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 * sys_sched_yield() compat mode
48 * This option switches the agressive yield implementation of the
49 * old scheduler back on.
51 unsigned int __read_mostly sysctl_sched_compat_yield;
54 * SCHED_BATCH wake-up granularity.
55 * (default: 25 msec, units: nanoseconds)
57 * This option delays the preemption effects of decoupled workloads
58 * and reduces their over-scheduling. Synchronous workloads will still
59 * have immediate wakeup/sleep latencies.
61 unsigned int sysctl_sched_batch_wakeup_granularity __read_mostly = 25000000UL;
64 * SCHED_OTHER wake-up granularity.
65 * (default: 1 msec, units: nanoseconds)
67 * This option delays the preemption effects of decoupled workloads
68 * and reduces their over-scheduling. Synchronous workloads will still
69 * have immediate wakeup/sleep latencies.
71 unsigned int sysctl_sched_wakeup_granularity __read_mostly = 1000000UL;
73 unsigned int sysctl_sched_stat_granularity __read_mostly;
76 * Initialized in sched_init_granularity() [to 5 times the base granularity]:
78 unsigned int sysctl_sched_runtime_limit __read_mostly;
81 * Debugging: various feature bits
84 SCHED_FEAT_FAIR_SLEEPERS = 1,
85 SCHED_FEAT_SLEEPER_AVG = 2,
86 SCHED_FEAT_SLEEPER_LOAD_AVG = 4,
87 SCHED_FEAT_PRECISE_CPU_LOAD = 8,
88 SCHED_FEAT_START_DEBIT = 16,
89 SCHED_FEAT_SKIP_INITIAL = 32,
92 unsigned int sysctl_sched_features __read_mostly =
93 SCHED_FEAT_FAIR_SLEEPERS *1 |
94 SCHED_FEAT_SLEEPER_AVG *0 |
95 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
96 SCHED_FEAT_PRECISE_CPU_LOAD *1 |
97 SCHED_FEAT_START_DEBIT *1 |
98 SCHED_FEAT_SKIP_INITIAL *0;
100 extern struct sched_class fair_sched_class;
102 /**************************************************************
103 * CFS operations on generic schedulable entities:
106 #ifdef CONFIG_FAIR_GROUP_SCHED
108 /* cpu runqueue to which this cfs_rq is attached */
109 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
114 /* currently running entity (if any) on this cfs_rq */
115 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
120 /* An entity is a task if it doesn't "own" a runqueue */
121 #define entity_is_task(se) (!se->my_q)
124 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
129 #else /* CONFIG_FAIR_GROUP_SCHED */
131 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
133 return container_of(cfs_rq, struct rq, cfs);
136 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
138 struct rq *rq = rq_of(cfs_rq);
140 if (unlikely(rq->curr->sched_class != &fair_sched_class))
143 return &rq->curr->se;
146 #define entity_is_task(se) 1
149 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
151 #endif /* CONFIG_FAIR_GROUP_SCHED */
153 static inline struct task_struct *task_of(struct sched_entity *se)
155 return container_of(se, struct task_struct, se);
159 /**************************************************************
160 * Scheduling class tree data structure manipulation methods:
164 * Enqueue an entity into the rb-tree:
167 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
169 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
170 struct rb_node *parent = NULL;
171 struct sched_entity *entry;
172 s64 key = se->fair_key;
176 * Find the right place in the rbtree:
180 entry = rb_entry(parent, struct sched_entity, run_node);
182 * We dont care about collisions. Nodes with
183 * the same key stay together.
185 if (key - entry->fair_key < 0) {
186 link = &parent->rb_left;
188 link = &parent->rb_right;
194 * Maintain a cache of leftmost tree entries (it is frequently
198 cfs_rq->rb_leftmost = &se->run_node;
200 rb_link_node(&se->run_node, parent, link);
201 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
202 update_load_add(&cfs_rq->load, se->load.weight);
203 cfs_rq->nr_running++;
206 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
210 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
212 if (cfs_rq->rb_leftmost == &se->run_node)
213 cfs_rq->rb_leftmost = rb_next(&se->run_node);
214 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
215 update_load_sub(&cfs_rq->load, se->load.weight);
216 cfs_rq->nr_running--;
219 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
222 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
224 return cfs_rq->rb_leftmost;
227 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
229 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
232 /**************************************************************
233 * Scheduling class statistics methods:
237 * Calculate the preemption granularity needed to schedule every
238 * runnable task once per sysctl_sched_latency amount of time.
