aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/sched_fair.c
diff options
context:
space:
mode:
authorPeter Zijlstra <a.p.zijlstra@chello.nl>2011-11-15 11:14:39 -0500
committerIngo Molnar <mingo@elte.hu>2011-11-17 06:20:22 -0500
commit391e43da797a96aeb65410281891f6d0b0e9611c (patch)
tree0ce6784525a5a8f75b377170cf1a7d60abccea29 /kernel/sched_fair.c
parent029632fbb7b7c9d85063cc9eb470de6c54873df3 (diff)
sched: Move all scheduler bits into kernel/sched/
There's too many sched*.[ch] files in kernel/, give them their own directory. (No code changed, other than Makefile glue added.) Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'kernel/sched_fair.c')
-rw-r--r--kernel/sched_fair.c5601
1 files changed, 0 insertions, 5601 deletions
diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
deleted file mode 100644
index cd3b64219d9f..000000000000
--- a/kernel/sched_fair.c
+++ /dev/null
@@ -1,5601 +0,0 @@
1/*
2 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
3 *
4 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
5 *
6 * Interactivity improvements by Mike Galbraith
7 * (C) 2007 Mike Galbraith <efault@gmx.de>
8 *
9 * Various enhancements by Dmitry Adamushko.
10 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
11 *
12 * Group scheduling enhancements by Srivatsa Vaddagiri
13 * Copyright IBM Corporation, 2007
14 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
15 *
16 * Scaled math optimizations by Thomas Gleixner
17 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
18 *
19 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
20 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
21 */
22
23#include <linux/latencytop.h>
24#include <linux/sched.h>
25#include <linux/cpumask.h>
26#include <linux/slab.h>
27#include <linux/profile.h>
28#include <linux/interrupt.h>
29
30#include <trace/events/sched.h>
31
32#include "sched.h"
33
34/*
35 * Targeted preemption latency for CPU-bound tasks:
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
37 *
38 * NOTE: this latency value is not the same as the concept of
39 * 'timeslice length' - timeslices in CFS are of variable length
40 * and have no persistent notion like in traditional, time-slice
41 * based scheduling concepts.
42 *
43 * (to see the precise effective timeslice length of your workload,
44 * run vmstat and monitor the context-switches (cs) field)
45 */
46unsigned int sysctl_sched_latency = 6000000ULL;
47unsigned int normalized_sysctl_sched_latency = 6000000ULL;
48
49/*
50 * The initial- and re-scaling of tunables is configurable
51 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
52 *
53 * Options are:
54 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
55 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
56 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
57 */
58enum sched_tunable_scaling sysctl_sched_tunable_scaling
59 = SCHED_TUNABLESCALING_LOG;
60
61/*
62 * Minimal preemption granularity for CPU-bound tasks:
63 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
64 */
65unsigned int sysctl_sched_min_granularity = 750000ULL;
66unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
67
68/*
69 * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
70 */
71static unsigned int sched_nr_latency = 8;
72
73/*
74 * After fork, child runs first. If set to 0 (default) then
75 * parent will (try to) run first.
76 */
77unsigned int sysctl_sched_child_runs_first __read_mostly;
78
79/*
80 * SCHED_OTHER wake-up granularity.
81 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 *
83 * This option delays the preemption effects of decoupled workloads
84 * and reduces their over-scheduling. Synchronous workloads will still
85 * have immediate wakeup/sleep latencies.
86 */
87unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
88unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
89
90const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91
92/*
93 * The exponential sliding window over which load is averaged for shares
94 * distribution.
95 * (default: 10msec)
96 */
97unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
98
99#ifdef CONFIG_CFS_BANDWIDTH
100/*
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
103 *
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
107 *
108 * default: 5 msec, units: microseconds
109 */
110unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
111#endif
112
113/*
114 * Increase the granularity value when there are more CPUs,
115 * because with more CPUs the 'effective latency' as visible
116 * to users decreases. But the relationship is not linear,
117 * so pick a second-best guess by going with the log2 of the
118 * number of CPUs.
119 *
120 * This idea comes from the SD scheduler of Con Kolivas:
121 */
122static int get_update_sysctl_factor(void)
123{
124 unsigned int cpus = min_t(int, num_online_cpus(), 8);
125 unsigned int factor;
126
127 switch (sysctl_sched_tunable_scaling) {
128 case SCHED_TUNABLESCALING_NONE:
129 factor = 1;
130 break;
131 case SCHED_TUNABLESCALING_LINEAR:
132 factor = cpus;
133 break;
134 case SCHED_TUNABLESCALING_LOG:
135 default:
136 factor = 1 + ilog2(cpus);
137 break;
138 }
139
140 return factor;
141}
142
143static void update_sysctl(void)
144{
145 unsigned int factor = get_update_sysctl_factor();
146
147#define SET_SYSCTL(name) \
148 (sysctl_##name = (factor) * normalized_sysctl_##name)
149 SET_SYSCTL(sched_min_granularity);
150 SET_SYSCTL(sched_latency);
151 SET_SYSCTL(sched_wakeup_granularity);
152#undef SET_SYSCTL
153}
154
155void sched_init_granularity(void)
156{
157 update_sysctl();
158}
159
160#if BITS_PER_LONG == 32
161# define WMULT_CONST (~0UL)
162#else
163# define WMULT_CONST (1UL << 32)
164#endif
165
166#define WMULT_SHIFT 32
167
168/*
169 * Shift right and round:
170 */
171#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
172
173/*
174 * delta *= weight / lw
175 */
176static unsigned long
177calc_delta_mine(unsigned long delta_exec, unsigned long weight,
178 struct load_weight *lw)
179{
180 u64 tmp;
181
182 /*
183 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
184 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
185 * 2^SCHED_LOAD_RESOLUTION.
186 */
187 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
188 tmp = (u64)delta_exec * scale_load_down(weight);
189 else
190 tmp = (u64)delta_exec;
191
192 if (!lw->inv_weight) {
193 unsigned long w = scale_load_down(lw->weight);
194
195 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
196 lw->inv_weight = 1;
197 else if (unlikely(!w))
198 lw->inv_weight = WMULT_CONST;
199 else
200 lw->inv_weight = WMULT_CONST / w;
201 }
202
203 /*
204 * Check whether we'd overflow the 64-bit multiplication:
205 */
206 if (unlikely(tmp > WMULT_CONST))
207 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
208 WMULT_SHIFT/2);
209 else
210 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
211
212 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
213}
214
215
216const struct sched_class fair_sched_class;
217
218/**************************************************************
219 * CFS operations on generic schedulable entities:
220 */
221
222#ifdef CONFIG_FAIR_GROUP_SCHED
223
224/* cpu runqueue to which this cfs_rq is attached */
225static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
226{
227 return cfs_rq->rq;
228}
229
230/* An entity is a task if it doesn't "own" a runqueue */
231#define entity_is_task(se) (!se->my_q)
232
233static inline struct task_struct *task_of(struct sched_entity *se)
234{
235#ifdef CONFIG_SCHED_DEBUG
236 WARN_ON_ONCE(!entity_is_task(se));
237#endif
238 return container_of(se, struct task_struct, se);
239}
240
241/* Walk up scheduling entities hierarchy */
242#define for_each_sched_entity(se) \
243 for (; se; se = se->parent)
244
245static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
246{
247 return p->se.cfs_rq;
248}
249
250/* runqueue on which this entity is (to be) queued */
251static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
252{
253 return se->cfs_rq;
254}
255
256/* runqueue "owned" by this group */
257static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
258{
259 return grp->my_q;
260}
261
262static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
263{
264 if (!cfs_rq->on_list) {
265 /*
266 * Ensure we either appear before our parent (if already
267 * enqueued) or force our parent to appear after us when it is
268 * enqueued. The fact that we always enqueue bottom-up
269 * reduces this to two cases.
270 */
271 if (cfs_rq->tg->parent &&
272 cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
273 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
274 &rq_of(cfs_rq)->leaf_cfs_rq_list);
275 } else {
276 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
277 &rq_of(cfs_rq)->leaf_cfs_rq_list);
278 }
279
280 cfs_rq->on_list = 1;
281 }
282}
283
284static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
285{
286 if (cfs_rq->on_list) {
287 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
288 cfs_rq->on_list = 0;
289 }
290}
291
292/* Iterate thr' all leaf cfs_rq's on a runqueue */
293#define for_each_leaf_cfs_rq(rq, cfs_rq) \
294 list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
295
296/* Do the two (enqueued) entities belong to the same group ? */
297static inline int
298is_same_group(struct sched_entity *se, struct sched_entity *pse)
299{
300 if (se->cfs_rq == pse->cfs_rq)
301 return 1;
302
303 return 0;
304}
305
306static inline struct sched_entity *parent_entity(struct sched_entity *se)
307{
308 return se->parent;
309}
310
311/* return depth at which a sched entity is present in the hierarchy */
312static inline int depth_se(struct sched_entity *se)
313{
314 int depth = 0;
315
316 for_each_sched_entity(se)
317 depth++;
318
319 return depth;
320}
321
322static void
323find_matching_se(struct sched_entity **se, struct sched_entity **pse)
324{
325 int se_depth, pse_depth;
326
327 /*
328 * preemption test can be made between sibling entities who are in the
329 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
330 * both tasks until we find their ancestors who are siblings of common
331 * parent.
332 */
333
334 /* First walk up until both entities are at same depth */
335 se_depth = depth_se(*se);
336 pse_depth = depth_se(*pse);
337
338 while (se_depth > pse_depth) {
339 se_depth--;
340 *se = parent_entity(*se);
341 }
342
343 while (pse_depth > se_depth) {
344 pse_depth--;
345 *pse = parent_entity(*pse);
346 }
347
348 while (!is_same_group(*se, *pse)) {
349 *se = parent_entity(*se);
350 *pse = parent_entity(*pse);
351 }
352}
353
354#else /* !CONFIG_FAIR_GROUP_SCHED */
355
356static inline struct task_struct *task_of(struct sched_entity *se)
357{
358 return container_of(se, struct task_struct, se);
359}
360
361static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
362{
363 return container_of(cfs_rq, struct rq, cfs);
364}
365
366#define entity_is_task(se) 1
367
368#define for_each_sched_entity(se) \
369 for (; se; se = NULL)
370
371static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
372{
373 return &task_rq(p)->cfs;
374}
375
376static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
377{
378 struct task_struct *p = task_of(se);
379 struct rq *rq = task_rq(p);
380
381 return &rq->cfs;
382}
383
384/* runqueue "owned" by this group */
385static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
386{
387 return NULL;
388}
389
390static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
391{
392}
393
394static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
395{
396}
397
398#define for_each_leaf_cfs_rq(rq, cfs_rq) \
399 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
400
401static inline int
402is_same_group(struct sched_entity *se, struct sched_entity *pse)
403{
404 return 1;
405}
406
407static inline struct sched_entity *parent_entity(struct sched_entity *se)
408{
409 return NULL;
410}
411
412static inline void
413find_matching_se(struct sched_entity **se, struct sched_entity **pse)
414{
415}
416
417#endif /* CONFIG_FAIR_GROUP_SCHED */
418
419static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
420 unsigned long delta_exec);
421
422/**************************************************************
423 * Scheduling class tree data structure manipulation methods:
424 */
425
426static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
427{
428 s64 delta = (s64)(vruntime - min_vruntime);
429 if (delta > 0)
430 min_vruntime = vruntime;
431
432 return min_vruntime;
433}
434
435static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
436{
437 s64 delta = (s64)(vruntime - min_vruntime);
438 if (delta < 0)
439 min_vruntime = vruntime;
440
441 return min_vruntime;
442}
443
444static inline int entity_before(struct sched_entity *a,
445 struct sched_entity *b)
446{
447 return (s64)(a->vruntime - b->vruntime) < 0;
448}
449
450static void update_min_vruntime(struct cfs_rq *cfs_rq)
451{
452 u64 vruntime = cfs_rq->min_vruntime;
453
454 if (cfs_rq->curr)
455 vruntime = cfs_rq->curr->vruntime;
456
457 if (cfs_rq->rb_leftmost) {
458 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
459 struct sched_entity,
460 run_node);
461
462 if (!cfs_rq->curr)
463 vruntime = se->vruntime;
464 else
465 vruntime = min_vruntime(vruntime, se->vruntime);
466 }
467
468 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
469#ifndef CONFIG_64BIT
470 smp_wmb();
471 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
472#endif
473}
474
475/*
476 * Enqueue an entity into the rb-tree:
477 */
478static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
479{
480 struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
481 struct rb_node *parent = NULL;
482 struct sched_entity *entry;
483 int leftmost = 1;
484
485 /*
486 * Find the right place in the rbtree:
487 */
488 while (*link) {
489 parent = *link;
490 entry = rb_entry(parent, struct sched_entity, run_node);
491 /*
492 * We dont care about collisions. Nodes with
493 * the same key stay together.
494 */
495 if (entity_before(se, entry)) {
496 link = &parent->rb_left;
497 } else {
498 link = &parent->rb_right;
499 leftmost = 0;
500 }
501 }
502
503 /*
504 * Maintain a cache of leftmost tree entries (it is frequently
505 * used):
506 */
507 if (leftmost)
508 cfs_rq->rb_leftmost = &se->run_node;
509
510 rb_link_node(&se->run_node, parent, link);
511 rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
512}
513
514static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
515{
516 if (cfs_rq->rb_leftmost == &se->run_node) {
517 struct rb_node *next_node;
518
519 next_node = rb_next(&se->run_node);
520 cfs_rq->rb_leftmost = next_node;
521 }
522
523 rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
524}
525
526struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
527{
528 struct rb_node *left = cfs_rq->rb_leftmost;
529
530 if (!left)
531 return NULL;
532
533 return rb_entry(left, struct sched_entity, run_node);
534}
535
536static struct sched_entity *__pick_next_entity(struct sched_entity *se)
537{
538 struct rb_node *next = rb_next(&se->run_node);
539
540 if (!next)
541 return NULL;
542
543 return rb_entry(next, struct sched_entity, run_node);
544}
545
546#ifdef CONFIG_SCHED_DEBUG
547struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
548{
549 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
550
551 if (!last)
552 return NULL;
553
554 return rb_entry(last, struct sched_entity, run_node);
555}
556
557/**************************************************************
558 * Scheduling class statistics methods:
559 */
560
561int sched_proc_update_handler(struct ctl_table *table, int write,
562 void __user *buffer, size_t *lenp,
563 loff_t *ppos)
564{
565 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
566 int factor = get_update_sysctl_factor();
567
568 if (ret || !write)
569 return ret;
570
571 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
572 sysctl_sched_min_granularity);
573
574#define WRT_SYSCTL(name) \
575 (normalized_sysctl_##name = sysctl_##name / (factor))
576 WRT_SYSCTL(sched_min_granularity);
577 WRT_SYSCTL(sched_latency);
578 WRT_SYSCTL(sched_wakeup_granularity);
579#undef WRT_SYSCTL
580
581 return 0;
582}
583#endif
584
585/*
586 * delta /= w
587 */
588static inline unsigned long
589calc_delta_fair(unsigned long delta, struct sched_entity *se)
590{
591 if (unlikely(se->load.weight != NICE_0_LOAD))
592 delta = calc_delta_mine(delta, NICE_0_LOAD, &se->load);
593
594 return delta;
595}
596
597/*
598 * The idea is to set a period in which each task runs once.
599 *
600 * When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
601 * this period because otherwise the slices get too small.
602 *
603 * p = (nr <= nl) ? l : l*nr/nl
604 */
605static u64 __sched_period(unsigned long nr_running)
606{
607 u64 period = sysctl_sched_latency;
608 unsigned long nr_latency = sched_nr_latency;
609
610 if (unlikely(nr_running > nr_latency)) {
611 period = sysctl_sched_min_granularity;
612 period *= nr_running;
613 }
614
615 return period;
616}
617
618/*
619 * We calculate the wall-time slice from the period by taking a part
620 * proportional to the weight.
621 *
622 * s = p*P[w/rw]
623 */
624static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
625{
626 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
627
628 for_each_sched_entity(se) {
629 struct load_weight *load;
630 struct load_weight lw;
631
632 cfs_rq = cfs_rq_of(se);
633 load = &cfs_rq->load;
634
635 if (unlikely(!se->on_rq)) {
636 lw = cfs_rq->load;
637
638 update_load_add(&lw, se->load.weight);
639 load = &lw;
640 }
641 slice = calc_delta_mine(slice, se->load.weight, load);
642 }
643 return slice;
644}
645
646/*
647 * We calculate the vruntime slice of a to be inserted task
648 *
649 * vs = s/w
650 */
651static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
652{
653 return calc_delta_fair(sched_slice(cfs_rq, se), se);
654}
655
656static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
657static void update_cfs_shares(struct cfs_rq *cfs_rq);
658
659/*
660 * Update the current task's runtime statistics. Skip current tasks that
661 * are not in our scheduling class.
662 */
663static inline void
664__update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
665 unsigned long delta_exec)
666{
667 unsigned long delta_exec_weighted;
668
669 schedstat_set(curr->statistics.exec_max,
670 max((u64)delta_exec, curr->statistics.exec_max));
671
672 curr->sum_exec_runtime += delta_exec;
673 schedstat_add(cfs_rq, exec_clock, delta_exec);
674 delta_exec_weighted = calc_delta_fair(delta_exec, curr);
675
676 curr->vruntime += delta_exec_weighted;
677 update_min_vruntime(cfs_rq);
678
679#if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
680 cfs_rq->load_unacc_exec_time += delta_exec;
681#endif
682}
683
684static void update_curr(struct cfs_rq *cfs_rq)
685{
686 struct sched_entity *curr = cfs_rq->curr;
687 u64 now = rq_of(cfs_rq)->clock_task;
688 unsigned long delta_exec;
689
690 if (unlikely(!curr))
691 return;
692
693 /*
694 * Get the amount of time the current task was running
695 * since the last time we changed load (this cannot
696 * overflow on 32 bits):
697 */
698 delta_exec = (unsigned long)(now - curr->exec_start);
699 if (!delta_exec)
700 return;
701
702 __update_curr(cfs_rq, curr, delta_exec);
703 curr->exec_start = now;
704
705 if (entity_is_task(curr)) {
706 struct task_struct *curtask = task_of(curr);
707
708 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
709 cpuacct_charge(curtask, delta_exec);
710 account_group_exec_runtime(curtask, delta_exec);
711 }
712
713 account_cfs_rq_runtime(cfs_rq, delta_exec);
714}
715
716static inline void
717update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
718{
719 schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
720}
721
722/*
723 * Task is being enqueued - update stats:
724 */
725static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
726{
727 /*
728 * Are we enqueueing a waiting task? (for current tasks
729 * a dequeue/enqueue event is a NOP)
730 */
731 if (se != cfs_rq->curr)
732 update_stats_wait_start(cfs_rq, se);
733}
734
735static void
736update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
737{
738 schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
739 rq_of(cfs_rq)->clock - se->statistics.wait_start));
740 schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
741 schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
742 rq_of(cfs_rq)->clock - se->statistics.wait_start);
743#ifdef CONFIG_SCHEDSTATS
744 if (entity_is_task(se)) {
745 trace_sched_stat_wait(task_of(se),
746 rq_of(cfs_rq)->clock - se->statistics.wait_start);
747 }
748#endif
749 schedstat_set(se->statistics.wait_start, 0);
750}
751
752static inline void
753update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
754{
755 /*
756 * Mark the end of the wait period if dequeueing a
757 * waiting task:
758 */
759 if (se != cfs_rq->curr)
760 update_stats_wait_end(cfs_rq, se);
761}
762
763/*
764 * We are picking a new current task - update its stats:
765 */
766static inline void
767update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
768{
769 /*
770 * We are starting a new run period:
771 */
772 se->exec_start = rq_of(cfs_rq)->clock_task;
773}
774
775/**************************************************
776 * Scheduling class queueing methods:
777 */
778
779#if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
780static void
781add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
782{
783 cfs_rq->task_weight += weight;
784}
785#else
786static inline void
787add_cfs_task_weight(struct cfs_rq *cfs_rq, unsigned long weight)
788{
789}
790#endif
791
792static void
793account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
794{
795 update_load_add(&cfs_rq->load, se->load.weight);
796 if (!parent_entity(se))
797 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
798 if (entity_is_task(se)) {
799 add_cfs_task_weight(cfs_rq, se->load.weight);
800 list_add(&se->group_node, &cfs_rq->tasks);
801 }
802 cfs_rq->nr_running++;
803}
804
805static void
806account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
807{
808 update_load_sub(&cfs_rq->load, se->load.weight);
809 if (!parent_entity(se))
810 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
811 if (entity_is_task(se)) {
812 add_cfs_task_weight(cfs_rq, -se->load.weight);
813 list_del_init(&se->group_node);
814 }
815 cfs_rq->nr_running--;
816}
817
818#ifdef CONFIG_FAIR_GROUP_SCHED
819/* we need this in update_cfs_load and load-balance functions below */
820static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
821# ifdef CONFIG_SMP
822static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
823 int global_update)
824{
825 struct task_group *tg = cfs_rq->tg;
826 long load_avg;
827
828 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
829 load_avg -= cfs_rq->load_contribution;
830
831 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
832 atomic_add(load_avg, &tg->load_weight);
833 cfs_rq->load_contribution += load_avg;
834 }
835}
836
837static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
838{
839 u64 period = sysctl_sched_shares_window;
840 u64 now, delta;
841 unsigned long load = cfs_rq->load.weight;
842
843 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
844 return;
845
846 now = rq_of(cfs_rq)->clock_task;
847 delta = now - cfs_rq->load_stamp;
848
849 /* truncate load history at 4 idle periods */
850 if (cfs_rq->load_stamp > cfs_rq->load_last &&
851 now - cfs_rq->load_last > 4 * period) {
852 cfs_rq->load_period = 0;
853 cfs_rq->load_avg = 0;
854 delta = period - 1;
855 }
856
857 cfs_rq->load_stamp = now;
858 cfs_rq->load_unacc_exec_time = 0;
859 cfs_rq->load_period += delta;
860 if (load) {
861 cfs_rq->load_last = now;
862 cfs_rq->load_avg += delta * load;
863 }
864
865 /* consider updating load contribution on each fold or truncate */
866 if (global_update || cfs_rq->load_period > period
867 || !cfs_rq->load_period)
868 update_cfs_rq_load_contribution(cfs_rq, global_update);
869
870 while (cfs_rq->load_period > period) {
871 /*
872 * Inline assembly required to prevent the compiler
873 * optimising this loop into a divmod call.
