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 /* An entity is a task if it doesn't "own" a runqueue */
123 #define entity_is_task(se) (!se->my_q)
125 #else /* CONFIG_FAIR_GROUP_SCHED */
127 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
129 return container_of(cfs_rq, struct rq, cfs);
132 #define entity_is_task(se) 1
134 #endif /* CONFIG_FAIR_GROUP_SCHED */
136 static inline struct task_struct *task_of(struct sched_entity *se)
138 return container_of(se, struct task_struct, se);
142 /**************************************************************
143 * Scheduling class tree data structure manipulation methods:
147 * Enqueue an entity into the rb-tree:
150 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
152 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
153 struct rb_node *parent = NULL;
154 struct sched_entity *entry;
155 s64 key = se->fair_key;
159 * Find the right place in the rbtree:
163 entry = rb_entry(parent, struct sched_entity, run_node);
165 * We dont care about collisions. Nodes with
166 * the same key stay together.
168 if (key - entry->fair_key < 0) {
169 link = &parent->rb_left;
171 link = &parent->rb_right;
177 * Maintain a cache of leftmost tree entries (it is frequently
181 cfs_rq->rb_leftmost = &se->run_node;
183 rb_link_node(&se->run_node, parent, link);
184 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
185 update_load_add(&cfs_rq->load, se->load.weight);
186 cfs_rq->nr_running++;
189 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
193 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
195 if (cfs_rq->rb_leftmost == &se->run_node)
196 cfs_rq->rb_leftmost = rb_next(&se->run_node);
197 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
198 update_load_sub(&cfs_rq->load, se->load.weight);
199 cfs_rq->nr_running--;
202 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
205 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
207 return cfs_rq->rb_leftmost;
210 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
212 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
215 /**************************************************************
216 * Scheduling class statistics methods:
220 * Calculate the preemption granularity needed to schedule every
221 * runnable task once per sysctl_sched_latency amount of time.
222 * (down to a sensible low limit on granularity)
224 * For example, if there are 2 tasks running and latency is 10 msecs,
225 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
226 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
227 * for each task. We do finer and finer scheduling up to until we
228 * reach the minimum granularity value.
230 * To achieve this we use the following dynamic-granularity rule:
232 * gran = lat/nr - lat/nr/nr
234 * This comes out of the following equations:
239 * kB2 = kB1 - d + d/nr
242 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
243 * '1' is start of time, '2' is end of time, 'd' is delay between
244 * 1 and 2 (during which task B was running), 'nr' is number of tasks
245 * running, 'lat' is the the period of each task. ('lat' is the
246 * sched_latency that we aim for.)
249 sched_granularity(struct cfs_rq *cfs_rq)
251 unsigned int gran = sysctl_sched_latency;
252 unsigned int nr = cfs_rq->nr_running;
255 gran = gran/nr - gran/nr/nr;
256 gran = max(gran, sysctl_sched_min_granularity);
263 * We rescale the rescheduling granularity of tasks according to their
264 * nice level, but only linearly, not exponentially:
267 niced_granularity(struct sched_entity *curr, unsigned long granularity)
271 if (likely(curr->load.weight == NICE_0_LOAD))
274 * Positive nice levels get the same granularity as nice-0:
276 if (likely(curr->load.weight < NICE_0_LOAD)) {
277 tmp = curr->load.weight * (u64)granularity;
278 return (long) (tmp >> NICE_0_SHIFT);
281 * Negative nice level tasks get linearly finer
284 tmp = curr->load.inv_weight * (u64)granularity;
287 * It will always fit into 'long':
289 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
293 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
295 long limit = sysctl_sched_runtime_limit;
298 * Niced tasks have the same history dynamic range as
301 if (unlikely(se->wait_runtime > limit)) {
302 se->wait_runtime = limit;
303 schedstat_inc(se, wait_runtime_overruns);
304 schedstat_inc(cfs_rq, wait_runtime_overruns);
306 if (unlikely(se->wait_runtime < -limit)) {
307 se->wait_runtime = -limit;
308 schedstat_inc(se, wait_runtime_underruns);
309 schedstat_inc(cfs_rq, wait_runtime_underruns);
314 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
316 se->wait_runtime += delta;
317 schedstat_add(se, sum_wait_runtime, delta);
318 limit_wait_runtime(cfs_rq, se);
322 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
324 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
325 __add_wait_runtime(cfs_rq, se, delta);
326 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
330 * Update the current task's runtime statistics. Skip current tasks that
331 * are not in our scheduling class.