239 * (down to a sensible low limit on granularity)
241 * For example, if there are 2 tasks running and latency is 10 msecs,
242 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
243 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
244 * for each task. We do finer and finer scheduling up to until we
245 * reach the minimum granularity value.
247 * To achieve this we use the following dynamic-granularity rule:
249 * gran = lat/nr - lat/nr/nr
251 * This comes out of the following equations:
256 * kB2 = kB1 - d + d/nr
259 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
260 * '1' is start of time, '2' is end of time, 'd' is delay between
261 * 1 and 2 (during which task B was running), 'nr' is number of tasks
262 * running, 'lat' is the the period of each task. ('lat' is the
263 * sched_latency that we aim for.)
266 sched_granularity(struct cfs_rq *cfs_rq)
268 unsigned int gran = sysctl_sched_latency;
269 unsigned int nr = cfs_rq->nr_running;
272 gran = gran/nr - gran/nr/nr;
273 gran = max(gran, sysctl_sched_min_granularity);
280 * We rescale the rescheduling granularity of tasks according to their
281 * nice level, but only linearly, not exponentially:
284 niced_granularity(struct sched_entity *curr, unsigned long granularity)
288 if (likely(curr->load.weight == NICE_0_LOAD))
291 * Positive nice levels get the same granularity as nice-0:
293 if (likely(curr->load.weight < NICE_0_LOAD)) {
294 tmp = curr->load.weight * (u64)granularity;
295 return (long) (tmp >> NICE_0_SHIFT);
298 * Negative nice level tasks get linearly finer
301 tmp = curr->load.inv_weight * (u64)granularity;
304 * It will always fit into 'long':
306 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
310 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
312 long limit = sysctl_sched_runtime_limit;
315 * Niced tasks have the same history dynamic range as
318 if (unlikely(se->wait_runtime > limit)) {
319 se->wait_runtime = limit;
320 schedstat_inc(se, wait_runtime_overruns);
321 schedstat_inc(cfs_rq, wait_runtime_overruns);
323 if (unlikely(se->wait_runtime < -limit)) {
324 se->wait_runtime = -limit;
325 schedstat_inc(se, wait_runtime_underruns);
326 schedstat_inc(cfs_rq, wait_runtime_underruns);
331 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
333 se->wait_runtime += delta;
334 schedstat_add(se, sum_wait_runtime, delta);
335 limit_wait_runtime(cfs_rq, se);
339 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
341 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
342 __add_wait_runtime(cfs_rq, se, delta);
343 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
347 * Update the current task's runtime statistics. Skip current tasks that
348 * are not in our scheduling class.
351 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr)
353 unsigned long delta, delta_exec, delta_fair, delta_mine;
354 struct load_weight *lw = &cfs_rq->load;
355 unsigned long load = lw->weight;
357 delta_exec = curr->delta_exec;
358 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
360 curr->sum_exec_runtime += delta_exec;
361 cfs_rq->exec_clock += delta_exec;
366 delta_fair = calc_delta_fair(delta_exec, lw);
367 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
369 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
370 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
371 delta = min(delta, (unsigned long)(
372 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
373 cfs_rq->sleeper_bonus -= delta;
377 cfs_rq->fair_clock += delta_fair;
379 * We executed delta_exec amount of time on the CPU,
380 * but we were only entitled to delta_mine amount of
381 * time during that period (if nr_running == 1 then
382 * the two values are equal)
383 * [Note: delta_mine - delta_exec is negative]:
385 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
388 static void update_curr(struct cfs_rq *cfs_rq)
390 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
391 unsigned long delta_exec;
397 * Get the amount of time the current task was running
398 * since the last time we changed load (this cannot
399 * overflow on 32 bits):
401 delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start);
403 curr->delta_exec += delta_exec;
405 if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
406 __update_curr(cfs_rq, curr);
407 curr->delta_exec = 0;
409 curr->exec_start = rq_of(cfs_rq)->clock;
413 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
415 se->wait_start_fair = cfs_rq->fair_clock;
416 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
420 * We calculate fair deltas here, so protect against the random effects
421 * of a multiplication overflow by capping it to the runtime limit:
423 #if BITS_PER_LONG == 32
424 static inline unsigned long
425 calc_weighted(unsigned long delta, unsigned long weight, int shift)
427 u64 tmp = (u64)delta * weight >> shift;
429 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
430 return sysctl_sched_runtime_limit*2;
434 static inline unsigned long
435 calc_weighted(unsigned long delta, unsigned long weight, int shift)
437 return delta * weight >> shift;
442 * Task is being enqueued - update stats:
444 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
449 * Are we enqueueing a waiting task? (for current tasks
450 * a dequeue/enqueue event is a NOP)
452 if (se != cfs_rq_curr(cfs_rq))
453 update_stats_wait_start(cfs_rq, se);
457 key = cfs_rq->fair_clock;
460 * Optimize the common nice 0 case:
462 if (likely(se->load.weight == NICE_0_LOAD)) {
463 key -= se->wait_runtime;
467 if (se->wait_runtime < 0) {
468 tmp = -se->wait_runtime;
469 key += (tmp * se->load.inv_weight) >>
470 (WMULT_SHIFT - NICE_0_SHIFT);
472 tmp = se->wait_runtime;
473 key -= (tmp * se->load.inv_weight) >>
474 (WMULT_SHIFT - NICE_0_SHIFT);
482 * Note: must be called with a freshly updated rq->fair_clock.
