Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4 *
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 *
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
9 *
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12 *
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16 *
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19 *
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
22 */
23#include "sched.h"
24
25#include <trace/events/sched.h>
26
27/*
28 * Targeted preemption latency for CPU-bound tasks:
29 *
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
34 *
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
37 *
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
39 */
40unsigned int sysctl_sched_latency = 6000000ULL;
41static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42
43/*
44 * The initial- and re-scaling of tunables is configurable
45 *
46 * Options are:
47 *
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
51 *
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
53 */
54enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
55
56/*
57 * Minimal preemption granularity for CPU-bound tasks:
58 *
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
60 */
61unsigned int sysctl_sched_min_granularity = 750000ULL;
62static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
63
64/*
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
66 */
67static unsigned int sched_nr_latency = 8;
68
69/*
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
72 */
73unsigned int sysctl_sched_child_runs_first __read_mostly;
74
75/*
76 * SCHED_OTHER wake-up granularity.
77 *
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
81 *
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
83 */
84unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
86
87const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
88
89#ifdef CONFIG_SMP
90/*
91 * For asym packing, by default the lower numbered CPU has higher priority.
92 */
93int __weak arch_asym_cpu_priority(int cpu)
94{
95 return -cpu;
96}
97
98/*
99 * The margin used when comparing utilization with CPU capacity.
100 *
101 * (default: ~20%)
102 */
103#define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
104
105#endif
106
107#ifdef CONFIG_CFS_BANDWIDTH
108/*
109 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
110 * each time a cfs_rq requests quota.
111 *
112 * Note: in the case that the slice exceeds the runtime remaining (either due
113 * to consumption or the quota being specified to be smaller than the slice)
114 * we will always only issue the remaining available time.
115 *
116 * (default: 5 msec, units: microseconds)
117 */
118unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
119#endif
120
121static inline void update_load_add(struct load_weight *lw, unsigned long inc)
122{
123 lw->weight += inc;
124 lw->inv_weight = 0;
125}
126
127static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
128{
129 lw->weight -= dec;
130 lw->inv_weight = 0;
131}
132
133static inline void update_load_set(struct load_weight *lw, unsigned long w)
134{
135 lw->weight = w;
136 lw->inv_weight = 0;
137}
138
139/*
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
144 * number of CPUs.
145 *
146 * This idea comes from the SD scheduler of Con Kolivas:
147 */
148static unsigned int get_update_sysctl_factor(void)
149{
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
151 unsigned int factor;
152
153 switch (sysctl_sched_tunable_scaling) {
154 case SCHED_TUNABLESCALING_NONE:
155 factor = 1;
156 break;
157 case SCHED_TUNABLESCALING_LINEAR:
158 factor = cpus;
159 break;
160 case SCHED_TUNABLESCALING_LOG:
161 default:
162 factor = 1 + ilog2(cpus);
163 break;
164 }
165
166 return factor;
167}
168
169static void update_sysctl(void)
170{
171 unsigned int factor = get_update_sysctl_factor();
172
173#define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity);
176 SET_SYSCTL(sched_latency);
177 SET_SYSCTL(sched_wakeup_granularity);
178#undef SET_SYSCTL
179}
180
181void sched_init_granularity(void)
182{
183 update_sysctl();
184}
185
186#define WMULT_CONST (~0U)
187#define WMULT_SHIFT 32
188
189static void __update_inv_weight(struct load_weight *lw)
190{
191 unsigned long w;
192
193 if (likely(lw->inv_weight))
194 return;
195
196 w = scale_load_down(lw->weight);
197
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
199 lw->inv_weight = 1;
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
202 else
203 lw->inv_weight = WMULT_CONST / w;
204}
205
206/*
207 * delta_exec * weight / lw.weight
208 * OR
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
210 *
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
214 *
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
217 */
218static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
219{
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
222
223 __update_inv_weight(lw);
224
225 if (unlikely(fact >> 32)) {
226 while (fact >> 32) {
227 fact >>= 1;
228 shift--;
229 }
230 }
231
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
234
235 while (fact >> 32) {
236 fact >>= 1;
237 shift--;
238 }
239
240 return mul_u64_u32_shr(delta_exec, fact, shift);
241}
242
243
244const struct sched_class fair_sched_class;
245
246/**************************************************************
247 * CFS operations on generic schedulable entities:
248 */
249
250#ifdef CONFIG_FAIR_GROUP_SCHED
251static inline struct task_struct *task_of(struct sched_entity *se)
252{
253 SCHED_WARN_ON(!entity_is_task(se));
254 return container_of(se, struct task_struct, se);
255}
256
257/* Walk up scheduling entities hierarchy */
258#define for_each_sched_entity(se) \
259 for (; se; se = se->parent)
260
261static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
262{
263 return p->se.cfs_rq;
264}
265
266/* runqueue on which this entity is (to be) queued */
267static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268{
269 return se->cfs_rq;
270}
271
272/* runqueue "owned" by this group */
273static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
274{
275 return grp->my_q;
276}
277
278static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
279{
280 if (!path)
281 return;
282
283 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
284 autogroup_path(cfs_rq->tg, path, len);
285 else if (cfs_rq && cfs_rq->tg->css.cgroup)
286 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
287 else
288 strlcpy(path, "(null)", len);
289}
290
291static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
292{
293 struct rq *rq = rq_of(cfs_rq);
294 int cpu = cpu_of(rq);
295
296 if (cfs_rq->on_list)
297 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
298
299 cfs_rq->on_list = 1;
300
301 /*
302 * Ensure we either appear before our parent (if already
303 * enqueued) or force our parent to appear after us when it is
304 * enqueued. The fact that we always enqueue bottom-up
305 * reduces this to two cases and a special case for the root
306 * cfs_rq. Furthermore, it also means that we will always reset
307 * tmp_alone_branch either when the branch is connected
308 * to a tree or when we reach the top of the tree
309 */
310 if (cfs_rq->tg->parent &&
311 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
312 /*
313 * If parent is already on the list, we add the child
314 * just before. Thanks to circular linked property of
315 * the list, this means to put the child at the tail
316 * of the list that starts by parent.
