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 const_debug unsigned int sysctl_sched_stat_granularity;
90 unsigned int sysctl_sched_runtime_limit __read_mostly;
93 * Debugging: various feature bits
96 SCHED_FEAT_FAIR_SLEEPERS = 1,
97 SCHED_FEAT_SLEEPER_AVG = 2,
98 SCHED_FEAT_SLEEPER_LOAD_AVG = 4,
99 SCHED_FEAT_PRECISE_CPU_LOAD = 8,
100 SCHED_FEAT_START_DEBIT = 16,
101 SCHED_FEAT_SKIP_INITIAL = 32,
104 const_debug unsigned int sysctl_sched_features =
105 SCHED_FEAT_FAIR_SLEEPERS *1 |
106 SCHED_FEAT_SLEEPER_AVG *0 |
107 SCHED_FEAT_SLEEPER_LOAD_AVG *1 |
108 SCHED_FEAT_PRECISE_CPU_LOAD *1 |
109 SCHED_FEAT_START_DEBIT *1 |
110 SCHED_FEAT_SKIP_INITIAL *0;
112 extern struct sched_class fair_sched_class;
114 /**************************************************************
115 * CFS operations on generic schedulable entities:
118 #ifdef CONFIG_FAIR_GROUP_SCHED
120 /* cpu runqueue to which this cfs_rq is attached */
121 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
126 /* currently running entity (if any) on this cfs_rq */
127 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
132 /* An entity is a task if it doesn't "own" a runqueue */
133 #define entity_is_task(se) (!se->my_q)
136 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
141 #else /* CONFIG_FAIR_GROUP_SCHED */
143 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
145 return container_of(cfs_rq, struct rq, cfs);
148 static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
150 struct rq *rq = rq_of(cfs_rq);
152 if (unlikely(rq->curr->sched_class != &fair_sched_class))
155 return &rq->curr->se;
158 #define entity_is_task(se) 1
161 set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
163 #endif /* CONFIG_FAIR_GROUP_SCHED */
165 static inline struct task_struct *task_of(struct sched_entity *se)
167 return container_of(se, struct task_struct, se);
171 /**************************************************************
172 * Scheduling class tree data structure manipulation methods:
176 * Enqueue an entity into the rb-tree:
179 __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
181 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
182 struct rb_node *parent = NULL;
183 struct sched_entity *entry;
184 s64 key = se->fair_key;
188 * Find the right place in the rbtree:
192 entry = rb_entry(parent, struct sched_entity, run_node);
194 * We dont care about collisions. Nodes with
195 * the same key stay together.
197 if (key - entry->fair_key < 0) {
198 link = &parent->rb_left;
200 link = &parent->rb_right;
206 * Maintain a cache of leftmost tree entries (it is frequently
210 cfs_rq->rb_leftmost = &se->run_node;
212 rb_link_node(&se->run_node, parent, link);
213 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
214 update_load_add(&cfs_rq->load, se->load.weight);
215 cfs_rq->nr_running++;
218 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
222 __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
224 if (cfs_rq->rb_leftmost == &se->run_node)
225 cfs_rq->rb_leftmost = rb_next(&se->run_node);
226 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
227 update_load_sub(&cfs_rq->load, se->load.weight);
228 cfs_rq->nr_running--;
231 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
234 static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
236 return cfs_rq->rb_leftmost;
239 static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
241 return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
244 /**************************************************************
245 * Scheduling class statistics methods:
249 * Calculate the preemption granularity needed to schedule every
250 * runnable task once per sysctl_sched_latency amount of time.
251 * (down to a sensible low limit on granularity)
253 * For example, if there are 2 tasks running and latency is 10 msecs,
254 * we switch tasks every 5 msecs. If we have 3 tasks running, we have
255 * to switch tasks every 3.33 msecs to get a 10 msecs observed latency
256 * for each task. We do finer and finer scheduling up to until we
257 * reach the minimum granularity value.
