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++;
198 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
202 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
204 if (cfs_rq->rb_leftmost == &se->run_node)
205 cfs_rq->rb_leftmost = rb_next(&se->run_node);
206 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
207 update_load_sub(&cfs_rq->load, se->load.weight);
208 cfs_rq->nr_running--;
211 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
214 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
216 return cfs_rq->rb_leftmost;
219 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
221 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
224 /**************************************************************
225 * Scheduling class statistics methods:
229 * Calculate the preemption granularity needed to schedule every
230 * runnable task once per sysctl_sched_latency amount of time.
231 * (down to a sensible low limit on granularity)
233 * For example, if there are 2 tasks running and latency is 10 msecs,
234 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
235 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
236 * for each task. We do finer and finer scheduling up to until we
237 * reach the minimum granularity value.
239 * To achieve this we use the following dynamic-granularity rule:
241 * gran = lat/nr - lat/nr/nr
243 * This comes out of the following equations:
248 * kB2 = kB1 - d + d/nr
251 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
252 * '1' is start of time, '2' is end of time, 'd' is delay between
253 * 1 and 2 (during which task B was running), 'nr' is number of tasks
254 * running, 'lat' is the the period of each task. ('lat' is the
255 * sched_latency that we aim for.)
258 sched_granularity(struct cfs_rq *cfs_rq)
260 unsigned int gran = sysctl_sched_latency;
261 unsigned int nr = cfs_rq->nr_running;
264 gran = gran/nr - gran/nr/nr;
265 gran = max(gran, sysctl_sched_min_granularity);
272 * We rescale the rescheduling granularity of tasks according to their
273 * nice level, but only linearly, not exponentially:
276 niced_granularity(struct sched_entity *curr, unsigned long granularity)
280 if (likely(curr->load.weight == NICE_0_LOAD))
283 * Positive nice levels get the same granularity as nice-0:
285 if (likely(curr->load.weight < NICE_0_LOAD)) {
286 tmp = curr->load.weight * (u64)granularity;
287 return (long) (tmp >> NICE_0_SHIFT);
290 * Negative nice level tasks get linearly finer
293 tmp = curr->load.inv_weight * (u64)granularity;
296 * It will always fit into 'long':
298 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
302 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
304 long limit = sysctl_sched_runtime_limit;
307 * Niced tasks have the same history dynamic range as
310 if (unlikely(se->wait_runtime > limit)) {
311 se->wait_runtime = limit;
312 schedstat_inc(se, wait_runtime_overruns);
313 schedstat_inc(cfs_rq, wait_runtime_overruns);
315 if (unlikely(se->wait_runtime < -limit)) {
316 se->wait_runtime = -limit;
317 schedstat_inc(se, wait_runtime_underruns);
318 schedstat_inc(cfs_rq, wait_runtime_underruns);
323 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
325 se->wait_runtime += delta;
326 schedstat_add(se, sum_wait_runtime, delta);
327 limit_wait_runtime(cfs_rq, se);
331 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
333 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
334 __add_wait_runtime(cfs_rq, se, delta);
335 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
339 * Update the current task's runtime statistics. Skip current tasks that
340 * are not in our scheduling class.