317 */
318 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
319 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
320 /*
321 * The branch is now connected to its tree so we can
322 * reset tmp_alone_branch to the beginning of the
323 * list.
324 */
325 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
326 return true;
327 }
328
329 if (!cfs_rq->tg->parent) {
330 /*
331 * cfs rq without parent should be put
332 * at the tail of the list.
333 */
334 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
335 &rq->leaf_cfs_rq_list);
336 /*
337 * We have reach the top of a tree so we can reset
338 * tmp_alone_branch to the beginning of the list.
339 */
340 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
341 return true;
342 }
343
344 /*
345 * The parent has not already been added so we want to
346 * make sure that it will be put after us.
347 * tmp_alone_branch points to the begin of the branch
348 * where we will add parent.
349 */
350 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
351 /*
352 * update tmp_alone_branch to points to the new begin
353 * of the branch
354 */
355 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
356 return false;
357}
358
359static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
360{
361 if (cfs_rq->on_list) {
362 struct rq *rq = rq_of(cfs_rq);
363
364 /*
365 * With cfs_rq being unthrottled/throttled during an enqueue,
366 * it can happen the tmp_alone_branch points the a leaf that
367 * we finally want to del. In this case, tmp_alone_branch moves
368 * to the prev element but it will point to rq->leaf_cfs_rq_list
369 * at the end of the enqueue.
370 */
371 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
372 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
373
374 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
375 cfs_rq->on_list = 0;
376 }
377}
378
379static inline void assert_list_leaf_cfs_rq(struct rq *rq)
380{
381 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
382}
383
384/* Iterate thr' all leaf cfs_rq's on a runqueue */
385#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
386 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
387 leaf_cfs_rq_list)
388
389/* Do the two (enqueued) entities belong to the same group ? */
390static inline struct cfs_rq *
391is_same_group(struct sched_entity *se, struct sched_entity *pse)
392{
393 if (se->cfs_rq == pse->cfs_rq)
394 return se->cfs_rq;
395
396 return NULL;
397}
398
399static inline struct sched_entity *parent_entity(struct sched_entity *se)
400{
401 return se->parent;
402}
403
404static void
405find_matching_se(struct sched_entity **se, struct sched_entity **pse)
406{
407 int se_depth, pse_depth;
408
409 /*
410 * preemption test can be made between sibling entities who are in the
411 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
412 * both tasks until we find their ancestors who are siblings of common
413 * parent.
414 */
415
416 /* First walk up until both entities are at same depth */
417 se_depth = (*se)->depth;
418 pse_depth = (*pse)->depth;
419
420 while (se_depth > pse_depth) {
421 se_depth--;
422 *se = parent_entity(*se);
423 }
424
425 while (pse_depth > se_depth) {
426 pse_depth--;
427 *pse = parent_entity(*pse);
428 }
429
430 while (!is_same_group(*se, *pse)) {
431 *se = parent_entity(*se);
432 *pse = parent_entity(*pse);
433 }
434}
435
436#else /* !CONFIG_FAIR_GROUP_SCHED */
437
438static inline struct task_struct *task_of(struct sched_entity *se)
439{
440 return container_of(se, struct task_struct, se);
441}
442
443#define for_each_sched_entity(se) \
444 for (; se; se = NULL)
445
446static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
447{
448 return &task_rq(p)->cfs;
449}
450
451static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
452{
453 struct task_struct *p = task_of(se);
454 struct rq *rq = task_rq(p);
455
456 return &rq->cfs;
457}
458
459/* runqueue "owned" by this group */
460static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
461{
462 return NULL;
463}
464
465static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
466{
467 if (path)
468 strlcpy(path, "(null)", len);
469}
470
471static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
472{
473 return true;
474}
475
476static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
477{
478}
479
480static inline void assert_list_leaf_cfs_rq(struct rq *rq)
481{
482}
483
484#define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
485 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
486
487static inline struct sched_entity *parent_entity(struct sched_entity *se)
488{
489 return NULL;
490}
491
492static inline void
493find_matching_se(struct sched_entity **se, struct sched_entity **pse)
494{
495}
496
497#endif /* CONFIG_FAIR_GROUP_SCHED */
498
499static __always_inline
500void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
501
502/**************************************************************
503 * Scheduling class tree data structure manipulation methods:
504 */
505
506static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
507{
508 s64 delta = (s64)(vruntime - max_vruntime);
509 if (delta > 0)
510 max_vruntime = vruntime;
511
512 return max_vruntime;
513}
514
515static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
516{
517 s64 delta = (s64)(vruntime - min_vruntime);
518 if (delta < 0)
519 min_vruntime = vruntime;
520
521 return min_vruntime;
522}
523
524static inline int entity_before(struct sched_entity *a,
525 struct sched_entity *b)
526{
527 return (s64)(a->vruntime - b->vruntime) < 0;
528}
529
530static void update_min_vruntime(struct cfs_rq *cfs_rq)
531{
532 struct sched_entity *curr = cfs_rq->curr;
533 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
534
535 u64 vruntime = cfs_rq->min_vruntime;
536
537 if (curr) {
538 if (curr->on_rq)
539 vruntime = curr->vruntime;
540 else
541 curr = NULL;
542 }
543
544 if (leftmost) { /* non-empty tree */
545 struct sched_entity *se;
546 se = rb_entry(leftmost, struct sched_entity, run_node);
547
548 if (!curr)
549 vruntime = se->vruntime;
550 else
551 vruntime = min_vruntime(vruntime, se->vruntime);
552 }
553
554 /* ensure we never gain time by being placed backwards. */
555 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
556#ifndef CONFIG_64BIT
557 smp_wmb();
558 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
559#endif
560}
561
562/*
563 * Enqueue an entity into the rb-tree:
564 */
565static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
566{
567 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
568 struct rb_node *parent = NULL;
569 struct sched_entity *entry;
570 bool leftmost = true;
571
572 /*
573 * Find the right place in the rbtree:
574 */
575 while (*link) {
576 parent = *link;
577 entry = rb_entry(parent, struct sched_entity, run_node);
578 /*
579 * We dont care about collisions. Nodes with
580 * the same key stay together.