334 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
335 unsigned long delta_exec)
337 unsigned long delta, delta_fair, delta_mine;
338 struct load_weight *lw = &cfs_rq->load;
339 unsigned long load = lw->weight;
341 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
343 curr->sum_exec_runtime += delta_exec;
344 cfs_rq->exec_clock += delta_exec;
349 delta_fair = calc_delta_fair(delta_exec, lw);
350 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
352 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
353 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
354 delta = min(delta, (unsigned long)(
355 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
356 cfs_rq->sleeper_bonus -= delta;
360 cfs_rq->fair_clock += delta_fair;
362 * We executed delta_exec amount of time on the CPU,
363 * but we were only entitled to delta_mine amount of
364 * time during that period (if nr_running == 1 then
365 * the two values are equal)
366 * [Note: delta_mine - delta_exec is negative]:
368 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
371 static void update_curr(struct cfs_rq *cfs_rq)
373 struct sched_entity *curr = cfs_rq->curr;
374 u64 now = rq_of(cfs_rq)->clock;
375 unsigned long delta_exec;
381 * Get the amount of time the current task was running
382 * since the last time we changed load (this cannot
383 * overflow on 32 bits):
385 delta_exec = (unsigned long)(now - curr->exec_start);
387 __update_curr(cfs_rq, curr, delta_exec);
388 curr->exec_start = now;
392 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
394 se->wait_start_fair = cfs_rq->fair_clock;
395 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
399 * We calculate fair deltas here, so protect against the random effects
400 * of a multiplication overflow by capping it to the runtime limit:
402 #if BITS_PER_LONG == 32
403 static inline unsigned long
404 calc_weighted(unsigned long delta, unsigned long weight, int shift)
406 u64 tmp = (u64)delta * weight >> shift;
408 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
409 return sysctl_sched_runtime_limit*2;
413 static inline unsigned long
414 calc_weighted(unsigned long delta, unsigned long weight, int shift)
416 return delta * weight >> shift;
421 * Task is being enqueued - update stats:
423 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
428 * Are we enqueueing a waiting task? (for current tasks
429 * a dequeue/enqueue event is a NOP)
431 if (se != cfs_rq->curr)
432 update_stats_wait_start(cfs_rq, se);
436 key = cfs_rq->fair_clock;
439 * Optimize the common nice 0 case:
441 if (likely(se->load.weight == NICE_0_LOAD)) {
442 key -= se->wait_runtime;
446 if (se->wait_runtime < 0) {
447 tmp = -se->wait_runtime;
448 key += (tmp * se->load.inv_weight) >>
449 (WMULT_SHIFT - NICE_0_SHIFT);
451 tmp = se->wait_runtime;
452 key -= (tmp * se->load.inv_weight) >>
453 (WMULT_SHIFT - NICE_0_SHIFT);
461 * Note: must be called with a freshly updated rq->fair_clock.