259 * To achieve this we use the following dynamic-granularity rule:
261 * gran = lat/nr - lat/nr/nr
263 * This comes out of the following equations:
268 * kB2 = kB1 - d + d/nr
271 * Where 'k' is key, 'A' is task A (waiting), 'B' is task B (running),
272 * '1' is start of time, '2' is end of time, 'd' is delay between
273 * 1 and 2 (during which task B was running), 'nr' is number of tasks
274 * running, 'lat' is the the period of each task. ('lat' is the
275 * sched_latency that we aim for.)
278 sched_granularity(struct cfs_rq *cfs_rq)
280 unsigned int gran = sysctl_sched_latency;
281 unsigned int nr = cfs_rq->nr_running;
284 gran = gran/nr - gran/nr/nr;
285 gran = max(gran, sysctl_sched_min_granularity);
292 * We rescale the rescheduling granularity of tasks according to their
293 * nice level, but only linearly, not exponentially:
296 niced_granularity(struct sched_entity *curr, unsigned long granularity)
300 if (likely(curr->load.weight == NICE_0_LOAD))
303 * Positive nice levels get the same granularity as nice-0:
305 if (likely(curr->load.weight < NICE_0_LOAD)) {
306 tmp = curr->load.weight * (u64)granularity;
307 return (long) (tmp >> NICE_0_SHIFT);
310 * Negative nice level tasks get linearly finer
313 tmp = curr->load.inv_weight * (u64)granularity;
316 * It will always fit into 'long':
318 return (long) (tmp >> (WMULT_SHIFT-NICE_0_SHIFT));
322 limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
324 long limit = sysctl_sched_runtime_limit;
327 * Niced tasks have the same history dynamic range as
330 if (unlikely(se->wait_runtime > limit)) {
331 se->wait_runtime = limit;
332 schedstat_inc(se, wait_runtime_overruns);
333 schedstat_inc(cfs_rq, wait_runtime_overruns);
335 if (unlikely(se->wait_runtime < -limit)) {
336 se->wait_runtime = -limit;
337 schedstat_inc(se, wait_runtime_underruns);
338 schedstat_inc(cfs_rq, wait_runtime_underruns);
343 __add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
345 se->wait_runtime += delta;
346 schedstat_add(se, sum_wait_runtime, delta);
347 limit_wait_runtime(cfs_rq, se);
351 add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
353 schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
354 __add_wait_runtime(cfs_rq, se, delta);
355 schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
359 * Update the current task's runtime statistics. Skip current tasks that
360 * are not in our scheduling class.
363 __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr)
365 unsigned long delta, delta_exec, delta_fair, delta_mine;
366 struct load_weight *lw = &cfs_rq->load;
367 unsigned long load = lw->weight;
369 delta_exec = curr->delta_exec;
370 schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
372 curr->sum_exec_runtime += delta_exec;
373 cfs_rq->exec_clock += delta_exec;
378 delta_fair = calc_delta_fair(delta_exec, lw);
379 delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
381 if (cfs_rq->sleeper_bonus > sysctl_sched_min_granularity) {
382 delta = min((u64)delta_mine, cfs_rq->sleeper_bonus);
383 delta = min(delta, (unsigned long)(
384 (long)sysctl_sched_runtime_limit - curr->wait_runtime));
385 cfs_rq->sleeper_bonus -= delta;
389 cfs_rq->fair_clock += delta_fair;
391 * We executed delta_exec amount of time on the CPU,
392 * but we were only entitled to delta_mine amount of
393 * time during that period (if nr_running == 1 then
394 * the two values are equal)
395 * [Note: delta_mine - delta_exec is negative]:
397 add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
400 static void update_curr(struct cfs_rq *cfs_rq)
402 struct sched_entity *curr = cfs_rq_curr(cfs_rq);
403 unsigned long delta_exec;
