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
);
401 * We calculate fair deltas here, so protect against the random effects
402 * of a multiplication overflow by capping it to the runtime limit:
404 #if BITS_PER_LONG == 32
405 static inline unsigned long
406 calc_weighted(unsigned long delta
, unsigned long weight
, int shift
)
408 u64 tmp
= (u64
)delta
* weight
>> shift
;
410 if (unlikely(tmp
> sysctl_sched_runtime_limit
*2))
411 return sysctl_sched_runtime_limit
*2;
415 static inline unsigned long
416 calc_weighted(unsigned long delta
, unsigned long weight
, int shift
)
418 return delta
* weight
>> shift
;
423 * Task is being enqueued - update stats:
425 static void update_stats_enqueue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
430 * Are we enqueueing a waiting task? (for current tasks
431 * a dequeue/enqueue event is a NOP)
433 if (se
!= cfs_rq
->curr
)
434 update_stats_wait_start(cfs_rq
, se
);
438 key
= cfs_rq
->fair_clock
;
441 * Optimize the common nice 0 case:
443 if (likely(se
->load
.weight
== NICE_0_LOAD
)) {
444 key
-= se
->wait_runtime
;
448 if (se
->wait_runtime
< 0) {
449 tmp
= -se
->wait_runtime
;
450 key
+= (tmp
* se
->load
.inv_weight
) >>
451 (WMULT_SHIFT
- NICE_0_SHIFT
);
453 tmp
= se
->wait_runtime
;
454 key
-= (tmp
* se
->load
.inv_weight
) >>
455 (WMULT_SHIFT
- NICE_0_SHIFT
);
463 * Note: must be called with a freshly updated rq->fair_clock.
466 __update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
467 unsigned long delta_fair
)
469 schedstat_set(se
->wait_max
, max(se
->wait_max
,
470 rq_of(cfs_rq
)->clock
- se
->wait_start
));
472 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
473 delta_fair
= calc_weighted(delta_fair
, se
->load
.weight
,
476 add_wait_runtime(cfs_rq
, se
, delta_fair
);
480 update_stats_wait_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
482 unsigned long delta_fair
;
484 if (unlikely(!se
->wait_start_fair
))
487 delta_fair
= (unsigned long)min((u64
)(2*sysctl_sched_runtime_limit
),
488 (u64
)(cfs_rq
->fair_clock
- se
->wait_start_fair
));
490 __update_stats_wait_end(cfs_rq
, se
, delta_fair
);
492 se
->wait_start_fair
= 0;
493 schedstat_set(se
->wait_start
, 0);
497 update_stats_dequeue(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
501 * Mark the end of the wait period if dequeueing a
504 if (se
!= cfs_rq
->curr
)
505 update_stats_wait_end(cfs_rq
, se
);
509 * We are picking a new current task - update its stats:
512 update_stats_curr_start(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
515 * We are starting a new run period:
517 se
->exec_start
= rq_of(cfs_rq
)->clock
;
521 * We are descheduling a task - update its stats:
524 update_stats_curr_end(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
529 /**************************************************
530 * Scheduling class queueing methods:
533 static void __enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
534 unsigned long delta_fair
)
536 unsigned long load
= cfs_rq
->load
.weight
;
540 * Do not boost sleepers if there's too much bonus 'in flight'
543 if (unlikely(cfs_rq
->sleeper_bonus
> sysctl_sched_runtime_limit
))
546 if (sched_feat(SLEEPER_LOAD_AVG
))
547 load
= rq_of(cfs_rq
)->cpu_load
[2];
550 * Fix up delta_fair with the effect of us running
551 * during the whole sleep period:
553 if (sched_feat(SLEEPER_AVG
))
554 delta_fair
= div64_likely32((u64
)delta_fair
