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 #define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
110 extern struct sched_class fair_sched_class;
112 /**************************************************************
113 * CFS operations on generic schedulable entities:
116 #ifdef CONFIG_FAIR_GROUP_SCHED
118 /* cpu runqueue to which this cfs_rq is attached */
119 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
124 /* An entity is a task if it doesn't "own" a runqueue */
125 #define entity_is_task(se) (!se->my_q)
127 #else /* CONFIG_FAIR_GROUP_SCHED */
129 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
131 return container_of(cfs_rq, struct rq, cfs);
134 #define entity_is_task(se) 1
136 #endif /* CONFIG_FAIR_GROUP_SCHED */
138 static inline struct task_struct *task_of(struct sched_entity *se)
140 return container_of(se, struct task_struct, se);
144 /**************************************************************
145 * Scheduling class tree data structure manipulation methods:
149 * Enqueue an entity into the rb-tree:
152 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
154 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
155 struct rb_node *parent = NULL;
156 struct sched_entity *entry;
157 s64 key = se->fair_key;
161 * Find the right place in the rbtree:
165 entry = rb_entry(parent, struct sched_entity, run_node);
167 * We dont care about collisions. Nodes with
168 * the same key stay together.
170 if (key - entry->fair_key < 0) {
171 link = &parent->rb_left;
173 link = &parent->rb_right;
179 * Maintain a cache of leftmost tree entries (it is frequently
183 cfs_rq->rb_leftmost = &se->run_node;
185 rb_link_node(&se->run_node, parent, link);
186 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
187 update_load_add(&cfs_rq->load, se->load.weight);
188 cfs_rq->nr_running++;
191 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
195 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
197 if (cfs_rq->rb_leftmost == &se->run_node)
198 cfs_rq->rb_leftmost = rb_next(&se->run_node);
199 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
200 update_load_sub(&cfs_rq->load, se->load.weight);
201 cfs_rq->nr_running--;
204 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
207 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
209 return cfs_rq->rb_leftmost;
212 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
214 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
217 /**************************************************************
218 * Scheduling class statistics methods:
222 * Calculate the preemption granularity needed to schedule every
223 * runnable task once per sysctl_sched_latency amount of time.
224 * (down to a sensible low limit on granularity)
226 * For example, if there are 2 tasks running and latency is 10 msecs,
227 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
228 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
229 * for each task. We do finer and finer scheduling up to until we
230 * reach the minimum granularity value.
232 * To achieve this we use the following dynamic-granularity rule:
234 * gran = lat/nr - lat/nr/nr
236 * This comes out of the following equations:
241 * kB2 = kB1 - d + d/nr
244 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
245 * '1' is start of time, '2' is end of time, 'd' is delay between
246 * 1 and 2 (during which task B was running), 'nr' is number of tasks
247 * running, 'lat' is the the period of each task. ('lat' is the
248 * sched_latency that we aim for.)
251 sched_granularity(struct cfs_rq *cfs_rq)
253 unsigned int gran = sysctl_sched_latency;
254 unsigned int nr = cfs_rq->nr_running;
257 gran = gran/nr - gran/nr/nr;
258 gran = max(gran, sysctl_sched_min_granularity);
265 * We rescale the rescheduling granularity of tasks according to their
266 * nice level, but only linearly, not exponentially:
269 niced_granularity(struct sched_entity *curr, unsigned long granularity)
273 if (likely(curr->load.weight == NICE_0_LOAD))
276 * Positive nice levels get the same granularity as nice-0:
278 if (likely(curr->load.weight < NICE_0_LOAD)) {
279 tmp = curr->load.weight * (u64)granularity;
280 return (long) (tmp >> NICE_0_SHIFT);
283 * Negative nice level tasks get linearly finer
286 tmp = curr->load.inv_weight * (u64)granularity;
289 * It will always fit into 'long':
291 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
295 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
297 long limit = sysctl_sched_runtime_limit;
300 * Niced tasks have the same history dynamic range as
303 if (unlikely(se->wait_runtime > limit)) {
304 se->wait_runtime = limit;
305 schedstat_inc(se, wait_runtime_overruns);
306 schedstat_inc(cfs_rq, wait_runtime_overruns);
308 if (unlikely(se->wait_runtime < -limit)) {
309 se->wait_runtime = -limit;
310 schedstat_inc(se, wait_runtime_underruns);
311 schedstat_inc(cfs_rq, wait_runtime_underruns);
316 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
318 se->wait_runtime += delta;
319 schedstat_add(se, sum_wait_runtime, delta);
320 limit_wait_runtime(cfs_rq, se);
324 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
326 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
327 __add_wait_runtime(cfs_rq, se, delta);
328 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
332 * Update the current task's runtime statistics. Skip current tasks that
333 * are not in our scheduling class.