409 * Get the amount of time the current task was running
410 * since the last time we changed load (this cannot
411 * overflow on 32 bits):
413 delta_exec = (unsigned long)(rq_of(cfs_rq)->clock - curr->exec_start);
415 curr->delta_exec += delta_exec;
417 if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
418 __update_curr(cfs_rq, curr);
419 curr->delta_exec = 0;
421 curr->exec_start = rq_of(cfs_rq)->clock;
425 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
427 se->wait_start_fair = cfs_rq->fair_clock;
428 schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
432 * We calculate fair deltas here, so protect against the random effects
433 * of a multiplication overflow by capping it to the runtime limit:
435 #if BITS_PER_LONG == 32
436 static inline unsigned long
437 calc_weighted(unsigned long delta, unsigned long weight, int shift)
439 u64 tmp = (u64)delta * weight >> shift;
441 if (unlikely(tmp > sysctl_sched_runtime_limit*2))
442 return sysctl_sched_runtime_limit*2;
446 static inline unsigned long
447 calc_weighted(unsigned long delta, unsigned long weight, int shift)
449 return delta * weight >> shift;
454 * Task is being enqueued - update stats:
456 static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
461 * Are we enqueueing a waiting task? (for current tasks
462 * a dequeue/enqueue event is a NOP)
464 if (se != cfs_rq_curr(cfs_rq))
465 update_stats_wait_start(cfs_rq, se);
469 key = cfs_rq->fair_clock;
472 * Optimize the common nice 0 case:
474 if (likely(se->load.weight == NICE_0_LOAD)) {
475 key -= se->wait_runtime;
479 if (se->wait_runtime < 0) {
480 tmp = -se->wait_runtime;
481 key += (tmp * se->load.inv_weight) >>
482 (WMULT_SHIFT - NICE_0_SHIFT);
484 tmp = se->wait_runtime;
485 key -= (tmp * se->load.inv_weight) >>
486 (WMULT_SHIFT - NICE_0_SHIFT);
494 * Note: must be called with a freshly updated rq->fair_clock.
497 __update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
499 unsigned long delta_fair = se->delta_fair_run;
501 schedstat_set(se->wait_max, max(se->wait_max,
502 rq_of(cfs_rq)->clock - se->wait_start));
504 if (unlikely(se->load.weight != NICE_0_LOAD))
505 delta_fair = calc_weighted(delta_fair, se->load.weight,
508 add_wait_runtime(cfs_rq, se, delta_fair);
512 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
514 unsigned long delta_fair;
516 if (unlikely(!se->wait_start_fair))
519 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
520 (u64)(cfs_rq->fair_clock - se->wait_start_fair));
522 se->delta_fair_run += delta_fair;
523 if (unlikely(abs(se->delta_fair_run) >=
524 sysctl_sched_stat_granularity)) {
525 __update_stats_wait_end(cfs_rq, se);
526 se->delta_fair_run = 0;
529 se->wait_start_fair = 0;
530 schedstat_set(se->wait_start, 0);
534 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
538 * Mark the end of the wait period if dequeueing a
541 if (se != cfs_rq_curr(cfs_rq))
542 update_stats_wait_end(cfs_rq, se);
546 * We are picking a new current task - update its stats:
549 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
552 * We are starting a new run period:
554 se->exec_start = rq_of(cfs_rq)->clock;
558 * We are descheduling a task - update its stats:
561 update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
566 /**************************************************
567 * Scheduling class queueing methods:
570 static void __enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
572 unsigned long load = cfs_rq->load.weight, delta_fair;
576 * Do not boost sleepers if there's too much bonus 'in flight'
579 if (unlikely(cfs_rq->sleeper_bonus > sysctl_sched_runtime_limit))
582 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
