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 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
26 #ifdef CONFIG_SCHED_DEBUG
27 # define const_debug __read_mostly
29 # define const_debug static const
33 * Targeted preemption latency for CPU-bound tasks:
34 * (default: 20ms, units: nanoseconds)
36 * NOTE: this latency value is not the same as the concept of
37 * 'timeslice length' - timeslices in CFS are of variable length.
38 * (to see the precise effective timeslice length of your workload,
39 * run vmstat and monitor the context-switches field)
41 * On SMP systems the value of this is multiplied by the log2 of the
42 * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
43 * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
44 * Targeted preemption latency for CPU-bound tasks:
46 const_debug unsigned int sysctl_sched_latency = 20000000ULL;
49 * After fork, child runs first. (default) If set to 0 then
50 * parent will (try to) run first.
52 const_debug unsigned int sysctl_sched_child_runs_first = 1;
55 * Minimal preemption granularity for CPU-bound tasks:
56 * (default: 2 msec, units: nanoseconds)
58 unsigned int sysctl_sched_min_granularity __read_mostly = 2000000ULL;
61 * sys_sched_yield() compat mode
63 * This option switches the agressive yield implementation of the
64 * old scheduler back on.
66 unsigned int __read_mostly sysctl_sched_compat_yield;
69 * SCHED_BATCH wake-up granularity.
70 * (default: 25 msec, units: nanoseconds)
72 * This option delays the preemption effects of decoupled workloads
73 * and reduces their over-scheduling. Synchronous workloads will still
74 * have immediate wakeup/sleep latencies.
76 const_debug unsigned int sysctl_sched_batch_wakeup_granularity = 25000000UL;
79 * SCHED_OTHER wake-up granularity.
80 * (default: 1 msec, units: nanoseconds)
82 * This option delays the preemption effects of decoupled workloads
83 * and reduces their over-scheduling. Synchronous workloads will still
84 * have immediate wakeup/sleep latencies.
86 const_debug unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88 unsigned int sysctl_sched_runtime_limit __read_mostly;
91 * Debugging: various feature bits
94 SCHED_FEAT_FAIR_SLEEPERS = 1,
95 SCHED_FEAT_SLEEPER_AVG = 2,
96 SCHED_FEAT_SLEEPER_LOAD_AVG = 4,
97 SCHED_FEAT_START_DEBIT = 8,
98 SCHED_FEAT_SKIP_INITIAL = 16,
101 const_debug unsigned int sysctl_sched_features =
102 SCHED_FEAT_FAIR_SLEEPERS *1 |
103 SCHED_FEAT_SLEEPER_AVG *0 |
104 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
105 SCHED_FEAT_START_DEBIT *1 |
106 SCHED_FEAT_SKIP_INITIAL *0;
108 extern struct sched_class fair_sched_class;
110 /**************************************************************
111 * CFS operations on generic schedulable entities:
114 #ifdef CONFIG_FAIR_GROUP_SCHED
116 /* cpu runqueue to which this cfs_rq is attached */
117 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
122 /* currently running entity (if any) on this cfs_rq */
123 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
128 /* An entity is a task if it doesn't "own" a runqueue */
129 #define entity_is_task(se) (!se->my_q)
132 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
137 #else /* CONFIG_FAIR_GROUP_SCHED */
139 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
141 return container_of(cfs_rq, struct rq, cfs);
144 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
146 struct rq *rq = rq_of(cfs_rq);
148 if (unlikely(rq->curr->sched_class != &fair_sched_class))
151 return &rq->curr->se;
154 #define entity_is_task(se) 1
157 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
159 #endif /* CONFIG_FAIR_GROUP_SCHED */
161 static inline struct task_struct *task_of(struct sched_entity *se)
163 return container_of(se, struct task_struct, se);
167 /**************************************************************
168 * Scheduling class tree data structure manipulation methods:
172 * Enqueue an entity into the rb-tree:
175 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
177 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
178 struct rb_node *parent = NULL;
179 struct sched_entity *entry;
180 s64 key = se->fair_key;
184 * Find the right place in the rbtree:
188 entry = rb_entry(parent, struct sched_entity, run_node);
190 * We dont care about collisions. Nodes with
191 * the same key stay together.