464 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se,
465 unsigned long delta_fair)
467 schedstat_set(se->wait_max, max(se->wait_max,
468 rq_of(cfs_rq)->clock - se->wait_start));
470 if (unlikely(se->load.weight != NICE_0_LOAD))
471 delta_fair = calc_weighted(delta_fair, se->load.weight,
474 add_wait_runtime(cfs_rq, se, delta_fair);
478 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
480 unsigned long delta_fair;
482 if (unlikely(!se->wait_start_fair))
485 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
486 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
488 __update_stats_wait_end(cfs_rq, se, delta_fair);
490 se->wait_start_fair = 0;
491 schedstat_set(se->wait_start, 0);
495 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
499 * Mark the end of the wait period if dequeueing a
502 if (se != cfs_rq->curr)
503 update_stats_wait_end(cfs_rq, se);
507 * We are picking a new current task - update its stats:
510 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
513 * We are starting a new run period:
515 se->exec_start = rq_of(cfs_rq)->clock;
519 * We are descheduling a task - update its stats:
522 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
527 /**************************************************
528 * Scheduling class queueing methods:
531 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
532 unsigned long delta_fair)
534 unsigned long load = cfs_rq->load.weight;
538 * Do not boost sleepers if there's too much bonus 'in flight'
541 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
544 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
545 load = rq_of(cfs_rq)->cpu_load[2];
548 * Fix up delta_fair with the effect of us running
549 * during the whole sleep period:
551 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
552 delta_fair = div64_likely32((u64)delta_fair * load,
553 load + se->load.weight);
555 if (unlikely(se->load.weight != NICE_0_LOAD))
556 delta_fair = calc_weighted(delta_fair, se->load.weight,
559 prev_runtime = se->wait_runtime;
560 __add_wait_runtime(cfs_rq, se, delta_fair);
561 delta_fair = se->wait_runtime - prev_runtime;
564 * Track the amount of bonus we've given to sleepers:
566 cfs_rq->sleeper_bonus += delta_fair;
569 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
571 struct task_struct *tsk = task_of(se);
572 unsigned long delta_fair;
574 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
575 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
578 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
579 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
581 __enqueue_sleeper(cfs_rq, se, delta_fair);
583 se->sleep_start_fair = 0;
585 #ifdef CONFIG_SCHEDSTATS
586 if (se->sleep_start) {
587 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
592 if (unlikely(delta > se->sleep_max))
593 se->sleep_max = delta;
596 se->sum_sleep_runtime += delta;
598 if (se->block_start) {
599 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
604 if (unlikely(delta > se->block_max))
605 se->block_max = delta;
608 se->sum_sleep_runtime += delta;
611 * Blocking time is in units of nanosecs, so shift by 20 to
612 * get a milliseconds-range estimation of the amount of
613 * time that the task spent sleeping:
615 if (unlikely(prof_on == SLEEP_PROFILING)) {
616 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
624 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
627 * Update the fair clock.
632 enqueue_sleeper(cfs_rq, se);
634 update_stats_enqueue(cfs_rq, se);
635 __enqueue_entity(cfs_rq, se);
639 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
641 update_stats_dequeue(cfs_rq, se);
643 se->sleep_start_fair = cfs_rq->fair_clock;
644 #ifdef CONFIG_SCHEDSTATS
645 if (entity_is_task(se)) {
646 struct task_struct *tsk = task_of(se);
648 if (tsk->state & TASK_INTERRUPTIBLE)
649 se->sleep_start = rq_of(cfs_rq)->clock;
650 if (tsk->state & TASK_UNINTERRUPTIBLE)
651 se->block_start = rq_of(cfs_rq)->clock;
655 __dequeue_entity(cfs_rq, se);
659 * Preempt the current task with a newly woken task if needed:
662 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
663 struct sched_entity *curr, unsigned long granularity)
665 s64 __delta = curr->fair_key - se->fair_key;
666 unsigned long ideal_runtime, delta_exec;
669 * ideal_runtime is compared against sum_exec_runtime, which is
670 * walltime, hence do not scale.
672 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
673 (unsigned long)sysctl_sched_min_granularity);
676 * If we executed more than what the latency constraint suggests,
677 * reduce the rescheduling granularity. This way the total latency
678 * of how much a task is not scheduled converges to
679 * sysctl_sched_latency:
681 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
682 if (delta_exec > ideal_runtime)
686 * Take scheduling granularity into account - do not
687 * preempt the current task unless the best task has
688 * a larger than sched_granularity fairness advantage:
690 * scale granularity as key space is in fair_clock.