581 */
582 if (entity_before(se, entry)) {
583 link = &parent->rb_left;
584 } else {
585 link = &parent->rb_right;
586 leftmost = false;
587 }
588 }
589
590 rb_link_node(&se->run_node, parent, link);
591 rb_insert_color_cached(&se->run_node,
592 &cfs_rq->tasks_timeline, leftmost);
593}
594
595static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
596{
597 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
598}
599
600struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
601{
602 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
603
604 if (!left)
605 return NULL;
606
607 return rb_entry(left, struct sched_entity, run_node);
608}
609
610static struct sched_entity *__pick_next_entity(struct sched_entity *se)
611{
612 struct rb_node *next = rb_next(&se->run_node);
613
614 if (!next)
615 return NULL;
616
617 return rb_entry(next, struct sched_entity, run_node);
618}
619
620#ifdef CONFIG_SCHED_DEBUG
621struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
622{
623 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
624
625 if (!last)
626 return NULL;
627
628 return rb_entry(last, struct sched_entity, run_node);
629}
630
631/**************************************************************
632 * Scheduling class statistics methods:
633 */
634
635int sched_proc_update_handler(struct ctl_table *table, int write,
636 void __user *buffer, size_t *lenp,
637 loff_t *ppos)
638{
639 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
640 unsigned int factor = get_update_sysctl_factor();
641
642 if (ret || !write)
643 return ret;
644
645 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
646 sysctl_sched_min_granularity);
647
648#define WRT_SYSCTL(name) \
649 (normalized_sysctl_##name = sysctl_##name / (factor))
650 WRT_SYSCTL(sched_min_granularity);
651 WRT_SYSCTL(sched_latency);
652 WRT_SYSCTL(sched_wakeup_granularity);
653#undef WRT_SYSCTL
654
655 return 0;
656}
657#endif
658
659/*
660 * delta /= w
661 */
662static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
663{
664 if (unlikely(se->load.weight != NICE_0_LOAD))
665 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
666
667 return delta;
668}
669
670/*
671 * The idea is to set a period in which each task runs once.
672 *
673 * When there are too many tasks (sched_nr_latency) we have to stretch
674 * this period because otherwise the slices get too small.
675 *
676 * p = (nr <= nl) ? l : l*nr/nl
677 */
678static u64 __sched_period(unsigned long nr_running)
679{
680 if (unlikely(nr_running > sched_nr_latency))
681 return nr_running * sysctl_sched_min_granularity;
682 else
683 return sysctl_sched_latency;
684}
685
686/*
687 * We calculate the wall-time slice from the period by taking a part
688 * proportional to the weight.
689 *
690 * s = p*P[w/rw]
691 */
692static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
693{
694 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
695
696 for_each_sched_entity(se) {
697 struct load_weight *load;
698 struct load_weight lw;
699
700 cfs_rq = cfs_rq_of(se);
701 load = &cfs_rq->load;
702
703 if (unlikely(!se->on_rq)) {
704 lw = cfs_rq->load;
705
706 update_load_add(&lw, se->load.weight);
707 load = &lw;
708 }
709 slice = __calc_delta(slice, se->load.weight, load);
710 }
711 return slice;
712}
713
714/*
715 * We calculate the vruntime slice of a to-be-inserted task.
716 *
717 * vs = s/w
718 */
719static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
720{
721 return calc_delta_fair(sched_slice(cfs_rq, se), se);
722}
723
724#include "pelt.h"
725#ifdef CONFIG_SMP
726
727static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
728static unsigned long task_h_load(struct task_struct *p);
729static unsigned long capacity_of(int cpu);
730
731/* Give new sched_entity start runnable values to heavy its load in infant time */
732void init_entity_runnable_average(struct sched_entity *se)
733{
734 struct sched_avg *sa = &se->avg;
735
736 memset(sa, 0, sizeof(*sa));
737
738 /*
739 * Tasks are initialized with full load to be seen as heavy tasks until
740 * they get a chance to stabilize to their real load level.
741 * Group entities are initialized with zero load to reflect the fact that
742 * nothing has been attached to the task group yet.
743 */
744 if (entity_is_task(se))
745 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
746
747 se->runnable_weight = se->load.weight;
748
749 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
750}
751
752static void attach_entity_cfs_rq(struct sched_entity *se);
753
754/*
755 * With new tasks being created, their initial util_avgs are extrapolated
756 * based on the cfs_rq's current util_avg:
757 *
758 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
759 *
760 * However, in many cases, the above util_avg does not give a desired
761 * value. Moreover, the sum of the util_avgs may be divergent, such
762 * as when the series is a harmonic series.
763 *
764 * To solve this problem, we also cap the util_avg of successive tasks to
765 * only 1/2 of the left utilization budget:
766 *
767 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
768 *
769 * where n denotes the nth task and cpu_scale the CPU capacity.
770 *
771 * For example, for a CPU with 1024 of capacity, a simplest series from
772 * the beginning would be like:
773 *
774 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
775 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
776 *
777 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
778 * if util_avg > util_avg_cap.
779 */
780void post_init_entity_util_avg(struct task_struct *p)
781{
782 struct sched_entity *se = &p->se;
783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
784 struct sched_avg *sa = &se->avg;
785 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
786 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
787
788 if (cap > 0) {
789 if (cfs_rq->avg.util_avg != 0) {
790 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
791 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
792
793 if (sa->util_avg > cap)
794 sa->util_avg = cap;
795 } else {
796 sa->util_avg = cap;
797 }
798 }
799
800 if (p->sched_class != &fair_sched_class) {
801 /*
802 * For !fair tasks do:
803 *
804 update_cfs_rq_load_avg(now, cfs_rq);
805 attach_entity_load_avg(cfs_rq, se, 0);
806 switched_from_fair(rq, p);
807 *
808 * such that the next switched_to_fair() has the
809 * expected state.
810 */
811 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
812 return;
813 }
814
815 attach_entity_cfs_rq(se);
816}
817
818#else /* !CONFIG_SMP */
819void init_entity_runnable_average(struct sched_entity *se)
820{
821}
822void post_init_entity_util_avg(struct task_struct *p)
823{
824}
825static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
826{
827}
828#endif /* CONFIG_SMP */
829
830/*
831 * Update the current task's runtime statistics.