485 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
487 unsigned long delta_fair = se->delta_fair_run;
489 schedstat_set(se->wait_max, max(se->wait_max,
490 rq_of(cfs_rq)->clock - se->wait_start));
492 if (unlikely(se->load.weight != NICE_0_LOAD))
493 delta_fair = calc_weighted(delta_fair, se->load.weight,
496 add_wait_runtime(cfs_rq, se, delta_fair);
500 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
502 unsigned long delta_fair;
504 if (unlikely(!se->wait_start_fair))
507 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
508 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
510 se->delta_fair_run += delta_fair;
511 if (unlikely(abs(se->delta_fair_run) >=
512 sysctl_sched_stat_granularity)) {
513 __update_stats_wait_end(cfs_rq, se);
514 se->delta_fair_run = 0;
517 se->wait_start_fair = 0;
518 schedstat_set(se->wait_start, 0);
522 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
526 * Mark the end of the wait period if dequeueing a
529 if (se != cfs_rq_curr(cfs_rq))
530 update_stats_wait_end(cfs_rq, se);
534 * We are picking a new current task - update its stats:
537 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
540 * We are starting a new run period:
542 se->exec_start = rq_of(cfs_rq)->clock;
546 * We are descheduling a task - update its stats:
549 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
554 /**************************************************
555 * Scheduling class queueing methods:
558 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
560 unsigned long load = cfs_rq->load.weight, delta_fair;
564 * Do not boost sleepers if there's too much bonus 'in flight'
567 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
570 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
571 load = rq_of(cfs_rq)->cpu_load[2];
573 delta_fair = se->delta_fair_sleep;
576 * Fix up delta_fair with the effect of us running
577 * during the whole sleep period:
579 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
580 delta_fair = div64_likely32((u64)delta_fair * load,
581 load + se->load.weight);
583 if (unlikely(se->load.weight != NICE_0_LOAD))
584 delta_fair = calc_weighted(delta_fair, se->load.weight,
587 prev_runtime = se->wait_runtime;
588 __add_wait_runtime(cfs_rq, se, delta_fair);
589 delta_fair = se->wait_runtime - prev_runtime;
592 * Track the amount of bonus we've given to sleepers:
594 cfs_rq->sleeper_bonus += delta_fair;
597 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
599 struct task_struct *tsk = task_of(se);
600 unsigned long delta_fair;
602 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
603 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
606 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
607 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
609 se->delta_fair_sleep += delta_fair;
610 if (unlikely(abs(se->delta_fair_sleep) >=
611 sysctl_sched_stat_granularity)) {
612 __enqueue_sleeper(cfs_rq, se);
613 se->delta_fair_sleep = 0;
616 se->sleep_start_fair = 0;
618 #ifdef CONFIG_SCHEDSTATS
619 if (se->sleep_start) {
620 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
625 if (unlikely(delta > se->sleep_max))
626 se->sleep_max = delta;
629 se->sum_sleep_runtime += delta;
631 if (se->block_start) {
632 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
637 if (unlikely(delta > se->block_max))
638 se->block_max = delta;
641 se->sum_sleep_runtime += delta;
647 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
650 * Update the fair clock.