692 if (__delta > niced_granularity(curr, granularity))
693 resched_task(rq_of(cfs_rq)->curr);
697 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
700 * Any task has to be enqueued before it get to execute on
701 * a CPU. So account for the time it spent waiting on the
702 * runqueue. (note, here we rely on pick_next_task() having
703 * done a put_prev_task_fair() shortly before this, which
704 * updated rq->fair_clock - used by update_stats_wait_end())
706 update_stats_wait_end(cfs_rq, se);
707 update_stats_curr_start(cfs_rq, se);
709 #ifdef CONFIG_SCHEDSTATS
711 * Track our maximum slice length, if the CPU's load is at
712 * least twice that of our own weight (i.e. dont track it
713 * when there are only lesser-weight tasks around):
715 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
716 se->slice_max = max(se->slice_max,
717 se->sum_exec_runtime - se->prev_sum_exec_runtime);
720 se->prev_sum_exec_runtime = se->sum_exec_runtime;
723 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
725 struct sched_entity *se = __pick_next_entity(cfs_rq);
727 set_next_entity(cfs_rq, se);
732 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
735 * If still on the runqueue then deactivate_task()
736 * was not called and update_curr() has to be done:
741 update_stats_curr_end(cfs_rq, prev);
744 update_stats_wait_start(cfs_rq, prev);
748 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
750 struct sched_entity *next;
753 * Dequeue and enqueue the task to update its
754 * position within the tree:
756 dequeue_entity(cfs_rq, curr, 0);
757 enqueue_entity(cfs_rq, curr, 0);
760 * Reschedule if another task tops the current one.
762 next = __pick_next_entity(cfs_rq);
766 __check_preempt_curr_fair(cfs_rq, next, curr,
767 sched_granularity(cfs_rq));
770 /**************************************************
771 * CFS operations on tasks:
774 #ifdef CONFIG_FAIR_GROUP_SCHED
776 /* Walk up scheduling entities hierarchy */
777 #define for_each_sched_entity(se) \
778 for (; se; se = se->parent)
780 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
785 /* runqueue on which this entity is (to be) queued */
786 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
791 /* runqueue "owned" by this group */
792 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
797 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
798 * another cpu ('this_cpu')
800 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
802 /* A later patch will take group into account */
803 return &cpu_rq(this_cpu)->cfs;
806 /* Iterate thr' all leaf cfs_rq's on a runqueue */
807 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
808 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
810 /* Do the two (enqueued) tasks belong to the same group ? */
811 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
813 if (curr->se.cfs_rq == p->se.cfs_rq)
819 #else /* CONFIG_FAIR_GROUP_SCHED */
821 #define for_each_sched_entity(se) \
822 for (; se; se = NULL)
824 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
826 return &task_rq(p)->cfs;
829 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
831 struct task_struct *p = task_of(se);
832 struct rq *rq = task_rq(p);
837 /* runqueue "owned" by this group */
838 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
843 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
845 return &cpu_rq(this_cpu)->cfs;
848 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
849 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
851 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
856 #endif /* CONFIG_FAIR_GROUP_SCHED */
859 * The enqueue_task method is called before nr_running is
860 * increased. Here we update the fair scheduling stats and
861 * then put the task into the rbtree:
863 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
865 struct cfs_rq *cfs_rq;
866 struct sched_entity *se = &p->se;
868 for_each_sched_entity(se) {
871 cfs_rq = cfs_rq_of(se);
872 enqueue_entity(cfs_rq, se, wakeup);
877 * The dequeue_task method is called before nr_running is
878 * decreased. We remove the task from the rbtree and
879 * update the fair scheduling stats:
881 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
883 struct cfs_rq *cfs_rq;
884 struct sched_entity *se = &p->se;
886 for_each_sched_entity(se) {
887 cfs_rq = cfs_rq_of(se);
888 dequeue_entity(cfs_rq, se, sleep);
889 /* Don't dequeue parent if it has other entities besides us */
890 if (cfs_rq->load.weight)
896 * sched_yield() support is very simple - we dequeue and enqueue.