193 if (key - entry->fair_key < 0) {
194 link = &parent->rb_left;
196 link = &parent->rb_right;
202 * Maintain a cache of leftmost tree entries (it is frequently
206 cfs_rq->rb_leftmost = &se->run_node;
208 rb_link_node(&se->run_node, parent, link);
209 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
210 update_load_add(&cfs_rq->load, se->load.weight);
211 cfs_rq->nr_running++;
214 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
218 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
220 if (cfs_rq->rb_leftmost == &se->run_node)
221 cfs_rq->rb_leftmost = rb_next(&se->run_node);
222 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
223 update_load_sub(&cfs_rq->load, se->load.weight);
224 cfs_rq->nr_running--;
227 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
230 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
232 return cfs_rq->rb_leftmost;
235 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
237 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
240 /**************************************************************
241 * Scheduling class statistics methods:
245 * Calculate the preemption granularity needed to schedule every
246 * runnable task once per sysctl_sched_latency amount of time.
247 * (down to a sensible low limit on granularity)
249 * For example, if there are 2 tasks running and latency is 10 msecs,
250 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
251 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
252 * for each task. We do finer and finer scheduling up to until we
253 * reach the minimum granularity value.
255 * To achieve this we use the following dynamic-granularity rule:
257 * gran = lat/nr - lat/nr/nr
259 * This comes out of the following equations:
264 * kB2 = kB1 - d + d/nr
267 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
268 * '1' is start of time, '2' is end of time, 'd' is delay between
269 * 1 and 2 (during which task B was running), 'nr' is number of tasks
270 * running, 'lat' is the the period of each task. ('lat' is the
271 * sched_latency that we aim for.)
274 sched_granularity(struct cfs_rq *cfs_rq)
276 unsigned int gran = sysctl_sched_latency;
277 unsigned int nr = cfs_rq->nr_running;
280 gran = gran/nr - gran/nr/nr;
281 gran = max(gran, sysctl_sched_min_granularity);
288 * We rescale the rescheduling granularity of tasks according to their
289 * nice level, but only linearly, not exponentially:
292 niced_granularity(struct sched_entity *curr, unsigned long granularity)
296 if (likely(curr->load.weight == NICE_0_LOAD))
299 * Positive nice levels get the same granularity as nice-0:
301 if (likely(curr->load.weight < NICE_0_LOAD)) {
302 tmp = curr->load.weight * (u64)granularity;
303 return (long) (tmp >> NICE_0_SHIFT);
306 * Negative nice level tasks get linearly finer
309 tmp = curr->load.inv_weight * (u64)granularity;
312 * It will always fit into 'long':
314 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
318 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
320 long limit = sysctl_sched_runtime_limit;
323 * Niced tasks have the same history dynamic range as
326 if (unlikely(se->wait_runtime > limit)) {
327 se->wait_runtime = limit;
328 schedstat_inc(se, wait_runtime_overruns);
329 schedstat_inc(cfs_rq, wait_runtime_overruns);
331 if (unlikely(se->wait_runtime < -limit)) {
332 se->wait_runtime = -limit;
333 schedstat_inc(se, wait_runtime_underruns);
334 schedstat_inc(cfs_rq, wait_runtime_underruns);
339 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
341 se->wait_runtime += delta;
342 schedstat_add(se, sum_wait_runtime, delta);
343 limit_wait_runtime(cfs_rq, se);
347 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
349 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
350 __add_wait_runtime(cfs_rq, se, delta);
351 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
355 * Update the current task's runtime statistics. Skip current tasks that
356 * are not in our scheduling class.