874 * See __iter_div_u64_rem() for another example of this.
875 */
876 asm("" : "+rm" (cfs_rq->load_period));
877 cfs_rq->load_period /= 2;
878 cfs_rq->load_avg /= 2;
879 }
880
881 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
882 list_del_leaf_cfs_rq(cfs_rq);
883}
884
885static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
886{
887 long tg_weight;
888
889 /*
890 * Use this CPU's actual weight instead of the last load_contribution
891 * to gain a more accurate current total weight. See
892 * update_cfs_rq_load_contribution().
893 */
894 tg_weight = atomic_read(&tg->load_weight);
895 tg_weight -= cfs_rq->load_contribution;
896 tg_weight += cfs_rq->load.weight;
897
898 return tg_weight;
899}
900
901static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
902{
903 long tg_weight, load, shares;
904
905 tg_weight = calc_tg_weight(tg, cfs_rq);
906 load = cfs_rq->load.weight;
907
908 shares = (tg->shares * load);
909 if (tg_weight)
910 shares /= tg_weight;
911
912 if (shares < MIN_SHARES)
913 shares = MIN_SHARES;
914 if (shares > tg->shares)
915 shares = tg->shares;
916
917 return shares;
918}
919
920static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
921{
922 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
923 update_cfs_load(cfs_rq, 0);
924 update_cfs_shares(cfs_rq);
925 }
926}
927# else /* CONFIG_SMP */
928static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
929{
930}
931
932static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
933{
934 return tg->shares;
935}
936
937static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
938{
939}
940# endif /* CONFIG_SMP */
941static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
942 unsigned long weight)
943{
944 if (se->on_rq) {
945 /* commit outstanding execution time */
946 if (cfs_rq->curr == se)
947 update_curr(cfs_rq);
948 account_entity_dequeue(cfs_rq, se);
949 }
950
951 update_load_set(&se->load, weight);
952
953 if (se->on_rq)
954 account_entity_enqueue(cfs_rq, se);
955}
956
957static void update_cfs_shares(struct cfs_rq *cfs_rq)
958{
959 struct task_group *tg;
960 struct sched_entity *se;
961 long shares;
962
963 tg = cfs_rq->tg;
964 se = tg->se[cpu_of(rq_of(cfs_rq))];
965 if (!se || throttled_hierarchy(cfs_rq))
966 return;
967#ifndef CONFIG_SMP
968 if (likely(se->load.weight == tg->shares))
969 return;
970#endif
971 shares = calc_cfs_shares(cfs_rq, tg);
972
973 reweight_entity(cfs_rq_of(se), se, shares);
974}
975#else /* CONFIG_FAIR_GROUP_SCHED */
976static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
977{
978}
979
980static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
981{
982}
983
984static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
985{
986}
987#endif /* CONFIG_FAIR_GROUP_SCHED */
988
989static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
990{
991#ifdef CONFIG_SCHEDSTATS
992 struct task_struct *tsk = NULL;
993
994 if (entity_is_task(se))
995 tsk = task_of(se);
996
997 if (se->statistics.sleep_start) {
998 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
999
1000 if ((s64)delta < 0)
1001 delta = 0;
1002
1003 if (unlikely(delta > se->statistics.sleep_max))
1004 se->statistics.sleep_max = delta;
1005
1006 se->statistics.sleep_start = 0;
1007 se->statistics.sum_sleep_runtime += delta;
1008
1009 if (tsk) {
1010 account_scheduler_latency(tsk, delta >> 10, 1);
1011 trace_sched_stat_sleep(tsk, delta);
1012 }
1013 }
1014 if (se->statistics.block_start) {
1015 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1016
1017 if ((s64)delta < 0)
1018 delta = 0;
1019
1020 if (unlikely(delta > se->statistics.block_max))
1021 se->statistics.block_max = delta;
1022
1023 se->statistics.block_start = 0;
1024 se->statistics.sum_sleep_runtime += delta;
1025
1026 if (tsk) {
1027 if (tsk->in_iowait) {
1028 se->statistics.iowait_sum += delta;
1029 se->statistics.iowait_count++;
1030 trace_sched_stat_iowait(tsk, delta);
1031 }
1032
1033 /*
1034 * Blocking time is in units of nanosecs, so shift by
1035 * 20 to get a milliseconds-range estimation of the
1036 * amount of time that the task spent sleeping:
1037 */
1038 if (unlikely(prof_on == SLEEP_PROFILING)) {
1039 profile_hits(SLEEP_PROFILING,
1040 (void *)get_wchan(tsk),
1041 delta >> 20);
1042 }
1043 account_scheduler_latency(tsk, delta >> 10, 0);
1044 }
1045 }
1046#endif
1047}
1048
1049static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1050{
1051#ifdef CONFIG_SCHED_DEBUG
1052 s64 d = se->vruntime - cfs_rq->min_vruntime;
1053
1054 if (d < 0)
1055 d = -d;
1056
1057 if (d > 3*sysctl_sched_latency)
1058 schedstat_inc(cfs_rq, nr_spread_over);
1059#endif
1060}
1061
1062static void
1063place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1064{
1065 u64 vruntime = cfs_rq->min_vruntime;
1066
1067 /*
1068 * The 'current' period is already promised to the current tasks,
1069 * however the extra weight of the new task will slow them down a
1070 * little, place the new task so that it fits in the slot that
1071 * stays open at the end.
1072 */
1073 if (initial && sched_feat(START_DEBIT))
1074 vruntime += sched_vslice(cfs_rq, se);
1075
1076 /* sleeps up to a single latency don't count. */
1077 if (!initial) {
1078 unsigned long thresh = sysctl_sched_latency;
1079
1080 /*
1081 * Halve their sleep time's effect, to allow
1082 * for a gentler effect of sleepers:
1083 */
1084 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1085 thresh >>= 1;
1086
1087 vruntime -= thresh;
1088 }
1089
1090 /* ensure we never gain time by being placed backwards. */
1091 vruntime = max_vruntime(se->vruntime, vruntime);
1092
1093 se->vruntime = vruntime;
1094}
1095
1096static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1097
1098static void
1099enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1100{
1101 /*
1102 * Update the normalized vruntime before updating min_vruntime
1103 * through callig update_curr().
1104 */
1105 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1106 se->vruntime += cfs_rq->min_vruntime;
1107
1108 /*
1109 * Update run-time statistics of the 'current'.
1110 */
1111 update_curr(cfs_rq);
1112 update_cfs_load(cfs_rq, 0);
1113 account_entity_enqueue(cfs_rq, se);
1114 update_cfs_shares(cfs_rq);
1115
1116 if (flags & ENQUEUE_WAKEUP) {
1117 place_entity(cfs_rq, se, 0);
1118 enqueue_sleeper(cfs_rq, se);
1119 }
1120
1121 update_stats_enqueue(cfs_rq, se);
1122 check_spread(cfs_rq, se);
1123 if (se != cfs_rq->curr)
1124 __enqueue_entity(cfs_rq, se);
1125 se->on_rq = 1;
1126
1127 if (cfs_rq->nr_running == 1) {
1128 list_add_leaf_cfs_rq(cfs_rq);
1129 check_enqueue_throttle(cfs_rq);
1130 }
1131}
1132
1133static void __clear_buddies_last(struct sched_entity *se)
1134{
1135 for_each_sched_entity(se) {
1136 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1137 if (cfs_rq->last == se)
1138 cfs_rq->last = NULL;
1139 else
1140 break;
1141 }
1142}
1143
1144static void __clear_buddies_next(struct sched_entity *se)
1145{
1146 for_each_sched_entity(se) {
1147 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1148 if (cfs_rq->next == se)
1149 cfs_rq->next = NULL;
1150 else
1151 break;
1152 }
1153}
1154
1155static void __clear_buddies_skip(struct sched_entity *se)
1156{
1157 for_each_sched_entity(se) {
1158 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1159 if (cfs_rq->skip == se)
1160 cfs_rq->skip = NULL;
1161 else
1162 break;
1163 }
1164}
1165
1166static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1167{
1168 if (cfs_rq->last == se)
1169 __clear_buddies_last(se);
1170
1171 if (cfs_rq->next == se)
1172 __clear_buddies_next(se);
1173
1174 if (cfs_rq->skip == se)
1175 __clear_buddies_skip(se);
1176}
1177
1178static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1179
1180static void
1181dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1182{
1183 /*
1184 * Update run-time statistics of the 'current'.
1185 */
1186 update_curr(cfs_rq);
1187
1188 update_stats_dequeue(cfs_rq, se);
1189 if (flags & DEQUEUE_SLEEP) {
1190#ifdef CONFIG_SCHEDSTATS
1191 if (entity_is_task(se)) {
1192 struct task_struct *tsk = task_of(se);
1193
1194 if (tsk->state & TASK_INTERRUPTIBLE)
1195 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1196 if (tsk->state & TASK_UNINTERRUPTIBLE)
1197 se->statistics.block_start = rq_of(cfs_rq)->clock;
1198 }
1199#endif
1200 }
1201
1202 clear_buddies(cfs_rq, se);
1203
1204 if (se != cfs_rq->curr)
1205 __dequeue_entity(cfs_rq, se);
1206 se->on_rq = 0;
1207 update_cfs_load(cfs_rq, 0);
1208 account_entity_dequeue(cfs_rq, se);
1209
1210 /*
1211 * Normalize the entity after updating the min_vruntime because the
1212 * update can refer to the ->curr item and we need to reflect this
1213 * movement in our normalized position.
1214 */
1215 if (!(flags & DEQUEUE_SLEEP))
1216 se->vruntime -= cfs_rq->min_vruntime;
1217
1218 /* return excess runtime on last dequeue */
1219 return_cfs_rq_runtime(cfs_rq);
1220
1221 update_min_vruntime(cfs_rq);
1222 update_cfs_shares(cfs_rq);
1223}
1224
1225/*
1226 * Preempt the current task with a newly woken task if needed:
1227 */
1228static void
1229check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1230{
1231 unsigned long ideal_runtime, delta_exec;
1232 struct sched_entity *se;
1233 s64 delta;
1234
1235 ideal_runtime = sched_slice(cfs_rq, curr);
1236 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1237 if (delta_exec > ideal_runtime) {
1238 resched_task(rq_of(cfs_rq)->curr);
1239 /*
1240 * The current task ran long enough, ensure it doesn't get
1241 * re-elected due to buddy favours.
1242 */
1243 clear_buddies(cfs_rq, curr);
1244 return;
1245 }
1246
1247 /*
1248 * Ensure that a task that missed wakeup preemption by a
1249 * narrow margin doesn't have to wait for a full slice.
1250 * This also mitigates buddy induced latencies under load.
1251 */
1252 if (delta_exec < sysctl_sched_min_granularity)
1253 return;
1254
1255 se = __pick_first_entity(cfs_rq);
1256 delta = curr->vruntime - se->vruntime;
1257
1258 if (delta < 0)
1259 return;
1260
1261 if (delta > ideal_runtime)
1262 resched_task(rq_of(cfs_rq)->curr);
1263}
1264
1265static void
1266set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1267{
1268 /* 'current' is not kept within the tree. */
1269 if (se->on_rq) {
1270 /*
1271 * Any task has to be enqueued before it get to execute on
1272 * a CPU. So account for the time it spent waiting on the
1273 * runqueue.
1274 */
1275 update_stats_wait_end(cfs_rq, se);
1276 __dequeue_entity(cfs_rq, se);
1277 }
1278
1279 update_stats_curr_start(cfs_rq, se);
1280 cfs_rq->curr = se;
1281#ifdef CONFIG_SCHEDSTATS
1282 /*
1283 * Track our maximum slice length, if the CPU's load is at
1284 * least twice that of our own weight (i.e. dont track it
1285 * when there are only lesser-weight tasks around):
1286 */
1287 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1288 se->statistics.slice_max = max(se->statistics.slice_max,
1289 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1290 }
1291#endif
1292 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1293}
1294
1295static int
1296wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1297
1298/*
1299 * Pick the next process, keeping these things in mind, in this order:
1300 * 1) keep things fair between processes/task groups
1301 * 2) pick the "next" process, since someone really wants that to run
1302 * 3) pick the "last" process, for cache locality
1303 * 4) do not run the "skip" process, if something else is available
1304 */
1305static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1306{
1307 struct sched_entity *se = __pick_first_entity(cfs_rq);
1308 struct sched_entity *left = se;
1309
1310 /*
1311 * Avoid running the skip buddy, if running something else can
1312 * be done without getting too unfair.
1313 */
1314 if (cfs_rq->skip == se) {
1315 struct sched_entity *second = __pick_next_entity(se);
1316 if (second && wakeup_preempt_entity(second, left) < 1)
1317 se = second;
1318 }
1319
1320 /*
1321 * Prefer last buddy, try to return the CPU to a preempted task.
1322 */
1323 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1324 se = cfs_rq->last;
1325
1326 /*
1327 * Someone really wants this to run. If it's not unfair, run it.
1328 */
1329 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1330 se = cfs_rq->next;
1331
1332 clear_buddies(cfs_rq, se);
1333
1334 return se;
1335}
1336
1337static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1338
1339static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1340{
1341 /*
1342 * If still on the runqueue then deactivate_task()
1343 * was not called and update_curr() has to be done:
1344 */
1345 if (prev->on_rq)
1346 update_curr(cfs_rq);
1347
1348 /* throttle cfs_rqs exceeding runtime */
1349 check_cfs_rq_runtime(cfs_rq);
1350
1351 check_spread(cfs_rq, prev);
1352 if (prev->on_rq) {
1353 update_stats_wait_start(cfs_rq, prev);
1354 /* Put 'current' back into the tree. */
1355 __enqueue_entity(cfs_rq, prev);
1356 }
1357 cfs_rq->curr = NULL;
1358}
1359
1360static void
1361entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1362{
1363 /*
1364 * Update run-time statistics of the 'current'.
1365 */
1366 update_curr(cfs_rq);
1367
1368 /*
1369 * Update share accounting for long-running entities.
1370 */
1371 update_entity_shares_tick(cfs_rq);
1372
1373#ifdef CONFIG_SCHED_HRTICK
1374 /*
1375 * queued ticks are scheduled to match the slice, so don't bother
1376 * validating it and just reschedule.
1377 */
1378 if (queued) {
1379 resched_task(rq_of(cfs_rq)->curr);
1380 return;
1381 }
1382 /*
1383 * don't let the period tick interfere with the hrtick preemption
1384 */
1385 if (!sched_feat(DOUBLE_TICK) &&
1386 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1387 return;
1388#endif
1389
1390 if (cfs_rq->nr_running > 1)
1391 check_preempt_tick(cfs_rq, curr);
1392}
1393
1394
1395/**************************************************
1396 * CFS bandwidth control machinery
1397 */
1398
1399#ifdef CONFIG_CFS_BANDWIDTH
1400
1401#ifdef HAVE_JUMP_LABEL
1402static struct jump_label_key __cfs_bandwidth_used;
1403
1404static inline bool cfs_bandwidth_used(void)
1405{
1406 return static_branch(&__cfs_bandwidth_used);
1407}
1408
1409void account_cfs_bandwidth_used(int enabled, int was_enabled)
1410{
1411 /* only need to count groups transitioning between enabled/!enabled */
1412 if (enabled && !was_enabled)
1413 jump_label_inc(&__cfs_bandwidth_used);
1414 else if (!enabled && was_enabled)
1415 jump_label_dec(&__cfs_bandwidth_used);
1416}
1417#else /* HAVE_JUMP_LABEL */
1418static bool cfs_bandwidth_used(void)
1419{
1420 return true;
1421}
1422
1423void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1424#endif /* HAVE_JUMP_LABEL */
1425
1426/*
1427 * default period for cfs group bandwidth.
1428 * default: 0.1s, units: nanoseconds
1429 */
1430static inline u64 default_cfs_period(void)
1431{
1432 return 100000000ULL;
1433}
1434
1435static inline u64 sched_cfs_bandwidth_slice(void)
1436{
1437 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1438}
1439
1440/*
1441 * Replenish runtime according to assigned quota and update expiration time.
1442 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1443 * additional synchronization around rq->lock.
1444 *
1445 * requires cfs_b->lock
1446 */
1447void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1448{
1449 u64 now;
1450
1451 if (cfs_b->quota == RUNTIME_INF)
1452 return;
1453
1454 now = sched_clock_cpu(smp_processor_id());
1455 cfs_b->runtime = cfs_b->quota;
1456 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1457}
1458
1459static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1460{
1461 return &tg->cfs_bandwidth;
1462}
1463
1464/* returns 0 on failure to allocate runtime */
1465static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1466{
1467 struct task_group *tg = cfs_rq->tg;
1468 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1469 u64 amount = 0, min_amount, expires;
1470
1471 /* note: this is a positive sum as runtime_remaining <= 0 */
1472 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1473
1474 raw_spin_lock(&cfs_b->lock);
1475 if (cfs_b->quota == RUNTIME_INF)
1476 amount = min_amount;
1477 else {
1478 /*
1479 * If the bandwidth pool has become inactive, then at least one
1480 * period must have elapsed since the last consumption.
1481 * Refresh the global state and ensure bandwidth timer becomes
1482 * active.
1483 */
1484 if (!cfs_b->timer_active) {
1485 __refill_cfs_bandwidth_runtime(cfs_b);
1486 __start_cfs_bandwidth(cfs_b);
1487 }
1488
1489 if (cfs_b->runtime > 0) {
1490 amount = min(cfs_b->runtime, min_amount);
1491 cfs_b->runtime -= amount;
1492 cfs_b->idle = 0;
1493 }
1494 }
1495 expires = cfs_b->runtime_expires;
1496 raw_spin_unlock(&cfs_b->lock);
1497
1498 cfs_rq->runtime_remaining += amount;
1499 /*
1500 * we may have advanced our local expiration to account for allowed
1501 * spread between our sched_clock and the one on which runtime was
1502 * issued.
1503 */
1504 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1505 cfs_rq->runtime_expires = expires;
1506
1507 return cfs_rq->runtime_remaining > 0;
1508}
1509
1510/*
1511 * Note: This depends on the synchronization provided by sched_clock and the
1512 * fact that rq->clock snapshots this value.
1513 */
1514static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1515{
1516 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1517 struct rq *rq = rq_of(cfs_rq);
1518
1519 /* if the deadline is ahead of our clock, nothing to do */
1520 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1521 return;
1522
1523 if (cfs_rq->runtime_remaining < 0)
1524 return;
1525
1526 /*
1527 * If the local deadline has passed we have to consider the
1528 * possibility that our sched_clock is 'fast' and the global deadline
1529 * has not truly expired.
