clock;
653 #ifdef CONFIG_SCHEDSTATS
654 if (entity_is_task(se)) {
655 struct task_struct *tsk = task_of(se);
657 if (tsk->state & TASK_INTERRUPTIBLE)
658 se->sleep_start = rq_of(cfs_rq)->clock;
659 if (tsk->state & TASK_UNINTERRUPTIBLE)
660 se->block_start = rq_of(cfs_rq)->clock;
662 cfs_rq->wait_runtime -= se->wait_runtime;
665 __dequeue_entity(cfs_rq, se);
669 * Preempt the current task with a newly woken task if needed:
672 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
673 struct sched_entity *curr, unsigned long granularity)
675 s64 __delta = curr->fair_key - se->fair_key;
678 * Take scheduling granularity into account - do not
679 * preempt the current task unless the best task has
680 * a larger than sched_granularity fairness advantage:
682 if (__delta > niced_granularity(curr, granularity))
683 resched_task(rq_of(cfs_rq)->curr);
687 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
690 * Any task has to be enqueued before it get to execute on
691 * a CPU. So account for the time it spent waiting on the
692 * runqueue. (note, here we rely on pick_next_task() having
693 * done a put_prev_task_fair() shortly before this, which
694 * updated rq->fair_clock - used by update_stats_wait_end())
696 update_stats_wait_end(cfs_rq, se);
697 update_stats_curr_start(cfs_rq, se);
698 set_cfs_rq_curr(cfs_rq, se);
701 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
703 struct sched_entity *se = __pick_next_entity(cfs_rq);
705 set_next_entity(cfs_rq, se);
710 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
713 * If still on the runqueue then deactivate_task()
714 * was not called and update_curr() has to be done:
719 update_stats_curr_end(cfs_rq, prev);
722 update_stats_wait_start(cfs_rq, prev);
723 set_cfs_rq_curr(cfs_rq, NULL);
726 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
728 struct sched_entity *next;
731 * Dequeue and enqueue the task to update its
732 * position within the tree:
734 dequeue_entity(cfs_rq, curr, 0);
735 enqueue_entity(cfs_rq, curr, 0);
738 * Reschedule if another task tops the current one.
740 next = __pick_next_entity(cfs_rq);
744 __check_preempt_curr_fair(cfs_rq, next, curr,
745 sched_granularity(cfs_rq));
748 /**************************************************
749 * CFS operations on tasks:
752 #ifdef CONFIG_FAIR_GROUP_SCHED
754 /* Walk up scheduling entities hierarchy */
755 #define for_each_sched_entity(se) \
756 for (; se; se = se->parent)
758 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
763 /* runqueue on which this entity is (to be) queued */
764 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
769 /* runqueue "owned" by this group */
770 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
775 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
776 * another cpu ('this_cpu')
778 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
780 /* A later patch will take group into account */
781 return &cpu_rq(this_cpu)->cfs;
784 /* Iterate thr' all leaf cfs_rq's on a runqueue */
785 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
786 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
788 /* Do the two (enqueued) tasks belong to the same group ? */
789 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
791 if (curr->se.cfs_rq == p->se.cfs_rq)
797 #else /* CONFIG_FAIR_GROUP_SCHED */
799 #define for_each_sched_entity(se) \
800 for (; se; se = NULL)
802 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
804 return &task_rq(p)->cfs;
807 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
809 struct task_struct *p = task_of(se);
810 struct rq *rq = task_rq(p);
815 /* runqueue "owned" by this group */
816 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
821 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
823 return &cpu_rq(this_cpu)->cfs;
826 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
827 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
829 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
834 #endif /* CONFIG_FAIR_GROUP_SCHED */
837 * The enqueue_task method is called before nr_running is
838 * increased. Here we update the fair scheduling stats and
839 * then put the task into the rbtree:
841 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
843 struct cfs_rq *cfs_rq;
844 struct sched_entity *se = &p->se;
846 for_each_sched_entity(se) {
849 cfs_rq = cfs_rq_of(se);
850 enqueue_entity(cfs_rq, se, wakeup);
855 * The dequeue_task method is called before nr_running is
856 * decreased. We remove the task from the rbtree and
857 * update the fair scheduling stats:
859 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
861 struct cfs_rq *cfs_rq;
862 struct sched_entity *se = &p->se;
864 for_each_sched_entity(se) {
865 cfs_rq = cfs_rq_of(se);
866 dequeue_entity(cfs_rq, se, sleep);
867 /* Don't dequeue parent if it has other entities besides us */
868 if (cfs_rq->load.weight)
874 * sched_yield() support is very simple - we dequeue and enqueue
876 static void yield_task_fair(struct rq *rq, struct task_struct *p)
878 struct cfs_rq *cfs_rq = task_cfs_rq(p);
880 __update_rq_clock(rq);
882 * Dequeue and enqueue the task to update its
883 * position within the tree:
885 dequeue_entity(cfs_rq, &p->se, 0);
886 enqueue_entity(cfs_rq, &p->se, 0);
890 * Preempt the current task with a newly woken task if needed:
892 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
894 struct task_struct *curr = rq->curr;
895 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
898 if (unlikely(rt_prio(p->prio))) {
905 gran = sysctl_sched_wakeup_granularity;
907 * Batch tasks prefer throughput over latency:
909 if (unlikely(p->policy == SCHED_BATCH))
910 gran = sysctl_sched_batch_wakeup_granularity;
912 if (is_same_group(curr, p))
913 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