* load
,
555 load
+ se
->load
.weight
);
557 if (unlikely(se
->load
.weight
!= NICE_0_LOAD
))
558 delta_fair
= calc_weighted(delta_fair
, se
->load
.weight
,
561 prev_runtime
= se
->wait_runtime
;
562 __add_wait_runtime(cfs_rq
, se
, delta_fair
);
563 delta_fair
= se
->wait_runtime
- prev_runtime
;
566 * Track the amount of bonus we've given to sleepers:
568 cfs_rq
->sleeper_bonus
+= delta_fair
;
571 static void enqueue_sleeper(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
573 struct task_struct
*tsk
= task_of(se
);
574 unsigned long delta_fair
;
576 if ((entity_is_task(se
) && tsk
->policy
== SCHED_BATCH
) ||
577 !sched_feat(FAIR_SLEEPERS
))
580 delta_fair
= (unsigned long)min((u64
)(2*sysctl_sched_runtime_limit
),
581 (u64
)(cfs_rq
->fair_clock
- se
->sleep_start_fair
));
583 __enqueue_sleeper(cfs_rq
, se
, delta_fair
);
585 se
->sleep_start_fair
= 0;
587 #ifdef CONFIG_SCHEDSTATS
588 if (se
->sleep_start
) {
589 u64 delta
= rq_of(cfs_rq
)->clock
- se
->sleep_start
;
594 if (unlikely(delta
> se
->sleep_max
))
595 se
->sleep_max
= delta
;
598 se
->sum_sleep_runtime
+= delta
;
600 if (se
->block_start
) {
601 u64 delta
= rq_of(cfs_rq
)->clock
- se
->block_start
;
606 if (unlikely(delta
> se
->block_max
))
607 se
->block_max
= delta
;
610 se
->sum_sleep_runtime
+= delta
;
613 * Blocking time is in units of nanosecs, so shift by 20 to
614 * get a milliseconds-range estimation of the amount of
615 * time that the task spent sleeping:
617 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
618 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(tsk
),
626 enqueue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int wakeup
)
629 * Update the fair clock.
634 enqueue_sleeper(cfs_rq
, se
);
636 update_stats_enqueue(cfs_rq
, se
);
637 __enqueue_entity(cfs_rq
, se
);
641 dequeue_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
, int sleep
)
643 update_stats_dequeue(cfs_rq
, se
);
645 se
->sleep_start_fair
= cfs_rq
->fair_clock
;
646 #ifdef CONFIG_SCHEDSTATS
647 if (entity_is_task(se
)) {
648 struct task_struct
*tsk
= task_of(se
);
650 if (tsk
->state
& TASK_INTERRUPTIBLE
)
651 se
->sleep_start
= rq_of(cfs_rq
)->clock
;
652 if (tsk
->state
& TASK_UNINTERRUPTIBLE
)
653 se
->block_start
= rq_of(cfs_rq
)->clock
;
657 __dequeue_entity(cfs_rq
, se
);
661 * Preempt the current task with a newly woken task if needed:
664 __check_preempt_curr_fair(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
,
665 struct sched_entity
*curr
, unsigned long granularity
)
667 s64 __delta
= curr
->fair_key
- se
->fair_key
;
668 unsigned long ideal_runtime
, delta_exec
;
671 * ideal_runtime is compared against sum_exec_runtime, which is
672 * walltime, hence do not scale.
674 ideal_runtime
= max(sysctl_sched_latency
/ cfs_rq
->nr_running
,
675 (unsigned long)sysctl_sched_min_granularity
);
678 * If we executed more than what the latency constraint suggests,
679 * reduce the rescheduling granularity. This way the total latency
680 * of how much a task is not scheduled converges to
681 * sysctl_sched_latency:
683 delta_exec
= curr
->sum_exec_runtime
- curr
->prev_sum_exec_runtime
;
684 if (delta_exec
> ideal_runtime
)
688 * Take scheduling granularity into account - do not
689 * preempt the current task unless the best task has
690 * a larger than sched_granularity fairness advantage:
692 * scale granularity as key space is in fair_clock.