343 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr)
345 unsigned long delta, delta_exec, delta_fair, delta_mine;
346 struct load_weight *lw = &cfs_rq->load;
347 unsigned long load = lw->weight;
349 delta_exec = curr->delta_exec;
350 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
352 curr->sum_exec_runtime += delta_exec;
353 cfs_rq->exec_clock += delta_exec;
358 delta_fair = calc_delta_fair(delta_exec, lw);
359 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
361 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
362 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
363 delta = min(delta, (unsigned long)(
364 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
365 cfs_rq->sleeper_bonus -= delta;
369 cfs_rq->fair_clock += delta_fair;
371 * We executed delta_exec amount of time on the CPU,
372 * but we were only entitled to delta_mine amount of
373 * time during that period (if nr_running == 1 then
374 * the two values are equal)
375 * [Note: delta_mine - delta_exec is negative]:
377 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
380 static void update_curr(struct cfs_rq *cfs_rq)
382 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
383 unsigned long delta_exec;
389 * Get the amount of time the current task was running
390 * since the last time we changed load (this cannot
391 * overflow on 32 bits):
393 delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start);
395 curr->delta_exec += delta_exec;
397 if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
398 __update_curr(cfs_rq, curr);
399 curr->delta_exec = 0;
401 curr->exec_start = rq_of(cfs_rq)->clock;
405 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
407 se->wait_start_fair = cfs_rq->fair_clock;
408 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
412 * We calculate fair deltas here, so protect against the random effects
413 * of a multiplication overflow by capping it to the runtime limit:
415 #if BITS_PER_LONG == 32
416 static inline unsigned long
417 calc_weighted(unsigned long delta, unsigned long weight, int shift)
419 u64 tmp = (u64)delta * weight >> shift;
421 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
422 return sysctl_sched_runtime_limit*2;
426 static inline unsigned long
427 calc_weighted(unsigned long delta, unsigned long weight, int shift)
429 return delta * weight >> shift;
434 * Task is being enqueued - update stats:
436 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
441 * Are we enqueueing a waiting task? (for current tasks
442 * a dequeue/enqueue event is a NOP)
444 if (se != cfs_rq_curr(cfs_rq))
445 update_stats_wait_start(cfs_rq, se);
449 key = cfs_rq->fair_clock;
452 * Optimize the common nice 0 case:
454 if (likely(se->load.weight == NICE_0_LOAD)) {
455 key -= se->wait_runtime;
459 if (se->wait_runtime < 0) {
460 tmp = -se->wait_runtime;
461 key += (tmp * se->load.inv_weight) >>
462 (WMULT_SHIFT - NICE_0_SHIFT);
464 tmp = se->wait_runtime;
465 key -= (tmp * se->load.inv_weight) >>
466 (WMULT_SHIFT - NICE_0_SHIFT);
474 * Note: must be called with a freshly updated rq->fair_clock.
477 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
479 unsigned long delta_fair = se->delta_fair_run;
481 schedstat_set(se->wait_max, max(se->wait_max,
482 rq_of(cfs_rq)->clock - se->wait_start));
484 if (unlikely(se->load.weight != NICE_0_LOAD))
485 delta_fair = calc_weighted(delta_fair, se->load.weight,
488 add_wait_runtime(cfs_rq, se, delta_fair);
492 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
494 unsigned long delta_fair;
496 if (unlikely(!se->wait_start_fair))
499 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
500 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
502 se->delta_fair_run += delta_fair;
503 if (unlikely(abs(se->delta_fair_run) >=
504 sysctl_sched_stat_granularity)) {
505 __update_stats_wait_end(cfs_rq, se);
506 se->delta_fair_run = 0;
509 se->wait_start_fair = 0;
510 schedstat_set(se->wait_start, 0);
514 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
518 * Mark the end of the wait period if dequeueing a
521 if (se != cfs_rq_curr(cfs_rq))
522 update_stats_wait_end(cfs_rq, se);
526 * We are picking a new current task - update its stats:
529 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
532 * We are starting a new run period:
534 se->exec_start = rq_of(cfs_rq)->clock;
538 * We are descheduling a task - update its stats:
541 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
546 /**************************************************
547 * Scheduling class queueing methods:
550 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
552 unsigned long load = cfs_rq->load.weight, delta_fair;
556 * Do not boost sleepers if there's too much bonus 'in flight'
559 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
562 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
563 load = rq_of(cfs_rq)->cpu_load[2];
565 delta_fair = se->delta_fair_sleep;
568 * Fix up delta_fair with the effect of us running
569 * during the whole sleep period:
571 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
572 delta_fair = div64_likely32((u64)delta_fair * load,
573 load + se->load.weight);
575 if (unlikely(se->load.weight != NICE_0_LOAD))
576 delta_fair = calc_weighted(delta_fair, se->load.weight,
579 prev_runtime = se->wait_runtime;
580 __add_wait_runtime(cfs_rq, se, delta_fair);
581 delta_fair = se->wait_runtime - prev_runtime;
584 * Track the amount of bonus we've given to sleepers:
586 cfs_rq->sleeper_bonus += delta_fair;
589 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
591 struct task_struct *tsk = task_of(se);
592 unsigned long delta_fair;
594 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
595 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
598 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
599 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
601 se->delta_fair_sleep += delta_fair;
602 if (unlikely(abs(se->delta_fair_sleep) >=
603 sysctl_sched_stat_granularity)) {
604 __enqueue_sleeper(cfs_rq, se);
605 se->delta_fair_sleep = 0;
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;
639 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
642 * Update the fair clock.