359 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
360 unsigned long delta_exec)
362 unsigned long delta, delta_fair, delta_mine;
363 struct load_weight *lw = &cfs_rq->load;
364 unsigned long load = lw->weight;
366 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
368 curr->sum_exec_runtime += delta_exec;
369 cfs_rq->exec_clock += delta_exec;
374 delta_fair = calc_delta_fair(delta_exec, lw);
375 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
377 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
378 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
379 delta = min(delta, (unsigned long)(
380 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
381 cfs_rq->sleeper_bonus -= delta;
385 cfs_rq->fair_clock += delta_fair;
387 * We executed delta_exec amount of time on the CPU,
388 * but we were only entitled to delta_mine amount of
389 * time during that period (if nr_running == 1 then
390 * the two values are equal)
391 * [Note: delta_mine - delta_exec is negative]:
393 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
396 static void update_curr(struct cfs_rq *cfs_rq)
398 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
399 u64 now = rq_of(cfs_rq)->clock;
400 unsigned long delta_exec;
406 * Get the amount of time the current task was running
407 * since the last time we changed load (this cannot
408 * overflow on 32 bits):
410 delta_exec = (unsigned long)(now - curr->exec_start);
412 __update_curr(cfs_rq, curr, delta_exec);
413 curr->exec_start = now;
417 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
419 se->wait_start_fair = cfs_rq->fair_clock;
420 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
424 * We calculate fair deltas here, so protect against the random effects
425 * of a multiplication overflow by capping it to the runtime limit:
427 #if BITS_PER_LONG == 32
428 static inline unsigned long
429 calc_weighted(unsigned long delta, unsigned long weight, int shift)
431 u64 tmp = (u64)delta * weight >> shift;
433 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
434 return sysctl_sched_runtime_limit*2;
438 static inline unsigned long
439 calc_weighted(unsigned long delta, unsigned long weight, int shift)
441 return delta * weight >> shift;
446 * Task is being enqueued - update stats:
448 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
453 * Are we enqueueing a waiting task? (for current tasks
454 * a dequeue/enqueue event is a NOP)
456 if (se != cfs_rq_curr(cfs_rq))
457 update_stats_wait_start(cfs_rq, se);
461 key = cfs_rq->fair_clock;
464 * Optimize the common nice 0 case:
466 if (likely(se->load.weight == NICE_0_LOAD)) {
467 key -= se->wait_runtime;
471 if (se->wait_runtime < 0) {
472 tmp = -se->wait_runtime;
473 key += (tmp * se->load.inv_weight) >>
474 (WMULT_SHIFT - NICE_0_SHIFT);
476 tmp = se->wait_runtime;
477 key -= (tmp * se->load.inv_weight) >>
478 (WMULT_SHIFT - NICE_0_SHIFT);
486 * Note: must be called with a freshly updated rq->fair_clock.