694 if (__delta
> niced_granularity(curr
, granularity
))
695 resched_task(rq_of(cfs_rq
)->curr
);
699 set_next_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*se
)
702 * Any task has to be enqueued before it get to execute on
703 * a CPU. So account for the time it spent waiting on the
704 * runqueue. (note, here we rely on pick_next_task() having
705 * done a put_prev_task_fair() shortly before this, which
706 * updated rq->fair_clock - used by update_stats_wait_end())
708 update_stats_wait_end(cfs_rq
, se
);
709 update_stats_curr_start(cfs_rq
, se
);
711 #ifdef CONFIG_SCHEDSTATS
713 * Track our maximum slice length, if the CPU's load is at
714 * least twice that of our own weight (i.e. dont track it
715 * when there are only lesser-weight tasks around):
717 if (rq_of(cfs_rq
)->ls
.load
.weight
>= 2*se
->load
.weight
) {
718 se
->slice_max
= max(se
->slice_max
,
719 se
->sum_exec_runtime
- se
->prev_sum_exec_runtime
);
722 se
->prev_sum_exec_runtime
= se
->sum_exec_runtime
;
725 static struct sched_entity
*pick_next_entity(struct cfs_rq
*cfs_rq
)
727 struct sched_entity
*se
= __pick_next_entity(cfs_rq
);
729 set_next_entity(cfs_rq
, se
);
734 static void put_prev_entity(struct cfs_rq
*cfs_rq
, struct sched_entity
*prev
)
737 * If still on the runqueue then deactivate_task()
738 * was not called and update_curr() has to be done:
743 update_stats_curr_end(cfs_rq
, prev
);
746 update_stats_wait_start(cfs_rq
, prev
);
750 static void entity_tick(struct cfs_rq
*cfs_rq
, struct sched_entity
*curr
)
752 struct sched_entity
*next
;
755 * Dequeue and enqueue the task to update its
756 * position within the tree:
758 dequeue_entity(cfs_rq
, curr
, 0);
759 enqueue_entity(cfs_rq
, curr
, 0);
762 * Reschedule if another task tops the current one.
764 next
= __pick_next_entity(cfs_rq
);
768 __check_preempt_curr_fair(cfs_rq
, next
, curr
,
769 sched_granularity(cfs_rq
));
772 /**************************************************
773 * CFS operations on tasks:
776 #ifdef CONFIG_FAIR_GROUP_SCHED
778 /* Walk up scheduling entities hierarchy */
779 #define for_each_sched_entity(se) \
780 for (; se; se = se->parent)
782 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
787 /* runqueue on which this entity is (to be) queued */
788 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
793 /* runqueue "owned" by this group */
794 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
799 /* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
800 * another cpu ('this_cpu')
802 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
804 /* A later patch will take group into account */
805 return &cpu_rq(this_cpu
)->cfs
;
808 /* Iterate thr' all leaf cfs_rq's on a runqueue */
809 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
810 list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
812 /* Do the two (enqueued) tasks belong to the same group ? */
813 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
815 if (curr
->se
.cfs_rq
== p
->se
.cfs_rq
)
821 #else /* CONFIG_FAIR_GROUP_SCHED */
823 #define for_each_sched_entity(se) \
824 for (; se; se = NULL)
826 static inline struct cfs_rq
*task_cfs_rq(struct task_struct
*p
)
828 return &task_rq(p
)->cfs
;
831 static inline struct cfs_rq
*cfs_rq_of(struct sched_entity
*se
)
833 struct task_struct
*p
= task_of(se
);
834 struct rq
*rq
= task_rq(p
);
839 /* runqueue "owned" by this group */
840 static inline struct cfs_rq
*group_cfs_rq(struct sched_entity
*grp
)
845 static inline struct cfs_rq
*cpu_cfs_rq(struct cfs_rq
*cfs_rq
, int this_cpu
)