1530 *
1531 * Fortunately we can check determine whether this the case by checking
1532 * whether the global deadline has advanced.
1533 */
1534
1535 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1536 /* extend local deadline, drift is bounded above by 2 ticks */
1537 cfs_rq->runtime_expires += TICK_NSEC;
1538 } else {
1539 /* global deadline is ahead, expiration has passed */
1540 cfs_rq->runtime_remaining = 0;
1541 }
1542}
1543
1544static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1545 unsigned long delta_exec)
1546{
1547 /* dock delta_exec before expiring quota (as it could span periods) */
1548 cfs_rq->runtime_remaining -= delta_exec;
1549 expire_cfs_rq_runtime(cfs_rq);
1550
1551 if (likely(cfs_rq->runtime_remaining > 0))
1552 return;
1553
1554 /*
1555 * if we're unable to extend our runtime we resched so that the active
1556 * hierarchy can be throttled
1557 */
1558 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1559 resched_task(rq_of(cfs_rq)->curr);
1560}
1561
1562static __always_inline void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1563 unsigned long delta_exec)
1564{
1565 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1566 return;
1567
1568 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1569}
1570
1571static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1572{
1573 return cfs_bandwidth_used() && cfs_rq->throttled;
1574}
1575
1576/* check whether cfs_rq, or any parent, is throttled */
1577static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1578{
1579 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1580}
1581
1582/*
1583 * Ensure that neither of the group entities corresponding to src_cpu or
1584 * dest_cpu are members of a throttled hierarchy when performing group
1585 * load-balance operations.
1586 */
1587static inline int throttled_lb_pair(struct task_group *tg,
1588 int src_cpu, int dest_cpu)
1589{
1590 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1591
1592 src_cfs_rq = tg->cfs_rq[src_cpu];
1593 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1594
1595 return throttled_hierarchy(src_cfs_rq) ||
1596 throttled_hierarchy(dest_cfs_rq);
1597}
1598
1599/* updated child weight may affect parent so we have to do this bottom up */
1600static int tg_unthrottle_up(struct task_group *tg, void *data)
1601{
1602 struct rq *rq = data;
1603 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1604
1605 cfs_rq->throttle_count--;
1606#ifdef CONFIG_SMP
1607 if (!cfs_rq->throttle_count) {
1608 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1609
1610 /* leaving throttled state, advance shares averaging windows */
1611 cfs_rq->load_stamp += delta;
1612 cfs_rq->load_last += delta;
1613
1614 /* update entity weight now that we are on_rq again */
1615 update_cfs_shares(cfs_rq);
1616 }
1617#endif
1618
1619 return 0;
1620}
1621
1622static int tg_throttle_down(struct task_group *tg, void *data)
1623{
1624 struct rq *rq = data;
1625 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1626
1627 /* group is entering throttled state, record last load */
1628 if (!cfs_rq->throttle_count)
1629 update_cfs_load(cfs_rq, 0);
1630 cfs_rq->throttle_count++;
1631
1632 return 0;
1633}
1634
1635static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1636{
1637 struct rq *rq = rq_of(cfs_rq);
1638 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1639 struct sched_entity *se;
1640 long task_delta, dequeue = 1;
1641
1642 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1643
1644 /* account load preceding throttle */
1645 rcu_read_lock();
1646 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1647 rcu_read_unlock();
1648
1649 task_delta = cfs_rq->h_nr_running;
1650 for_each_sched_entity(se) {
1651 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1652 /* throttled entity or throttle-on-deactivate */
1653 if (!se->on_rq)
1654 break;
1655
1656 if (dequeue)
1657 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1658 qcfs_rq->h_nr_running -= task_delta;
1659
1660 if (qcfs_rq->load.weight)
1661 dequeue = 0;
1662 }
1663
1664 if (!se)
1665 rq->nr_running -= task_delta;
1666
1667 cfs_rq->throttled = 1;
1668 cfs_rq->throttled_timestamp = rq->clock;
1669 raw_spin_lock(&cfs_b->lock);
1670 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1671 raw_spin_unlock(&cfs_b->lock);
1672}
1673
1674void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1675{
1676 struct rq *rq = rq_of(cfs_rq);
1677 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1678 struct sched_entity *se;
1679 int enqueue = 1;
1680 long task_delta;
1681
1682 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1683
1684 cfs_rq->throttled = 0;
1685 raw_spin_lock(&cfs_b->lock);
1686 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1687 list_del_rcu(&cfs_rq->throttled_list);
1688 raw_spin_unlock(&cfs_b->lock);
1689 cfs_rq->throttled_timestamp = 0;
1690
1691 update_rq_clock(rq);
1692 /* update hierarchical throttle state */
1693 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1694
1695 if (!cfs_rq->load.weight)
1696 return;
1697
1698 task_delta = cfs_rq->h_nr_running;
1699 for_each_sched_entity(se) {
1700 if (se->on_rq)
1701 enqueue = 0;
1702
1703 cfs_rq = cfs_rq_of(se);
1704 if (enqueue)
1705 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1706 cfs_rq->h_nr_running += task_delta;
1707
1708 if (cfs_rq_throttled(cfs_rq))
1709 break;
1710 }
1711
1712 if (!se)
1713 rq->nr_running += task_delta;
1714
1715 /* determine whether we need to wake up potentially idle cpu */
1716 if (rq->curr == rq->idle && rq->cfs.nr_running)
1717 resched_task(rq->curr);
1718}
1719
1720static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1721 u64 remaining, u64 expires)
1722{
1723 struct cfs_rq *cfs_rq;
1724 u64 runtime = remaining;
1725
1726 rcu_read_lock();
1727 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1728 throttled_list) {
1729 struct rq *rq = rq_of(cfs_rq);
1730
1731 raw_spin_lock(&rq->lock);
1732 if (!cfs_rq_throttled(cfs_rq))
1733 goto next;
1734
1735 runtime = -cfs_rq->runtime_remaining + 1;
1736 if (runtime > remaining)
1737 runtime = remaining;
1738 remaining -= runtime;
1739
1740 cfs_rq->runtime_remaining += runtime;
1741 cfs_rq->runtime_expires = expires;
1742
1743 /* we check whether we're throttled above */
1744 if (cfs_rq->runtime_remaining > 0)
1745 unthrottle_cfs_rq(cfs_rq);
1746
1747next:
1748 raw_spin_unlock(&rq->lock);
1749
1750 if (!remaining)
1751 break;
1752 }
1753 rcu_read_unlock();
1754
1755 return remaining;
1756}
1757
1758/*
1759 * Responsible for refilling a task_group's bandwidth and unthrottling its
1760 * cfs_rqs as appropriate. If there has been no activity within the last
1761 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1762 * used to track this state.
1763 */
1764static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1765{
1766 u64 runtime, runtime_expires;
1767 int idle = 1, throttled;
1768
1769 raw_spin_lock(&cfs_b->lock);
1770 /* no need to continue the timer with no bandwidth constraint */
1771 if (cfs_b->quota == RUNTIME_INF)
1772 goto out_unlock;
1773
1774 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1775 /* idle depends on !throttled (for the case of a large deficit) */
1776 idle = cfs_b->idle && !throttled;
1777 cfs_b->nr_periods += overrun;
1778
1779 /* if we're going inactive then everything else can be deferred */
1780 if (idle)
1781 goto out_unlock;
1782
1783 __refill_cfs_bandwidth_runtime(cfs_b);
1784
1785 if (!throttled) {
1786 /* mark as potentially idle for the upcoming period */
1787 cfs_b->idle = 1;
1788 goto out_unlock;
1789 }
1790
1791 /* account preceding periods in which throttling occurred */
1792 cfs_b->nr_throttled += overrun;
1793
1794 /*
1795 * There are throttled entities so we must first use the new bandwidth
1796 * to unthrottle them before making it generally available. This
1797 * ensures that all existing debts will be paid before a new cfs_rq is
1798 * allowed to run.
1799 */
1800 runtime = cfs_b->runtime;
1801 runtime_expires = cfs_b->runtime_expires;
1802 cfs_b->runtime = 0;
1803
1804 /*
1805 * This check is repeated as we are holding onto the new bandwidth
1806 * while we unthrottle. This can potentially race with an unthrottled
1807 * group trying to acquire new bandwidth from the global pool.
1808 */
1809 while (throttled && runtime > 0) {
1810 raw_spin_unlock(&cfs_b->lock);
1811 /* we can't nest cfs_b->lock while distributing bandwidth */
1812 runtime = distribute_cfs_runtime(cfs_b, runtime,
1813 runtime_expires);
1814 raw_spin_lock(&cfs_b->lock);
1815
1816 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1817 }
1818
1819 /* return (any) remaining runtime */
1820 cfs_b->runtime = runtime;
1821 /*
1822 * While we are ensured activity in the period following an
1823 * unthrottle, this also covers the case in which the new bandwidth is
1824 * insufficient to cover the existing bandwidth deficit. (Forcing the
1825 * timer to remain active while there are any throttled entities.)
1826 */
1827 cfs_b->idle = 0;
1828out_unlock:
1829 if (idle)
1830 cfs_b->timer_active = 0;
1831 raw_spin_unlock(&cfs_b->lock);
1832
1833 return idle;
1834}
1835
1836/* a cfs_rq won't donate quota below this amount */
1837static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1838/* minimum remaining period time to redistribute slack quota */
1839static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1840/* how long we wait to gather additional slack before distributing */
1841static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1842
1843/* are we near the end of the current quota period? */
1844static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
1845{
1846 struct hrtimer *refresh_timer = &cfs_b->period_timer;
1847 u64 remaining;
1848
1849 /* if the call-back is running a quota refresh is already occurring */
1850 if (hrtimer_callback_running(refresh_timer))
1851 return 1;
1852
1853 /* is a quota refresh about to occur? */
1854 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
1855 if (remaining < min_expire)
1856 return 1;
1857
1858 return 0;
1859}
1860
1861static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
1862{
1863 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
1864
1865 /* if there's a quota refresh soon don't bother with slack */
1866 if (runtime_refresh_within(cfs_b, min_left))
1867 return;
1868
1869 start_bandwidth_timer(&cfs_b->slack_timer,
1870 ns_to_ktime(cfs_bandwidth_slack_period));
1871}
1872
1873/* we know any runtime found here is valid as update_curr() precedes return */
1874static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1875{
1876 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1877 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
1878
1879 if (slack_runtime <= 0)
1880 return;
1881
1882 raw_spin_lock(&cfs_b->lock);
1883 if (cfs_b->quota != RUNTIME_INF &&
1884 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
1885 cfs_b->runtime += slack_runtime;
1886
1887 /* we are under rq->lock, defer unthrottling using a timer */
1888 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
1889 !list_empty(&cfs_b->throttled_cfs_rq))
1890 start_cfs_slack_bandwidth(cfs_b);
1891 }
1892 raw_spin_unlock(&cfs_b->lock);
1893
1894 /* even if it's not valid for return we don't want to try again */
1895 cfs_rq->runtime_remaining -= slack_runtime;
1896}
1897
1898static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1899{
1900 if (!cfs_bandwidth_used())
1901 return;
1902
1903 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
1904 return;
1905
1906 __return_cfs_rq_runtime(cfs_rq);
1907}
1908
1909/*
1910 * This is done with a timer (instead of inline with bandwidth return) since
1911 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
1912 */
1913static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
1914{
1915 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
1916 u64 expires;
1917
1918 /* confirm we're still not at a refresh boundary */
1919 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
1920 return;
1921
1922 raw_spin_lock(&cfs_b->lock);
1923 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
1924 runtime = cfs_b->runtime;
1925 cfs_b->runtime = 0;
1926 }
1927 expires = cfs_b->runtime_expires;
1928 raw_spin_unlock(&cfs_b->lock);
1929
1930 if (!runtime)
1931 return;
1932
1933 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
1934
1935 raw_spin_lock(&cfs_b->lock);
1936 if (expires == cfs_b->runtime_expires)
1937 cfs_b->runtime = runtime;
1938 raw_spin_unlock(&cfs_b->lock);
1939}
1940
1941/*
1942 * When a group wakes up we want to make sure that its quota is not already
1943 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1944 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1945 */
1946static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
1947{
1948 if (!cfs_bandwidth_used())
1949 return;
1950
1951 /* an active group must be handled by the update_curr()->put() path */
1952 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
1953 return;
1954
1955 /* ensure the group is not already throttled */
1956 if (cfs_rq_throttled(cfs_rq))
1957 return;
1958
1959 /* update runtime allocation */
1960 account_cfs_rq_runtime(cfs_rq, 0);
1961 if (cfs_rq->runtime_remaining <= 0)
1962 throttle_cfs_rq(cfs_rq);
1963}
1964
1965/* conditionally throttle active cfs_rq's from put_prev_entity() */
1966static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1967{
1968 if (!cfs_bandwidth_used())
1969 return;
1970
1971 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
1972 return;
1973
1974 /*
1975 * it's possible for a throttled entity to be forced into a running
1976 * state (e.g. set_curr_task), in this case we're finished.
1977 */
1978 if (cfs_rq_throttled(cfs_rq))
1979 return;
1980
1981 throttle_cfs_rq(cfs_rq);
1982}
1983
1984static inline u64 default_cfs_period(void);
1985static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
1986static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
1987
1988static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
1989{
1990 struct cfs_bandwidth *cfs_b =
1991 container_of(timer, struct cfs_bandwidth, slack_timer);
1992 do_sched_cfs_slack_timer(cfs_b);
1993
1994 return HRTIMER_NORESTART;
1995}
1996
1997static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
1998{
1999 struct cfs_bandwidth *cfs_b =
2000 container_of(timer, struct cfs_bandwidth, period_timer);
2001 ktime_t now;
2002 int overrun;
2003 int idle = 0;
2004
2005 for (;;) {
2006 now = hrtimer_cb_get_time(timer);
2007 overrun = hrtimer_forward(timer, now, cfs_b->period);
2008
2009 if (!overrun)
2010 break;
2011
2012 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2013 }
2014
2015 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2016}
2017
2018void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2019{
2020 raw_spin_lock_init(&cfs_b->lock);
2021 cfs_b->runtime = 0;
2022 cfs_b->quota = RUNTIME_INF;
2023 cfs_b->period = ns_to_ktime(default_cfs_period());
2024
2025 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2026 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2027 cfs_b->period_timer.function = sched_cfs_period_timer;
2028 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2029 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2030}
2031
2032static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2033{
2034 cfs_rq->runtime_enabled = 0;
2035 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2036}
2037
2038/* requires cfs_b->lock, may release to reprogram timer */
2039void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2040{
2041 /*
2042 * The timer may be active because we're trying to set a new bandwidth
2043 * period or because we're racing with the tear-down path
2044 * (timer_active==0 becomes visible before the hrtimer call-back
2045 * terminates). In either case we ensure that it's re-programmed
2046 */
2047 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2048 raw_spin_unlock(&cfs_b->lock);
2049 /* ensure cfs_b->lock is available while we wait */
2050 hrtimer_cancel(&cfs_b->period_timer);
2051
2052 raw_spin_lock(&cfs_b->lock);
2053 /* if someone else restarted the timer then we're done */
2054 if (cfs_b->timer_active)
2055 return;
2056 }
2057
2058 cfs_b->timer_active = 1;
2059 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2060}
2061
2062static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2063{
2064 hrtimer_cancel(&cfs_b->period_timer);
2065 hrtimer_cancel(&cfs_b->slack_timer);
2066}
2067
2068void unthrottle_offline_cfs_rqs(struct rq *rq)
2069{
2070 struct cfs_rq *cfs_rq;
2071
2072 for_each_leaf_cfs_rq(rq, cfs_rq) {
2073 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2074
2075 if (!cfs_rq->runtime_enabled)
2076 continue;
2077
2078 /*
2079 * clock_task is not advancing so we just need to make sure
2080 * there's some valid quota amount
2081 */
2082 cfs_rq->runtime_remaining = cfs_b->quota;
2083 if (cfs_rq_throttled(cfs_rq))
2084 unthrottle_cfs_rq(cfs_rq);
2085 }
2086}
2087
2088#else /* CONFIG_CFS_BANDWIDTH */
2089static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
2090 unsigned long delta_exec) {}
2091static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2092static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2093static void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2094
2095static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2096{
2097 return 0;
2098}
2099
2100static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2101{
2102 return 0;
2103}
2104
2105static inline int throttled_lb_pair(struct task_group *tg,
2106 int src_cpu, int dest_cpu)
2107{
2108 return 0;
2109}
2110
2111void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2112
2113#ifdef CONFIG_FAIR_GROUP_SCHED
2114static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2115#endif
2116
2117static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2118{
2119 return NULL;
2120}
2121static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2122void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2123
2124#endif /* CONFIG_CFS_BANDWIDTH */
2125
2126/**************************************************
2127 * CFS operations on tasks:
2128 */
2129
2130#ifdef CONFIG_SCHED_HRTICK
2131static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2132{
2133 struct sched_entity *se = &p->se;
2134 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2135
2136 WARN_ON(task_rq(p) != rq);
2137
2138 if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
2139 u64 slice = sched_slice(cfs_rq, se);
2140 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2141 s64 delta = slice - ran;
2142
2143 if (delta < 0) {
2144 if (rq->curr == p)
2145 resched_task(p);
2146 return;
2147 }
2148
2149 /*
2150 * Don't schedule slices shorter than 10000ns, that just
2151 * doesn't make sense. Rely on vruntime for fairness.
2152 */
2153 if (rq->curr != p)
2154 delta = max_t(s64, 10000LL, delta);
2155
2156 hrtick_start(rq, delta);
2157 }
2158}
2159
2160/*
2161 * called from enqueue/dequeue and updates the hrtick when the
2162 * current task is from our class and nr_running is low enough
2163 * to matter.
2164 */
2165static void hrtick_update(struct rq *rq)
2166{
2167 struct task_struct *curr = rq->curr;
2168
2169 if (curr->sched_class != &fair_sched_class)
2170 return;
2171
2172 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2173 hrtick_start_fair(rq, curr);
2174}
2175#else /* !CONFIG_SCHED_HRTICK */
2176static inline void
2177hrtick_start_fair(struct rq *rq, struct task_struct *p)
2178{
2179}
2180
2181static inline void hrtick_update(struct rq *rq)
2182{
2183}
2184#endif
2185
2186/*
2187 * The enqueue_task method is called before nr_running is
2188 * increased. Here we update the fair scheduling stats and
2189 * then put the task into the rbtree:
2190 */
2191static void
2192enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2193{
2194 struct cfs_rq *cfs_rq;
2195 struct sched_entity *se = &p->se;
2196
2197 for_each_sched_entity(se) {
2198 if (se->on_rq)
2199 break;
2200 cfs_rq = cfs_rq_of(se);
2201 enqueue_entity(cfs_rq, se, flags);
2202
2203 /*
2204 * end evaluation on encountering a throttled cfs_rq
2205 *
2206 * note: in the case of encountering a throttled cfs_rq we will
2207 * post the final h_nr_running increment below.
2208 */
2209 if (cfs_rq_throttled(cfs_rq))
2210 break;
2211 cfs_rq->h_nr_running++;
2212
2213 flags = ENQUEUE_WAKEUP;
2214 }
2215
2216 for_each_sched_entity(se) {
2217 cfs_rq = cfs_rq_of(se);
2218 cfs_rq->h_nr_running++;
2219
2220 if (cfs_rq_throttled(cfs_rq))
2221 break;
2222
2223 update_cfs_load(cfs_rq, 0);
2224 update_cfs_shares(cfs_rq);
2225 }
2226
2227 if (!se)
2228 inc_nr_running(rq);
2229 hrtick_update(rq);
2230}
2231
2232static void set_next_buddy(struct sched_entity *se);
2233
2234/*
2235 * The dequeue_task method is called before nr_running is
2236 * decreased. We remove the task from the rbtree and
2237 * update the fair scheduling stats:
2238 */
2239static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2240{
2241 struct cfs_rq *cfs_rq;
2242 struct sched_entity *se = &p->se;
2243 int task_sleep = flags & DEQUEUE_SLEEP;
2244
2245 for_each_sched_entity(se) {
2246 cfs_rq = cfs_rq_of(se);
2247 dequeue_entity(cfs_rq, se, flags);
2248
2249 /*
2250 * end evaluation on encountering a throttled cfs_rq
2251 *
2252 * note: in the case of encountering a throttled cfs_rq we will
2253 * post the final h_nr_running decrement below.