655 enqueue_sleeper(cfs_rq, se);
657 update_stats_enqueue(cfs_rq, se);
658 __enqueue_entity(cfs_rq, se);
662 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
664 update_stats_dequeue(cfs_rq, se);
666 se->sleep_start_fair = cfs_rq->fair_clock;
667 #ifdef CONFIG_SCHEDSTATS
668 if (entity_is_task(se)) {
669 struct task_struct *tsk = task_of(se);
671 if (tsk->state & TASK_INTERRUPTIBLE)
672 se->sleep_start = rq_of(cfs_rq)->clock;
673 if (tsk->state & TASK_UNINTERRUPTIBLE)
674 se->block_start = rq_of(cfs_rq)->clock;
678 __dequeue_entity(cfs_rq, se);
682 * Preempt the current task with a newly woken task if needed:
685 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
686 struct sched_entity *curr, unsigned long granularity)
688 s64 __delta = curr->fair_key - se->fair_key;
689 unsigned long ideal_runtime, delta_exec;
692 * ideal_runtime is compared against sum_exec_runtime, which is
693 * walltime, hence do not scale.
695 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
696 (unsigned long)sysctl_sched_min_granularity);
699 * If we executed more than what the latency constraint suggests,
700 * reduce the rescheduling granularity. This way the total latency
701 * of how much a task is not scheduled converges to
702 * sysctl_sched_latency:
704 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
705 if (delta_exec > ideal_runtime)
709 * Take scheduling granularity into account - do not
710 * preempt the current task unless the best task has
711 * a larger than sched_granularity fairness advantage:
713 * scale granularity as key space is in fair_clock.
715 if (__delta > niced_granularity(curr, granularity))
716 resched_task(rq_of(cfs_rq)->curr);
720 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
723 * Any task has to be enqueued before it get to execute on
724 * a CPU. So account for the time it spent waiting on the
725 * runqueue. (note, here we rely on pick_next_task() having
726 * done a put_prev_task_fair() shortly before this, which
727 * updated rq->fair_clock - used by update_stats_wait_end())
729 update_stats_wait_end(cfs_rq, se);
730 update_stats_curr_start(cfs_rq, se);
731 set_cfs_rq_curr(cfs_rq, se);
732 se->prev_sum_exec_runtime = se->sum_exec_runtime;
735 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
737 struct sched_entity *se = __pick_next_entity(cfs_rq);
739 set_next_entity(cfs_rq, se);
744 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
747 * If still on the runqueue then deactivate_task()
748 * was not called and update_curr() has to be done:
753 update_stats_curr_end(cfs_rq, prev);
756 update_stats_wait_start(cfs_rq, prev);
757 set_cfs_rq_curr(cfs_rq, NULL);
760 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
762 struct sched_entity *next;
765 * Dequeue and enqueue the task to update its
766 * position within the tree:
768 dequeue_entity(cfs_rq, curr, 0);
769 enqueue_entity(cfs_rq, curr, 0);
772 * Reschedule if another task tops the current one.
774 next = __pick_next_entity(cfs_rq);
778 __check_preempt_curr_fair(cfs_rq, next, curr,
779 sched_granularity(cfs_rq));
782 /**************************************************
783 * CFS operations on tasks:
786 #ifdef CONFIG_FAIR_GROUP_SCHED
788 /* Walk up scheduling entities hierarchy */
789 #define for_each_sched_entity(se) \
790 for (; se; se = se->parent)
792 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
797 /* runqueue on which this entity is (to be) queued */
798 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
803 /* runqueue "owned" by this group */
804 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
809 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
810 * another cpu ('this_cpu')
812 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
814 /* A later patch will take group into account */
815 return &cpu_rq(this_cpu)->cfs;
818 /* Iterate thr' all leaf cfs_rq's on a runqueue */
819 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
820 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
822 /* Do the two (enqueued) tasks belong to the same group ? */
823 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
825 if (curr->se.cfs_rq == p->se.cfs_rq)
831 #else /* CONFIG_FAIR_GROUP_SCHED */
833 #define for_each_sched_entity(se) \
834 for (; se; se = NULL)
836 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
838 return &task_rq(p)->cfs;
841 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
843 struct task_struct *p = task_of(se);
844 struct rq *rq = task_rq(p);
849 /* runqueue "owned" by this group */
850 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
855 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
857 return &cpu_rq(this_cpu)->cfs;
860 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
861 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
863 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
868 #endif /* CONFIG_FAIR_GROUP_SCHED */
871 * The enqueue_task method is called before nr_running is
872 * increased. Here we update the fair scheduling stats and
873 * then put the task into the rbtree:
875 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
877 struct cfs_rq *cfs_rq;
878 struct sched_entity *se = &p->se;
880 for_each_sched_entity(se) {
883 cfs_rq = cfs_rq_of(se);
884 enqueue_entity(cfs_rq, se, wakeup);
889 * The dequeue_task method is called before nr_running is
890 * decreased. We remove the task from the rbtree and
891 * update the fair scheduling stats:
893 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
895 struct cfs_rq *cfs_rq;
896 struct sched_entity *se = &p->se;
898 for_each_sched_entity(se) {
899 cfs_rq = cfs_rq_of(se);
900 dequeue_entity(cfs_rq, se, sleep);
901 /* Don't dequeue parent if it has other entities besides us */
902 if (cfs_rq->load.weight)
908 * sched_yield() support is very simple - we dequeue and enqueue.