336 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
337 unsigned long delta_exec)
339 unsigned long delta, delta_fair, delta_mine;
340 struct load_weight *lw = &cfs_rq->load;
341 unsigned long load = lw->weight;
343 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
345 curr->sum_exec_runtime += delta_exec;
346 cfs_rq->exec_clock += delta_exec;
351 delta_fair = calc_delta_fair(delta_exec, lw);
352 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
354 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
355 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
356 delta = min(delta, (unsigned long)(
357 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
358 cfs_rq->sleeper_bonus -= delta;
362 cfs_rq->fair_clock += delta_fair;
364 * We executed delta_exec amount of time on the CPU,
365 * but we were only entitled to delta_mine amount of
366 * time during that period (if nr_running == 1 then
367 * the two values are equal)
368 * [Note: delta_mine - delta_exec is negative]:
370 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
373 static void update_curr(struct cfs_rq *cfs_rq)
375 struct sched_entity *curr = cfs_rq->curr;
376 u64 now = rq_of(cfs_rq)->clock;
377 unsigned long delta_exec;
383 * Get the amount of time the current task was running
384 * since the last time we changed load (this cannot
385 * overflow on 32 bits):
387 delta_exec = (unsigned long)(now - curr->exec_start);
389 __update_curr(cfs_rq, curr, delta_exec);
390 curr->exec_start = now;
394 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
396 se->wait_start_fair = cfs_rq->fair_clock;
397 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
400 static inline unsigned long
401 calc_weighted(unsigned long delta, struct sched_entity *se)
403 unsigned long weight = se->load.weight;
405 if (unlikely(weight != NICE_0_LOAD))
406 return (u64)delta * se->load.weight >> NICE_0_SHIFT;
412 * Task is being enqueued - update stats:
414 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
419 * Are we enqueueing a waiting task? (for current tasks
420 * a dequeue/enqueue event is a NOP)
422 if (se != cfs_rq->curr)
423 update_stats_wait_start(cfs_rq, se);
427 key = cfs_rq->fair_clock;
430 * Optimize the common nice 0 case:
432 if (likely(se->load.weight == NICE_0_LOAD)) {
433 key -= se->wait_runtime;
437 if (se->wait_runtime < 0) {
438 tmp = -se->wait_runtime;
439 key += (tmp * se->load.inv_weight) >>
440 (WMULT_SHIFT - NICE_0_SHIFT);
442 tmp = se->wait_runtime;
443 key -= (tmp * se->load.inv_weight) >>
444 (WMULT_SHIFT - NICE_0_SHIFT);
452 * Note: must be called with a freshly updated rq->fair_clock.