2254 */
2255 if (cfs_rq_throttled(cfs_rq))
2256 break;
2257 cfs_rq->h_nr_running--;
2258
2259 /* Don't dequeue parent if it has other entities besides us */
2260 if (cfs_rq->load.weight) {
2261 /*
2262 * Bias pick_next to pick a task from this cfs_rq, as
2263 * p is sleeping when it is within its sched_slice.
2264 */
2265 if (task_sleep && parent_entity(se))
2266 set_next_buddy(parent_entity(se));
2267
2268 /* avoid re-evaluating load for this entity */
2269 se = parent_entity(se);
2270 break;
2271 }
2272 flags |= DEQUEUE_SLEEP;
2273 }
2274
2275 for_each_sched_entity(se) {
2276 cfs_rq = cfs_rq_of(se);
2277 cfs_rq->h_nr_running--;
2278
2279 if (cfs_rq_throttled(cfs_rq))
2280 break;
2281
2282 update_cfs_load(cfs_rq, 0);
2283 update_cfs_shares(cfs_rq);
2284 }
2285
2286 if (!se)
2287 dec_nr_running(rq);
2288 hrtick_update(rq);
2289}
2290
2291#ifdef CONFIG_SMP
2292/* Used instead of source_load when we know the type == 0 */
2293static unsigned long weighted_cpuload(const int cpu)
2294{
2295 return cpu_rq(cpu)->load.weight;
2296}
2297
2298/*
2299 * Return a low guess at the load of a migration-source cpu weighted
2300 * according to the scheduling class and "nice" value.
2301 *
2302 * We want to under-estimate the load of migration sources, to
2303 * balance conservatively.
2304 */
2305static unsigned long source_load(int cpu, int type)
2306{
2307 struct rq *rq = cpu_rq(cpu);
2308 unsigned long total = weighted_cpuload(cpu);
2309
2310 if (type == 0 || !sched_feat(LB_BIAS))
2311 return total;
2312
2313 return min(rq->cpu_load[type-1], total);
2314}
2315
2316/*
2317 * Return a high guess at the load of a migration-target cpu weighted
2318 * according to the scheduling class and "nice" value.
2319 */
2320static unsigned long target_load(int cpu, int type)
2321{
2322 struct rq *rq = cpu_rq(cpu);
2323 unsigned long total = weighted_cpuload(cpu);
2324
2325 if (type == 0 || !sched_feat(LB_BIAS))
2326 return total;
2327
2328 return max(rq->cpu_load[type-1], total);
2329}
2330
2331static unsigned long power_of(int cpu)
2332{
2333 return cpu_rq(cpu)->cpu_power;
2334}
2335
2336static unsigned long cpu_avg_load_per_task(int cpu)
2337{
2338 struct rq *rq = cpu_rq(cpu);
2339 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2340
2341 if (nr_running)
2342 return rq->load.weight / nr_running;
2343
2344 return 0;
2345}
2346
2347
2348static void task_waking_fair(struct task_struct *p)
2349{
2350 struct sched_entity *se = &p->se;
2351 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2352 u64 min_vruntime;
2353
2354#ifndef CONFIG_64BIT
2355 u64 min_vruntime_copy;
2356
2357 do {
2358 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2359 smp_rmb();
2360 min_vruntime = cfs_rq->min_vruntime;
2361 } while (min_vruntime != min_vruntime_copy);
2362#else
2363 min_vruntime = cfs_rq->min_vruntime;
2364#endif
2365
2366 se->vruntime -= min_vruntime;
2367}
2368
2369#ifdef CONFIG_FAIR_GROUP_SCHED
2370/*
2371 * effective_load() calculates the load change as seen from the root_task_group
2372 *
2373 * Adding load to a group doesn't make a group heavier, but can cause movement
2374 * of group shares between cpus. Assuming the shares were perfectly aligned one
2375 * can calculate the shift in shares.
2376 *
2377 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2378 * on this @cpu and results in a total addition (subtraction) of @wg to the
2379 * total group weight.
2380 *
2381 * Given a runqueue weight distribution (rw_i) we can compute a shares
2382 * distribution (s_i) using:
2383 *
2384 * s_i = rw_i / \Sum rw_j (1)
2385 *
2386 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2387 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2388 * shares distribution (s_i):
2389 *
2390 * rw_i = { 2, 4, 1, 0 }
2391 * s_i = { 2/7, 4/7, 1/7, 0 }
2392 *
2393 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2394 * task used to run on and the CPU the waker is running on), we need to
2395 * compute the effect of waking a task on either CPU and, in case of a sync
2396 * wakeup, compute the effect of the current task going to sleep.
2397 *
2398 * So for a change of @wl to the local @cpu with an overall group weight change
2399 * of @wl we can compute the new shares distribution (s'_i) using:
2400 *
2401 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2402 *
2403 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2404 * differences in waking a task to CPU 0. The additional task changes the
2405 * weight and shares distributions like:
2406 *
2407 * rw'_i = { 3, 4, 1, 0 }
2408 * s'_i = { 3/8, 4/8, 1/8, 0 }
2409 *
2410 * We can then compute the difference in effective weight by using:
2411 *
2412 * dw_i = S * (s'_i - s_i) (3)
2413 *
2414 * Where 'S' is the group weight as seen by its parent.
2415 *
2416 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2417 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2418 * 4/7) times the weight of the group.
2419 */
2420static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2421{
2422 struct sched_entity *se = tg->se[cpu];
2423
2424 if (!tg->parent) /* the trivial, non-cgroup case */
2425 return wl;
2426
2427 for_each_sched_entity(se) {
2428 long w, W;
2429
2430 tg = se->my_q->tg;
2431
2432 /*
2433 * W = @wg + \Sum rw_j
2434 */
2435 W = wg + calc_tg_weight(tg, se->my_q);
2436
2437 /*
2438 * w = rw_i + @wl
2439 */
2440 w = se->my_q->load.weight + wl;
2441
2442 /*
2443 * wl = S * s'_i; see (2)
2444 */
2445 if (W > 0 && w < W)
2446 wl = (w * tg->shares) / W;
2447 else
2448 wl = tg->shares;
2449
2450 /*
2451 * Per the above, wl is the new se->load.weight value; since
2452 * those are clipped to [MIN_SHARES, ...) do so now. See
2453 * calc_cfs_shares().
2454 */
2455 if (wl < MIN_SHARES)
2456 wl = MIN_SHARES;
2457
2458 /*
2459 * wl = dw_i = S * (s'_i - s_i); see (3)
2460 */
2461 wl -= se->load.weight;
2462
2463 /*
2464 * Recursively apply this logic to all parent groups to compute
2465 * the final effective load change on the root group. Since
2466 * only the @tg group gets extra weight, all parent groups can
2467 * only redistribute existing shares. @wl is the shift in shares
2468 * resulting from this level per the above.
2469 */
2470 wg = 0;
2471 }
2472
2473 return wl;
2474}
2475#else
2476
2477static inline unsigned long effective_load(struct task_group *tg, int cpu,
2478 unsigned long wl, unsigned long wg)
2479{
2480 return wl;
2481}
2482
2483#endif
2484
2485static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2486{
2487 s64 this_load, load;
2488 int idx, this_cpu, prev_cpu;
2489 unsigned long tl_per_task;
2490 struct task_group *tg;
2491 unsigned long weight;
2492 int balanced;
2493
2494 idx = sd->wake_idx;
2495 this_cpu = smp_processor_id();
2496 prev_cpu = task_cpu(p);
2497 load = source_load(prev_cpu, idx);
2498 this_load = target_load(this_cpu, idx);
2499
2500 /*
2501 * If sync wakeup then subtract the (maximum possible)
2502 * effect of the currently running task from the load
2503 * of the current CPU:
2504 */
2505 if (sync) {
2506 tg = task_group(current);
2507 weight = current->se.load.weight;
2508
2509 this_load += effective_load(tg, this_cpu, -weight, -weight);
2510 load += effective_load(tg, prev_cpu, 0, -weight);
2511 }
2512
2513 tg = task_group(p);
2514 weight = p->se.load.weight;
2515
2516 /*
2517 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2518 * due to the sync cause above having dropped this_load to 0, we'll
2519 * always have an imbalance, but there's really nothing you can do
2520 * about that, so that's good too.
2521 *
2522 * Otherwise check if either cpus are near enough in load to allow this
2523 * task to be woken on this_cpu.
2524 */
2525 if (this_load > 0) {
2526 s64 this_eff_load, prev_eff_load;
2527
2528 this_eff_load = 100;
2529 this_eff_load *= power_of(prev_cpu);
2530 this_eff_load *= this_load +
2531 effective_load(tg, this_cpu, weight, weight);
2532
2533 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2534 prev_eff_load *= power_of(this_cpu);
2535 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2536
2537 balanced = this_eff_load <= prev_eff_load;
2538 } else
2539 balanced = true;
2540
2541 /*
2542 * If the currently running task will sleep within
2543 * a reasonable amount of time then attract this newly
2544 * woken task:
2545 */
2546 if (sync && balanced)
2547 return 1;
2548
2549 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2550 tl_per_task = cpu_avg_load_per_task(this_cpu);
2551
2552 if (balanced ||
2553 (this_load <= load &&
2554 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2555 /*
2556 * This domain has SD_WAKE_AFFINE and
2557 * p is cache cold in this domain, and
2558 * there is no bad imbalance.
2559 */
2560 schedstat_inc(sd, ttwu_move_affine);
2561 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2562
2563 return 1;
2564 }
2565 return 0;
2566}
2567
2568/*
2569 * find_idlest_group finds and returns the least busy CPU group within the
2570 * domain.
2571 */
2572static struct sched_group *
2573find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2574 int this_cpu, int load_idx)
2575{
2576 struct sched_group *idlest = NULL, *group = sd->groups;
2577 unsigned long min_load = ULONG_MAX, this_load = 0;
2578 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2579
2580 do {
2581 unsigned long load, avg_load;
2582 int local_group;
2583 int i;
2584
2585 /* Skip over this group if it has no CPUs allowed */
2586 if (!cpumask_intersects(sched_group_cpus(group),
2587 tsk_cpus_allowed(p)))
2588 continue;
2589
2590 local_group = cpumask_test_cpu(this_cpu,
2591 sched_group_cpus(group));
2592
2593 /* Tally up the load of all CPUs in the group */
2594 avg_load = 0;
2595
2596 for_each_cpu(i, sched_group_cpus(group)) {
2597 /* Bias balancing toward cpus of our domain */
2598 if (local_group)
2599 load = source_load(i, load_idx);
2600 else
2601 load = target_load(i, load_idx);
2602
2603 avg_load += load;
2604 }
2605
2606 /* Adjust by relative CPU power of the group */
2607 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2608
2609 if (local_group) {
2610 this_load = avg_load;
2611 } else if (avg_load < min_load) {
2612 min_load = avg_load;
2613 idlest = group;
2614 }
2615 } while (group = group->next, group != sd->groups);
2616
2617 if (!idlest || 100*this_load < imbalance*min_load)
2618 return NULL;
2619 return idlest;
2620}
2621
2622/*
2623 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2624 */
2625static int
2626find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2627{
2628 unsigned long load, min_load = ULONG_MAX;
2629 int idlest = -1;
2630 int i;
2631
2632 /* Traverse only the allowed CPUs */
2633 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2634 load = weighted_cpuload(i);
2635
2636 if (load < min_load || (load == min_load && i == this_cpu)) {
2637 min_load = load;
2638 idlest = i;
2639 }
2640 }
2641
2642 return idlest;
2643}
2644
2645/*
2646 * Try and locate an idle CPU in the sched_domain.
2647 */
2648static int select_idle_sibling(struct task_struct *p, int target)
2649{
2650 int cpu = smp_processor_id();
2651 int prev_cpu = task_cpu(p);
2652 struct sched_domain *sd;
2653 struct sched_group *sg;
2654 int i, smt = 0;
2655
2656 /*
2657 * If the task is going to be woken-up on this cpu and if it is
2658 * already idle, then it is the right target.
2659 */
2660 if (target == cpu && idle_cpu(cpu))
2661 return cpu;
2662
2663 /*
2664 * If the task is going to be woken-up on the cpu where it previously
2665 * ran and if it is currently idle, then it the right target.
2666 */
2667 if (target == prev_cpu && idle_cpu(prev_cpu))
2668 return prev_cpu;
2669
2670 /*
2671 * Otherwise, iterate the domains and find an elegible idle cpu.
2672 */
2673 rcu_read_lock();
2674again:
2675 for_each_domain(target, sd) {
2676 if (!smt && (sd->flags & SD_SHARE_CPUPOWER))
2677 continue;
2678
2679 if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) {
2680 if (!smt) {
2681 smt = 1;
2682 goto again;
2683 }
2684 break;
2685 }
2686
2687 sg = sd->groups;
2688 do {
2689 if (!cpumask_intersects(sched_group_cpus(sg),
2690 tsk_cpus_allowed(p)))
2691 goto next;
2692
2693 for_each_cpu(i, sched_group_cpus(sg)) {
2694 if (!idle_cpu(i))
2695 goto next;
2696 }
2697
2698 target = cpumask_first_and(sched_group_cpus(sg),
2699 tsk_cpus_allowed(p));
2700 goto done;
2701next:
2702 sg = sg->next;
2703 } while (sg != sd->groups);
2704 }
2705done:
2706 rcu_read_unlock();
2707
2708 return target;
2709}
2710
2711/*
2712 * sched_balance_self: balance the current task (running on cpu) in domains
2713 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2714 * SD_BALANCE_EXEC.
2715 *
2716 * Balance, ie. select the least loaded group.
2717 *
2718 * Returns the target CPU number, or the same CPU if no balancing is needed.
2719 *
2720 * preempt must be disabled.
2721 */
2722static int
2723select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2724{
2725 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2726 int cpu = smp_processor_id();
2727 int prev_cpu = task_cpu(p);
2728 int new_cpu = cpu;
2729 int want_affine = 0;
2730 int want_sd = 1;
2731 int sync = wake_flags & WF_SYNC;
2732
2733 if (sd_flag & SD_BALANCE_WAKE) {
2734 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2735 want_affine = 1;
2736 new_cpu = prev_cpu;
2737 }
2738
2739 rcu_read_lock();
2740 for_each_domain(cpu, tmp) {
2741 if (!(tmp->flags & SD_LOAD_BALANCE))
2742 continue;
2743
2744 /*
2745 * If power savings logic is enabled for a domain, see if we
2746 * are not overloaded, if so, don't balance wider.
2747 */
2748 if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
2749 unsigned long power = 0;
2750 unsigned long nr_running = 0;
2751 unsigned long capacity;
2752 int i;
2753
2754 for_each_cpu(i, sched_domain_span(tmp)) {
2755 power += power_of(i);
2756 nr_running += cpu_rq(i)->cfs.nr_running;
2757 }
2758
2759 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2760
2761 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2762 nr_running /= 2;
2763
2764 if (nr_running < capacity)
2765 want_sd = 0;
2766 }
2767
2768 /*
2769 * If both cpu and prev_cpu are part of this domain,
2770 * cpu is a valid SD_WAKE_AFFINE target.
2771 */
2772 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2773 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2774 affine_sd = tmp;
2775 want_affine = 0;
2776 }
2777
2778 if (!want_sd && !want_affine)
2779 break;
2780
2781 if (!(tmp->flags & sd_flag))
2782 continue;
2783
2784 if (want_sd)
2785 sd = tmp;
2786 }
2787
2788 if (affine_sd) {
2789 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2790 prev_cpu = cpu;
2791
2792 new_cpu = select_idle_sibling(p, prev_cpu);
2793 goto unlock;
2794 }
2795
2796 while (sd) {
2797 int load_idx = sd->forkexec_idx;
2798 struct sched_group *group;
2799 int weight;
2800
2801 if (!(sd->flags & sd_flag)) {
2802 sd = sd->child;
2803 continue;
2804 }
2805
2806 if (sd_flag & SD_BALANCE_WAKE)
2807 load_idx = sd->wake_idx;
2808
2809 group = find_idlest_group(sd, p, cpu, load_idx);
2810 if (!group) {
2811 sd = sd->child;
2812 continue;
2813 }
2814
2815 new_cpu = find_idlest_cpu(group, p, cpu);
2816 if (new_cpu == -1 || new_cpu == cpu) {
2817 /* Now try balancing at a lower domain level of cpu */
2818 sd = sd->child;
2819 continue;
2820 }
2821
2822 /* Now try balancing at a lower domain level of new_cpu */
2823 cpu = new_cpu;
2824 weight = sd->span_weight;
2825 sd = NULL;
2826 for_each_domain(cpu, tmp) {
2827 if (weight <= tmp->span_weight)
2828 break;
2829 if (tmp->flags & sd_flag)
2830 sd = tmp;
2831 }
2832 /* while loop will break here if sd == NULL */
2833 }
2834unlock:
2835 rcu_read_unlock();
2836
2837 return new_cpu;
2838}
2839#endif /* CONFIG_SMP */
2840
2841static unsigned long
2842wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2843{
2844 unsigned long gran = sysctl_sched_wakeup_granularity;
2845
2846 /*
2847 * Since its curr running now, convert the gran from real-time
2848 * to virtual-time in his units.
2849 *
2850 * By using 'se' instead of 'curr' we penalize light tasks, so
2851 * they get preempted easier. That is, if 'se' < 'curr' then
2852 * the resulting gran will be larger, therefore penalizing the
2853 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2854 * be smaller, again penalizing the lighter task.
2855 *
2856 * This is especially important for buddies when the leftmost
2857 * task is higher priority than the buddy.
2858 */
2859 return calc_delta_fair(gran, se);
2860}
2861
2862/*
2863 * Should 'se' preempt 'curr'.
2864 *
2865 * |s1
2866 * |s2
2867 * |s3
2868 * g
2869 * |<--->|c
2870 *
2871 * w(c, s1) = -1
2872 * w(c, s2) = 0
2873 * w(c, s3) = 1
2874 *
2875 */
2876static int
2877wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2878{
2879 s64 gran, vdiff = curr->vruntime - se->vruntime;
2880
2881 if (vdiff <= 0)
2882 return -1;
2883
2884 gran = wakeup_gran(curr, se);
2885 if (vdiff > gran)
2886 return 1;
2887
2888 return 0;
2889}
2890
2891static void set_last_buddy(struct sched_entity *se)
2892{
2893 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2894 return;
2895
2896 for_each_sched_entity(se)
2897 cfs_rq_of(se)->last = se;
2898}
2899
2900static void set_next_buddy(struct sched_entity *se)
2901{
2902 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2903 return;
2904
2905 for_each_sched_entity(se)
2906 cfs_rq_of(se)->next = se;
2907}
2908
2909static void set_skip_buddy(struct sched_entity *se)
2910{
2911 for_each_sched_entity(se)
2912 cfs_rq_of(se)->skip = se;
2913}
2914
2915/*
2916 * Preempt the current task with a newly woken task if needed:
2917 */
2918static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2919{
2920 struct task_struct *curr = rq->curr;
2921 struct sched_entity *se = &curr->se, *pse = &p->se;
2922 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2923 int scale = cfs_rq->nr_running >= sched_nr_latency;
2924 int next_buddy_marked = 0;
2925
2926 if (unlikely(se == pse))
2927 return;
2928
2929 /*
2930 * This is possible from callers such as pull_task(), in which we
2931 * unconditionally check_prempt_curr() after an enqueue (which may have
2932 * lead to a throttle). This both saves work and prevents false
2933 * next-buddy nomination below.
2934 */
2935 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
2936 return;
2937
2938 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2939 set_next_buddy(pse);
2940 next_buddy_marked = 1;
2941 }
2942
2943 /*
2944 * We can come here with TIF_NEED_RESCHED already set from new task
2945 * wake up path.
2946 *
2947 * Note: this also catches the edge-case of curr being in a throttled
2948 * group (e.g. via set_curr_task), since update_curr() (in the
2949 * enqueue of curr) will have resulted in resched being set. This
2950 * prevents us from potentially nominating it as a false LAST_BUDDY
2951 * below.