489 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se,
490 unsigned long delta_fair)
492 schedstat_set(se->wait_max, max(se->wait_max,
493 rq_of(cfs_rq)->clock - se->wait_start));
495 if (unlikely(se->load.weight != NICE_0_LOAD))
496 delta_fair = calc_weighted(delta_fair, se->load.weight,
499 add_wait_runtime(cfs_rq, se, delta_fair);
503 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
505 unsigned long delta_fair;
507 if (unlikely(!se->wait_start_fair))
510 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
511 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
513 __update_stats_wait_end(cfs_rq, se, delta_fair);
515 se->wait_start_fair = 0;
516 schedstat_set(se->wait_start, 0);
520 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
524 * Mark the end of the wait period if dequeueing a
527 if (se != cfs_rq_curr(cfs_rq))
528 update_stats_wait_end(cfs_rq, se);
532 * We are picking a new current task - update its stats:
535 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
538 * We are starting a new run period:
540 se->exec_start = rq_of(cfs_rq)->clock;
544 * We are descheduling a task - update its stats:
547 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
552 /**************************************************
553 * Scheduling class queueing methods:
556 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
557 unsigned long delta_fair)
559 unsigned long load = cfs_rq->load.weight;
563 * Do not boost sleepers if there's too much bonus 'in flight'
566 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
569 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
570 load = rq_of(cfs_rq)->cpu_load[2];
573 * Fix up delta_fair with the effect of us running
574 * during the whole sleep period:
576 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
577 delta_fair = div64_likely32((u64)delta_fair * load,
578 load + se->load.weight);
580 if (unlikely(se->load.weight != NICE_0_LOAD))
581 delta_fair = calc_weighted(delta_fair, se->load.weight,
584 prev_runtime = se->wait_runtime;
585 __add_wait_runtime(cfs_rq, se, delta_fair);
586 delta_fair = se->wait_runtime - prev_runtime;
589 * Track the amount of bonus we've given to sleepers:
591 cfs_rq->sleeper_bonus += delta_fair;
594 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
596 struct task_struct *tsk = task_of(se);
597 unsigned long delta_fair;
599 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
600 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
603 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
604 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
606 __enqueue_sleeper(cfs_rq, se, delta_fair);
608 se->sleep_start_fair = 0;
610 #ifdef CONFIG_SCHEDSTATS
611 if (se->sleep_start) {
612 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
617 if (unlikely(delta > se->sleep_max))
618 se->sleep_max = delta;
621 se->sum_sleep_runtime += delta;
623 if (se->block_start) {
624 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
629 if (unlikely(delta > se->block_max))
630 se->block_max = delta;
633 se->sum_sleep_runtime += delta;
636 * Blocking time is in units of nanosecs, so shift by 20 to
637 * get a milliseconds-range estimation of the amount of
638 * time that the task spent sleeping:
640 if (unlikely(prof_on == SLEEP_PROFILING)) {
641 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
649 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
652 * Update the fair clock.
657 enqueue_sleeper(cfs_rq, se);
659 update_stats_enqueue(cfs_rq, se);
660 __enqueue_entity(cfs_rq, se);
664 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
666 update_stats_dequeue(cfs_rq, se);
668 se->sleep_start_fair = cfs_rq->fair_clock;
669 #ifdef CONFIG_SCHEDSTATS
670 if (entity_is_task(se)) {
671 struct task_struct *tsk = task_of(se);
673 if (tsk->state & TASK_INTERRUPTIBLE)
674 se->sleep_start = rq_of(cfs_rq)->clock;
675 if (tsk->state & TASK_UNINTERRUPTIBLE)
676 se->block_start = rq_of(cfs_rq)->clock;
680 __dequeue_entity(cfs_rq, se);
684 * Preempt the current task with a newly woken task if needed:
687 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
688 struct sched_entity *curr, unsigned long granularity)
690 s64 __delta = curr->fair_key - se->fair_key;
691 unsigned long ideal_runtime, delta_exec;
694 * ideal_runtime is compared against sum_exec_runtime, which is
695 * walltime, hence do not scale.
697 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
698 (unsigned long)sysctl_sched_min_granularity);
701 * If we executed more than what the latency constraint suggests,
702 * reduce the rescheduling granularity. This way the total latency
703 * of how much a task is not scheduled converges to
704 * sysctl_sched_latency:
706 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
707 if (delta_exec > ideal_runtime)
711 * Take scheduling granularity into account - do not
712 * preempt the current task unless the best task has
713 * a larger than sched_granularity fairness advantage:
715 * scale granularity as key space is in fair_clock.