910 * If compat_yield is turned on then we requeue to the end of the tree.
912 static void yield_task_fair(struct rq *rq, struct task_struct *p)
914 struct cfs_rq *cfs_rq = task_cfs_rq(p);
915 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
916 struct sched_entity *rightmost, *se = &p->se;
917 struct rb_node *parent;
920 * Are we the only task in the tree?
922 if (unlikely(cfs_rq->nr_running == 1))
925 if (likely(!sysctl_sched_compat_yield)) {
926 __update_rq_clock(rq);
928 * Dequeue and enqueue the task to update its
929 * position within the tree:
931 dequeue_entity(cfs_rq, &p->se, 0);
932 enqueue_entity(cfs_rq, &p->se, 0);
937 * Find the rightmost entry in the rbtree:
941 link = &parent->rb_right;
944 rightmost = rb_entry(parent, struct sched_entity, run_node);
946 * Already in the rightmost position?
948 if (unlikely(rightmost == se))
952 * Minimally necessary key value to be last in the tree:
954 se->fair_key = rightmost->fair_key + 1;
956 if (cfs_rq->rb_leftmost == &se->run_node)
957 cfs_rq->rb_leftmost = rb_next(&se->run_node);
959 * Relink the task to the rightmost position:
961 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
962 rb_link_node(&se->run_node, parent, link);
963 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
967 * Preempt the current task with a newly woken task if needed:
969 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
971 struct task_struct *curr = rq->curr;
972 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
975 if (unlikely(rt_prio(p->prio))) {
982 gran = sysctl_sched_wakeup_granularity;
984 * Batch tasks prefer throughput over latency:
986 if (unlikely(p->policy == SCHED_BATCH))
987 gran = sysctl_sched_batch_wakeup_granularity;
989 if (is_same_group(curr, p))
990 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
993 static struct task_struct *pick_next_task_fair(struct rq *rq)
995 struct cfs_rq *cfs_rq = &rq->cfs;
996 struct sched_entity *se;
998 if (unlikely(!cfs_rq->nr_running))
1002 se = pick_next_entity(cfs_rq);
1003 cfs_rq = group_cfs_rq(se);
1010 * Account for a descheduled task:
1012 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1014 struct sched_entity *se = &prev->se;
1015 struct cfs_rq *cfs_rq;
1017 for_each_sched_entity(se) {
1018 cfs_rq = cfs_rq_of(se);
1019 put_prev_entity(cfs_rq, se);
1023 /**************************************************
1024 * Fair scheduling class load-balancing methods:
1028 * Load-balancing iterator. Note: while the runqueue stays locked
1029 * during the whole iteration, the current task might be
1030 * dequeued so the iterator has to be dequeue-safe. Here we
1031 * achieve that by always pre-iterating before returning
1034 static inline struct task_struct *
1035 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1037 struct task_struct *p;
1042 p = rb_entry(curr, struct task_struct, se.run_node);
1043 cfs_rq->rb_load_balance_curr = rb_next(curr);
1048 static struct task_struct *load_balance_start_fair(void *arg)
1050 struct cfs_rq *cfs_rq = arg;
1052 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1055 static struct task_struct *load_balance_next_fair(void *arg)
1057 struct cfs_rq *cfs_rq = arg;
1059 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1062 #ifdef CONFIG_FAIR_GROUP_SCHED
1063 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1065 struct sched_entity *curr;
1066 struct task_struct *p;
1068 if (!cfs_rq->nr_running)
1071 curr = __pick_next_entity(cfs_rq);
1078 static unsigned long
1079 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1080 unsigned long max_nr_move, unsigned long max_load_move,
1081 struct sched_domain *sd, enum cpu_idle_type idle,
1082 int *all_pinned, int *this_best_prio)
1084 struct cfs_rq *busy_cfs_rq;
1085 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1086 long rem_load_move = max_load_move;