2952 */
2953 if (test_tsk_need_resched(curr))
2954 return;
2955
2956 /* Idle tasks are by definition preempted by non-idle tasks. */
2957 if (unlikely(curr->policy == SCHED_IDLE) &&
2958 likely(p->policy != SCHED_IDLE))
2959 goto preempt;
2960
2961 /*
2962 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2963 * is driven by the tick):
2964 */
2965 if (unlikely(p->policy != SCHED_NORMAL))
2966 return;
2967
2968 find_matching_se(&se, &pse);
2969 update_curr(cfs_rq_of(se));
2970 BUG_ON(!pse);
2971 if (wakeup_preempt_entity(se, pse) == 1) {
2972 /*
2973 * Bias pick_next to pick the sched entity that is
2974 * triggering this preemption.
2975 */
2976 if (!next_buddy_marked)
2977 set_next_buddy(pse);
2978 goto preempt;
2979 }
2980
2981 return;
2982
2983preempt:
2984 resched_task(curr);
2985 /*
2986 * Only set the backward buddy when the current task is still
2987 * on the rq. This can happen when a wakeup gets interleaved
2988 * with schedule on the ->pre_schedule() or idle_balance()
2989 * point, either of which can * drop the rq lock.
2990 *
2991 * Also, during early boot the idle thread is in the fair class,
2992 * for obvious reasons its a bad idea to schedule back to it.
2993 */
2994 if (unlikely(!se->on_rq || curr == rq->idle))
2995 return;
2996
2997 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2998 set_last_buddy(se);
2999}
3000
3001static struct task_struct *pick_next_task_fair(struct rq *rq)
3002{
3003 struct task_struct *p;
3004 struct cfs_rq *cfs_rq = &rq->cfs;
3005 struct sched_entity *se;
3006
3007 if (!cfs_rq->nr_running)
3008 return NULL;
3009
3010 do {
3011 se = pick_next_entity(cfs_rq);
3012 set_next_entity(cfs_rq, se);
3013 cfs_rq = group_cfs_rq(se);
3014 } while (cfs_rq);
3015
3016 p = task_of(se);
3017 hrtick_start_fair(rq, p);
3018
3019 return p;
3020}
3021
3022/*
3023 * Account for a descheduled task:
3024 */
3025static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3026{
3027 struct sched_entity *se = &prev->se;
3028 struct cfs_rq *cfs_rq;
3029
3030 for_each_sched_entity(se) {
3031 cfs_rq = cfs_rq_of(se);
3032 put_prev_entity(cfs_rq, se);
3033 }
3034}
3035
3036/*
3037 * sched_yield() is very simple
3038 *
3039 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3040 */
3041static void yield_task_fair(struct rq *rq)
3042{
3043 struct task_struct *curr = rq->curr;
3044 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3045 struct sched_entity *se = &curr->se;
3046
3047 /*
3048 * Are we the only task in the tree?
3049 */
3050 if (unlikely(rq->nr_running == 1))
3051 return;
3052
3053 clear_buddies(cfs_rq, se);
3054
3055 if (curr->policy != SCHED_BATCH) {
3056 update_rq_clock(rq);
3057 /*
3058 * Update run-time statistics of the 'current'.
3059 */
3060 update_curr(cfs_rq);
3061 }
3062
3063 set_skip_buddy(se);
3064}
3065
3066static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3067{
3068 struct sched_entity *se = &p->se;
3069
3070 /* throttled hierarchies are not runnable */
3071 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3072 return false;
3073
3074 /* Tell the scheduler that we'd really like pse to run next. */
3075 set_next_buddy(se);
3076
3077 yield_task_fair(rq);
3078
3079 return true;
3080}
3081
3082#ifdef CONFIG_SMP
3083/**************************************************
3084 * Fair scheduling class load-balancing methods:
3085 */
3086
3087/*
3088 * pull_task - move a task from a remote runqueue to the local runqueue.
3089 * Both runqueues must be locked.
3090 */
3091static void pull_task(struct rq *src_rq, struct task_struct *p,
3092 struct rq *this_rq, int this_cpu)
3093{
3094 deactivate_task(src_rq, p, 0);
3095 set_task_cpu(p, this_cpu);
3096 activate_task(this_rq, p, 0);
3097 check_preempt_curr(this_rq, p, 0);
3098}
3099
3100/*
3101 * Is this task likely cache-hot:
3102 */
3103static int
3104task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3105{
3106 s64 delta;
3107
3108 if (p->sched_class != &fair_sched_class)
3109 return 0;
3110
3111 if (unlikely(p->policy == SCHED_IDLE))
3112 return 0;
3113
3114 /*
3115 * Buddy candidates are cache hot:
3116 */
3117 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3118 (&p->se == cfs_rq_of(&p->se)->next ||
3119 &p->se == cfs_rq_of(&p->se)->last))
3120 return 1;
3121
3122 if (sysctl_sched_migration_cost == -1)
3123 return 1;
3124 if (sysctl_sched_migration_cost == 0)
3125 return 0;
3126
3127 delta = now - p->se.exec_start;
3128
3129 return delta < (s64)sysctl_sched_migration_cost;
3130}
3131
3132/*
3133 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3134 */
3135static
3136int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3137 struct sched_domain *sd, enum cpu_idle_type idle,
3138 int *all_pinned)
3139{
3140 int tsk_cache_hot = 0;
3141 /*
3142 * We do not migrate tasks that are:
3143 * 1) running (obviously), or
3144 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3145 * 3) are cache-hot on their current CPU.
3146 */
3147 if (!cpumask_test_cpu(this_cpu, tsk_cpus_allowed(p))) {
3148 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3149 return 0;
3150 }
3151 *all_pinned = 0;
3152
3153 if (task_running(rq, p)) {
3154 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
3155 return 0;
3156 }
3157
3158 /*
3159 * Aggressive migration if:
3160 * 1) task is cache cold, or
3161 * 2) too many balance attempts have failed.
3162 */
3163
3164 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
3165 if (!tsk_cache_hot ||
3166 sd->nr_balance_failed > sd->cache_nice_tries) {
3167#ifdef CONFIG_SCHEDSTATS
3168 if (tsk_cache_hot) {
3169 schedstat_inc(sd, lb_hot_gained[idle]);
3170 schedstat_inc(p, se.statistics.nr_forced_migrations);
3171 }
3172#endif
3173 return 1;
3174 }
3175
3176 if (tsk_cache_hot) {
3177 schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
3178 return 0;
3179 }
3180 return 1;
3181}
3182
3183/*
3184 * move_one_task tries to move exactly one task from busiest to this_rq, as
3185 * part of active balancing operations within "domain".
3186 * Returns 1 if successful and 0 otherwise.
3187 *
3188 * Called with both runqueues locked.
3189 */
3190static int
3191move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3192 struct sched_domain *sd, enum cpu_idle_type idle)
3193{
3194 struct task_struct *p, *n;
3195 struct cfs_rq *cfs_rq;
3196 int pinned = 0;
3197
3198 for_each_leaf_cfs_rq(busiest, cfs_rq) {
3199 list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
3200 if (throttled_lb_pair(task_group(p),
3201 busiest->cpu, this_cpu))
3202 break;
3203
3204 if (!can_migrate_task(p, busiest, this_cpu,
3205 sd, idle, &pinned))
3206 continue;
3207
3208 pull_task(busiest, p, this_rq, this_cpu);
3209 /*
3210 * Right now, this is only the second place pull_task()
3211 * is called, so we can safely collect pull_task()
3212 * stats here rather than inside pull_task().
3213 */
3214 schedstat_inc(sd, lb_gained[idle]);
3215 return 1;
3216 }
3217 }
3218
3219 return 0;
3220}
3221
3222static unsigned long
3223balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3224 unsigned long max_load_move, struct sched_domain *sd,
3225 enum cpu_idle_type idle, int *all_pinned,
3226 struct cfs_rq *busiest_cfs_rq)
3227{
3228 int loops = 0, pulled = 0;
3229 long rem_load_move = max_load_move;
3230 struct task_struct *p, *n;
3231
3232 if (max_load_move == 0)
3233 goto out;
3234
3235 list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
3236 if (loops++ > sysctl_sched_nr_migrate)
3237 break;
3238
3239 if ((p->se.load.weight >> 1) > rem_load_move ||
3240 !can_migrate_task(p, busiest, this_cpu, sd, idle,
3241 all_pinned))
3242 continue;
3243
3244 pull_task(busiest, p, this_rq, this_cpu);
3245 pulled++;
3246 rem_load_move -= p->se.load.weight;
3247
3248#ifdef CONFIG_PREEMPT
3249 /*
3250 * NEWIDLE balancing is a source of latency, so preemptible
3251 * kernels will stop after the first task is pulled to minimize
3252 * the critical section.
3253 */
3254 if (idle == CPU_NEWLY_IDLE)
3255 break;
3256#endif
3257
3258 /*
3259 * We only want to steal up to the prescribed amount of
3260 * weighted load.
3261 */
3262 if (rem_load_move <= 0)
3263 break;
3264 }
3265out:
3266 /*
3267 * Right now, this is one of only two places pull_task() is called,
3268 * so we can safely collect pull_task() stats here rather than
3269 * inside pull_task().
3270 */
3271 schedstat_add(sd, lb_gained[idle], pulled);
3272
3273 return max_load_move - rem_load_move;
3274}
3275
3276#ifdef CONFIG_FAIR_GROUP_SCHED
3277/*
3278 * update tg->load_weight by folding this cpu's load_avg
3279 */
3280static int update_shares_cpu(struct task_group *tg, int cpu)
3281{
3282 struct cfs_rq *cfs_rq;
3283 unsigned long flags;
3284 struct rq *rq;
3285
3286 if (!tg->se[cpu])
3287 return 0;
3288
3289 rq = cpu_rq(cpu);
3290 cfs_rq = tg->cfs_rq[cpu];
3291
3292 raw_spin_lock_irqsave(&rq->lock, flags);
3293
3294 update_rq_clock(rq);
3295 update_cfs_load(cfs_rq, 1);
3296
3297 /*
3298 * We need to update shares after updating tg->load_weight in
3299 * order to adjust the weight of groups with long running tasks.
3300 */
3301 update_cfs_shares(cfs_rq);
3302
3303 raw_spin_unlock_irqrestore(&rq->lock, flags);
3304
3305 return 0;
3306}
3307
3308static void update_shares(int cpu)
3309{
3310 struct cfs_rq *cfs_rq;
3311 struct rq *rq = cpu_rq(cpu);
3312
3313 rcu_read_lock();
3314 /*
3315 * Iterates the task_group tree in a bottom up fashion, see
3316 * list_add_leaf_cfs_rq() for details.
3317 */
3318 for_each_leaf_cfs_rq(rq, cfs_rq) {
3319 /* throttled entities do not contribute to load */
3320 if (throttled_hierarchy(cfs_rq))
3321 continue;
3322
3323 update_shares_cpu(cfs_rq->tg, cpu);
3324 }
3325 rcu_read_unlock();
3326}
3327
3328/*
3329 * Compute the cpu's hierarchical load factor for each task group.
3330 * This needs to be done in a top-down fashion because the load of a child
3331 * group is a fraction of its parents load.
3332 */
3333static int tg_load_down(struct task_group *tg, void *data)
3334{
3335 unsigned long load;
3336 long cpu = (long)data;
3337
3338 if (!tg->parent) {
3339 load = cpu_rq(cpu)->load.weight;
3340 } else {
3341 load = tg->parent->cfs_rq[cpu]->h_load;
3342 load *= tg->se[cpu]->load.weight;
3343 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3344 }
3345
3346 tg->cfs_rq[cpu]->h_load = load;
3347
3348 return 0;
3349}
3350
3351static void update_h_load(long cpu)
3352{
3353 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3354}
3355
3356static unsigned long
3357load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
3358 unsigned long max_load_move,
3359 struct sched_domain *sd, enum cpu_idle_type idle,
3360 int *all_pinned)
3361{
3362 long rem_load_move = max_load_move;
3363 struct cfs_rq *busiest_cfs_rq;
3364
3365 rcu_read_lock();
3366 update_h_load(cpu_of(busiest));
3367
3368 for_each_leaf_cfs_rq(busiest, busiest_cfs_rq) {
3369 unsigned long busiest_h_load = busiest_cfs_rq->h_load;
3370 unsigned long busiest_weight = busiest_cfs_rq->load.weight;
3371 u64 rem_load, moved_load;
3372
3373 /*
3374 * empty group or part of a throttled hierarchy
3375 */
3376 if (!busiest_cfs_rq->task_weight ||
3377 throttled_lb_pair(busiest_cfs_rq->tg, cpu_of(busiest), this_cpu))
3378 continue;
3379
3380 rem_load = (u64)rem_load_move * busiest_weight;
3381 rem_load = div_u64(rem_load, busiest_h_load + 1);
3382
3383 moved_load = balance_tasks(this_rq, this_cpu, busiest,
3384 rem_load, sd, idle, all_pinned,
3385 busiest_cfs_rq);
3386
3387 if (!moved_load)
3388 continue;
3389
3390 moved_load *= busiest_h_load;
3391 moved_load = div_u64(moved_load, busiest_weight + 1);
3392
3393 rem_load_move -= moved_load;
3394 if (rem_load_move < 0)
3395 break;
3396 }
3397 rcu_read_unlock();
3398
3399 return max_load_move - rem_load_move;
3400}
3401#else
3402static inline void update_shares(int cpu)
3403{
3404}
3405
3406static unsigned long
3407load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
3408 unsigned long max_load_move,
3409 struct sched_domain *sd, enum cpu_idle_type idle,
3410 int *all_pinned)
3411{
3412 return balance_tasks(this_rq, this_cpu, busiest,
3413 max_load_move, sd, idle, all_pinned,
3414 &busiest->cfs);
3415}
3416#endif
3417
3418/*
3419 * move_tasks tries to move up to max_load_move weighted load from busiest to
3420 * this_rq, as part of a balancing operation within domain "sd".
3421 * Returns 1 if successful and 0 otherwise.
3422 *
3423 * Called with both runqueues locked.
3424 */
3425static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3426 unsigned long max_load_move,
3427 struct sched_domain *sd, enum cpu_idle_type idle,
3428 int *all_pinned)
3429{
3430 unsigned long total_load_moved = 0, load_moved;
3431
3432 do {
3433 load_moved = load_balance_fair(this_rq, this_cpu, busiest,
3434 max_load_move - total_load_moved,
3435 sd, idle, all_pinned);
3436
3437 total_load_moved += load_moved;
3438
3439#ifdef CONFIG_PREEMPT
3440 /*
3441 * NEWIDLE balancing is a source of latency, so preemptible
3442 * kernels will stop after the first task is pulled to minimize
3443 * the critical section.
3444 */
3445 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3446 break;
3447
3448 if (raw_spin_is_contended(&this_rq->lock) ||
3449 raw_spin_is_contended(&busiest->lock))
3450 break;
3451#endif
3452 } while (load_moved && max_load_move > total_load_moved);
3453
3454 return total_load_moved > 0;
3455}
3456
3457/********** Helpers for find_busiest_group ************************/
3458/*
3459 * sd_lb_stats - Structure to store the statistics of a sched_domain
3460 * during load balancing.
3461 */
3462struct sd_lb_stats {
3463 struct sched_group *busiest; /* Busiest group in this sd */
3464 struct sched_group *this; /* Local group in this sd */
3465 unsigned long total_load; /* Total load of all groups in sd */
3466 unsigned long total_pwr; /* Total power of all groups in sd */
3467 unsigned long avg_load; /* Average load across all groups in sd */
3468
3469 /** Statistics of this group */
3470 unsigned long this_load;
3471 unsigned long this_load_per_task;
3472 unsigned long this_nr_running;
3473 unsigned long this_has_capacity;
3474 unsigned int this_idle_cpus;
3475
3476 /* Statistics of the busiest group */
3477 unsigned int busiest_idle_cpus;
3478 unsigned long max_load;
3479 unsigned long busiest_load_per_task;
3480 unsigned long busiest_nr_running;
3481 unsigned long busiest_group_capacity;
3482 unsigned long busiest_has_capacity;
3483 unsigned int busiest_group_weight;
3484
3485 int group_imb; /* Is there imbalance in this sd */
3486#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3487 int power_savings_balance; /* Is powersave balance needed for this sd */
3488 struct sched_group *group_min; /* Least loaded group in sd */
3489 struct sched_group *group_leader; /* Group which relieves group_min */
3490 unsigned long min_load_per_task; /* load_per_task in group_min */
3491 unsigned long leader_nr_running; /* Nr running of group_leader */
3492 unsigned long min_nr_running; /* Nr running of group_min */
3493#endif
3494};
3495
3496/*
3497 * sg_lb_stats - stats of a sched_group required for load_balancing
3498 */
3499struct sg_lb_stats {
3500 unsigned long avg_load; /*Avg load across the CPUs of the group */
3501 unsigned long group_load; /* Total load over the CPUs of the group */
3502 unsigned long sum_nr_running; /* Nr tasks running in the group */
3503 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3504 unsigned long group_capacity;
3505 unsigned long idle_cpus;
3506 unsigned long group_weight;
3507 int group_imb; /* Is there an imbalance in the group ? */
3508 int group_has_capacity; /* Is there extra capacity in the group? */
3509};
3510
3511/**
3512 * get_sd_load_idx - Obtain the load index for a given sched domain.
3513 * @sd: The sched_domain whose load_idx is to be obtained.
3514 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3515 */
3516static inline int get_sd_load_idx(struct sched_domain *sd,
3517 enum cpu_idle_type idle)
3518{
3519 int load_idx;
3520
3521 switch (idle) {
3522 case CPU_NOT_IDLE:
3523 load_idx = sd->busy_idx;
3524 break;
3525
3526 case CPU_NEWLY_IDLE:
3527 load_idx = sd->newidle_idx;
3528 break;
3529 default:
3530 load_idx = sd->idle_idx;
3531 break;
3532 }
3533
3534 return load_idx;
3535}
3536
3537
3538#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3539/**
3540 * init_sd_power_savings_stats - Initialize power savings statistics for
3541 * the given sched_domain, during load balancing.
3542 *
3543 * @sd: Sched domain whose power-savings statistics are to be initialized.
3544 * @sds: Variable containing the statistics for sd.
3545 * @idle: Idle status of the CPU at which we're performing load-balancing.
3546 */
3547static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3548 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3549{
3550 /*
3551 * Busy processors will not participate in power savings
3552 * balance.
3553 */
3554 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3555 sds->power_savings_balance = 0;
3556 else {
3557 sds->power_savings_balance = 1;
3558 sds->min_nr_running = ULONG_MAX;
3559 sds->leader_nr_running = 0;
3560 }
3561}
3562
3563/**
3564 * update_sd_power_savings_stats - Update the power saving stats for a
3565 * sched_domain while performing load balancing.
3566 *
3567 * @group: sched_group belonging to the sched_domain under consideration.
3568 * @sds: Variable containing the statistics of the sched_domain
3569 * @local_group: Does group contain the CPU for which we're performing
3570 * load balancing ?
3571 * @sgs: Variable containing the statistics of the group.
3572 */
3573static inline void update_sd_power_savings_stats(struct sched_group *group,
3574 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3575{
3576
3577 if (!sds->power_savings_balance)
3578 return;
3579
3580 /*
3581 * If the local group is idle or completely loaded
3582 * no need to do power savings balance at this domain
3583 */
3584 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3585 !sds->this_nr_running))
3586 sds->power_savings_balance = 0;
3587
3588 /*
3589 * If a group is already running at full capacity or idle,
3590 * don't include that group in power savings calculations
3591 */
3592 if (!sds->power_savings_balance ||
3593 sgs->sum_nr_running >= sgs->group_capacity ||
3594 !sgs->sum_nr_running)
3595 return;
3596
3597 /*
3598 * Calculate the group which has the least non-idle load.
3599 * This is the group from where we need to pick up the load
3600 * for saving power
3601 */
3602 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3603 (sgs->sum_nr_running == sds->min_nr_running &&
3604 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3605 sds->group_min = group;
3606 sds->min_nr_running = sgs->sum_nr_running;
3607 sds->min_load_per_task = sgs->sum_weighted_load /
3608 sgs->sum_nr_running;
3609 }
3610
3611 /*
3612 * Calculate the group which is almost near its
3613 * capacity but still has some space to pick up some load
3614 * from other group and save more power
3615 */
3616 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3617 return;
3618
3619 if (sgs->sum_nr_running > sds->leader_nr_running ||
3620 (sgs->sum_nr_running == sds->leader_nr_running &&
3621 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3622 sds->group_leader = group;
3623 sds->leader_nr_running = sgs->sum_nr_running;
3624 }
3625}
3626
3627/**
3628 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3629 * @sds: Variable containing the statistics of the sched_domain
3630 * under consideration.