898 * If compat_yield is turned on then we requeue to the end of the tree.
900 static void yield_task_fair(struct rq *rq, struct task_struct *p)
902 struct cfs_rq *cfs_rq = task_cfs_rq(p);
903 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
904 struct sched_entity *rightmost, *se = &p->se;
905 struct rb_node *parent;
908 * Are we the only task in the tree?
910 if (unlikely(cfs_rq->nr_running == 1))
913 if (likely(!sysctl_sched_compat_yield)) {
914 __update_rq_clock(rq);
916 * Dequeue and enqueue the task to update its
917 * position within the tree:
919 dequeue_entity(cfs_rq, &p->se, 0);
920 enqueue_entity(cfs_rq, &p->se, 0);
925 * Find the rightmost entry in the rbtree:
929 link = &parent->rb_right;
932 rightmost = rb_entry(parent, struct sched_entity, run_node);
934 * Already in the rightmost position?
936 if (unlikely(rightmost == se))
940 * Minimally necessary key value to be last in the tree:
942 se->fair_key = rightmost->fair_key + 1;
944 if (cfs_rq->rb_leftmost == &se->run_node)
945 cfs_rq->rb_leftmost = rb_next(&se->run_node);
947 * Relink the task to the rightmost position:
949 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
950 rb_link_node(&se->run_node, parent, link);
951 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
955 * Preempt the current task with a newly woken task if needed:
957 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
959 struct task_struct *curr = rq->curr;
960 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
963 if (unlikely(rt_prio(p->prio))) {
970 gran = sysctl_sched_wakeup_granularity;
972 * Batch tasks prefer throughput over latency:
974 if (unlikely(p->policy == SCHED_BATCH))
975 gran = sysctl_sched_batch_wakeup_granularity;
977 if (is_same_group(curr, p))
978 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
981 static struct task_struct *pick_next_task_fair(struct rq *rq)
983 struct cfs_rq *cfs_rq = &rq->cfs;
984 struct sched_entity *se;
986 if (unlikely(!cfs_rq->nr_running))
990 se = pick_next_entity(cfs_rq);
991 cfs_rq = group_cfs_rq(se);
998 * Account for a descheduled task:
1000 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1002 struct sched_entity *se = &prev->se;
1003 struct cfs_rq *cfs_rq;
1005 for_each_sched_entity(se) {
1006 cfs_rq = cfs_rq_of(se);
1007 put_prev_entity(cfs_rq, se);
1011 /**************************************************
1012 * Fair scheduling class load-balancing methods:
1016 * Load-balancing iterator. Note: while the runqueue stays locked
1017 * during the whole iteration, the current task might be
1018 * dequeued so the iterator has to be dequeue-safe. Here we
1019 * achieve that by always pre-iterating before returning
1022 static inline struct task_struct *
1023 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1025 struct task_struct *p;
1030 p = rb_entry(curr, struct task_struct, se.run_node);
1031 cfs_rq->rb_load_balance_curr = rb_next(curr);
1036 static struct task_struct *load_balance_start_fair(void *arg)
1038 struct cfs_rq *cfs_rq = arg;
1040 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1043 static struct task_struct *load_balance_next_fair(void *arg)
1045 struct cfs_rq *cfs_rq = arg;
1047 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1050 #ifdef CONFIG_FAIR_GROUP_SCHED
1051 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1053 struct sched_entity *curr;
1054 struct task_struct *p;
1056 if (!cfs_rq->nr_running)
1059 curr = __pick_next_entity(cfs_rq);
1066 static unsigned long
1067 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1068 unsigned long max_nr_move, unsigned long max_load_move,
1069 struct sched_domain *sd, enum cpu_idle_type idle,
1070 int *all_pinned, int *this_best_prio)
1072 struct cfs_rq *busy_cfs_rq;
1073 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1074 long rem_load_move = max_load_move;
1075 struct rq_iterator cfs_rq_iterator;