847 return &cpu_rq(this_cpu
)->cfs
;
850 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
851 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
853 static inline int is_same_group(struct task_struct
*curr
, struct task_struct
*p
)
858 #endif /* CONFIG_FAIR_GROUP_SCHED */
861 * The enqueue_task method is called before nr_running is
862 * increased. Here we update the fair scheduling stats and
863 * then put the task into the rbtree:
865 static void enqueue_task_fair(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
867 struct cfs_rq
*cfs_rq
;
868 struct sched_entity
*se
= &p
->se
;
870 for_each_sched_entity(se
) {
873 cfs_rq
= cfs_rq_of(se
);
874 enqueue_entity(cfs_rq
, se
, wakeup
);
879 * The dequeue_task method is called before nr_running is
880 * decreased. We remove the task from the rbtree and
881 * update the fair scheduling stats:
883 static void dequeue_task_fair(struct rq
*rq
, struct task_struct
*p
, int sleep
)
885 struct cfs_rq
*cfs_rq
;
886 struct sched_entity
*se
= &p
->se
;
888 for_each_sched_entity(se
) {
889 cfs_rq
= cfs_rq_of(se
);
890 dequeue_entity(cfs_rq
, se
, sleep
);
891 /* Don't dequeue parent if it has other entities besides us */
892 if (cfs_rq
->load
.weight
)
898 * sched_yield() support is very simple - we dequeue and enqueue.
900 * If compat_yield is turned on then we requeue to the end of the tree.
902 static void yield_task_fair(struct rq
*rq
, struct task_struct
*p
)
904 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
905 struct rb_node
**link
= &cfs_rq
->tasks_timeline
.rb_node
;
906 struct sched_entity
*rightmost
, *se
= &p
->se
;
907 struct rb_node
*parent
;
910 * Are we the only task in the tree?
912 if (unlikely(cfs_rq
->nr_running
== 1))
915 if (likely(!sysctl_sched_compat_yield
)) {
916 __update_rq_clock(rq
);
918 * Dequeue and enqueue the task to update its
919 * position within the tree:
921 dequeue_entity(cfs_rq
, &p
->se
, 0);
922 enqueue_entity(cfs_rq
, &p
->se
, 0);
927 * Find the rightmost entry in the rbtree:
931 link
= &parent
->rb_right
;
934 rightmost
= rb_entry(parent
, struct sched_entity
, run_node
);
936 * Already in the rightmost position?
938 if (unlikely(rightmost
== se
))
942 * Minimally necessary key value to be last in the tree:
944 se
->fair_key
= rightmost
->fair_key
+ 1;
946 if (cfs_rq
->rb_leftmost
== &se
->run_node
)
947 cfs_rq
->rb_leftmost
= rb_next(&se
->run_node
);
949 * Relink the task to the rightmost position:
951 rb_erase(&se
->run_node
, &cfs_rq
->tasks_timeline
);
952 rb_link_node(&se
->run_node
, parent
, link
);
953 rb_insert_color(&se
->run_node
, &cfs_rq
->tasks_timeline
);
957 * Preempt the current task with a newly woken task if needed:
959 static void check_preempt_curr_fair(struct rq
*rq
, struct task_struct
*p
)
961 struct task_struct
*curr
= rq
->curr
;
962 struct cfs_rq
*cfs_rq
= task_cfs_rq(curr
);
965 if (unlikely(rt_prio(p
->prio
))) {
972 gran
= sysctl_sched_wakeup_granularity
;
974 * Batch tasks prefer throughput over latency:
976 if (unlikely(p
->policy
== SCHED_BATCH
))
977 gran
= sysctl_sched_batch_wakeup_granularity
;
979 if (is_same_group(curr
, p
))
980 __check_preempt_curr_fair(cfs_rq
, &p
->se
, &curr
->se
, gran
);
983 static struct task_struct
*pick_next_task_fair(struct rq
*rq
)
985 struct cfs_rq
*cfs_rq
= &rq
->cfs
;
986 struct sched_entity
*se
;
988 if (unlikely(!cfs_rq
->nr_running
))
992 se
= pick_next_entity(cfs_rq
);
993 cfs_rq
= group_cfs_rq(se
);
1000 * Account for a descheduled task:
1002 static void put_prev_task_fair(struct rq
*rq
, struct task_struct
*prev
)