455 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se,
456 unsigned long delta_fair)
458 schedstat_set(se->wait_max, max(se->wait_max,
459 rq_of(cfs_rq)->clock - se->wait_start));
461 delta_fair = calc_weighted(delta_fair, se);
463 add_wait_runtime(cfs_rq, se, delta_fair);
467 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
469 unsigned long delta_fair;
471 if (unlikely(!se->wait_start_fair))
474 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
475 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
477 __update_stats_wait_end(cfs_rq, se, delta_fair);
479 se->wait_start_fair = 0;
480 schedstat_set(se->wait_start, 0);
484 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
488 * Mark the end of the wait period if dequeueing a
491 if (se != cfs_rq->curr)
492 update_stats_wait_end(cfs_rq, se);
496 * We are picking a new current task - update its stats:
499 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
502 * We are starting a new run period:
504 se->exec_start = rq_of(cfs_rq)->clock;
508 * We are descheduling a task - update its stats:
511 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
516 /**************************************************
517 * Scheduling class queueing methods:
520 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se,
521 unsigned long delta_fair)
523 unsigned long load = cfs_rq->load.weight;
527 * Do not boost sleepers if there's too much bonus 'in flight'
530 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
533 if (sched_feat(SLEEPER_LOAD_AVG))
534 load = rq_of(cfs_rq)->cpu_load[2];
537 * Fix up delta_fair with the effect of us running
538 * during the whole sleep period:
540 if (sched_feat(SLEEPER_AVG))
541 delta_fair = div64_likely32((u64)delta_fair * load,
542 load + se->load.weight);
544 delta_fair = calc_weighted(delta_fair, se);
546 prev_runtime = se->wait_runtime;
547 __add_wait_runtime(cfs_rq, se, delta_fair);
548 delta_fair = se->wait_runtime - prev_runtime;
551 * Track the amount of bonus we've given to sleepers:
553 cfs_rq->sleeper_bonus += delta_fair;
556 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
558 struct task_struct *tsk = task_of(se);
559 unsigned long delta_fair;
561 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
562 !sched_feat(FAIR_SLEEPERS))
565 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
566 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
568 __enqueue_sleeper(cfs_rq, se, delta_fair);
570 se->sleep_start_fair = 0;
572 #ifdef CONFIG_SCHEDSTATS
573 if (se->sleep_start) {
574 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
579 if (unlikely(delta > se->sleep_max))
580 se->sleep_max = delta;
583 se->sum_sleep_runtime += delta;
585 if (se->block_start) {
586 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
591 if (unlikely(delta > se->block_max))
592 se->block_max = delta;
595 se->sum_sleep_runtime += delta;
598 * Blocking time is in units of nanosecs, so shift by 20 to
599 * get a milliseconds-range estimation of the amount of
600 * time that the task spent sleeping:
602 if (unlikely(prof_on == SLEEP_PROFILING)) {
603 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
611 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
614 * Update the fair clock.
619 enqueue_sleeper(cfs_rq, se);
621 update_stats_enqueue(cfs_rq, se);
622 __enqueue_entity(cfs_rq, se);
626 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
628 update_stats_dequeue(cfs_rq, se);
630 se->sleep_start_fair = cfs_rq->fair_clock;
631 #ifdef CONFIG_SCHEDSTATS
632 if (entity_is_task(se)) {
633 struct task_struct *tsk = task_of(se);
635 if (tsk->state & TASK_INTERRUPTIBLE)
636 se->sleep_start = rq_of(cfs_rq)->clock;
637 if (tsk->state & TASK_UNINTERRUPTIBLE)
638 se->block_start = rq_of(cfs_rq)->clock;
642 __dequeue_entity(cfs_rq, se);
646 * Preempt the current task with a newly woken task if needed:
649 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
650 struct sched_entity *curr, unsigned long granularity)
652 s64 __delta = curr->fair_key - se->fair_key;
653 unsigned long ideal_runtime, delta_exec;
656 * ideal_runtime is compared against sum_exec_runtime, which is
657 * walltime, hence do not scale.
659 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
660 (unsigned long)sysctl_sched_min_granularity);
663 * If we executed more than what the latency constraint suggests,
664 * reduce the rescheduling granularity. This way the total latency
665 * of how much a task is not scheduled converges to
666 * sysctl_sched_latency:
668 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
669 if (delta_exec > ideal_runtime)
673 * Take scheduling granularity into account - do not
674 * preempt the current task unless the best task has
675 * a larger than sched_granularity fairness advantage:
677 * scale granularity as key space is in fair_clock.