832 */
833static void update_curr(struct cfs_rq *cfs_rq)
834{
835 struct sched_entity *curr = cfs_rq->curr;
836 u64 now = rq_clock_task(rq_of(cfs_rq));
837 u64 delta_exec;
838
839 if (unlikely(!curr))
840 return;
841
842 delta_exec = now - curr->exec_start;
843 if (unlikely((s64)delta_exec <= 0))
844 return;
845
846 curr->exec_start = now;
847
848 schedstat_set(curr->statistics.exec_max,
849 max(delta_exec, curr->statistics.exec_max));
850
851 curr->sum_exec_runtime += delta_exec;
852 schedstat_add(cfs_rq->exec_clock, delta_exec);
853
854 curr->vruntime += calc_delta_fair(delta_exec, curr);
855 update_min_vruntime(cfs_rq);
856
857 if (entity_is_task(curr)) {
858 struct task_struct *curtask = task_of(curr);
859
860 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
861 cgroup_account_cputime(curtask, delta_exec);
862 account_group_exec_runtime(curtask, delta_exec);
863 }
864
865 account_cfs_rq_runtime(cfs_rq, delta_exec);
866}
867
868static void update_curr_fair(struct rq *rq)
869{
870 update_curr(cfs_rq_of(&rq->curr->se));
871}
872
873static inline void
874update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
875{
876 u64 wait_start, prev_wait_start;
877
878 if (!schedstat_enabled())
879 return;
880
881 wait_start = rq_clock(rq_of(cfs_rq));
882 prev_wait_start = schedstat_val(se->statistics.wait_start);
883
884 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
885 likely(wait_start > prev_wait_start))
886 wait_start -= prev_wait_start;
887
888 __schedstat_set(se->statistics.wait_start, wait_start);
889}
890
891static inline void
892update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
893{
894 struct task_struct *p;
895 u64 delta;
896
897 if (!schedstat_enabled())
898 return;
899
900 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
901
902 if (entity_is_task(se)) {
903 p = task_of(se);
904 if (task_on_rq_migrating(p)) {
905 /*
906 * Preserve migrating task's wait time so wait_start
907 * time stamp can be adjusted to accumulate wait time
908 * prior to migration.
909 */
910 __schedstat_set(se->statistics.wait_start, delta);
911 return;
912 }
913 trace_sched_stat_wait(p, delta);
914 }
915
916 __schedstat_set(se->statistics.wait_max,
917 max(schedstat_val(se->statistics.wait_max), delta));
918 __schedstat_inc(se->statistics.wait_count);
919 __schedstat_add(se->statistics.wait_sum, delta);
920 __schedstat_set(se->statistics.wait_start, 0);
921}
922
923static inline void
924update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
925{
926 struct task_struct *tsk = NULL;
927 u64 sleep_start, block_start;
928
929 if (!schedstat_enabled())
930 return;
931
932 sleep_start = schedstat_val(se->statistics.sleep_start);
933 block_start = schedstat_val(se->statistics.block_start);
934
935 if (entity_is_task(se))
936 tsk = task_of(se);
937
938 if (sleep_start) {
939 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
940
941 if ((s64)delta < 0)
942 delta = 0;
943
944 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
945 __schedstat_set(se->statistics.sleep_max, delta);
946
947 __schedstat_set(se->statistics.sleep_start, 0);
948 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
949
950 if (tsk) {
951 account_scheduler_latency(tsk, delta >> 10, 1);
952 trace_sched_stat_sleep(tsk, delta);
953 }
954 }
955 if (block_start) {
956 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
957
958 if ((s64)delta < 0)
959 delta = 0;
960
961 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
962 __schedstat_set(se->statistics.block_max, delta);
963
964 __schedstat_set(se->statistics.block_start, 0);
965 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
966
967 if (tsk) {
968 if (tsk->in_iowait) {
969 __schedstat_add(se->statistics.iowait_sum, delta);
970 __schedstat_inc(se->statistics.iowait_count);
971 trace_sched_stat_iowait(tsk, delta);
972 }
973
974 trace_sched_stat_blocked(tsk, delta);
975
976 /*
977 * Blocking time is in units of nanosecs, so shift by
978 * 20 to get a milliseconds-range estimation of the
979 * amount of time that the task spent sleeping:
980 */
981 if (unlikely(prof_on == SLEEP_PROFILING)) {
982 profile_hits(SLEEP_PROFILING,
983 (void *)get_wchan(tsk),
984 delta >> 20);
985 }
986 account_scheduler_latency(tsk, delta >> 10, 0);
987 }
988 }
989}
990
991/*
992 * Task is being enqueued - update stats:
993 */
994static inline void
995update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
996{
997 if (!schedstat_enabled())
998 return;
999
1000 /*
1001 * Are we enqueueing a waiting task? (for current tasks
1002 * a dequeue/enqueue event is a NOP)
1003 */
1004 if (se != cfs_rq->curr)
1005 update_stats_wait_start(cfs_rq, se);
1006
1007 if (flags & ENQUEUE_WAKEUP)
1008 update_stats_enqueue_sleeper(cfs_rq, se);
1009}
1010
1011static inline void
1012update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1013{
1014
1015 if (!schedstat_enabled())
1016 return;
1017
1018 /*
1019 * Mark the end of the wait period if dequeueing a
1020 * waiting task:
1021 */
1022 if (se != cfs_rq->curr)
1023 update_stats_wait_end(cfs_rq, se);
1024
1025 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1026 struct task_struct *tsk = task_of(se);
1027
1028 if (tsk->state & TASK_INTERRUPTIBLE)
1029 __schedstat_set(se->statistics.sleep_start,
1030 rq_clock(rq_of(cfs_rq)));
1031 if (tsk->state & TASK_UNINTERRUPTIBLE)
1032 __schedstat_set(se->statistics.block_start,
1033 rq_clock(rq_of(cfs_rq)));
1034 }
1035}
1036
1037/*
1038 * We are picking a new current task - update its stats:
1039 */
1040static inline void
1041update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1042{
1043 /*
1044 * We are starting a new run period:
1045 */
1046 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1047}
1048
1049/**************************************************
1050 * Scheduling class queueing methods:
1051 */
1052
1053#ifdef CONFIG_NUMA_BALANCING
1054/*
1055 * Approximate time to scan a full NUMA task in ms. The task scan period is
1056 * calculated based on the tasks virtual memory size and
1057 * numa_balancing_scan_size.