916 static struct task_struct *pick_next_task_fair(struct rq *rq)
918 struct cfs_rq *cfs_rq = &rq->cfs;
919 struct sched_entity *se;
921 if (unlikely(!cfs_rq->nr_running))
925 se = pick_next_entity(cfs_rq);
926 cfs_rq = group_cfs_rq(se);
933 * Account for a descheduled task:
935 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
937 struct sched_entity *se = &prev->se;
938 struct cfs_rq *cfs_rq;
940 for_each_sched_entity(se) {
941 cfs_rq = cfs_rq_of(se);
942 put_prev_entity(cfs_rq, se);
946 /**************************************************
947 * Fair scheduling class load-balancing methods:
951 * Load-balancing iterator. Note: while the runqueue stays locked
952 * during the whole iteration, the current task might be
953 * dequeued so the iterator has to be dequeue-safe. Here we
954 * achieve that by always pre-iterating before returning
957 static inline struct task_struct *
958 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
960 struct task_struct *p;
965 p = rb_entry(curr, struct task_struct, se.run_node);
966 cfs_rq->rb_load_balance_curr = rb_next(curr);
971 static struct task_struct *load_balance_start_fair(void *arg)
973 struct cfs_rq *cfs_rq = arg;
975 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
978 static struct task_struct *load_balance_next_fair(void *arg)
980 struct cfs_rq *cfs_rq = arg;
982 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
985 #ifdef CONFIG_FAIR_GROUP_SCHED
986 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
988 struct sched_entity *curr;
989 struct task_struct *p;
991 if (!cfs_rq->nr_running)
994 curr = __pick_next_entity(cfs_rq);
1001 static unsigned long
1002 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1003 unsigned long max_nr_move, unsigned long max_load_move,
1004 struct sched_domain *sd, enum cpu_idle_type idle,
1005 int *all_pinned, int *this_best_prio)
1007 struct cfs_rq *busy_cfs_rq;
1008 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1009 long rem_load_move = max_load_move;
1010 struct rq_iterator cfs_rq_iterator;
1012 cfs_rq_iterator.start = load_balance_start_fair;
1013 cfs_rq_iterator.next = load_balance_next_fair;
1015 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1016 #ifdef CONFIG_FAIR_GROUP_SCHED
1017 struct cfs_rq *this_cfs_rq;
1019 unsigned long maxload;
1021 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1023 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1024 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1028 /* Don't pull more than imbalance/2 */
1030 maxload = min(rem_load_move, imbalance);
1032 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1034 # define maxload rem_load_move
1036 /* pass busy_cfs_rq argument into
1037 * load_balance_[start|next]_fair iterators
1039 cfs_rq_iterator.arg = busy_cfs_rq;
1040 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1041 max_nr_move, maxload, sd, idle, all_pinned,
1042 &load_moved, this_best_prio, &cfs_rq_iterator);
1044 total_nr_moved += nr_moved;
1045 max_nr_move -= nr_moved;
1046 rem_load_move -= load_moved;
1048 if (max_nr_move <= 0 || rem_load_move <= 0)
1052 return max_load_move - rem_load_move;
1056 * scheduler tick hitting a task of our scheduling class:
1058 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1060 struct cfs_rq *cfs_rq;
1061 struct sched_entity *se = &curr->se;
1063 for_each_sched_entity(se) {
1064 cfs_rq = cfs_rq_of(se);
1065 entity_tick(cfs_rq, se);
1070 * Share the fairness runtime between parent and child, thus the
1071 * total amount of pressure for CPU stays equal - new tasks
1072 * get a chance to run but frequent forkers are not allowed to
1073 * monopolize the CPU. Note: the parent runqueue is locked,
1074 * the child is not running yet.
1076 static void task_new_fair(struct rq *rq, struct task_struct *p)
1078 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1079 struct sched_entity *se = &p->se;
1081 sched_info_queued(p);
1083 update_stats_enqueue(cfs_rq, se);
1085 * Child runs first: we let it run before the parent
1086 * until it reschedules once. We set up the key so that
1087 * it will preempt the parent:
1089 p->se.fair_key = current->se.fair_key -
1090 niced_granularity(&rq->curr->se, sched_granularity(cfs_rq)) - 1;
1092 * The first wait is dominated by the child-runs-first logic,
1093 * so do not credit it with that waiting time yet:
1095 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1096 p->se.wait_start_fair = 0;
1099 * The statistical average of wait_runtime is about
1100 * -granularity/2, so initialize the task with that:
1102 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1103 p->se.wait_runtime = -(sched_granularity(cfs_rq) / 2);
1105 __enqueue_entity(cfs_rq, se);
1108 #ifdef CONFIG_FAIR_GROUP_SCHED
1109 /* Account for a task changing its policy or group.
1111 * This routine is mostly called to set cfs_rq->curr field when a task
1112 * migrates between groups/classes.
1114 static void set_curr_task_fair(struct rq *rq)
1116 struct sched_entity *se = &rq->curr->se;
1118 for_each_sched_entity(se)
1119 set_next_entity(cfs_rq_of(se), se);
1122 static void set_curr_task_fair(struct rq *rq)
1128 * All the scheduling class methods:
1130 struct sched_class fair_sched_class __read_mostly = {
1131 .enqueue_task = enqueue_task_fair,
1132 .dequeue_task = dequeue_task_fair,
1133 .yield_task = yield_task_fair,
1135 .check_preempt_curr = check_preempt_curr_fair,
1137 .pick_next_task = pick_next_task_fair,
1138 .put_prev_task = put_prev_task_fair,
1140 .load_balance = load_balance_fair,
1142 .set_curr_task = set_curr_task_fair,
1143 .task_tick = task_tick_fair,
1144 .task_new = task_new_fair,
1147 #ifdef CONFIG_SCHED_DEBUG
1148 static void print_cfs_stats(struct seq_file *m, int cpu)
1150 struct cfs_rq *cfs_rq;
1152 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1153 print_cfs_rq(m, cpu, cfs_rq);