583 load = rq_of(cfs_rq)->cpu_load[2];
585 delta_fair = se->delta_fair_sleep;
588 * Fix up delta_fair with the effect of us running
589 * during the whole sleep period:
591 if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
592 delta_fair = div64_likely32((u64)delta_fair * load,
593 load + se->load.weight);
595 if (unlikely(se->load.weight != NICE_0_LOAD))
596 delta_fair = calc_weighted(delta_fair, se->load.weight,
599 prev_runtime = se->wait_runtime;
600 __add_wait_runtime(cfs_rq, se, delta_fair);
601 delta_fair = se->wait_runtime - prev_runtime;
604 * Track the amount of bonus we've given to sleepers:
606 cfs_rq->sleeper_bonus += delta_fair;
609 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
611 struct task_struct *tsk = task_of(se);
612 unsigned long delta_fair;
614 if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
615 !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
618 delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
619 (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
621 se->delta_fair_sleep += delta_fair;
622 if (unlikely(abs(se->delta_fair_sleep) >=
623 sysctl_sched_stat_granularity)) {
624 __enqueue_sleeper(cfs_rq, se);
625 se->delta_fair_sleep = 0;
628 se->sleep_start_fair = 0;
630 #ifdef CONFIG_SCHEDSTATS
631 if (se->sleep_start) {
632 u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
637 if (unlikely(delta > se->sleep_max))
638 se->sleep_max = delta;
641 se->sum_sleep_runtime += delta;
643 if (se->block_start) {
644 u64 delta = rq_of(cfs_rq)->clock - se->block_start;
649 if (unlikely(delta > se->block_max))
650 se->block_max = delta;
653 se->sum_sleep_runtime += delta;
656 * Blocking time is in units of nanosecs, so shift by 20 to
657 * get a milliseconds-range estimation of the amount of
658 * time that the task spent sleeping:
660 if (unlikely(prof_on == SLEEP_PROFILING)) {
661 profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
669 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
672 * Update the fair clock.
677 enqueue_sleeper(cfs_rq, se);
679 update_stats_enqueue(cfs_rq, se);
680 __enqueue_entity(cfs_rq, se);
684 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
686 update_stats_dequeue(cfs_rq, se);
688 se->sleep_start_fair = cfs_rq->fair_clock;
689 #ifdef CONFIG_SCHEDSTATS
690 if (entity_is_task(se)) {
691 struct task_struct *tsk = task_of(se);
693 if (tsk->state & TASK_INTERRUPTIBLE)
694 se->sleep_start = rq_of(cfs_rq)->clock;
695 if (tsk->state & TASK_UNINTERRUPTIBLE)
696 se->block_start = rq_of(cfs_rq)->clock;
700 __dequeue_entity(cfs_rq, se);
704 * Preempt the current task with a newly woken task if needed:
707 __check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
708 struct sched_entity *curr, unsigned long granularity)
710 s64 __delta = curr->fair_key - se->fair_key;
711 unsigned long ideal_runtime, delta_exec;
714 * ideal_runtime is compared against sum_exec_runtime, which is
715 * walltime, hence do not scale.
717 ideal_runtime = max(sysctl_sched_latency / cfs_rq->nr_running,
718 (unsigned long)sysctl_sched_min_granularity);
721 * If we executed more than what the latency constraint suggests,
722 * reduce the rescheduling granularity. This way the total latency
723 * of how much a task is not scheduled converges to
724 * sysctl_sched_latency:
726 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
727 if (delta_exec > ideal_runtime)
731 * Take scheduling granularity into account - do not
732 * preempt the current task unless the best task has
733 * a larger than sched_granularity fairness advantage:
735 * scale granularity as key space is in fair_clock.