647 enqueue_sleeper(cfs_rq, se);
649 update_stats_enqueue(cfs_rq, se);
650 __enqueue_entity(cfs_rq, se);
654 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
656 update_stats_dequeue(cfs_rq, se);
658 se->sleep_start_fair = cfs_rq->fair_clock;
659 #ifdef CONFIG_SCHEDSTATS
660 if (entity_is_task(se)) {
661 struct task_struct *tsk = task_of(se);
663 if (tsk->state & TASK_INTERRUPTIBLE)
664 se->sleep_start = rq_of(cfs_rq)->clock;
665 if (tsk->state & TASK_UNINTERRUPTIBLE)
666 se->block_start = rq_of(cfs_rq)->clock;
670 __dequeue_entity(cfs_rq, se);
674 * Preempt the current task with a newly woken task if needed:
677 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
678 struct sched_entity *curr, unsigned long granularity)
680 s64 __delta = curr->fair_key - se->fair_key;
683 * Take scheduling granularity into account - do not
684 * preempt the current task unless the best task has
685 * a larger than sched_granularity fairness advantage:
687 if (__delta > niced_granularity(curr, granularity)) {
688 resched_task(rq_of(cfs_rq)->curr);
695 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
698 * Any task has to be enqueued before it get to execute on
699 * a CPU. So account for the time it spent waiting on the
700 * runqueue. (note, here we rely on pick_next_task() having
701 * done a put_prev_task_fair() shortly before this, which
702 * updated rq->fair_clock - used by update_stats_wait_end())
704 update_stats_wait_end(cfs_rq, se);
705 update_stats_curr_start(cfs_rq, se);
706 set_cfs_rq_curr(cfs_rq, se);
709 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
711 struct sched_entity *se = __pick_next_entity(cfs_rq);
713 set_next_entity(cfs_rq, se);
718 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
721 * If still on the runqueue then deactivate_task()
722 * was not called and update_curr() has to be done:
727 update_stats_curr_end(cfs_rq, prev);
730 update_stats_wait_start(cfs_rq, prev);
731 set_cfs_rq_curr(cfs_rq, NULL);
734 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
736 unsigned long gran, ideal_runtime, delta_exec;
737 struct sched_entity *next;
740 * Dequeue and enqueue the task to update its
741 * position within the tree:
743 dequeue_entity(cfs_rq, curr, 0);
744 enqueue_entity(cfs_rq, curr, 0);
747 * Reschedule if another task tops the current one.
749 next = __pick_next_entity(cfs_rq);
753 gran = sched_granularity(cfs_rq);
754 ideal_runtime = niced_granularity(curr,
755 max(sysctl_sched_latency / cfs_rq->nr_running,
756 (unsigned long)sysctl_sched_min_granularity));
758 * If we executed more than what the latency constraint suggests,
759 * reduce the rescheduling granularity. This way the total latency
760 * of how much a task is not scheduled converges to
761 * sysctl_sched_latency:
763 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
764 if (delta_exec > ideal_runtime)
767 if (__check_preempt_curr_fair(cfs_rq, next, curr, gran))
768 curr->prev_sum_exec_runtime = curr->sum_exec_runtime;
771 /**************************************************
772 * CFS operations on tasks:
775 #ifdef CONFIG_FAIR_GROUP_SCHED
777 /* Walk up scheduling entities hierarchy */
778 #define for_each_sched_entity(se) \
779 for (; se; se = se->parent)
781 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
786 /* runqueue on which this entity is (to be) queued */
787 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
792 /* runqueue "owned" by this group */
793 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
798 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
799 * another cpu ('this_cpu')
801 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
803 /* A later patch will take group into account */
804 return &cpu_rq(this_cpu)->cfs;
807 /* Iterate thr' all leaf cfs_rq's on a runqueue */
808 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
809 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
811 /* Do the two (enqueued) tasks belong to the same group ? */
812 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
814 if (curr->se.cfs_rq == p->se.cfs_rq)
820 #else /* CONFIG_FAIR_GROUP_SCHED */
822 #define for_each_sched_entity(se) \