1058 */
1059unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1060unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1061
1062/* Portion of address space to scan in MB */
1063unsigned int sysctl_numa_balancing_scan_size = 256;
1064
1065/* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1066unsigned int sysctl_numa_balancing_scan_delay = 1000;
1067
1068struct numa_group {
1069 refcount_t refcount;
1070
1071 spinlock_t lock; /* nr_tasks, tasks */
1072 int nr_tasks;
1073 pid_t gid;
1074 int active_nodes;
1075
1076 struct rcu_head rcu;
1077 unsigned long total_faults;
1078 unsigned long max_faults_cpu;
1079 /*
1080 * Faults_cpu is used to decide whether memory should move
1081 * towards the CPU. As a consequence, these stats are weighted
1082 * more by CPU use than by memory faults.
1083 */
1084 unsigned long *faults_cpu;
1085 unsigned long faults[0];
1086};
1087
1088/*
1089 * For functions that can be called in multiple contexts that permit reading
1090 * ->numa_group (see struct task_struct for locking rules).
1091 */
1092static struct numa_group *deref_task_numa_group(struct task_struct *p)
1093{
1094 return rcu_dereference_check(p->numa_group, p == current ||
1095 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1096}
1097
1098static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1099{
1100 return rcu_dereference_protected(p->numa_group, p == current);
1101}
1102
1103static inline unsigned long group_faults_priv(struct numa_group *ng);
1104static inline unsigned long group_faults_shared(struct numa_group *ng);
1105
1106static unsigned int task_nr_scan_windows(struct task_struct *p)
1107{
1108 unsigned long rss = 0;
1109 unsigned long nr_scan_pages;
1110
1111 /*
1112 * Calculations based on RSS as non-present and empty pages are skipped
1113 * by the PTE scanner and NUMA hinting faults should be trapped based
1114 * on resident pages
1115 */
1116 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1117 rss = get_mm_rss(p->mm);
1118 if (!rss)
1119 rss = nr_scan_pages;
1120
1121 rss = round_up(rss, nr_scan_pages);
1122 return rss / nr_scan_pages;
1123}
1124
1125/* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1126#define MAX_SCAN_WINDOW 2560
1127
1128static unsigned int task_scan_min(struct task_struct *p)
1129{
1130 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1131 unsigned int scan, floor;
1132 unsigned int windows = 1;
1133
1134 if (scan_size < MAX_SCAN_WINDOW)
1135 windows = MAX_SCAN_WINDOW / scan_size;
1136 floor = 1000 / windows;
1137
1138 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1139 return max_t(unsigned int, floor, scan);
1140}
1141
1142static unsigned int task_scan_start(struct task_struct *p)
1143{
1144 unsigned long smin = task_scan_min(p);
1145 unsigned long period = smin;
1146 struct numa_group *ng;
1147
1148 /* Scale the maximum scan period with the amount of shared memory. */
1149 rcu_read_lock();
1150 ng = rcu_dereference(p->numa_group);
1151 if (ng) {
1152 unsigned long shared = group_faults_shared(ng);
1153 unsigned long private = group_faults_priv(ng);
1154
1155 period *= refcount_read(&ng->refcount);
1156 period *= shared + 1;
1157 period /= private + shared + 1;
1158 }
1159 rcu_read_unlock();
1160
1161 return max(smin, period);
1162}
1163
1164static unsigned int task_scan_max(struct task_struct *p)
1165{
1166 unsigned long smin = task_scan_min(p);
1167 unsigned long smax;
1168 struct numa_group *ng;
1169
1170 /* Watch for min being lower than max due to floor calculations */
1171 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1172
1173 /* Scale the maximum scan period with the amount of shared memory. */
1174 ng = deref_curr_numa_group(p);
1175 if (ng) {
1176 unsigned long shared = group_faults_shared(ng);
1177 unsigned long private = group_faults_priv(ng);
1178 unsigned long period = smax;
1179
1180 period *= refcount_read(&ng->refcount);
1181 period *= shared + 1;
1182 period /= private + shared + 1;
1183
1184 smax = max(smax, period);
1185 }
1186
1187 return max(smin, smax);
1188}
1189
1190static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1191{
1192 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1193 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1194}
1195
1196static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1197{
1198 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1199 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1200}
1201
1202/* Shared or private faults. */
1203#define NR_NUMA_HINT_FAULT_TYPES 2
1204
1205/* Memory and CPU locality */
1206#define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1207
1208/* Averaged statistics, and temporary buffers. */
1209#define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1210
1211pid_t task_numa_group_id(struct task_struct *p)
1212{
1213 struct numa_group *ng;
1214 pid_t gid = 0;
1215
1216 rcu_read_lock();
1217 ng = rcu_dereference(p->numa_group);
1218 if (ng)
1219 gid = ng->gid;
1220 rcu_read_unlock();
1221
1222 return gid;
1223}
1224
1225/*
1226 * The averaged statistics, shared & private, memory & CPU,
1227 * occupy the first half of the array. The second half of the
1228 * array is for current counters, which are averaged into the
1229 * first set by task_numa_placement.