717 if (__delta > niced_granularity(curr, granularity))
718 resched_task(rq_of(cfs_rq)->curr);
722 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
725 * Any task has to be enqueued before it get to execute on
726 * a CPU. So account for the time it spent waiting on the
727 * runqueue. (note, here we rely on pick_next_task() having
728 * done a put_prev_task_fair() shortly before this, which
729 * updated rq->fair_clock - used by update_stats_wait_end())
731 update_stats_wait_end(cfs_rq, se);
732 update_stats_curr_start(cfs_rq, se);
733 set_cfs_rq_curr(cfs_rq, se);
734 #ifdef CONFIG_SCHEDSTATS
736 * Track our maximum slice length, if the CPU's load is at
737 * least twice that of our own weight (i.e. dont track it
738 * when there are only lesser-weight tasks around):
740 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
741 se->slice_max = max(se->slice_max,
742 se->sum_exec_runtime - se->prev_sum_exec_runtime);
745 se->prev_sum_exec_runtime = se->sum_exec_runtime;
748 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
750 struct sched_entity *se = __pick_next_entity(cfs_rq);
752 set_next_entity(cfs_rq, se);
757 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
760 * If still on the runqueue then deactivate_task()
761 * was not called and update_curr() has to be done:
766 update_stats_curr_end(cfs_rq, prev);
769 update_stats_wait_start(cfs_rq, prev);
770 set_cfs_rq_curr(cfs_rq, NULL);
773 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
775 struct sched_entity *next;
778 * Dequeue and enqueue the task to update its
779 * position within the tree:
781 dequeue_entity(cfs_rq, curr, 0);
782 enqueue_entity(cfs_rq, curr, 0);
785 * Reschedule if another task tops the current one.
787 next = __pick_next_entity(cfs_rq);
791 __check_preempt_curr_fair(cfs_rq, next, curr,
792 sched_granularity(cfs_rq));
795 /**************************************************
796 * CFS operations on tasks:
799 #ifdef CONFIG_FAIR_GROUP_SCHED
801 /* Walk up scheduling entities hierarchy */
802 #define for_each_sched_entity(se) \
803 for (; se; se = se->parent)
805 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
810 /* runqueue on which this entity is (to be) queued */
811 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
816 /* runqueue "owned" by this group */
817 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
822 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
823 * another cpu ('this_cpu')
825 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
827 /* A later patch will take group into account */
828 return &cpu_rq(this_cpu)->cfs;
831 /* Iterate thr' all leaf cfs_rq's on a runqueue */
832 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
833 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
835 /* Do the two (enqueued) tasks belong to the same group ? */
836 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
838 if (curr->se.cfs_rq == p->se.cfs_rq)
844 #else /* CONFIG_FAIR_GROUP_SCHED */
846 #define for_each_sched_entity(se) \
847 for (; se; se = NULL)
849 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
851 return &task_rq(p)->cfs;
854 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
856 struct task_struct *p = task_of(se);
857 struct rq *rq = task_rq(p);
862 /* runqueue "owned" by this group */
863 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
868 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
870 return &cpu_rq(this_cpu)->cfs;
873 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
874 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
876 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
881 #endif /* CONFIG_FAIR_GROUP_SCHED */
884 * The enqueue_task method is called before nr_running is
885 * increased. Here we update the fair scheduling stats and
886 * then put the task into the rbtree:
888 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
890 struct cfs_rq *cfs_rq;
891 struct sched_entity *se = &p->se;
893 for_each_sched_entity(se) {
896 cfs_rq = cfs_rq_of(se);
897 enqueue_entity(cfs_rq, se, wakeup);
902 * The dequeue_task method is called before nr_running is
903 * decreased. We remove the task from the rbtree and
904 * update the fair scheduling stats:
906 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
908 struct cfs_rq *cfs_rq;
909 struct sched_entity *se = &p->se;
911 for_each_sched_entity(se) {
912 cfs_rq = cfs_rq_of(se);
913 dequeue_entity(cfs_rq, se, sleep);
914 /* Don't dequeue parent if it has other entities besides us */
915 if (cfs_rq->load.weight)
921 * sched_yield() support is very simple - we dequeue and enqueue.
923 * If compat_yield is turned on then we requeue to the end of the tree.