679 if (__delta > niced_granularity(curr, granularity))
680 resched_task(rq_of(cfs_rq)->curr);
684 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
687 * Any task has to be enqueued before it get to execute on
688 * a CPU. So account for the time it spent waiting on the
689 * runqueue. (note, here we rely on pick_next_task() having
690 * done a put_prev_task_fair() shortly before this, which
691 * updated rq->fair_clock - used by update_stats_wait_end())
693 update_stats_wait_end(cfs_rq, se);
694 update_stats_curr_start(cfs_rq, se);
696 #ifdef CONFIG_SCHEDSTATS
698 * Track our maximum slice length, if the CPU's load is at
699 * least twice that of our own weight (i.e. dont track it
700 * when there are only lesser-weight tasks around):
702 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
703 se->slice_max = max(se->slice_max,
704 se->sum_exec_runtime - se->prev_sum_exec_runtime);
707 se->prev_sum_exec_runtime = se->sum_exec_runtime;
710 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
712 struct sched_entity *se = __pick_next_entity(cfs_rq);
714 set_next_entity(cfs_rq, se);
719 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
722 * If still on the runqueue then deactivate_task()
723 * was not called and update_curr() has to be done:
728 update_stats_curr_end(cfs_rq, prev);
731 update_stats_wait_start(cfs_rq, prev);
735 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
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 __check_preempt_curr_fair(cfs_rq, next, curr,
754 sched_granularity(cfs_rq));
757 /**************************************************
758 * CFS operations on tasks:
761 #ifdef CONFIG_FAIR_GROUP_SCHED
763 /* Walk up scheduling entities hierarchy */
764 #define for_each_sched_entity(se) \
765 for (; se; se = se->parent)
767 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
772 /* runqueue on which this entity is (to be) queued */
773 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
778 /* runqueue "owned" by this group */
779 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
784 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
785 * another cpu ('this_cpu')
787 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
789 /* A later patch will take group into account */
790 return &cpu_rq(this_cpu)->cfs;
793 /* Iterate thr' all leaf cfs_rq's on a runqueue */
794 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
795 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
797 /* Do the two (enqueued) tasks belong to the same group ? */
798 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
800 if (curr->se.cfs_rq == p->se.cfs_rq)
806 #else /* CONFIG_FAIR_GROUP_SCHED */
808 #define for_each_sched_entity(se) \
809 for (; se; se = NULL)
811 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
813 return &task_rq(p)->cfs;
816 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
818 struct task_struct *p = task_of(se);
819 struct rq *rq = task_rq(p);
824 /* runqueue "owned" by this group */
825 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
830 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
832 return &cpu_rq(this_cpu)->cfs;
835 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
836 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
838 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
843 #endif /* CONFIG_FAIR_GROUP_SCHED */
846 * The enqueue_task method is called before nr_running is
847 * increased. Here we update the fair scheduling stats and
848 * then put the task into the rbtree:
850 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
852 struct cfs_rq *cfs_rq;
853 struct sched_entity *se = &p->se;
855 for_each_sched_entity(se) {
858 cfs_rq = cfs_rq_of(se);
859 enqueue_entity(cfs_rq, se, wakeup);
864 * The dequeue_task method is called before nr_running is
865 * decreased. We remove the task from the rbtree and
866 * update the fair scheduling stats:
868 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
870 struct cfs_rq *cfs_rq;
871 struct sched_entity *se = &p->se;
873 for_each_sched_entity(se) {
874 cfs_rq = cfs_rq_of(se);
875 dequeue_entity(cfs_rq, se, sleep);
876 /* Don't dequeue parent if it has other entities besides us */
877 if (cfs_rq->load.weight)
883 * sched_yield() support is very simple - we dequeue and enqueue.