3631 * @this_cpu: Cpu at which we're currently performing load-balancing.
3632 * @imbalance: Variable to store the imbalance.
3633 *
3634 * Description:
3635 * Check if we have potential to perform some power-savings balance.
3636 * If yes, set the busiest group to be the least loaded group in the
3637 * sched_domain, so that it's CPUs can be put to idle.
3638 *
3639 * Returns 1 if there is potential to perform power-savings balance.
3640 * Else returns 0.
3641 */
3642static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3643 int this_cpu, unsigned long *imbalance)
3644{
3645 if (!sds->power_savings_balance)
3646 return 0;
3647
3648 if (sds->this != sds->group_leader ||
3649 sds->group_leader == sds->group_min)
3650 return 0;
3651
3652 *imbalance = sds->min_load_per_task;
3653 sds->busiest = sds->group_min;
3654
3655 return 1;
3656
3657}
3658#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3659static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3660 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3661{
3662 return;
3663}
3664
3665static inline void update_sd_power_savings_stats(struct sched_group *group,
3666 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3667{
3668 return;
3669}
3670
3671static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3672 int this_cpu, unsigned long *imbalance)
3673{
3674 return 0;
3675}
3676#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3677
3678
3679unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3680{
3681 return SCHED_POWER_SCALE;
3682}
3683
3684unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3685{
3686 return default_scale_freq_power(sd, cpu);
3687}
3688
3689unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3690{
3691 unsigned long weight = sd->span_weight;
3692 unsigned long smt_gain = sd->smt_gain;
3693
3694 smt_gain /= weight;
3695
3696 return smt_gain;
3697}
3698
3699unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3700{
3701 return default_scale_smt_power(sd, cpu);
3702}
3703
3704unsigned long scale_rt_power(int cpu)
3705{
3706 struct rq *rq = cpu_rq(cpu);
3707 u64 total, available;
3708
3709 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3710
3711 if (unlikely(total < rq->rt_avg)) {
3712 /* Ensures that power won't end up being negative */
3713 available = 0;
3714 } else {
3715 available = total - rq->rt_avg;
3716 }
3717
3718 if (unlikely((s64)total < SCHED_POWER_SCALE))
3719 total = SCHED_POWER_SCALE;
3720
3721 total >>= SCHED_POWER_SHIFT;
3722
3723 return div_u64(available, total);
3724}
3725
3726static void update_cpu_power(struct sched_domain *sd, int cpu)
3727{
3728 unsigned long weight = sd->span_weight;
3729 unsigned long power = SCHED_POWER_SCALE;
3730 struct sched_group *sdg = sd->groups;
3731
3732 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3733 if (sched_feat(ARCH_POWER))
3734 power *= arch_scale_smt_power(sd, cpu);
3735 else
3736 power *= default_scale_smt_power(sd, cpu);
3737
3738 power >>= SCHED_POWER_SHIFT;
3739 }
3740
3741 sdg->sgp->power_orig = power;
3742
3743 if (sched_feat(ARCH_POWER))
3744 power *= arch_scale_freq_power(sd, cpu);
3745 else
3746 power *= default_scale_freq_power(sd, cpu);
3747
3748 power >>= SCHED_POWER_SHIFT;
3749
3750 power *= scale_rt_power(cpu);
3751 power >>= SCHED_POWER_SHIFT;
3752
3753 if (!power)
3754 power = 1;
3755
3756 cpu_rq(cpu)->cpu_power = power;
3757 sdg->sgp->power = power;
3758}
3759
3760void update_group_power(struct sched_domain *sd, int cpu)
3761{
3762 struct sched_domain *child = sd->child;
3763 struct sched_group *group, *sdg = sd->groups;
3764 unsigned long power;
3765
3766 if (!child) {
3767 update_cpu_power(sd, cpu);
3768 return;
3769 }
3770
3771 power = 0;
3772
3773 group = child->groups;
3774 do {
3775 power += group->sgp->power;
3776 group = group->next;
3777 } while (group != child->groups);
3778
3779 sdg->sgp->power = power;
3780}
3781
3782/*
3783 * Try and fix up capacity for tiny siblings, this is needed when
3784 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3785 * which on its own isn't powerful enough.
3786 *
3787 * See update_sd_pick_busiest() and check_asym_packing().
3788 */
3789static inline int
3790fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3791{
3792 /*
3793 * Only siblings can have significantly less than SCHED_POWER_SCALE
3794 */
3795 if (!(sd->flags & SD_SHARE_CPUPOWER))
3796 return 0;
3797
3798 /*
3799 * If ~90% of the cpu_power is still there, we're good.
3800 */
3801 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3802 return 1;
3803
3804 return 0;
3805}
3806
3807/**
3808 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3809 * @sd: The sched_domain whose statistics are to be updated.
3810 * @group: sched_group whose statistics are to be updated.
3811 * @this_cpu: Cpu for which load balance is currently performed.
3812 * @idle: Idle status of this_cpu
3813 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3814 * @local_group: Does group contain this_cpu.
3815 * @cpus: Set of cpus considered for load balancing.
3816 * @balance: Should we balance.
3817 * @sgs: variable to hold the statistics for this group.
3818 */
3819static inline void update_sg_lb_stats(struct sched_domain *sd,
3820 struct sched_group *group, int this_cpu,
3821 enum cpu_idle_type idle, int load_idx,
3822 int local_group, const struct cpumask *cpus,
3823 int *balance, struct sg_lb_stats *sgs)
3824{
3825 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3826 int i;
3827 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3828 unsigned long avg_load_per_task = 0;
3829
3830 if (local_group)
3831 balance_cpu = group_first_cpu(group);
3832
3833 /* Tally up the load of all CPUs in the group */
3834 max_cpu_load = 0;
3835 min_cpu_load = ~0UL;
3836 max_nr_running = 0;
3837
3838 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3839 struct rq *rq = cpu_rq(i);
3840
3841 /* Bias balancing toward cpus of our domain */
3842 if (local_group) {
3843 if (idle_cpu(i) && !first_idle_cpu) {
3844 first_idle_cpu = 1;
3845 balance_cpu = i;
3846 }
3847
3848 load = target_load(i, load_idx);
3849 } else {
3850 load = source_load(i, load_idx);
3851 if (load > max_cpu_load) {
3852 max_cpu_load = load;
3853 max_nr_running = rq->nr_running;
3854 }
3855 if (min_cpu_load > load)
3856 min_cpu_load = load;
3857 }
3858
3859 sgs->group_load += load;
3860 sgs->sum_nr_running += rq->nr_running;
3861 sgs->sum_weighted_load += weighted_cpuload(i);
3862 if (idle_cpu(i))
3863 sgs->idle_cpus++;
3864 }
3865
3866 /*
3867 * First idle cpu or the first cpu(busiest) in this sched group
3868 * is eligible for doing load balancing at this and above
3869 * domains. In the newly idle case, we will allow all the cpu's
3870 * to do the newly idle load balance.
3871 */
3872 if (idle != CPU_NEWLY_IDLE && local_group) {
3873 if (balance_cpu != this_cpu) {
3874 *balance = 0;
3875 return;
3876 }
3877 update_group_power(sd, this_cpu);
3878 }
3879
3880 /* Adjust by relative CPU power of the group */
3881 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3882
3883 /*
3884 * Consider the group unbalanced when the imbalance is larger
3885 * than the average weight of a task.
3886 *
3887 * APZ: with cgroup the avg task weight can vary wildly and
3888 * might not be a suitable number - should we keep a
3889 * normalized nr_running number somewhere that negates
3890 * the hierarchy?
3891 */
3892 if (sgs->sum_nr_running)
3893 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3894
3895 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
3896 sgs->group_imb = 1;
3897
3898 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3899 SCHED_POWER_SCALE);
3900 if (!sgs->group_capacity)
3901 sgs->group_capacity = fix_small_capacity(sd, group);
3902 sgs->group_weight = group->group_weight;
3903
3904 if (sgs->group_capacity > sgs->sum_nr_running)
3905 sgs->group_has_capacity = 1;
3906}
3907
3908/**
3909 * update_sd_pick_busiest - return 1 on busiest group
3910 * @sd: sched_domain whose statistics are to be checked
3911 * @sds: sched_domain statistics
3912 * @sg: sched_group candidate to be checked for being the busiest
3913 * @sgs: sched_group statistics
3914 * @this_cpu: the current cpu
3915 *
3916 * Determine if @sg is a busier group than the previously selected
3917 * busiest group.
3918 */
3919static bool update_sd_pick_busiest(struct sched_domain *sd,
3920 struct sd_lb_stats *sds,
3921 struct sched_group *sg,
3922 struct sg_lb_stats *sgs,
3923 int this_cpu)
3924{
3925 if (sgs->avg_load <= sds->max_load)
3926 return false;
3927
3928 if (sgs->sum_nr_running > sgs->group_capacity)
3929 return true;
3930
3931 if (sgs->group_imb)
3932 return true;
3933
3934 /*
3935 * ASYM_PACKING needs to move all the work to the lowest
3936 * numbered CPUs in the group, therefore mark all groups
3937 * higher than ourself as busy.
3938 */
3939 if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3940 this_cpu < group_first_cpu(sg)) {
3941 if (!sds->busiest)
3942 return true;
3943
3944 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3945 return true;
3946 }
3947
3948 return false;
3949}
3950
3951/**
3952 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3953 * @sd: sched_domain whose statistics are to be updated.
3954 * @this_cpu: Cpu for which load balance is currently performed.
3955 * @idle: Idle status of this_cpu
3956 * @cpus: Set of cpus considered for load balancing.
3957 * @balance: Should we balance.
3958 * @sds: variable to hold the statistics for this sched_domain.
3959 */
3960static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3961 enum cpu_idle_type idle, const struct cpumask *cpus,
3962 int *balance, struct sd_lb_stats *sds)
3963{
3964 struct sched_domain *child = sd->child;
3965 struct sched_group *sg = sd->groups;
3966 struct sg_lb_stats sgs;
3967 int load_idx, prefer_sibling = 0;
3968
3969 if (child && child->flags & SD_PREFER_SIBLING)
3970 prefer_sibling = 1;
3971
3972 init_sd_power_savings_stats(sd, sds, idle);
3973 load_idx = get_sd_load_idx(sd, idle);
3974
3975 do {
3976 int local_group;
3977
3978 local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
3979 memset(&sgs, 0, sizeof(sgs));
3980 update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
3981 local_group, cpus, balance, &sgs);
3982
3983 if (local_group && !(*balance))
3984 return;
3985
3986 sds->total_load += sgs.group_load;
3987 sds->total_pwr += sg->sgp->power;
3988
3989 /*
3990 * In case the child domain prefers tasks go to siblings
3991 * first, lower the sg capacity to one so that we'll try
3992 * and move all the excess tasks away. We lower the capacity
3993 * of a group only if the local group has the capacity to fit
3994 * these excess tasks, i.e. nr_running < group_capacity. The
3995 * extra check prevents the case where you always pull from the
3996 * heaviest group when it is already under-utilized (possible
3997 * with a large weight task outweighs the tasks on the system).
3998 */
3999 if (prefer_sibling && !local_group && sds->this_has_capacity)
4000 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4001
4002 if (local_group) {
4003 sds->this_load = sgs.avg_load;
4004 sds->this = sg;
4005 sds->this_nr_running = sgs.sum_nr_running;
4006 sds->this_load_per_task = sgs.sum_weighted_load;
4007 sds->this_has_capacity = sgs.group_has_capacity;
4008 sds->this_idle_cpus = sgs.idle_cpus;
4009 } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
4010 sds->max_load = sgs.avg_load;
4011 sds->busiest = sg;
4012 sds->busiest_nr_running = sgs.sum_nr_running;
4013 sds->busiest_idle_cpus = sgs.idle_cpus;
4014 sds->busiest_group_capacity = sgs.group_capacity;
4015 sds->busiest_load_per_task = sgs.sum_weighted_load;
4016 sds->busiest_has_capacity = sgs.group_has_capacity;
4017 sds->busiest_group_weight = sgs.group_weight;
4018 sds->group_imb = sgs.group_imb;
4019 }
4020
4021 update_sd_power_savings_stats(sg, sds, local_group, &sgs);
4022 sg = sg->next;
4023 } while (sg != sd->groups);
4024}
4025
4026/**
4027 * check_asym_packing - Check to see if the group is packed into the
4028 * sched doman.
4029 *
4030 * This is primarily intended to used at the sibling level. Some
4031 * cores like POWER7 prefer to use lower numbered SMT threads. In the
4032 * case of POWER7, it can move to lower SMT modes only when higher
4033 * threads are idle. When in lower SMT modes, the threads will
4034 * perform better since they share less core resources. Hence when we
4035 * have idle threads, we want them to be the higher ones.
4036 *
4037 * This packing function is run on idle threads. It checks to see if
4038 * the busiest CPU in this domain (core in the P7 case) has a higher
4039 * CPU number than the packing function is being run on. Here we are
4040 * assuming lower CPU number will be equivalent to lower a SMT thread
4041 * number.
4042 *
4043 * Returns 1 when packing is required and a task should be moved to
4044 * this CPU. The amount of the imbalance is returned in *imbalance.
4045 *
4046 * @sd: The sched_domain whose packing is to be checked.
4047 * @sds: Statistics of the sched_domain which is to be packed
4048 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
4049 * @imbalance: returns amount of imbalanced due to packing.
4050 */
4051static int check_asym_packing(struct sched_domain *sd,
4052 struct sd_lb_stats *sds,
4053 int this_cpu, unsigned long *imbalance)
4054{
4055 int busiest_cpu;
4056
4057 if (!(sd->flags & SD_ASYM_PACKING))
4058 return 0;
4059
4060 if (!sds->busiest)
4061 return 0;
4062
4063 busiest_cpu = group_first_cpu(sds->busiest);
4064 if (this_cpu > busiest_cpu)
4065 return 0;
4066
4067 *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
4068 SCHED_POWER_SCALE);
4069 return 1;
4070}
4071
4072/**
4073 * fix_small_imbalance - Calculate the minor imbalance that exists
4074 * amongst the groups of a sched_domain, during
4075 * load balancing.
4076 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4077 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
4078 * @imbalance: Variable to store the imbalance.
4079 */
4080static inline void fix_small_imbalance(struct sd_lb_stats *sds,
4081 int this_cpu, unsigned long *imbalance)
4082{
4083 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4084 unsigned int imbn = 2;
4085 unsigned long scaled_busy_load_per_task;
4086
4087 if (sds->this_nr_running) {
4088 sds->this_load_per_task /= sds->this_nr_running;
4089 if (sds->busiest_load_per_task >
4090 sds->this_load_per_task)
4091 imbn = 1;
4092 } else
4093 sds->this_load_per_task =
4094 cpu_avg_load_per_task(this_cpu);
4095
4096 scaled_busy_load_per_task = sds->busiest_load_per_task
4097 * SCHED_POWER_SCALE;
4098 scaled_busy_load_per_task /= sds->busiest->sgp->power;
4099
4100 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4101 (scaled_busy_load_per_task * imbn)) {
4102 *imbalance = sds->busiest_load_per_task;
4103 return;
4104 }
4105
4106 /*
4107 * OK, we don't have enough imbalance to justify moving tasks,
4108 * however we may be able to increase total CPU power used by
4109 * moving them.
4110 */
4111
4112 pwr_now += sds->busiest->sgp->power *
4113 min(sds->busiest_load_per_task, sds->max_load);
4114 pwr_now += sds->this->sgp->power *
4115 min(sds->this_load_per_task, sds->this_load);
4116 pwr_now /= SCHED_POWER_SCALE;
4117
4118 /* Amount of load we'd subtract */
4119 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4120 sds->busiest->sgp->power;
4121 if (sds->max_load > tmp)
4122 pwr_move += sds->busiest->sgp->power *
4123 min(sds->busiest_load_per_task, sds->max_load - tmp);
4124
4125 /* Amount of load we'd add */
4126 if (sds->max_load * sds->busiest->sgp->power <
4127 sds->busiest_load_per_task * SCHED_POWER_SCALE)
4128 tmp = (sds->max_load * sds->busiest->sgp->power) /
4129 sds->this->sgp->power;
4130 else
4131 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
4132 sds->this->sgp->power;
4133 pwr_move += sds->this->sgp->power *
4134 min(sds->this_load_per_task, sds->this_load + tmp);
4135 pwr_move /= SCHED_POWER_SCALE;
4136
4137 /* Move if we gain throughput */
4138 if (pwr_move > pwr_now)
4139 *imbalance = sds->busiest_load_per_task;
4140}
4141
4142/**
4143 * calculate_imbalance - Calculate the amount of imbalance present within the
4144 * groups of a given sched_domain during load balance.
4145 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4146 * @this_cpu: Cpu for which currently load balance is being performed.
4147 * @imbalance: The variable to store the imbalance.
4148 */
4149static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4150 unsigned long *imbalance)
4151{
4152 unsigned long max_pull, load_above_capacity = ~0UL;
4153
4154 sds->busiest_load_per_task /= sds->busiest_nr_running;
4155 if (sds->group_imb) {
4156 sds->busiest_load_per_task =
4157 min(sds->busiest_load_per_task, sds->avg_load);
4158 }
4159
4160 /*
4161 * In the presence of smp nice balancing, certain scenarios can have
4162 * max load less than avg load(as we skip the groups at or below
4163 * its cpu_power, while calculating max_load..)
4164 */
4165 if (sds->max_load < sds->avg_load) {
4166 *imbalance = 0;
4167 return fix_small_imbalance(sds, this_cpu, imbalance);
4168 }
4169
4170 if (!sds->group_imb) {
4171 /*
4172 * Don't want to pull so many tasks that a group would go idle.
4173 */
4174 load_above_capacity = (sds->busiest_nr_running -
4175 sds->busiest_group_capacity);
4176
4177 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4178
4179 load_above_capacity /= sds->busiest->sgp->power;
4180 }
4181
4182 /*
4183 * We're trying to get all the cpus to the average_load, so we don't
4184 * want to push ourselves above the average load, nor do we wish to
4185 * reduce the max loaded cpu below the average load. At the same time,
4186 * we also don't want to reduce the group load below the group capacity
4187 * (so that we can implement power-savings policies etc). Thus we look
4188 * for the minimum possible imbalance.
4189 * Be careful of negative numbers as they'll appear as very large values
4190 * with unsigned longs.
4191 */
4192 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4193
4194 /* How much load to actually move to equalise the imbalance */
4195 *imbalance = min(max_pull * sds->busiest->sgp->power,
4196 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4197 / SCHED_POWER_SCALE;
4198
4199 /*
4200 * if *imbalance is less than the average load per runnable task
4201 * there is no guarantee that any tasks will be moved so we'll have
4202 * a think about bumping its value to force at least one task to be
4203 * moved
4204 */
4205 if (*imbalance < sds->busiest_load_per_task)
4206 return fix_small_imbalance(sds, this_cpu, imbalance);
4207
4208}
4209
4210/******* find_busiest_group() helpers end here *********************/
4211
4212/**
4213 * find_busiest_group - Returns the busiest group within the sched_domain
4214 * if there is an imbalance. If there isn't an imbalance, and
4215 * the user has opted for power-savings, it returns a group whose
4216 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4217 * such a group exists.
4218 *
4219 * Also calculates the amount of weighted load which should be moved
4220 * to restore balance.
4221 *
4222 * @sd: The sched_domain whose busiest group is to be returned.
4223 * @this_cpu: The cpu for which load balancing is currently being performed.
4224 * @imbalance: Variable which stores amount of weighted load which should
4225 * be moved to restore balance/put a group to idle.
4226 * @idle: The idle status of this_cpu.
4227 * @cpus: The set of CPUs under consideration for load-balancing.
4228 * @balance: Pointer to a variable indicating if this_cpu
4229 * is the appropriate cpu to perform load balancing at this_level.
4230 *
4231 * Returns: - the busiest group if imbalance exists.
4232 * - If no imbalance and user has opted for power-savings balance,
4233 * return the least loaded group whose CPUs can be
4234 * put to idle by rebalancing its tasks onto our group.
4235 */
4236static struct sched_group *
4237find_busiest_group(struct sched_domain *sd, int this_cpu,
4238 unsigned long *imbalance, enum cpu_idle_type idle,
4239 const struct cpumask *cpus, int *balance)
4240{
4241 struct sd_lb_stats sds;
4242
4243 memset(&sds, 0, sizeof(sds));
4244
4245 /*
4246 * Compute the various statistics relavent for load balancing at
4247 * this level.