823 for (; se; se = NULL)
825 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
827 return &task_rq(p)->cfs;
830 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
832 struct task_struct *p = task_of(se);
833 struct rq *rq = task_rq(p);
838 /* runqueue "owned" by this group */
839 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
844 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
846 return &cpu_rq(this_cpu)->cfs;
849 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
850 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
852 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
857 #endif /* CONFIG_FAIR_GROUP_SCHED */
860 * The enqueue_task method is called before nr_running is
861 * increased. Here we update the fair scheduling stats and
862 * then put the task into the rbtree:
864 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
866 struct cfs_rq *cfs_rq;
867 struct sched_entity *se = &p->se;
869 for_each_sched_entity(se) {
872 cfs_rq = cfs_rq_of(se);
873 enqueue_entity(cfs_rq, se, wakeup);
878 * The dequeue_task method is called before nr_running is
879 * decreased. We remove the task from the rbtree and
880 * update the fair scheduling stats:
882 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
884 struct cfs_rq *cfs_rq;
885 struct sched_entity *se = &p->se;
887 for_each_sched_entity(se) {
888 cfs_rq = cfs_rq_of(se);
889 dequeue_entity(cfs_rq, se, sleep);
890 /* Don't dequeue parent if it has other entities besides us */
891 if (cfs_rq->load.weight)
897 * sched_yield() support is very simple - we dequeue and enqueue
899 static void yield_task_fair(struct rq *rq, struct task_struct *p)
901 struct cfs_rq *cfs_rq = task_cfs_rq(p);
903 __update_rq_clock(rq);
905 * Dequeue and enqueue the task to update its
906 * position within the tree:
908 dequeue_entity(cfs_rq, &p->se, 0);
909 enqueue_entity(cfs_rq, &p->se, 0);
913 * Preempt the current task with a newly woken task if needed:
915 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
917 struct task_struct *curr = rq->curr;
918 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
921 if (unlikely(rt_prio(p->prio))) {
928 gran = sysctl_sched_wakeup_granularity;
930 * Batch tasks prefer throughput over latency:
932 if (unlikely(p->policy == SCHED_BATCH))
933 gran = sysctl_sched_batch_wakeup_granularity;
935 if (is_same_group(curr, p))
936 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
939 static struct task_struct *pick_next_task_fair(struct rq *rq)
941 struct cfs_rq *cfs_rq = &rq->cfs;
942 struct sched_entity *se;
944 if (unlikely(!cfs_rq->nr_running))
948 se = pick_next_entity(cfs_rq);
949 cfs_rq = group_cfs_rq(se);
956 * Account for a descheduled task:
958 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
960 struct sched_entity *se = &prev->se;
961 struct cfs_rq *cfs_rq;
963 for_each_sched_entity(se) {
964 cfs_rq = cfs_rq_of(se);
965 put_prev_entity(cfs_rq, se);
969 /**************************************************
970 * Fair scheduling class load-balancing methods:
974 * Load-balancing iterator. Note: while the runqueue stays locked
975 * during the whole iteration, the current task might be
976 * dequeued so the iterator has to be dequeue-safe. Here we
977 * achieve that by always pre-iterating before returning
980 static inline struct task_struct *
981 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
983 struct task_struct *p;
988 p = rb_entry(curr, struct task_struct, se.run_node);
989 cfs_rq->rb_load_balance_curr = rb_next(curr);
994 static struct task_struct *load_balance_start_fair(void *arg)
996 struct cfs_rq *cfs_rq = arg;
998 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1001 static struct task_struct *load_balance_next_fair(void *arg)
1003 struct cfs_rq *cfs_rq = arg;
1005 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1008 #ifdef CONFIG_FAIR_GROUP_SCHED
1009 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1011 struct sched_entity *curr;
1012 struct task_struct *p;
1014 if (!cfs_rq->nr_running)
1017 curr = __pick_next_entity(cfs_rq);
1024 static unsigned long
1025 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1026 unsigned long max_nr_move, unsigned long max_load_move,