1230 */
1231static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1232{
1233 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1234}
1235
1236static inline unsigned long task_faults(struct task_struct *p, int nid)
1237{
1238 if (!p->numa_faults)
1239 return 0;
1240
1241 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1242 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1243}
1244
1245static inline unsigned long group_faults(struct task_struct *p, int nid)
1246{
1247 struct numa_group *ng = deref_task_numa_group(p);
1248
1249 if (!ng)
1250 return 0;
1251
1252 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1253 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1254}
1255
1256static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1257{
1258 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1259 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1260}
1261
1262static inline unsigned long group_faults_priv(struct numa_group *ng)
1263{
1264 unsigned long faults = 0;
1265 int node;
1266
1267 for_each_online_node(node) {
1268 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1269 }
1270
1271 return faults;
1272}
1273
1274static inline unsigned long group_faults_shared(struct numa_group *ng)
1275{
1276 unsigned long faults = 0;
1277 int node;
1278
1279 for_each_online_node(node) {
1280 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1281 }
1282
1283 return faults;
1284}
1285
1286/*
1287 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1288 * considered part of a numa group's pseudo-interleaving set. Migrations
1289 * between these nodes are slowed down, to allow things to settle down.
1290 */
1291#define ACTIVE_NODE_FRACTION 3
1292
1293static bool numa_is_active_node(int nid, struct numa_group *ng)
1294{
1295 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1296}
1297
1298/* Handle placement on systems where not all nodes are directly connected. */
1299static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1300 int maxdist, bool task)
1301{
1302 unsigned long score = 0;
1303 int node;
1304
1305 /*
1306 * All nodes are directly connected, and the same distance
1307 * from each other. No need for fancy placement algorithms.
1308 */
1309 if (sched_numa_topology_type == NUMA_DIRECT)
1310 return 0;
1311
1312 /*
1313 * This code is called for each node, introducing N^2 complexity,
1314 * which should be ok given the number of nodes rarely exceeds 8.
1315 */
1316 for_each_online_node(node) {
1317 unsigned long faults;
1318 int dist = node_distance(nid, node);
1319
1320 /*
1321 * The furthest away nodes in the system are not interesting
1322 * for placement; nid was already counted.
1323 */
1324 if (dist == sched_max_numa_distance || node == nid)
1325 continue;
1326
1327 /*
1328 * On systems with a backplane NUMA topology, compare groups
1329 * of nodes, and move tasks towards the group with the most
1330 * memory accesses. When comparing two nodes at distance
1331 * "hoplimit", only nodes closer by than "hoplimit" are part
1332 * of each group. Skip other nodes.
1333 */
1334 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1335 dist >= maxdist)
1336 continue;
1337
1338 /* Add up the faults from nearby nodes. */
1339 if (task)
1340 faults = task_faults(p, node);
1341 else
1342 faults = group_faults(p, node);
1343
1344 /*
1345 * On systems with a glueless mesh NUMA topology, there are
1346 * no fixed "groups of nodes". Instead, nodes that are not
1347 * directly connected bounce traffic through intermediate
1348 * nodes; a numa_group can occupy any set of nodes.
1349 * The further away a node is, the less the faults count.
1350 * This seems to result in good task placement.
1351 */
1352 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1353 faults *= (sched_max_numa_distance - dist);
1354 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1355 }
1356
1357 score += faults;
1358 }
1359
1360 return score;
1361}
1362
1363/*
1364 * These return the fraction of accesses done by a particular task, or
1365 * task group, on a particular numa node. The group weight is given a
1366 * larger multiplier, in order to group tasks together that are almost
1367 * evenly spread out between numa nodes.
1368 */
1369static inline unsigned long task_weight(struct task_struct *p, int nid,
1370 int dist)
1371{
1372 unsigned long faults, total_faults;
1373
1374 if (!p->numa_faults)
1375 return 0;
1376
1377 total_faults = p->total_numa_faults;
1378
1379 if (!total_faults)
1380 return 0;
1381
1382 faults = task_faults(p, nid);
1383 faults += score_nearby_nodes(p, nid, dist, true);
1384
1385 return 1000 * faults / total_faults;
1386}
1387
1388static inline unsigned long group_weight(struct task_struct *p, int nid,
1389 int dist)
1390{
1391 struct numa_group *ng = deref_task_numa_group(p);
1392 unsigned long faults, total_faults;
1393
1394 if (!ng)
1395 return 0;
1396
1397 total_faults = ng->total_faults;
1398
1399 if (!total_faults)
1400 return 0;
1401
1402 faults = group_faults(p, nid);
1403 faults += score_nearby_nodes(p, nid, dist, false);
1404
1405 return 1000 * faults / total_faults;
1406}
1407
1408bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1409 int src_nid, int dst_cpu)
1410{
1411 struct numa_group *ng = deref_curr_numa_group(p);
1412 int dst_nid = cpu_to_node(dst_cpu);
1413 int last_cpupid, this_cpupid;
1414
1415 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1416 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1417
1418 /*
1419 * Allow first faults or private faults to migrate immediately early in
1420 * the lifetime of a task. The magic number 4 is based on waiting for
1421 * two full passes of the "multi-stage node selection" test that is
1422 * executed below.
1423 */
1424 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1425 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1426 return true;
1427
1428 /*
1429 * Multi-stage node selection is used in conjunction with a periodic
1430 * migration fault to build a temporal task<->page relation. By using
1431 * a two-stage filter we remove short/unlikely relations.
1432 *
1433 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1434 * a task's usage of a particular page (n_p) per total usage of this
1435 * page (n_t) (in a given time-span) to a probability.
1436 *
1437 * Our periodic faults will sample this probability and getting the
1438 * same result twice in a row, given these samples are fully
1439 * independent, is then given by P(n)^2, provided our sample period
1440 * is sufficiently short compared to the usage pattern.
1441 *
1442 * This quadric squishes small probabilities, making it less likely we
1443 * act on an unlikely task<->page relation.
1444 */
1445 if (!cpupid_pid_unset(last_cpupid) &&
1446 cpupid_to_nid(last_cpupid) != dst_nid)
1447 return false;
1448
1449 /* Always allow migrate on private faults */
1450 if (cpupid_match_pid(p, last_cpupid))
1451 return true;
1452
1453 /* A shared fault, but p->numa_group has not been set up yet. */
1454 if (!ng)
1455 return true;
1456
1457 /*
1458 * Destination node is much more heavily used than the source
1459 * node? Allow migration.