1004 struct sched_entity
*se
= &prev
->se
;
1005 struct cfs_rq
*cfs_rq
;
1007 for_each_sched_entity(se
) {
1008 cfs_rq
= cfs_rq_of(se
);
1009 put_prev_entity(cfs_rq
, se
);
1013 /**************************************************
1014 * Fair scheduling class load-balancing methods:
1018 * Load-balancing iterator. Note: while the runqueue stays locked
1019 * during the whole iteration, the current task might be
1020 * dequeued so the iterator has to be dequeue-safe. Here we
1021 * achieve that by always pre-iterating before returning
1024 static inline struct task_struct
*
1025 __load_balance_iterator(struct cfs_rq
*cfs_rq
, struct rb_node
*curr
)
1027 struct task_struct
*p
;
1032 p
= rb_entry(curr
, struct task_struct
, se
.run_node
);
1033 cfs_rq
->rb_load_balance_curr
= rb_next(curr
);
1038 static struct task_struct
*load_balance_start_fair(void *arg
)
1040 struct cfs_rq
*cfs_rq
= arg
;
1042 return __load_balance_iterator(cfs_rq
, first_fair(cfs_rq
));
1045 static struct task_struct
*load_balance_next_fair(void *arg
)
1047 struct cfs_rq
*cfs_rq
= arg
;
1049 return __load_balance_iterator(cfs_rq
, cfs_rq
->rb_load_balance_curr
);
1052 #ifdef CONFIG_FAIR_GROUP_SCHED
1053 static int cfs_rq_best_prio(struct cfs_rq
*cfs_rq
)
1055 struct sched_entity
*curr
;
1056 struct task_struct
*p
;
1058 if (!cfs_rq
->nr_running
)
1061 curr
= __pick_next_entity(cfs_rq
);
1068 static unsigned long
1069 load_balance_fair(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1070 unsigned long max_nr_move
, unsigned long max_load_move
,
1071 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1072 int *all_pinned
, int *this_best_prio
)
1074 struct cfs_rq
*busy_cfs_rq
;
1075 unsigned long load_moved
, total_nr_moved
= 0, nr_moved
;
1076 long rem_load_move
= max_load_move
;
1077 struct rq_iterator cfs_rq_iterator
;
1079 cfs_rq_iterator
.start
= load_balance_start_fair
;
1080 cfs_rq_iterator
.next
= load_balance_next_fair
;
1082 for_each_leaf_cfs_rq(busiest
, busy_cfs_rq
) {
1083 #ifdef CONFIG_FAIR_GROUP_SCHED
1084 struct cfs_rq
*this_cfs_rq
;
1086 unsigned long maxload
;
1088 this_cfs_rq
= cpu_cfs_rq(busy_cfs_rq
, this_cpu
);
1090 imbalance
= busy_cfs_rq
->load
.weight
- this_cfs_rq
->load
.weight
;
1091 /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
1095 /* Don't pull more than imbalance/2 */
1097 maxload
= min(rem_load_move
, imbalance
);
1099 *this_best_prio
= cfs_rq_best_prio(this_cfs_rq
);
1101 # define maxload rem_load_move
1103 /* pass busy_cfs_rq argument into
1104 * load_balance_[start|next]_fair iterators
1106 cfs_rq_iterator
.arg
= busy_cfs_rq
;
1107 nr_moved
= balance_tasks(this_rq
, this_cpu
, busiest
,
1108 max_nr_move
, maxload
, sd
, idle
, all_pinned
,
1109 &load_moved
, this_best_prio
, &cfs_rq_iterator
);
1111 total_nr_moved
+= nr_moved
;
1112 max_nr_move
-= nr_moved
;
1113 rem_load_move
-= load_moved
;
1115 if (max_nr_move
<= 0 || rem_load_move
<= 0)
1119 return max_load_move
- rem_load_move
;
1123 * scheduler tick hitting a task of our scheduling class:
1125 static void task_tick_fair(struct rq
*rq
, struct task_struct
*curr
)
1127 struct cfs_rq
*cfs_rq
;
1128 struct sched_entity
*se
= &curr
->se
;
1130 for_each_sched_entity(se
) {
1131 cfs_rq
= cfs_rq_of(se
);
1132 entity_tick(cfs_rq
, se
);
1137 * Share the fairness runtime between parent and child, thus the
1138 * total amount of pressure for CPU stays equal - new tasks
1139 * get a chance to run but frequent forkers are not allowed to
1140 * monopolize the CPU. Note: the parent runqueue is locked,
1141 * the child is not running yet.