925 static void yield_task_fair(struct rq *rq, struct task_struct *p)
927 struct cfs_rq *cfs_rq = task_cfs_rq(p);
928 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
929 struct sched_entity *rightmost, *se = &p->se;
930 struct rb_node *parent;
933 * Are we the only task in the tree?
935 if (unlikely(cfs_rq->nr_running == 1))
938 if (likely(!sysctl_sched_compat_yield)) {
939 __update_rq_clock(rq);
941 * Dequeue and enqueue the task to update its
942 * position within the tree:
944 dequeue_entity(cfs_rq, &p->se, 0);
945 enqueue_entity(cfs_rq, &p->se, 0);
950 * Find the rightmost entry in the rbtree:
954 link = &parent->rb_right;
957 rightmost = rb_entry(parent, struct sched_entity, run_node);
959 * Already in the rightmost position?
961 if (unlikely(rightmost == se))
965 * Minimally necessary key value to be last in the tree:
967 se->fair_key = rightmost->fair_key + 1;
969 if (cfs_rq->rb_leftmost == &se->run_node)
970 cfs_rq->rb_leftmost = rb_next(&se->run_node);
972 * Relink the task to the rightmost position:
974 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
975 rb_link_node(&se->run_node, parent, link);
976 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
980 * Preempt the current task with a newly woken task if needed:
982 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
984 struct task_struct *curr = rq->curr;
985 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
988 if (unlikely(rt_prio(p->prio))) {
995 gran = sysctl_sched_wakeup_granularity;
997 * Batch tasks prefer throughput over latency:
999 if (unlikely(p->policy == SCHED_BATCH))
1000 gran = sysctl_sched_batch_wakeup_granularity;
1002 if (is_same_group(curr, p))
1003 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
1006 static struct task_struct *pick_next_task_fair(struct rq *rq)
1008 struct cfs_rq *cfs_rq = &rq->cfs;
1009 struct sched_entity *se;
1011 if (unlikely(!cfs_rq->nr_running))
1015 se = pick_next_entity(cfs_rq);
1016 cfs_rq = group_cfs_rq(se);
1023 * Account for a descheduled task:
1025 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1027 struct sched_entity *se = &prev->se;
1028 struct cfs_rq *cfs_rq;
1030 for_each_sched_entity(se) {
1031 cfs_rq = cfs_rq_of(se);
1032 put_prev_entity(cfs_rq, se);
1036 /**************************************************
1037 * Fair scheduling class load-balancing methods:
1041 * Load-balancing iterator. Note: while the runqueue stays locked
1042 * during the whole iteration, the current task might be
1043 * dequeued so the iterator has to be dequeue-safe. Here we
1044 * achieve that by always pre-iterating before returning
1047 static inline struct task_struct *
1048 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1050 struct task_struct *p;
1055 p = rb_entry(curr, struct task_struct, se.run_node);
1056 cfs_rq->rb_load_balance_curr = rb_next(curr);
1061 static struct task_struct *load_balance_start_fair(void *arg)
1063 struct cfs_rq *cfs_rq = arg;
1065 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1068 static struct task_struct *load_balance_next_fair(void *arg)
1070 struct cfs_rq *cfs_rq = arg;
1072 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1075 #ifdef CONFIG_FAIR_GROUP_SCHED
1076 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1078 struct sched_entity *curr;
1079 struct task_struct *p;
1081 if (!cfs_rq->nr_running)
1084 curr = __pick_next_entity(cfs_rq);
1091 static unsigned long
1092 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1093 unsigned long max_nr_move, unsigned long max_load_move,
1094 struct sched_domain *sd, enum cpu_idle_type idle,
1095 int *all_pinned, int *this_best_prio)
1097 struct cfs_rq *busy_cfs_rq;
1098 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1099 long rem_load_move = max_load_move;
1100 struct rq_iterator cfs_rq_iterator;
1102 cfs_rq_iterator.start = load_balance_start_fair;