1460 */
1461 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1462 ACTIVE_NODE_FRACTION)
1463 return true;
1464
1465 /*
1466 * Distribute memory according to CPU & memory use on each node,
1467 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1468 *
1469 * faults_cpu(dst) 3 faults_cpu(src)
1470 * --------------- * - > ---------------
1471 * faults_mem(dst) 4 faults_mem(src)
1472 */
1473 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1474 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1475}
1476
1477static unsigned long cpu_runnable_load(struct rq *rq);
1478
1479/* Cached statistics for all CPUs within a node */
1480struct numa_stats {
1481 unsigned long load;
1482
1483 /* Total compute capacity of CPUs on a node */
1484 unsigned long compute_capacity;
1485};
1486
1487/*
1488 * XXX borrowed from update_sg_lb_stats
1489 */
1490static void update_numa_stats(struct numa_stats *ns, int nid)
1491{
1492 int cpu;
1493
1494 memset(ns, 0, sizeof(*ns));
1495 for_each_cpu(cpu, cpumask_of_node(nid)) {
1496 struct rq *rq = cpu_rq(cpu);
1497
1498 ns->load += cpu_runnable_load(rq);
1499 ns->compute_capacity += capacity_of(cpu);
1500 }
1501
1502}
1503
1504struct task_numa_env {
1505 struct task_struct *p;
1506
1507 int src_cpu, src_nid;
1508 int dst_cpu, dst_nid;
1509
1510 struct numa_stats src_stats, dst_stats;
1511
1512 int imbalance_pct;
1513 int dist;
1514
1515 struct task_struct *best_task;
1516 long best_imp;
1517 int best_cpu;
1518};
1519
1520static void task_numa_assign(struct task_numa_env *env,
1521 struct task_struct *p, long imp)
1522{
1523 struct rq *rq = cpu_rq(env->dst_cpu);
1524
1525 /* Bail out if run-queue part of active NUMA balance. */
1526 if (xchg(&rq->numa_migrate_on, 1))
1527 return;
1528
1529 /*
1530 * Clear previous best_cpu/rq numa-migrate flag, since task now
1531 * found a better CPU to move/swap.
1532 */
1533 if (env->best_cpu != -1) {
1534 rq = cpu_rq(env->best_cpu);
1535 WRITE_ONCE(rq->numa_migrate_on, 0);
1536 }
1537
1538 if (env->best_task)
1539 put_task_struct(env->best_task);
1540 if (p)
1541 get_task_struct(p);
1542
1543 env->best_task = p;
1544 env->best_imp = imp;
1545 env->best_cpu = env->dst_cpu;
1546}
1547
1548static bool load_too_imbalanced(long src_load, long dst_load,
1549 struct task_numa_env *env)
1550{
1551 long imb, old_imb;
1552 long orig_src_load, orig_dst_load;
1553 long src_capacity, dst_capacity;
1554
1555 /*
1556 * The load is corrected for the CPU capacity available on each node.
1557 *
1558 * src_load dst_load
1559 * ------------ vs ---------
1560 * src_capacity dst_capacity
1561 */
1562 src_capacity = env->src_stats.compute_capacity;
1563 dst_capacity = env->dst_stats.compute_capacity;
1564
1565 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1566
1567 orig_src_load = env->src_stats.load;
1568 orig_dst_load = env->dst_stats.load;
1569
1570 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1571
1572 /* Would this change make things worse? */
1573 return (imb > old_imb);
1574}
1575
1576/*
1577 * Maximum NUMA importance can be 1998 (2*999);
1578 * SMALLIMP @ 30 would be close to 1998/64.
1579 * Used to deter task migration.
1580 */
1581#define SMALLIMP 30
1582
1583/*
1584 * This checks if the overall compute and NUMA accesses of the system would
1585 * be improved if the source tasks was migrated to the target dst_cpu taking
1586 * into account that it might be best if task running on the dst_cpu should
1587 * be exchanged with the source task
1588 */
1589static void task_numa_compare(struct task_numa_env *env,
1590 long taskimp, long groupimp, bool maymove)
1591{
1592 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1593 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1594 long imp = p_ng ? groupimp : taskimp;
1595 struct task_struct *cur;
1596 long src_load, dst_load;
1597 int dist = env->dist;
1598 long moveimp = imp;
1599 long load;
1600
1601 if (READ_ONCE(dst_rq->numa_migrate_on))
1602 return;
1603
1604 rcu_read_lock();
1605 cur = rcu_dereference(dst_rq->curr);
1606 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1607 cur = NULL;
1608
1609 /*
1610 * Because we have preemption enabled we can get migrated around and
1611 * end try selecting ourselves (current == env->p) as a swap candidate.
1612 */
1613 if (cur == env->p)
1614 goto unlock;
1615
1616 if (!cur) {
1617 if (maymove && moveimp >= env->best_imp)
1618 goto assign;
1619 else
1620 goto unlock;
1621 }
1622
1623 /*
1624 * "imp" is the fault differential for the source task between the
1625 * source and destination node. Calculate the total differential for
1626 * the source task and potential destination task. The more negative
1627 * the value is, the more remote accesses that would be expected to
1628 * be incurred if the tasks were swapped.
1629 */
1630 /* Skip this swap candidate if cannot move to the source cpu */
1631 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1632 goto unlock;
1633
1634 /*
1635 * If dst and source tasks are in the same NUMA group, or not
1636 * in any group then look only at task weights.
1637 */
1638 cur_ng = rcu_dereference(cur->numa_group);
1639 if (cur_ng == p_ng) {
1640 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1641 task_weight(cur, env->dst_nid, dist);
1642 /*
1643 * Add some hysteresis to prevent swapping the
1644 * tasks within a group over tiny differences.
1645 */
1646 if (cur_ng)
1647 imp -= imp / 16;
1648 } else {
1649 /*
1650 * Compare the group weights. If a task is all by itself
1651 * (not part of a group), use the task weight instead.
1652 */
1653 if (cur_ng && p_ng)
1654 imp += group_weight(cur, env->src_nid, dist) -
1655 group_weight(cur, env->dst_nid, dist);
1656 else
1657 imp += task_weight(cur, env->src_nid, dist) -
1658 task_weight(cur, env->dst_nid, dist);
1659 }
1660
1661 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1662 imp = moveimp;
1663 cur = NULL;
1664 goto assign;
1665 }
1666
1667 /*
1668 * If the NUMA importance is less than SMALLIMP,
1669 * task migration might only result in ping pong
1670 * of tasks and also hurt performance due to cache
1671 * misses.