1143 static void task_new_fair(struct rq
*rq
, struct task_struct
*p
)
1145 struct cfs_rq
*cfs_rq
= task_cfs_rq(p
);
1146 struct sched_entity
*se
= &p
->se
, *curr
= cfs_rq
->curr
;
1148 sched_info_queued(p
);
1150 update_curr(cfs_rq
);
1151 update_stats_enqueue(cfs_rq
, se
);
1153 * Child runs first: we let it run before the parent
1154 * until it reschedules once. We set up the key so that
1155 * it will preempt the parent:
1157 se
->fair_key
= curr
->fair_key
-
1158 niced_granularity(curr
, sched_granularity(cfs_rq
)) - 1;
1160 * The first wait is dominated by the child-runs-first logic,
1161 * so do not credit it with that waiting time yet:
1163 if (sched_feat(SKIP_INITIAL
))
1164 se
->wait_start_fair
= 0;
1167 * The statistical average of wait_runtime is about
1168 * -granularity/2, so initialize the task with that:
1170 if (sched_feat(START_DEBIT
))
1171 se
->wait_runtime
= -(sched_granularity(cfs_rq
) / 2);
1173 __enqueue_entity(cfs_rq
, se
);
1174 resched_task(rq
->curr
);
1177 #ifdef CONFIG_FAIR_GROUP_SCHED
1178 /* Account for a task changing its policy or group.
1180 * This routine is mostly called to set cfs_rq->curr field when a task
1181 * migrates between groups/classes.
1183 static void set_curr_task_fair(struct rq
*rq
)
1185 struct sched_entity
*se
= &rq
->curr
->se
;
1187 for_each_sched_entity(se
)
1188 set_next_entity(cfs_rq_of(se
), se
);
1191 static void set_curr_task_fair(struct rq
*rq
)
1197 * All the scheduling class methods:
1199 struct sched_class fair_sched_class __read_mostly
= {
1200 .enqueue_task
= enqueue_task_fair
,
1201 .dequeue_task
= dequeue_task_fair
,
1202 .yield_task
= yield_task_fair
,
1204 .check_preempt_curr
= check_preempt_curr_fair
,
1206 .pick_next_task
= pick_next_task_fair
,
1207 .put_prev_task
= put_prev_task_fair
,
1209 .load_balance
= load_balance_fair
,
1211 .set_curr_task
= set_curr_task_fair
,
1212 .task_tick
= task_tick_fair
,
1213 .task_new
= task_new_fair
,
1216 #ifdef CONFIG_SCHED_DEBUG
1217 static void print_cfs_stats(struct seq_file
*m
, int cpu
)
1219 struct cfs_rq
*cfs_rq
;
1221 for_each_leaf_cfs_rq(cpu_rq(cpu
), cfs_rq
)
1222 print_cfs_rq(m
, cpu
, cfs_rq
);