1103 cfs_rq_iterator.next = load_balance_next_fair;
1105 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1106 #ifdef CONFIG_FAIR_GROUP_SCHED
1107 struct cfs_rq *this_cfs_rq;
1109 unsigned long maxload;
1111 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1113 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1114 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1118 /* Don't pull more than imbalance/2 */
1120 maxload = min(rem_load_move, imbalance);
1122 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1124 # define maxload rem_load_move
1126 /* pass busy_cfs_rq argument into
1127 * load_balance_[start|next]_fair iterators
1129 cfs_rq_iterator.arg = busy_cfs_rq;
1130 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1131 max_nr_move, maxload, sd, idle, all_pinned,
1132 &load_moved, this_best_prio, &cfs_rq_iterator);
1134 total_nr_moved += nr_moved;
1135 max_nr_move -= nr_moved;
1136 rem_load_move -= load_moved;
1138 if (max_nr_move <= 0 || rem_load_move <= 0)
1142 return max_load_move - rem_load_move;
1146 * scheduler tick hitting a task of our scheduling class:
1148 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1150 struct cfs_rq *cfs_rq;
1151 struct sched_entity *se = &curr->se;
1153 for_each_sched_entity(se) {
1154 cfs_rq = cfs_rq_of(se);
1155 entity_tick(cfs_rq, se);
1160 * Share the fairness runtime between parent and child, thus the
1161 * total amount of pressure for CPU stays equal - new tasks
1162 * get a chance to run but frequent forkers are not allowed to
1163 * monopolize the CPU. Note: the parent runqueue is locked,
1164 * the child is not running yet.
1166 static void task_new_fair(struct rq *rq, struct task_struct *p)
1168 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1169 struct sched_entity *se = &p->se, *curr = cfs_rq_curr(cfs_rq);
1171 sched_info_queued(p);
1173 update_curr(cfs_rq);
1174 update_stats_enqueue(cfs_rq, se);
1176 * Child runs first: we let it run before the parent
1177 * until it reschedules once. We set up the key so that
1178 * it will preempt the parent:
1180 se->fair_key = curr->fair_key -
1181 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1183 * The first wait is dominated by the child-runs-first logic,
1184 * so do not credit it with that waiting time yet:
1186 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1187 se->wait_start_fair = 0;
1190 * The statistical average of wait_runtime is about
1191 * -granularity/2, so initialize the task with that:
1193 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1194 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1196 __enqueue_entity(cfs_rq, se);
1197 resched_task(rq->curr);
1200 #ifdef CONFIG_FAIR_GROUP_SCHED
1201 /* Account for a task changing its policy or group.
1203 * This routine is mostly called to set cfs_rq->curr field when a task
1204 * migrates between groups/classes.
1206 static void set_curr_task_fair(struct rq *rq)
1208 struct sched_entity *se = &rq->curr->se;
1210 for_each_sched_entity(se)
1211 set_next_entity(cfs_rq_of(se), se);
1214 static void set_curr_task_fair(struct rq *rq)
1220 * All the scheduling class methods:
1222 struct sched_class fair_sched_class __read_mostly = {
1223 .enqueue_task = enqueue_task_fair,
1224 .dequeue_task = dequeue_task_fair,
1225 .yield_task = yield_task_fair,
1227 .check_preempt_curr = check_preempt_curr_fair,
1229 .pick_next_task = pick_next_task_fair,
1230 .put_prev_task = put_prev_task_fair,
1232 .load_balance = load_balance_fair,
1234 .set_curr_task = set_curr_task_fair,
1235 .task_tick = task_tick_fair,
1236 .task_new = task_new_fair,
1239 #ifdef CONFIG_SCHED_DEBUG
1240 static void print_cfs_stats(struct seq_file *m, int cpu)
1242 struct cfs_rq *cfs_rq;
1244 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1245 print_cfs_rq(m, cpu, cfs_rq);