1087 struct rq_iterator cfs_rq_iterator;
1089 cfs_rq_iterator.start = load_balance_start_fair;
1090 cfs_rq_iterator.next = load_balance_next_fair;
1092 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1093 #ifdef CONFIG_FAIR_GROUP_SCHED
1094 struct cfs_rq *this_cfs_rq;
1096 unsigned long maxload;
1098 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1100 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1101 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1105 /* Don't pull more than imbalance/2 */
1107 maxload = min(rem_load_move, imbalance);
1109 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1111 # define maxload rem_load_move
1113 /* pass busy_cfs_rq argument into
1114 * load_balance_[start|next]_fair iterators
1116 cfs_rq_iterator.arg = busy_cfs_rq;
1117 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1118 max_nr_move, maxload, sd, idle, all_pinned,
1119 &load_moved, this_best_prio, &cfs_rq_iterator);
1121 total_nr_moved += nr_moved;
1122 max_nr_move -= nr_moved;
1123 rem_load_move -= load_moved;
1125 if (max_nr_move <= 0 || rem_load_move <= 0)
1129 return max_load_move - rem_load_move;
1133 * scheduler tick hitting a task of our scheduling class:
1135 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1137 struct cfs_rq *cfs_rq;
1138 struct sched_entity *se = &curr->se;
1140 for_each_sched_entity(se) {
1141 cfs_rq = cfs_rq_of(se);
1142 entity_tick(cfs_rq, se);
1147 * Share the fairness runtime between parent and child, thus the
1148 * total amount of pressure for CPU stays equal - new tasks
1149 * get a chance to run but frequent forkers are not allowed to
1150 * monopolize the CPU. Note: the parent runqueue is locked,
1151 * the child is not running yet.
1153 static void task_new_fair(struct rq *rq, struct task_struct *p)
1155 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1156 struct sched_entity *se = &p->se, *curr = cfs_rq_curr(cfs_rq);
1158 sched_info_queued(p);
1160 update_curr(cfs_rq);
1161 update_stats_enqueue(cfs_rq, se);
1163 * Child runs first: we let it run before the parent
1164 * until it reschedules once. We set up the key so that
1165 * it will preempt the parent:
1167 se->fair_key = curr->fair_key -
1168 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1170 * The first wait is dominated by the child-runs-first logic,
1171 * so do not credit it with that waiting time yet:
1173 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1174 se->wait_start_fair = 0;
1177 * The statistical average of wait_runtime is about
1178 * -granularity/2, so initialize the task with that:
1180 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1181 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1183 __enqueue_entity(cfs_rq, se);
1186 #ifdef CONFIG_FAIR_GROUP_SCHED
1187 /* Account for a task changing its policy or group.
1189 * This routine is mostly called to set cfs_rq->curr field when a task
1190 * migrates between groups/classes.
1192 static void set_curr_task_fair(struct rq *rq)
1194 struct sched_entity *se = &rq->curr->se;
1196 for_each_sched_entity(se)
1197 set_next_entity(cfs_rq_of(se), se);
1200 static void set_curr_task_fair(struct rq *rq)
1206 * All the scheduling class methods:
1208 struct sched_class fair_sched_class __read_mostly = {
1209 .enqueue_task = enqueue_task_fair,
1210 .dequeue_task = dequeue_task_fair,
1211 .yield_task = yield_task_fair,
1213 .check_preempt_curr = check_preempt_curr_fair,
1215 .pick_next_task = pick_next_task_fair,
1216 .put_prev_task = put_prev_task_fair,
1218 .load_balance = load_balance_fair,
1220 .set_curr_task = set_curr_task_fair,
1221 .task_tick = task_tick_fair,
1222 .task_new = task_new_fair,
1225 #ifdef CONFIG_SCHED_DEBUG
1226 static void print_cfs_stats(struct seq_file *m, int cpu)
1228 struct cfs_rq *cfs_rq;
1230 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1231 print_cfs_rq(m, cpu, cfs_rq);