885 * If compat_yield is turned on then we requeue to the end of the tree.
887 static void yield_task_fair(struct rq *rq, struct task_struct *p)
889 struct cfs_rq *cfs_rq = task_cfs_rq(p);
890 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
891 struct sched_entity *rightmost, *se = &p->se;
892 struct rb_node *parent;
895 * Are we the only task in the tree?
897 if (unlikely(cfs_rq->nr_running == 1))
900 if (likely(!sysctl_sched_compat_yield)) {
901 __update_rq_clock(rq);
903 * Dequeue and enqueue the task to update its
904 * position within the tree:
906 dequeue_entity(cfs_rq, &p->se, 0);
907 enqueue_entity(cfs_rq, &p->se, 0);
912 * Find the rightmost entry in the rbtree:
916 link = &parent->rb_right;
919 rightmost = rb_entry(parent, struct sched_entity, run_node);
921 * Already in the rightmost position?
923 if (unlikely(rightmost == se))
927 * Minimally necessary key value to be last in the tree:
929 se->fair_key = rightmost->fair_key + 1;
931 if (cfs_rq->rb_leftmost == &se->run_node)
932 cfs_rq->rb_leftmost = rb_next(&se->run_node);
934 * Relink the task to the rightmost position:
936 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
937 rb_link_node(&se->run_node, parent, link);
938 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
942 * Preempt the current task with a newly woken task if needed:
944 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
946 struct task_struct *curr = rq->curr;
947 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
950 if (unlikely(rt_prio(p->prio))) {
957 gran = sysctl_sched_wakeup_granularity;
959 * Batch tasks prefer throughput over latency:
961 if (unlikely(p->policy == SCHED_BATCH))
962 gran = sysctl_sched_batch_wakeup_granularity;
964 if (is_same_group(curr, p))
965 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
968 static struct task_struct *pick_next_task_fair(struct rq *rq)
970 struct cfs_rq *cfs_rq = &rq->cfs;
971 struct sched_entity *se;
973 if (unlikely(!cfs_rq->nr_running))
977 se = pick_next_entity(cfs_rq);
978 cfs_rq = group_cfs_rq(se);
985 * Account for a descheduled task:
987 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
989 struct sched_entity *se = &prev->se;
990 struct cfs_rq *cfs_rq;
992 for_each_sched_entity(se) {
993 cfs_rq = cfs_rq_of(se);
994 put_prev_entity(cfs_rq, se);
998 /**************************************************
999 * Fair scheduling class load-balancing methods:
1003 * Load-balancing iterator. Note: while the runqueue stays locked
1004 * during the whole iteration, the current task might be
1005 * dequeued so the iterator has to be dequeue-safe. Here we
1006 * achieve that by always pre-iterating before returning
1009 static inline struct task_struct *
1010 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1012 struct task_struct *p;
1017 p = rb_entry(curr, struct task_struct, se.run_node);
1018 cfs_rq->rb_load_balance_curr = rb_next(curr);
1023 static struct task_struct *load_balance_start_fair(void *arg)
1025 struct cfs_rq *cfs_rq = arg;
1027 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1030 static struct task_struct *load_balance_next_fair(void *arg)
1032 struct cfs_rq *cfs_rq = arg;
1034 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1037 #ifdef CONFIG_FAIR_GROUP_SCHED
1038 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1040 struct sched_entity *curr;
1041 struct task_struct *p;
1043 if (!cfs_rq->nr_running)
1046 curr = __pick_next_entity(cfs_rq);
1053 static unsigned long
1054 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1055 unsigned long max_nr_move, unsigned long max_load_move,
1056 struct sched_domain *sd, enum cpu_idle_type idle,
1057 int *all_pinned, int *this_best_prio)
1059 struct cfs_rq *busy_cfs_rq;
1060 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1061 long rem_load_move = max_load_move;