737 if (__delta > niced_granularity(curr, granularity))
738 resched_task(rq_of(cfs_rq)->curr);
742 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 * Any task has to be enqueued before it get to execute on
746 * a CPU. So account for the time it spent waiting on the
747 * runqueue. (note, here we rely on pick_next_task() having
748 * done a put_prev_task_fair() shortly before this, which
749 * updated rq->fair_clock - used by update_stats_wait_end())
751 update_stats_wait_end(cfs_rq, se);
752 update_stats_curr_start(cfs_rq, se);
753 set_cfs_rq_curr(cfs_rq, se);
754 #ifdef CONFIG_SCHEDSTATS
756 * Track our maximum slice length, if the CPU's load is at
757 * least twice that of our own weight (i.e. dont track it
758 * when there are only lesser-weight tasks around):
760 if (rq_of(cfs_rq)->ls.load.weight >= 2*se->load.weight) {
761 se->slice_max = max(se->slice_max,
762 se->sum_exec_runtime - se->prev_sum_exec_runtime);
765 se->prev_sum_exec_runtime = se->sum_exec_runtime;
768 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
770 struct sched_entity *se = __pick_next_entity(cfs_rq);
772 set_next_entity(cfs_rq, se);
777 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
780 * If still on the runqueue then deactivate_task()
781 * was not called and update_curr() has to be done:
786 update_stats_curr_end(cfs_rq, prev);
789 update_stats_wait_start(cfs_rq, prev);
790 set_cfs_rq_curr(cfs_rq, NULL);
793 static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
795 struct sched_entity *next;
798 * Dequeue and enqueue the task to update its
799 * position within the tree:
801 dequeue_entity(cfs_rq, curr, 0);
802 enqueue_entity(cfs_rq, curr, 0);
805 * Reschedule if another task tops the current one.
807 next = __pick_next_entity(cfs_rq);
811 __check_preempt_curr_fair(cfs_rq, next, curr,
812 sched_granularity(cfs_rq));
815 /**************************************************
816 * CFS operations on tasks:
819 #ifdef CONFIG_FAIR_GROUP_SCHED
821 /* Walk up scheduling entities hierarchy */
822 #define for_each_sched_entity(se) \
823 for (; se; se = se->parent)
825 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
830 /* runqueue on which this entity is (to be) queued */
831 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
836 /* runqueue "owned" by this group */
837 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
842 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
843 * another cpu ('this_cpu')
845 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
847 /* A later patch will take group into account */
848 return &cpu_rq(this_cpu)->cfs;
851 /* Iterate thr' all leaf cfs_rq's on a runqueue */
852 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
853 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
855 /* Do the two (enqueued) tasks belong to the same group ? */
856 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
858 if (curr->se.cfs_rq == p->se.cfs_rq)
864 #else /* CONFIG_FAIR_GROUP_SCHED */
866 #define for_each_sched_entity(se) \
867 for (; se; se = NULL)
869 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
871 return &task_rq(p)->cfs;
874 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
876 struct task_struct *p = task_of(se);
877 struct rq *rq = task_rq(p);
882 /* runqueue "owned" by this group */
883 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
888 static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
890 return &cpu_rq(this_cpu)->cfs;
893 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
894 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
896 static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
901 #endif /* CONFIG_FAIR_GROUP_SCHED */
904 * The enqueue_task method is called before nr_running is
905 * increased. Here we update the fair scheduling stats and
906 * then put the task into the rbtree:
908 static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
910 struct cfs_rq *cfs_rq;
911 struct sched_entity *se = &p->se;
913 for_each_sched_entity(se) {
916 cfs_rq = cfs_rq_of(se);
917 enqueue_entity(cfs_rq, se, wakeup);
922 * The dequeue_task method is called before nr_running is
923 * decreased. We remove the task from the rbtree and
924 * update the fair scheduling stats:
926 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
928 struct cfs_rq *cfs_rq;
929 struct sched_entity *se = &p->se;
931 for_each_sched_entity(se) {
932 cfs_rq = cfs_rq_of(se);
933 dequeue_entity(cfs_rq, se, sleep);
934 /* Don't dequeue parent if it has other entities besides us */
935 if (cfs_rq->load.weight)
941 * sched_yield() support is very simple - we dequeue and enqueue.
943 * If compat_yield is turned on then we requeue to the end of the tree.
945 static void yield_task_fair(struct rq *rq, struct task_struct *p)
947 struct cfs_rq *cfs_rq = task_cfs_rq(p);
948 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
949 struct sched_entity *rightmost, *se = &p->se;
950 struct rb_node *parent;
953 * Are we the only task in the tree?