1077 cfs_rq_iterator.start = load_balance_start_fair;
1078 cfs_rq_iterator.next = load_balance_next_fair;
1080 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1081 #ifdef CONFIG_FAIR_GROUP_SCHED
1082 struct cfs_rq *this_cfs_rq;
1084 unsigned long maxload;
1086 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1088 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1089 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1093 /* Don't pull more than imbalance/2 */
1095 maxload = min(rem_load_move, imbalance);
1097 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1099 # define maxload rem_load_move
1101 /* pass busy_cfs_rq argument into
1102 * load_balance_[start|next]_fair iterators
1104 cfs_rq_iterator.arg = busy_cfs_rq;
1105 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1106 max_nr_move, maxload, sd, idle, all_pinned,
1107 &load_moved, this_best_prio, &cfs_rq_iterator);
1109 total_nr_moved += nr_moved;
1110 max_nr_move -= nr_moved;
1111 rem_load_move -= load_moved;
1113 if (max_nr_move <= 0 || rem_load_move <= 0)
1117 return max_load_move - rem_load_move;
1121 * scheduler tick hitting a task of our scheduling class:
1123 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1125 struct cfs_rq *cfs_rq;
1126 struct sched_entity *se = &curr->se;
1128 for_each_sched_entity(se) {
1129 cfs_rq = cfs_rq_of(se);
1130 entity_tick(cfs_rq, se);
1135 * Share the fairness runtime between parent and child, thus the
1136 * total amount of pressure for CPU stays equal - new tasks
1137 * get a chance to run but frequent forkers are not allowed to
1138 * monopolize the CPU. Note: the parent runqueue is locked,
1139 * the child is not running yet.
1141 static void task_new_fair(struct rq *rq, struct task_struct *p)
1143 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1144 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1146 sched_info_queued(p);
1148 update_curr(cfs_rq);
1149 update_stats_enqueue(cfs_rq, se);
1151 * Child runs first: we let it run before the parent
1152 * until it reschedules once. We set up the key so that
1153 * it will preempt the parent:
1155 se->fair_key = curr->fair_key -
1156 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1158 * The first wait is dominated by the child-runs-first logic,
1159 * so do not credit it with that waiting time yet:
1161 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1162 se->wait_start_fair = 0;
1165 * The statistical average of wait_runtime is about
1166 * -granularity/2, so initialize the task with that:
1168 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1169 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1171 __enqueue_entity(cfs_rq, se);
1172 resched_task(rq->curr);
1175 #ifdef CONFIG_FAIR_GROUP_SCHED
1176 /* Account for a task changing its policy or group.
1178 * This routine is mostly called to set cfs_rq->curr field when a task
1179 * migrates between groups/classes.
1181 static void set_curr_task_fair(struct rq *rq)
1183 struct sched_entity *se = &rq->curr->se;
1185 for_each_sched_entity(se)
1186 set_next_entity(cfs_rq_of(se), se);
1189 static void set_curr_task_fair(struct rq *rq)
1195 * All the scheduling class methods:
1197 struct sched_class fair_sched_class __read_mostly = {
1198 .enqueue_task = enqueue_task_fair,
1199 .dequeue_task = dequeue_task_fair,
1200 .yield_task = yield_task_fair,
1202 .check_preempt_curr = check_preempt_curr_fair,
1204 .pick_next_task = pick_next_task_fair,
1205 .put_prev_task = put_prev_task_fair,
1207 .load_balance = load_balance_fair,
1209 .set_curr_task = set_curr_task_fair,
1210 .task_tick = task_tick_fair,
1211 .task_new = task_new_fair,
1214 #ifdef CONFIG_SCHED_DEBUG
1215 static void print_cfs_stats(struct seq_file *m, int cpu)
1217 struct cfs_rq *cfs_rq;
1219 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1220 print_cfs_rq(m, cpu, cfs_rq);