4248 */
4249 update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
4250
4251 /*
4252 * this_cpu is not the appropriate cpu to perform load balancing at
4253 * this level.
4254 */
4255 if (!(*balance))
4256 goto ret;
4257
4258 if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
4259 check_asym_packing(sd, &sds, this_cpu, imbalance))
4260 return sds.busiest;
4261
4262 /* There is no busy sibling group to pull tasks from */
4263 if (!sds.busiest || sds.busiest_nr_running == 0)
4264 goto out_balanced;
4265
4266 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4267
4268 /*
4269 * If the busiest group is imbalanced the below checks don't
4270 * work because they assumes all things are equal, which typically
4271 * isn't true due to cpus_allowed constraints and the like.
4272 */
4273 if (sds.group_imb)
4274 goto force_balance;
4275
4276 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4277 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4278 !sds.busiest_has_capacity)
4279 goto force_balance;
4280
4281 /*
4282 * If the local group is more busy than the selected busiest group
4283 * don't try and pull any tasks.
4284 */
4285 if (sds.this_load >= sds.max_load)
4286 goto out_balanced;
4287
4288 /*
4289 * Don't pull any tasks if this group is already above the domain
4290 * average load.
4291 */
4292 if (sds.this_load >= sds.avg_load)
4293 goto out_balanced;
4294
4295 if (idle == CPU_IDLE) {
4296 /*
4297 * This cpu is idle. If the busiest group load doesn't
4298 * have more tasks than the number of available cpu's and
4299 * there is no imbalance between this and busiest group
4300 * wrt to idle cpu's, it is balanced.
4301 */
4302 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4303 sds.busiest_nr_running <= sds.busiest_group_weight)
4304 goto out_balanced;
4305 } else {
4306 /*
4307 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4308 * imbalance_pct to be conservative.
4309 */
4310 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4311 goto out_balanced;
4312 }
4313
4314force_balance:
4315 /* Looks like there is an imbalance. Compute it */
4316 calculate_imbalance(&sds, this_cpu, imbalance);
4317 return sds.busiest;
4318
4319out_balanced:
4320 /*
4321 * There is no obvious imbalance. But check if we can do some balancing
4322 * to save power.
4323 */
4324 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4325 return sds.busiest;
4326ret:
4327 *imbalance = 0;
4328 return NULL;
4329}
4330
4331/*
4332 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4333 */
4334static struct rq *
4335find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
4336 enum cpu_idle_type idle, unsigned long imbalance,
4337 const struct cpumask *cpus)
4338{
4339 struct rq *busiest = NULL, *rq;
4340 unsigned long max_load = 0;
4341 int i;
4342
4343 for_each_cpu(i, sched_group_cpus(group)) {
4344 unsigned long power = power_of(i);
4345 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4346 SCHED_POWER_SCALE);
4347 unsigned long wl;
4348
4349 if (!capacity)
4350 capacity = fix_small_capacity(sd, group);
4351
4352 if (!cpumask_test_cpu(i, cpus))
4353 continue;
4354
4355 rq = cpu_rq(i);
4356 wl = weighted_cpuload(i);
4357
4358 /*
4359 * When comparing with imbalance, use weighted_cpuload()
4360 * which is not scaled with the cpu power.
4361 */
4362 if (capacity && rq->nr_running == 1 && wl > imbalance)
4363 continue;
4364
4365 /*
4366 * For the load comparisons with the other cpu's, consider
4367 * the weighted_cpuload() scaled with the cpu power, so that
4368 * the load can be moved away from the cpu that is potentially
4369 * running at a lower capacity.
4370 */
4371 wl = (wl * SCHED_POWER_SCALE) / power;
4372
4373 if (wl > max_load) {
4374 max_load = wl;
4375 busiest = rq;
4376 }
4377 }
4378
4379 return busiest;
4380}
4381
4382/*
4383 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4384 * so long as it is large enough.
4385 */
4386#define MAX_PINNED_INTERVAL 512
4387
4388/* Working cpumask for load_balance and load_balance_newidle. */
4389DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4390
4391static int need_active_balance(struct sched_domain *sd, int idle,
4392 int busiest_cpu, int this_cpu)
4393{
4394 if (idle == CPU_NEWLY_IDLE) {
4395
4396 /*
4397 * ASYM_PACKING needs to force migrate tasks from busy but
4398 * higher numbered CPUs in order to pack all tasks in the
4399 * lowest numbered CPUs.
4400 */
4401 if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
4402 return 1;
4403
4404 /*
4405 * The only task running in a non-idle cpu can be moved to this
4406 * cpu in an attempt to completely freeup the other CPU
4407 * package.
4408 *
4409 * The package power saving logic comes from
4410 * find_busiest_group(). If there are no imbalance, then
4411 * f_b_g() will return NULL. However when sched_mc={1,2} then
4412 * f_b_g() will select a group from which a running task may be
4413 * pulled to this cpu in order to make the other package idle.
4414 * If there is no opportunity to make a package idle and if
4415 * there are no imbalance, then f_b_g() will return NULL and no
4416 * action will be taken in load_balance_newidle().
4417 *
4418 * Under normal task pull operation due to imbalance, there
4419 * will be more than one task in the source run queue and
4420 * move_tasks() will succeed. ld_moved will be true and this
4421 * active balance code will not be triggered.
4422 */
4423 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4424 return 0;
4425 }
4426
4427 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4428}
4429
4430static int active_load_balance_cpu_stop(void *data);
4431
4432/*
4433 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4434 * tasks if there is an imbalance.
4435 */
4436static int load_balance(int this_cpu, struct rq *this_rq,
4437 struct sched_domain *sd, enum cpu_idle_type idle,
4438 int *balance)
4439{
4440 int ld_moved, all_pinned = 0, active_balance = 0;
4441 struct sched_group *group;
4442 unsigned long imbalance;
4443 struct rq *busiest;
4444 unsigned long flags;
4445 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4446
4447 cpumask_copy(cpus, cpu_active_mask);
4448
4449 schedstat_inc(sd, lb_count[idle]);
4450
4451redo:
4452 group = find_busiest_group(sd, this_cpu, &imbalance, idle,
4453 cpus, balance);
4454
4455 if (*balance == 0)
4456 goto out_balanced;
4457
4458 if (!group) {
4459 schedstat_inc(sd, lb_nobusyg[idle]);
4460 goto out_balanced;
4461 }
4462
4463 busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
4464 if (!busiest) {
4465 schedstat_inc(sd, lb_nobusyq[idle]);
4466 goto out_balanced;
4467 }
4468
4469 BUG_ON(busiest == this_rq);
4470
4471 schedstat_add(sd, lb_imbalance[idle], imbalance);
4472
4473 ld_moved = 0;
4474 if (busiest->nr_running > 1) {
4475 /*
4476 * Attempt to move tasks. If find_busiest_group has found
4477 * an imbalance but busiest->nr_running <= 1, the group is
4478 * still unbalanced. ld_moved simply stays zero, so it is
4479 * correctly treated as an imbalance.
4480 */
4481 all_pinned = 1;
4482 local_irq_save(flags);
4483 double_rq_lock(this_rq, busiest);
4484 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4485 imbalance, sd, idle, &all_pinned);
4486 double_rq_unlock(this_rq, busiest);
4487 local_irq_restore(flags);
4488
4489 /*
4490 * some other cpu did the load balance for us.
4491 */
4492 if (ld_moved && this_cpu != smp_processor_id())
4493 resched_cpu(this_cpu);
4494
4495 /* All tasks on this runqueue were pinned by CPU affinity */
4496 if (unlikely(all_pinned)) {
4497 cpumask_clear_cpu(cpu_of(busiest), cpus);
4498 if (!cpumask_empty(cpus))
4499 goto redo;
4500 goto out_balanced;
4501 }
4502 }
4503
4504 if (!ld_moved) {
4505 schedstat_inc(sd, lb_failed[idle]);
4506 /*
4507 * Increment the failure counter only on periodic balance.
4508 * We do not want newidle balance, which can be very
4509 * frequent, pollute the failure counter causing
4510 * excessive cache_hot migrations and active balances.
4511 */
4512 if (idle != CPU_NEWLY_IDLE)
4513 sd->nr_balance_failed++;
4514
4515 if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
4516 raw_spin_lock_irqsave(&busiest->lock, flags);
4517
4518 /* don't kick the active_load_balance_cpu_stop,
4519 * if the curr task on busiest cpu can't be
4520 * moved to this_cpu
4521 */
4522 if (!cpumask_test_cpu(this_cpu,
4523 tsk_cpus_allowed(busiest->curr))) {
4524 raw_spin_unlock_irqrestore(&busiest->lock,
4525 flags);
4526 all_pinned = 1;
4527 goto out_one_pinned;
4528 }
4529
4530 /*
4531 * ->active_balance synchronizes accesses to
4532 * ->active_balance_work. Once set, it's cleared
4533 * only after active load balance is finished.
4534 */
4535 if (!busiest->active_balance) {
4536 busiest->active_balance = 1;
4537 busiest->push_cpu = this_cpu;
4538 active_balance = 1;
4539 }
4540 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4541
4542 if (active_balance)
4543 stop_one_cpu_nowait(cpu_of(busiest),
4544 active_load_balance_cpu_stop, busiest,
4545 &busiest->active_balance_work);
4546
4547 /*
4548 * We've kicked active balancing, reset the failure
4549 * counter.
4550 */
4551 sd->nr_balance_failed = sd->cache_nice_tries+1;
4552 }
4553 } else
4554 sd->nr_balance_failed = 0;
4555
4556 if (likely(!active_balance)) {
4557 /* We were unbalanced, so reset the balancing interval */
4558 sd->balance_interval = sd->min_interval;
4559 } else {
4560 /*
4561 * If we've begun active balancing, start to back off. This
4562 * case may not be covered by the all_pinned logic if there
4563 * is only 1 task on the busy runqueue (because we don't call
4564 * move_tasks).
4565 */
4566 if (sd->balance_interval < sd->max_interval)
4567 sd->balance_interval *= 2;
4568 }
4569
4570 goto out;
4571
4572out_balanced:
4573 schedstat_inc(sd, lb_balanced[idle]);
4574
4575 sd->nr_balance_failed = 0;
4576
4577out_one_pinned:
4578 /* tune up the balancing interval */
4579 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4580 (sd->balance_interval < sd->max_interval))
4581 sd->balance_interval *= 2;
4582
4583 ld_moved = 0;
4584out:
4585 return ld_moved;
4586}
4587
4588/*
4589 * idle_balance is called by schedule() if this_cpu is about to become
4590 * idle. Attempts to pull tasks from other CPUs.
4591 */
4592void idle_balance(int this_cpu, struct rq *this_rq)
4593{
4594 struct sched_domain *sd;
4595 int pulled_task = 0;
4596 unsigned long next_balance = jiffies + HZ;
4597
4598 this_rq->idle_stamp = this_rq->clock;
4599
4600 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4601 return;
4602
4603 /*
4604 * Drop the rq->lock, but keep IRQ/preempt disabled.
4605 */
4606 raw_spin_unlock(&this_rq->lock);
4607
4608 update_shares(this_cpu);
4609 rcu_read_lock();
4610 for_each_domain(this_cpu, sd) {
4611 unsigned long interval;
4612 int balance = 1;
4613
4614 if (!(sd->flags & SD_LOAD_BALANCE))
4615 continue;
4616
4617 if (sd->flags & SD_BALANCE_NEWIDLE) {
4618 /* If we've pulled tasks over stop searching: */
4619 pulled_task = load_balance(this_cpu, this_rq,
4620 sd, CPU_NEWLY_IDLE, &balance);
4621 }
4622
4623 interval = msecs_to_jiffies(sd->balance_interval);
4624 if (time_after(next_balance, sd->last_balance + interval))
4625 next_balance = sd->last_balance + interval;
4626 if (pulled_task) {
4627 this_rq->idle_stamp = 0;
4628 break;
4629 }
4630 }
4631 rcu_read_unlock();
4632
4633 raw_spin_lock(&this_rq->lock);
4634
4635 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4636 /*
4637 * We are going idle. next_balance may be set based on
4638 * a busy processor. So reset next_balance.
4639 */
4640 this_rq->next_balance = next_balance;
4641 }
4642}
4643
4644/*
4645 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4646 * running tasks off the busiest CPU onto idle CPUs. It requires at
4647 * least 1 task to be running on each physical CPU where possible, and
4648 * avoids physical / logical imbalances.
4649 */
4650static int active_load_balance_cpu_stop(void *data)
4651{
4652 struct rq *busiest_rq = data;
4653 int busiest_cpu = cpu_of(busiest_rq);
4654 int target_cpu = busiest_rq->push_cpu;
4655 struct rq *target_rq = cpu_rq(target_cpu);
4656 struct sched_domain *sd;
4657
4658 raw_spin_lock_irq(&busiest_rq->lock);
4659
4660 /* make sure the requested cpu hasn't gone down in the meantime */
4661 if (unlikely(busiest_cpu != smp_processor_id() ||
4662 !busiest_rq->active_balance))
4663 goto out_unlock;
4664
4665 /* Is there any task to move? */
4666 if (busiest_rq->nr_running <= 1)
4667 goto out_unlock;
4668
4669 /*
4670 * This condition is "impossible", if it occurs
4671 * we need to fix it. Originally reported by
4672 * Bjorn Helgaas on a 128-cpu setup.
4673 */
4674 BUG_ON(busiest_rq == target_rq);
4675
4676 /* move a task from busiest_rq to target_rq */
4677 double_lock_balance(busiest_rq, target_rq);
4678
4679 /* Search for an sd spanning us and the target CPU. */
4680 rcu_read_lock();
4681 for_each_domain(target_cpu, sd) {
4682 if ((sd->flags & SD_LOAD_BALANCE) &&
4683 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4684 break;
4685 }
4686
4687 if (likely(sd)) {
4688 schedstat_inc(sd, alb_count);
4689
4690 if (move_one_task(target_rq, target_cpu, busiest_rq,
4691 sd, CPU_IDLE))
4692 schedstat_inc(sd, alb_pushed);
4693 else
4694 schedstat_inc(sd, alb_failed);
4695 }
4696 rcu_read_unlock();
4697 double_unlock_balance(busiest_rq, target_rq);
4698out_unlock:
4699 busiest_rq->active_balance = 0;
4700 raw_spin_unlock_irq(&busiest_rq->lock);
4701 return 0;
4702}
4703
4704#ifdef CONFIG_NO_HZ
4705/*
4706 * idle load balancing details
4707 * - One of the idle CPUs nominates itself as idle load_balancer, while
4708 * entering idle.
4709 * - This idle load balancer CPU will also go into tickless mode when
4710 * it is idle, just like all other idle CPUs
4711 * - When one of the busy CPUs notice that there may be an idle rebalancing
4712 * needed, they will kick the idle load balancer, which then does idle
4713 * load balancing for all the idle CPUs.
4714 */
4715static struct {
4716 atomic_t load_balancer;
4717 atomic_t first_pick_cpu;
4718 atomic_t second_pick_cpu;
4719 cpumask_var_t idle_cpus_mask;
4720 cpumask_var_t grp_idle_mask;
4721 unsigned long next_balance; /* in jiffy units */
4722} nohz ____cacheline_aligned;
4723
4724int get_nohz_load_balancer(void)
4725{
4726 return atomic_read(&nohz.load_balancer);
4727}
4728
4729#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4730/**
4731 * lowest_flag_domain - Return lowest sched_domain containing flag.
4732 * @cpu: The cpu whose lowest level of sched domain is to
4733 * be returned.
4734 * @flag: The flag to check for the lowest sched_domain
4735 * for the given cpu.
4736 *
4737 * Returns the lowest sched_domain of a cpu which contains the given flag.
4738 */
4739static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4740{
4741 struct sched_domain *sd;
4742
4743 for_each_domain(cpu, sd)
4744 if (sd->flags & flag)
4745 break;
4746
4747 return sd;
4748}
4749
4750/**
4751 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4752 * @cpu: The cpu whose domains we're iterating over.
4753 * @sd: variable holding the value of the power_savings_sd
4754 * for cpu.
4755 * @flag: The flag to filter the sched_domains to be iterated.
4756 *
4757 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4758 * set, starting from the lowest sched_domain to the highest.
4759 */
4760#define for_each_flag_domain(cpu, sd, flag) \
4761 for (sd = lowest_flag_domain(cpu, flag); \
4762 (sd && (sd->flags & flag)); sd = sd->parent)
4763
4764/**
4765 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4766 * @ilb_group: group to be checked for semi-idleness
4767 *
4768 * Returns: 1 if the group is semi-idle. 0 otherwise.
4769 *
4770 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4771 * and atleast one non-idle CPU. This helper function checks if the given
4772 * sched_group is semi-idle or not.
4773 */
4774static inline int is_semi_idle_group(struct sched_group *ilb_group)
4775{
4776 cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
4777 sched_group_cpus(ilb_group));
4778
4779 /*
4780 * A sched_group is semi-idle when it has atleast one busy cpu
4781 * and atleast one idle cpu.
4782 */
4783 if (cpumask_empty(nohz.grp_idle_mask))
4784 return 0;
4785
4786 if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
4787 return 0;
4788
4789 return 1;
4790}
4791/**
4792 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4793 * @cpu: The cpu which is nominating a new idle_load_balancer.
4794 *
4795 * Returns: Returns the id of the idle load balancer if it exists,
4796 * Else, returns >= nr_cpu_ids.
4797 *
4798 * This algorithm picks the idle load balancer such that it belongs to a
4799 * semi-idle powersavings sched_domain. The idea is to try and avoid
4800 * completely idle packages/cores just for the purpose of idle load balancing
4801 * when there are other idle cpu's which are better suited for that job.
4802 */
4803static int find_new_ilb(int cpu)
4804{
4805 struct sched_domain *sd;
4806 struct sched_group *ilb_group;
4807 int ilb = nr_cpu_ids;
4808
4809 /*
4810 * Have idle load balancer selection from semi-idle packages only
4811 * when power-aware load balancing is enabled
4812 */
4813 if (!(sched_smt_power_savings || sched_mc_power_savings))
4814 goto out_done;
4815
4816 /*
4817 * Optimize for the case when we have no idle CPUs or only one
4818 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4819 */
4820 if (cpumask_weight(nohz.idle_cpus_mask) < 2)
4821 goto out_done;
4822
4823 rcu_read_lock();
4824 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4825 ilb_group = sd->groups;
4826
4827 do {
4828 if (is_semi_idle_group(ilb_group)) {
4829 ilb = cpumask_first(nohz.grp_idle_mask);
4830 goto unlock;
4831 }
4832
4833 ilb_group = ilb_group->next;
4834
4835 } while (ilb_group != sd->groups);
4836 }
4837unlock:
4838 rcu_read_unlock();
4839
4840out_done:
4841 return ilb;
4842}
4843#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4844static inline int find_new_ilb(int call_cpu)
4845{
4846 return nr_cpu_ids;
4847}
4848#endif
4849
4850/*
4851 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4852 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4853 * CPU (if there is one).
4854 */
4855static void nohz_balancer_kick(int cpu)
4856{
4857 int ilb_cpu;
4858
4859 nohz.next_balance++;
4860
4861 ilb_cpu = get_nohz_load_balancer();
4862
4863 if (ilb_cpu >= nr_cpu_ids) {
4864 ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
4865 if (ilb_cpu >= nr_cpu_ids)
4866 return;
4867 }
4868
4869 if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
4870 cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
4871
4872 smp_mb();
4873 /*
4874 * Use smp_send_reschedule() instead of resched_cpu().
4875 * This way we generate a sched IPI on the target cpu which
4876 * is idle. And the softirq performing nohz idle load balance
4877 * will be run before returning from the IPI.
4878 */
4879 smp_send_reschedule(ilb_cpu);
4880 }
4881 return;
4882}
4883
4884/*
4885 * This routine will try to nominate the ilb (idle load balancing)
4886 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4887 * load balancing on behalf of all those cpus.
4888 *
4889 * When the ilb owner becomes busy, we will not have new ilb owner until some
4890 * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
4891 * idle load balancing by kicking one of the idle CPUs.
4892 *
4893 * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
4894 * ilb owner CPU in future (when there is a need for idle load balancing on
4895 * behalf of all idle CPUs).
4896 */
4897void select_nohz_load_balancer(int stop_tick)
4898{
4899 int cpu = smp_processor_id();
4900
4901 if (stop_tick) {
4902 if (!cpu_active(cpu)) {
4903 if (atomic_read(&nohz.load_balancer) != cpu)
4904 return;
4905
4906 /*
4907 * If we are going offline and still the leader,
4908 * give up!