955 if (unlikely(cfs_rq->nr_running == 1))
958 if (likely(!sysctl_sched_compat_yield)) {
959 __update_rq_clock(rq);
961 * Dequeue and enqueue the task to update its
962 * position within the tree:
964 dequeue_entity(cfs_rq, &p->se, 0);
965 enqueue_entity(cfs_rq, &p->se, 0);
970 * Find the rightmost entry in the rbtree:
974 link = &parent->rb_right;
977 rightmost = rb_entry(parent, struct sched_entity, run_node);
979 * Already in the rightmost position?
981 if (unlikely(rightmost == se))
985 * Minimally necessary key value to be last in the tree:
987 se->fair_key = rightmost->fair_key + 1;
989 if (cfs_rq->rb_leftmost == &se->run_node)
990 cfs_rq->rb_leftmost = rb_next(&se->run_node);
992 * Relink the task to the rightmost position:
994 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
995 rb_link_node(&se->run_node, parent, link);
996 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
1000 * Preempt the current task with a newly woken task if needed:
1002 static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
1004 struct task_struct *curr = rq->curr;
1005 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
1008 if (unlikely(rt_prio(p->prio))) {
1009 update_rq_clock(rq);
1010 update_curr(cfs_rq);
1015 gran = sysctl_sched_wakeup_granularity;
1017 * Batch tasks prefer throughput over latency:
1019 if (unlikely(p->policy == SCHED_BATCH))
1020 gran = sysctl_sched_batch_wakeup_granularity;
1022 if (is_same_group(curr, p))
1023 __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
1026 static struct task_struct *pick_next_task_fair(struct rq *rq)
1028 struct cfs_rq *cfs_rq = &rq->cfs;
1029 struct sched_entity *se;
1031 if (unlikely(!cfs_rq->nr_running))
1035 se = pick_next_entity(cfs_rq);
1036 cfs_rq = group_cfs_rq(se);
1043 * Account for a descheduled task:
1045 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
1047 struct sched_entity *se = &prev->se;
1048 struct cfs_rq *cfs_rq;
1050 for_each_sched_entity(se) {
1051 cfs_rq = cfs_rq_of(se);
1052 put_prev_entity(cfs_rq, se);
1056 /**************************************************
1057 * Fair scheduling class load-balancing methods:
1061 * Load-balancing iterator. Note: while the runqueue stays locked
1062 * during the whole iteration, the current task might be
1063 * dequeued so the iterator has to be dequeue-safe. Here we
1064 * achieve that by always pre-iterating before returning
1067 static inline struct task_struct *
1068 __load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
1070 struct task_struct *p;
1075 p = rb_entry(curr, struct task_struct, se.run_node);
1076 cfs_rq->rb_load_balance_curr = rb_next(curr);
1081 static struct task_struct *load_balance_start_fair(void *arg)
1083 struct cfs_rq *cfs_rq = arg;
1085 return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
1088 static struct task_struct *load_balance_next_fair(void *arg)
1090 struct cfs_rq *cfs_rq = arg;
1092 return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
1095 #ifdef CONFIG_FAIR_GROUP_SCHED
1096 static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
1098 struct sched_entity *curr;
1099 struct task_struct *p;
1101 if (!cfs_rq->nr_running)
1104 curr = __pick_next_entity(cfs_rq);
1111 static unsigned long
1112 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
1113 unsigned long max_nr_move, unsigned long max_load_move,
1114 struct sched_domain *sd, enum cpu_idle_type idle,
1115 int *all_pinned, int *this_best_prio)
1117 struct cfs_rq *busy_cfs_rq;
1118 unsigned long load_moved, total_nr_moved = 0, nr_moved;
1119 long rem_load_move = max_load_move;
1120 struct rq_iterator cfs_rq_iterator;
1122 cfs_rq_iterator.start = load_balance_start_fair;
1123 cfs_rq_iterator.next = load_balance_next_fair;