1062 struct rq_iterator cfs_rq_iterator;
1064 cfs_rq_iterator.start = load_balance_start_fair;
1065 cfs_rq_iterator.next = load_balance_next_fair;
1067 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1068 #ifdef CONFIG_FAIR_GROUP_SCHED
1069 struct cfs_rq *this_cfs_rq;
1071 unsigned long maxload;
1073 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1075 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1076 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1080 /* Don't pull more than imbalance/2 */
1082 maxload = min(rem_load_move, imbalance);
1084 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1086 # define maxload rem_load_move
1088 /* pass busy_cfs_rq argument into
1089 * load_balance_[start|next]_fair iterators
1091 cfs_rq_iterator.arg = busy_cfs_rq;
1092 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1093 max_nr_move, maxload, sd, idle, all_pinned,
1094 &load_moved, this_best_prio, &cfs_rq_iterator);
1096 total_nr_moved += nr_moved;
1097 max_nr_move -= nr_moved;
1098 rem_load_move -= load_moved;
1100 if (max_nr_move <= 0 || rem_load_move <= 0)
1104 return max_load_move - rem_load_move;
1108 * scheduler tick hitting a task of our scheduling class:
1110 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1112 struct cfs_rq *cfs_rq;
1113 struct sched_entity *se = &curr->se;
1115 for_each_sched_entity(se) {
1116 cfs_rq = cfs_rq_of(se);
1117 entity_tick(cfs_rq, se);
1122 * Share the fairness runtime between parent and child, thus the
1123 * total amount of pressure for CPU stays equal - new tasks
1124 * get a chance to run but frequent forkers are not allowed to
1125 * monopolize the CPU. Note: the parent runqueue is locked,
1126 * the child is not running yet.
1128 static void task_new_fair(struct rq *rq, struct task_struct *p)
1130 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1131 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
1133 sched_info_queued(p);
1135 update_curr(cfs_rq);
1136 update_stats_enqueue(cfs_rq, se);
1138 * Child runs first: we let it run before the parent
1139 * until it reschedules once. We set up the key so that
1140 * it will preempt the parent:
1142 se->fair_key = curr->fair_key -
1143 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1145 * The first wait is dominated by the child-runs-first logic,
1146 * so do not credit it with that waiting time yet:
1148 if (sched_feat(SKIP_INITIAL))
1149 se->wait_start_fair = 0;
1152 * The statistical average of wait_runtime is about
1153 * -granularity/2, so initialize the task with that:
1155 if (sched_feat(START_DEBIT))
1156 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1158 __enqueue_entity(cfs_rq, se);
1159 resched_task(rq->curr);
1162 #ifdef CONFIG_FAIR_GROUP_SCHED
1163 /* Account for a task changing its policy or group.
1165 * This routine is mostly called to set cfs_rq->curr field when a task
1166 * migrates between groups/classes.
1168 static void set_curr_task_fair(struct rq *rq)
1170 struct sched_entity *se = &rq->curr->se;
1172 for_each_sched_entity(se)
1173 set_next_entity(cfs_rq_of(se), se);
1176 static void set_curr_task_fair(struct rq *rq)
1182 * All the scheduling class methods:
1184 struct sched_class fair_sched_class __read_mostly = {
1185 .enqueue_task = enqueue_task_fair,
1186 .dequeue_task = dequeue_task_fair,
1187 .yield_task = yield_task_fair,
1189 .check_preempt_curr = check_preempt_curr_fair,
1191 .pick_next_task = pick_next_task_fair,
1192 .put_prev_task = put_prev_task_fair,
1194 .load_balance = load_balance_fair,
1196 .set_curr_task = set_curr_task_fair,
1197 .task_tick = task_tick_fair,
1198 .task_new = task_new_fair,
1201 #ifdef CONFIG_SCHED_DEBUG
1202 static void print_cfs_stats(struct seq_file *m, int cpu)
1204 struct cfs_rq *cfs_rq;
1206 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1207 print_cfs_rq(m, cpu, cfs_rq);