1027 struct sched_domain *sd, enum cpu_idle_type idle,
1028 int *all_pinned, int *this_best_prio)
1030 struct cfs_rq *busy_cfs_rq;
1031 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1032 long rem_load_move = max_load_move;
1033 struct rq_iterator cfs_rq_iterator;
1035 cfs_rq_iterator.start = load_balance_start_fair;
1036 cfs_rq_iterator.next = load_balance_next_fair;
1038 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1039 #ifdef CONFIG_FAIR_GROUP_SCHED
1040 struct cfs_rq *this_cfs_rq;
1042 unsigned long maxload;
1044 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1046 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1047 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1051 /* Don't pull more than imbalance/2 */
1053 maxload = min(rem_load_move, imbalance);
1055 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1057 # define maxload rem_load_move
1059 /* pass busy_cfs_rq argument into
1060 * load_balance_[start|next]_fair iterators
1062 cfs_rq_iterator.arg = busy_cfs_rq;
1063 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1064 max_nr_move, maxload, sd, idle, all_pinned,
1065 &load_moved, this_best_prio, &cfs_rq_iterator);
1067 total_nr_moved += nr_moved;
1068 max_nr_move -= nr_moved;
1069 rem_load_move -= load_moved;
1071 if (max_nr_move <= 0 || rem_load_move <= 0)
1075 return max_load_move - rem_load_move;
1079 * scheduler tick hitting a task of our scheduling class:
1081 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1083 struct cfs_rq *cfs_rq;
1084 struct sched_entity *se = &curr->se;
1086 for_each_sched_entity(se) {
1087 cfs_rq = cfs_rq_of(se);
1088 entity_tick(cfs_rq, se);
1093 * Share the fairness runtime between parent and child, thus the
1094 * total amount of pressure for CPU stays equal - new tasks
1095 * get a chance to run but frequent forkers are not allowed to
1096 * monopolize the CPU. Note: the parent runqueue is locked,
1097 * the child is not running yet.
1099 static void task_new_fair(struct rq *rq, struct task_struct *p)
1101 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1102 struct sched_entity *se = &p->se, *curr = cfs_rq_curr(cfs_rq);
1104 sched_info_queued(p);
1106 update_curr(cfs_rq);
1107 update_stats_enqueue(cfs_rq, se);
1109 * Child runs first: we let it run before the parent
1110 * until it reschedules once. We set up the key so that
1111 * it will preempt the parent:
1113 se->fair_key = curr->fair_key -
1114 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1116 * The first wait is dominated by the child-runs-first logic,
1117 * so do not credit it with that waiting time yet:
1119 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1120 se->wait_start_fair = 0;
1123 * The statistical average of wait_runtime is about
1124 * -granularity/2, so initialize the task with that:
1126 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1127 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1129 __enqueue_entity(cfs_rq, se);
1132 #ifdef CONFIG_FAIR_GROUP_SCHED
1133 /* Account for a task changing its policy or group.
1135 * This routine is mostly called to set cfs_rq->curr field when a task
1136 * migrates between groups/classes.
1138 static void set_curr_task_fair(struct rq *rq)
1140 struct sched_entity *se = &rq->curr->se;
1142 for_each_sched_entity(se)
1143 set_next_entity(cfs_rq_of(se), se);
1146 static void set_curr_task_fair(struct rq *rq)
1152 * All the scheduling class methods:
1154 struct sched_class fair_sched_class __read_mostly = {
1155 .enqueue_task = enqueue_task_fair,
1156 .dequeue_task = dequeue_task_fair,
1157 .yield_task = yield_task_fair,
1159 .check_preempt_curr = check_preempt_curr_fair,
1161 .pick_next_task = pick_next_task_fair,
1162 .put_prev_task = put_prev_task_fair,
1164 .load_balance = load_balance_fair,
1166 .set_curr_task = set_curr_task_fair,
1167 .task_tick = task_tick_fair,
1168 .task_new = task_new_fair,
1171 #ifdef CONFIG_SCHED_DEBUG
1172 static void print_cfs_stats(struct seq_file *m, int cpu)
1174 struct cfs_rq *cfs_rq;
1176 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1177 print_cfs_rq(m, cpu, cfs_rq);