1125 for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
1126 #ifdef CONFIG_FAIR_GROUP_SCHED
1127 struct cfs_rq *this_cfs_rq;
1129 unsigned long maxload;
1131 this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
1133 imbalance = busy_cfs_rq->load.weight - this_cfs_rq->load.weight;
1134 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1138 /* Don't pull more than imbalance/2 */
1140 maxload = min(rem_load_move, imbalance);
1142 *this_best_prio = cfs_rq_best_prio(this_cfs_rq);
1144 # define maxload rem_load_move
1146 /* pass busy_cfs_rq argument into
1147 * load_balance_[start|next]_fair iterators
1149 cfs_rq_iterator.arg = busy_cfs_rq;
1150 nr_moved = balance_tasks(this_rq, this_cpu, busiest,
1151 max_nr_move, maxload, sd, idle, all_pinned,
1152 &load_moved, this_best_prio, &cfs_rq_iterator);
1154 total_nr_moved += nr_moved;
1155 max_nr_move -= nr_moved;
1156 rem_load_move -= load_moved;
1158 if (max_nr_move <= 0 || rem_load_move <= 0)
1162 return max_load_move - rem_load_move;
1166 * scheduler tick hitting a task of our scheduling class:
1168 static void task_tick_fair(struct rq *rq, struct task_struct *curr)
1170 struct cfs_rq *cfs_rq;
1171 struct sched_entity *se = &curr->se;
1173 for_each_sched_entity(se) {
1174 cfs_rq = cfs_rq_of(se);
1175 entity_tick(cfs_rq, se);
1180 * Share the fairness runtime between parent and child, thus the
1181 * total amount of pressure for CPU stays equal - new tasks
1182 * get a chance to run but frequent forkers are not allowed to
1183 * monopolize the CPU. Note: the parent runqueue is locked,
1184 * the child is not running yet.
1186 static void task_new_fair(struct rq *rq, struct task_struct *p)
1188 struct cfs_rq *cfs_rq = task_cfs_rq(p);
1189 struct sched_entity *se = &p->se, *curr = cfs_rq_curr(cfs_rq);
1191 sched_info_queued(p);
1193 update_curr(cfs_rq);
1194 update_stats_enqueue(cfs_rq, se);
1196 * Child runs first: we let it run before the parent
1197 * until it reschedules once. We set up the key so that
1198 * it will preempt the parent:
1200 se->fair_key = curr->fair_key -
1201 niced_granularity(curr, sched_granularity(cfs_rq)) - 1;
1203 * The first wait is dominated by the child-runs-first logic,
1204 * so do not credit it with that waiting time yet:
1206 if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
1207 se->wait_start_fair = 0;
1210 * The statistical average of wait_runtime is about
1211 * -granularity/2, so initialize the task with that:
1213 if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
1214 se->wait_runtime = -(sched_granularity(cfs_rq) / 2);
1216 __enqueue_entity(cfs_rq, se);
1217 resched_task(rq->curr);
1220 #ifdef CONFIG_FAIR_GROUP_SCHED
1221 /* Account for a task changing its policy or group.
1223 * This routine is mostly called to set cfs_rq->curr field when a task
1224 * migrates between groups/classes.
1226 static void set_curr_task_fair(struct rq *rq)
1228 struct sched_entity *se = &rq->curr->se;
1230 for_each_sched_entity(se)
1231 set_next_entity(cfs_rq_of(se), se);
1234 static void set_curr_task_fair(struct rq *rq)
1240 * All the scheduling class methods:
1242 struct sched_class fair_sched_class __read_mostly = {
1243 .enqueue_task = enqueue_task_fair,
1244 .dequeue_task = dequeue_task_fair,
1245 .yield_task = yield_task_fair,
1247 .check_preempt_curr = check_preempt_curr_fair,
1249 .pick_next_task = pick_next_task_fair,
1250 .put_prev_task = put_prev_task_fair,
1252 .load_balance = load_balance_fair,
1254 .set_curr_task = set_curr_task_fair,
1255 .task_tick = task_tick_fair,
1256 .task_new = task_new_fair,
1259 #ifdef CONFIG_SCHED_DEBUG
1260 static void print_cfs_stats(struct seq_file *m, int cpu)
1262 struct cfs_rq *cfs_rq;
1264 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
1265 print_cfs_rq(m, cpu, cfs_rq);