4909 */
4910 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4911 nr_cpu_ids) != cpu)
4912 BUG();
4913
4914 return;
4915 }
4916
4917 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4918
4919 if (atomic_read(&nohz.first_pick_cpu) == cpu)
4920 atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
4921 if (atomic_read(&nohz.second_pick_cpu) == cpu)
4922 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
4923
4924 if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
4925 int new_ilb;
4926
4927 /* make me the ilb owner */
4928 if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
4929 cpu) != nr_cpu_ids)
4930 return;
4931
4932 /*
4933 * Check to see if there is a more power-efficient
4934 * ilb.
4935 */
4936 new_ilb = find_new_ilb(cpu);
4937 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4938 atomic_set(&nohz.load_balancer, nr_cpu_ids);
4939 resched_cpu(new_ilb);
4940 return;
4941 }
4942 return;
4943 }
4944 } else {
4945 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
4946 return;
4947
4948 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4949
4950 if (atomic_read(&nohz.load_balancer) == cpu)
4951 if (atomic_cmpxchg(&nohz.load_balancer, cpu,
4952 nr_cpu_ids) != cpu)
4953 BUG();
4954 }
4955 return;
4956}
4957#endif
4958
4959static DEFINE_SPINLOCK(balancing);
4960
4961static unsigned long __read_mostly max_load_balance_interval = HZ/10;
4962
4963/*
4964 * Scale the max load_balance interval with the number of CPUs in the system.
4965 * This trades load-balance latency on larger machines for less cross talk.
4966 */
4967void update_max_interval(void)
4968{
4969 max_load_balance_interval = HZ*num_online_cpus()/10;
4970}
4971
4972/*
4973 * It checks each scheduling domain to see if it is due to be balanced,
4974 * and initiates a balancing operation if so.
4975 *
4976 * Balancing parameters are set up in arch_init_sched_domains.
4977 */
4978static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4979{
4980 int balance = 1;
4981 struct rq *rq = cpu_rq(cpu);
4982 unsigned long interval;
4983 struct sched_domain *sd;
4984 /* Earliest time when we have to do rebalance again */
4985 unsigned long next_balance = jiffies + 60*HZ;
4986 int update_next_balance = 0;
4987 int need_serialize;
4988
4989 update_shares(cpu);
4990
4991 rcu_read_lock();
4992 for_each_domain(cpu, sd) {
4993 if (!(sd->flags & SD_LOAD_BALANCE))
4994 continue;
4995
4996 interval = sd->balance_interval;
4997 if (idle != CPU_IDLE)
4998 interval *= sd->busy_factor;
4999
5000 /* scale ms to jiffies */
5001 interval = msecs_to_jiffies(interval);
5002 interval = clamp(interval, 1UL, max_load_balance_interval);
5003
5004 need_serialize = sd->flags & SD_SERIALIZE;
5005
5006 if (need_serialize) {
5007 if (!spin_trylock(&balancing))
5008 goto out;
5009 }
5010
5011 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5012 if (load_balance(cpu, rq, sd, idle, &balance)) {
5013 /*
5014 * We've pulled tasks over so either we're no
5015 * longer idle.
5016 */
5017 idle = CPU_NOT_IDLE;
5018 }
5019 sd->last_balance = jiffies;
5020 }
5021 if (need_serialize)
5022 spin_unlock(&balancing);
5023out:
5024 if (time_after(next_balance, sd->last_balance + interval)) {
5025 next_balance = sd->last_balance + interval;
5026 update_next_balance = 1;
5027 }
5028
5029 /*
5030 * Stop the load balance at this level. There is another
5031 * CPU in our sched group which is doing load balancing more
5032 * actively.
5033 */
5034 if (!balance)
5035 break;
5036 }
5037 rcu_read_unlock();
5038
5039 /*
5040 * next_balance will be updated only when there is a need.
5041 * When the cpu is attached to null domain for ex, it will not be
5042 * updated.
5043 */
5044 if (likely(update_next_balance))
5045 rq->next_balance = next_balance;
5046}
5047
5048#ifdef CONFIG_NO_HZ
5049/*
5050 * In CONFIG_NO_HZ case, the idle balance kickee will do the
5051 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5052 */
5053static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
5054{
5055 struct rq *this_rq = cpu_rq(this_cpu);
5056 struct rq *rq;
5057 int balance_cpu;
5058
5059 if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
5060 return;
5061
5062 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
5063 if (balance_cpu == this_cpu)
5064 continue;
5065
5066 /*
5067 * If this cpu gets work to do, stop the load balancing
5068 * work being done for other cpus. Next load
5069 * balancing owner will pick it up.
5070 */
5071 if (need_resched()) {
5072 this_rq->nohz_balance_kick = 0;
5073 break;
5074 }
5075
5076 raw_spin_lock_irq(&this_rq->lock);
5077 update_rq_clock(this_rq);
5078 update_cpu_load(this_rq);
5079 raw_spin_unlock_irq(&this_rq->lock);
5080
5081 rebalance_domains(balance_cpu, CPU_IDLE);
5082
5083 rq = cpu_rq(balance_cpu);
5084 if (time_after(this_rq->next_balance, rq->next_balance))
5085 this_rq->next_balance = rq->next_balance;
5086 }
5087 nohz.next_balance = this_rq->next_balance;
5088 this_rq->nohz_balance_kick = 0;
5089}
5090
5091/*
5092 * Current heuristic for kicking the idle load balancer
5093 * - first_pick_cpu is the one of the busy CPUs. It will kick
5094 * idle load balancer when it has more than one process active. This
5095 * eliminates the need for idle load balancing altogether when we have
5096 * only one running process in the system (common case).
5097 * - If there are more than one busy CPU, idle load balancer may have
5098 * to run for active_load_balance to happen (i.e., two busy CPUs are
5099 * SMT or core siblings and can run better if they move to different
5100 * physical CPUs). So, second_pick_cpu is the second of the busy CPUs
5101 * which will kick idle load balancer as soon as it has any load.
5102 */
5103static inline int nohz_kick_needed(struct rq *rq, int cpu)
5104{
5105 unsigned long now = jiffies;
5106 int ret;
5107 int first_pick_cpu, second_pick_cpu;
5108
5109 if (time_before(now, nohz.next_balance))
5110 return 0;
5111
5112 if (idle_cpu(cpu))
5113 return 0;
5114
5115 first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
5116 second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
5117
5118 if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
5119 second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
5120 return 0;
5121
5122 ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
5123 if (ret == nr_cpu_ids || ret == cpu) {
5124 atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
5125 if (rq->nr_running > 1)
5126 return 1;
5127 } else {
5128 ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
5129 if (ret == nr_cpu_ids || ret == cpu) {
5130 if (rq->nr_running)
5131 return 1;
5132 }
5133 }
5134 return 0;
5135}
5136#else
5137static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
5138#endif
5139
5140/*
5141 * run_rebalance_domains is triggered when needed from the scheduler tick.
5142 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
5143 */
5144static void run_rebalance_domains(struct softirq_action *h)
5145{
5146 int this_cpu = smp_processor_id();
5147 struct rq *this_rq = cpu_rq(this_cpu);
5148 enum cpu_idle_type idle = this_rq->idle_balance ?
5149 CPU_IDLE : CPU_NOT_IDLE;
5150
5151 rebalance_domains(this_cpu, idle);
5152
5153 /*
5154 * If this cpu has a pending nohz_balance_kick, then do the
5155 * balancing on behalf of the other idle cpus whose ticks are
5156 * stopped.
5157 */
5158 nohz_idle_balance(this_cpu, idle);
5159}
5160
5161static inline int on_null_domain(int cpu)
5162{
5163 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
5164}
5165
5166/*
5167 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5168 */
5169void trigger_load_balance(struct rq *rq, int cpu)
5170{
5171 /* Don't need to rebalance while attached to NULL domain */
5172 if (time_after_eq(jiffies, rq->next_balance) &&
5173 likely(!on_null_domain(cpu)))
5174 raise_softirq(SCHED_SOFTIRQ);
5175#ifdef CONFIG_NO_HZ
5176 else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
5177 nohz_balancer_kick(cpu);
5178#endif
5179}
5180
5181static void rq_online_fair(struct rq *rq)
5182{
5183 update_sysctl();
5184}
5185
5186static void rq_offline_fair(struct rq *rq)
5187{
5188 update_sysctl();
5189}
5190
5191#endif /* CONFIG_SMP */
5192
5193/*
5194 * scheduler tick hitting a task of our scheduling class:
5195 */
5196static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
5197{
5198 struct cfs_rq *cfs_rq;
5199 struct sched_entity *se = &curr->se;
5200
5201 for_each_sched_entity(se) {
5202 cfs_rq = cfs_rq_of(se);
5203 entity_tick(cfs_rq, se, queued);
5204 }
5205}
5206
5207/*
5208 * called on fork with the child task as argument from the parent's context
5209 * - child not yet on the tasklist
5210 * - preemption disabled
5211 */
5212static void task_fork_fair(struct task_struct *p)
5213{
5214 struct cfs_rq *cfs_rq = task_cfs_rq(current);
5215 struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
5216 int this_cpu = smp_processor_id();
5217 struct rq *rq = this_rq();
5218 unsigned long flags;
5219
5220 raw_spin_lock_irqsave(&rq->lock, flags);
5221
5222 update_rq_clock(rq);
5223
5224 if (unlikely(task_cpu(p) != this_cpu)) {
5225 rcu_read_lock();
5226 __set_task_cpu(p, this_cpu);
5227 rcu_read_unlock();
5228 }
5229
5230 update_curr(cfs_rq);
5231
5232 if (curr)
5233 se->vruntime = curr->vruntime;
5234 place_entity(cfs_rq, se, 1);
5235
5236 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
5237 /*
5238 * Upon rescheduling, sched_class::put_prev_task() will place
5239 * 'current' within the tree based on its new key value.
5240 */
5241 swap(curr->vruntime, se->vruntime);
5242 resched_task(rq->curr);
5243 }
5244
5245 se->vruntime -= cfs_rq->min_vruntime;
5246
5247 raw_spin_unlock_irqrestore(&rq->lock, flags);
5248}
5249
5250/*
5251 * Priority of the task has changed. Check to see if we preempt
5252 * the current task.
5253 */
5254static void
5255prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
5256{
5257 if (!p->se.on_rq)
5258 return;
5259
5260 /*
5261 * Reschedule if we are currently running on this runqueue and
5262 * our priority decreased, or if we are not currently running on
5263 * this runqueue and our priority is higher than the current's
5264 */
5265 if (rq->curr == p) {
5266 if (p->prio > oldprio)
5267 resched_task(rq->curr);
5268 } else
5269 check_preempt_curr(rq, p, 0);
5270}
5271
5272static void switched_from_fair(struct rq *rq, struct task_struct *p)
5273{
5274 struct sched_entity *se = &p->se;
5275 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5276
5277 /*
5278 * Ensure the task's vruntime is normalized, so that when its
5279 * switched back to the fair class the enqueue_entity(.flags=0) will
5280 * do the right thing.
5281 *
5282 * If it was on_rq, then the dequeue_entity(.flags=0) will already
5283 * have normalized the vruntime, if it was !on_rq, then only when
5284 * the task is sleeping will it still have non-normalized vruntime.
5285 */
5286 if (!se->on_rq && p->state != TASK_RUNNING) {
5287 /*
5288 * Fix up our vruntime so that the current sleep doesn't
5289 * cause 'unlimited' sleep bonus.
5290 */
5291 place_entity(cfs_rq, se, 0);
5292 se->vruntime -= cfs_rq->min_vruntime;
5293 }
5294}
5295
5296/*
5297 * We switched to the sched_fair class.
5298 */
5299static void switched_to_fair(struct rq *rq, struct task_struct *p)
5300{
5301 if (!p->se.on_rq)
5302 return;
5303
5304 /*
5305 * We were most likely switched from sched_rt, so
5306 * kick off the schedule if running, otherwise just see
5307 * if we can still preempt the current task.
5308 */
5309 if (rq->curr == p)
5310 resched_task(rq->curr);
5311 else
5312 check_preempt_curr(rq, p, 0);
5313}
5314
5315/* Account for a task changing its policy or group.
5316 *
5317 * This routine is mostly called to set cfs_rq->curr field when a task
5318 * migrates between groups/classes.
5319 */
5320static void set_curr_task_fair(struct rq *rq)
5321{
5322 struct sched_entity *se = &rq->curr->se;
5323
5324 for_each_sched_entity(se) {
5325 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5326
5327 set_next_entity(cfs_rq, se);
5328 /* ensure bandwidth has been allocated on our new cfs_rq */
5329 account_cfs_rq_runtime(cfs_rq, 0);
5330 }
5331}
5332
5333void init_cfs_rq(struct cfs_rq *cfs_rq)
5334{
5335 cfs_rq->tasks_timeline = RB_ROOT;
5336 INIT_LIST_HEAD(&cfs_rq->tasks);
5337 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5338#ifndef CONFIG_64BIT
5339 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5340#endif
5341}
5342
5343#ifdef CONFIG_FAIR_GROUP_SCHED
5344static void task_move_group_fair(struct task_struct *p, int on_rq)
5345{
5346 /*
5347 * If the task was not on the rq at the time of this cgroup movement
5348 * it must have been asleep, sleeping tasks keep their ->vruntime
5349 * absolute on their old rq until wakeup (needed for the fair sleeper
5350 * bonus in place_entity()).
5351 *
5352 * If it was on the rq, we've just 'preempted' it, which does convert
5353 * ->vruntime to a relative base.
5354 *
5355 * Make sure both cases convert their relative position when migrating
5356 * to another cgroup's rq. This does somewhat interfere with the
5357 * fair sleeper stuff for the first placement, but who cares.
5358 */
5359 if (!on_rq)
5360 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5361 set_task_rq(p, task_cpu(p));
5362 if (!on_rq)
5363 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5364}
5365
5366void free_fair_sched_group(struct task_group *tg)
5367{
5368 int i;
5369
5370 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5371
5372 for_each_possible_cpu(i) {
5373 if (tg->cfs_rq)
5374 kfree(tg->cfs_rq[i]);
5375 if (tg->se)
5376 kfree(tg->se[i]);
5377 }
5378
5379 kfree(tg->cfs_rq);
5380 kfree(tg->se);
5381}
5382
5383int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5384{
5385 struct cfs_rq *cfs_rq;
5386 struct sched_entity *se;
5387 int i;
5388
5389 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5390 if (!tg->cfs_rq)
5391 goto err;
5392 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5393 if (!tg->se)
5394 goto err;
5395
5396 tg->shares = NICE_0_LOAD;
5397
5398 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5399
5400 for_each_possible_cpu(i) {
5401 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5402 GFP_KERNEL, cpu_to_node(i));
5403 if (!cfs_rq)
5404 goto err;
5405
5406 se = kzalloc_node(sizeof(struct sched_entity),
5407 GFP_KERNEL, cpu_to_node(i));
5408 if (!se)
5409 goto err_free_rq;
5410
5411 init_cfs_rq(cfs_rq);
5412 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5413 }
5414
5415 return 1;
5416
5417err_free_rq:
5418 kfree(cfs_rq);
5419err:
5420 return 0;
5421}
5422
5423void unregister_fair_sched_group(struct task_group *tg, int cpu)
5424{
5425 struct rq *rq = cpu_rq(cpu);
5426 unsigned long flags;
5427
5428 /*
5429 * Only empty task groups can be destroyed; so we can speculatively
5430 * check on_list without danger of it being re-added.
5431 */
5432 if (!tg->cfs_rq[cpu]->on_list)
5433 return;
5434
5435 raw_spin_lock_irqsave(&rq->lock, flags);
5436 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5437 raw_spin_unlock_irqrestore(&rq->lock, flags);
5438}
5439
5440void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5441 struct sched_entity *se, int cpu,
5442 struct sched_entity *parent)
5443{
5444 struct rq *rq = cpu_rq(cpu);
5445
5446 cfs_rq->tg = tg;
5447 cfs_rq->rq = rq;
5448#ifdef CONFIG_SMP
5449 /* allow initial update_cfs_load() to truncate */
5450 cfs_rq->load_stamp = 1;
5451#endif
5452 init_cfs_rq_runtime(cfs_rq);
5453
5454 tg->cfs_rq[cpu] = cfs_rq;
5455 tg->se[cpu] = se;
5456
5457 /* se could be NULL for root_task_group */
5458 if (!se)
5459 return;
5460
5461 if (!parent)
5462 se->cfs_rq = &rq->cfs;
5463 else
5464 se->cfs_rq = parent->my_q;
5465
5466 se->my_q = cfs_rq;
5467 update_load_set(&se->load, 0);
5468 se->parent = parent;
5469}
5470
5471static DEFINE_MUTEX(shares_mutex);
5472
5473int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5474{
5475 int i;
5476 unsigned long flags;
5477
5478 /*
5479 * We can't change the weight of the root cgroup.
5480 */
5481 if (!tg->se[0])
5482 return -EINVAL;
5483
5484 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5485
5486 mutex_lock(&shares_mutex);
5487 if (tg->shares == shares)
5488 goto done;
5489
5490 tg->shares = shares;
5491 for_each_possible_cpu(i) {
5492 struct rq *rq = cpu_rq(i);
5493 struct sched_entity *se;
5494
5495 se = tg->se[i];
5496 /* Propagate contribution to hierarchy */
5497 raw_spin_lock_irqsave(&rq->lock, flags);
5498 for_each_sched_entity(se)
5499 update_cfs_shares(group_cfs_rq(se));
5500 raw_spin_unlock_irqrestore(&rq->lock, flags);
5501 }
5502
5503done:
5504 mutex_unlock(&shares_mutex);
5505 return 0;
5506}
5507#else /* CONFIG_FAIR_GROUP_SCHED */
5508
5509void free_fair_sched_group(struct task_group *tg) { }
5510
5511int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5512{
5513 return 1;
5514}
5515
5516void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5517
5518#endif /* CONFIG_FAIR_GROUP_SCHED */
5519
5520
5521static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5522{
5523 struct sched_entity *se = &task->se;
5524 unsigned int rr_interval = 0;
5525
5526 /*
5527 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5528 * idle runqueue:
5529 */
5530 if (rq->cfs.load.weight)
5531 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5532
5533 return rr_interval;
5534}
5535
5536/*
5537 * All the scheduling class methods:
5538 */
5539const struct sched_class fair_sched_class = {
5540 .next = &idle_sched_class,
5541 .enqueue_task = enqueue_task_fair,
5542 .dequeue_task = dequeue_task_fair,
5543 .yield_task = yield_task_fair,
5544 .yield_to_task = yield_to_task_fair,
5545
5546 .check_preempt_curr = check_preempt_wakeup,
5547
5548 .pick_next_task = pick_next_task_fair,
5549 .put_prev_task = put_prev_task_fair,
5550
5551#ifdef CONFIG_SMP
5552 .select_task_rq = select_task_rq_fair,
5553
5554 .rq_online = rq_online_fair,
5555 .rq_offline = rq_offline_fair,
5556
5557 .task_waking = task_waking_fair,
5558#endif
5559
5560 .set_curr_task = set_curr_task_fair,
5561 .task_tick = task_tick_fair,
5562 .task_fork = task_fork_fair,
5563
5564 .prio_changed = prio_changed_fair,
5565 .switched_from = switched_from_fair,
5566 .switched_to = switched_to_fair,
5567
5568 .get_rr_interval = get_rr_interval_fair,
5569
5570#ifdef CONFIG_FAIR_GROUP_SCHED
5571 .task_move_group = task_move_group_fair,
5572#endif
5573};
5574
5575#ifdef CONFIG_SCHED_DEBUG
5576void print_cfs_stats(struct seq_file *m, int cpu)
5577{
5578 struct cfs_rq *cfs_rq;
5579
5580 rcu_read_lock();
5581 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5582 print_cfs_rq(m, cpu, cfs_rq);
5583 rcu_read_unlock();
5584}
5585#endif
5586
5587__init void init_sched_fair_class(void)
5588{
5589#ifdef CONFIG_SMP
5590 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5591
5592#ifdef CONFIG_NO_HZ
5593 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5594 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
5595 atomic_set(&nohz.load_balancer, nr_cpu_ids);
5596 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
5597 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
5598#endif
5599#endif /* SMP */
5600
5601}
generated by cgit v1.2.2 (git 2.25.0) at 2025-09-05 19:16:01 -0400