1672 */
1673 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1674 goto unlock;
1675
1676 /*
1677 * In the overloaded case, try and keep the load balanced.
1678 */
1679 load = task_h_load(env->p) - task_h_load(cur);
1680 if (!load)
1681 goto assign;
1682
1683 dst_load = env->dst_stats.load + load;
1684 src_load = env->src_stats.load - load;
1685
1686 if (load_too_imbalanced(src_load, dst_load, env))
1687 goto unlock;
1688
1689assign:
1690 /*
1691 * One idle CPU per node is evaluated for a task numa move.
1692 * Call select_idle_sibling to maybe find a better one.
1693 */
1694 if (!cur) {
1695 /*
1696 * select_idle_siblings() uses an per-CPU cpumask that
1697 * can be used from IRQ context.
1698 */
1699 local_irq_disable();
1700 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1701 env->dst_cpu);
1702 local_irq_enable();
1703 }
1704
1705 task_numa_assign(env, cur, imp);
1706unlock:
1707 rcu_read_unlock();
1708}
1709
1710static void task_numa_find_cpu(struct task_numa_env *env,
1711 long taskimp, long groupimp)
1712{
1713 long src_load, dst_load, load;
1714 bool maymove = false;
1715 int cpu;
1716
1717 load = task_h_load(env->p);
1718 dst_load = env->dst_stats.load + load;
1719 src_load = env->src_stats.load - load;
1720
1721 /*
1722 * If the improvement from just moving env->p direction is better
1723 * than swapping tasks around, check if a move is possible.
1724 */
1725 maymove = !load_too_imbalanced(src_load, dst_load, env);
1726
1727 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1728 /* Skip this CPU if the source task cannot migrate */
1729 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1730 continue;
1731
1732 env->dst_cpu = cpu;
1733 task_numa_compare(env, taskimp, groupimp, maymove);
1734 }
1735}
1736
1737static int task_numa_migrate(struct task_struct *p)
1738{
1739 struct task_numa_env env = {
1740 .p = p,
1741
1742 .src_cpu = task_cpu(p),
1743 .src_nid = task_node(p),
1744
1745 .imbalance_pct = 112,
1746
1747 .best_task = NULL,
1748 .best_imp = 0,
1749 .best_cpu = -1,
1750 };
1751 unsigned long taskweight, groupweight;
1752 struct sched_domain *sd;
1753 long taskimp, groupimp;
1754 struct numa_group *ng;
1755 struct rq *best_rq;
1756 int nid, ret, dist;
1757
1758 /*
1759 * Pick the lowest SD_NUMA domain, as that would have the smallest
1760 * imbalance and would be the first to start moving tasks about.
1761 *
1762 * And we want to avoid any moving of tasks about, as that would create
1763 * random movement of tasks -- counter the numa conditions we're trying
1764 * to satisfy here.
1765 */
1766 rcu_read_lock();
1767 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1768 if (sd)
1769 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1770 rcu_read_unlock();
1771
1772 /*
1773 * Cpusets can break the scheduler domain tree into smaller
1774 * balance domains, some of which do not cross NUMA boundaries.
1775 * Tasks that are "trapped" in such domains cannot be migrated
1776 * elsewhere, so there is no point in (re)trying.
1777 */
1778 if (unlikely(!sd)) {
1779 sched_setnuma(p, task_node(p));
1780 return -EINVAL;
1781 }
1782
1783 env.dst_nid = p->numa_preferred_nid;
1784 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1785 taskweight = task_weight(p, env.src_nid, dist);
1786 groupweight = group_weight(p, env.src_nid, dist);
1787 update_numa_stats(&env.src_stats, env.src_nid);
1788 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1789 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1790 update_numa_stats(&env.dst_stats, env.dst_nid);
1791
1792 /* Try to find a spot on the preferred nid. */
1793 task_numa_find_cpu(&env, taskimp, groupimp);
1794
1795 /*
1796 * Look at other nodes in these cases:
1797 * - there is no space available on the preferred_nid
1798 * - the task is part of a numa_group that is interleaved across
1799 * multiple NUMA nodes; in order to better consolidate the group,
1800 * we need to check other locations.
1801 */
1802 ng = deref_curr_numa_group(p);
1803 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1804 for_each_online_node(nid) {
1805 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1806 continue;
1807
1808 dist = node_distance(env.src_nid, env.dst_nid);
1809 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1810 dist != env.dist) {
1811 taskweight = task_weight(p, env.src_nid, dist);
1812 groupweight = group_weight(p, env.src_nid, dist);
1813 }
1814
1815 /* Only consider nodes where both task and groups benefit */
1816 taskimp = task_weight(p, nid, dist) - taskweight;
1817 groupimp = group_weight(p, nid, dist) - groupweight;
1818 if (taskimp < 0 && groupimp < 0)
1819 continue;
1820
1821 env.dist = dist;
1822 env.dst_nid = nid;
1823 update_numa_stats(&env.dst_stats, env.dst_nid);
1824 task_numa_find_cpu(&env, taskimp, groupimp);
1825 }
1826 }
1827
1828 /*
1829 * If the task is part of a workload that spans multiple NUMA nodes,
1830 * and is migrating into one of the workload's active nodes, remember
1831 * this node as the task's preferred numa node, so the workload can
1832 * settle down.
1833 * A task that migrated to a second choice node will be better off
1834 * trying for a better one later. Do not set the preferred node here.
1835 */
1836 if (ng) {
1837 if (env.best_cpu == -1)
1838 nid = env.src_nid;
1839 else
1840 nid = cpu_to_node(env.best_cpu);
1841
1842 if (nid != p->numa_preferred_nid)
1843 sched_setnuma(p, nid);
1844 }
1845
1846 /* No better CPU than the current one was found. */
1847 if (env.best_cpu == -1)
1848 return -EAGAIN;
1849
1850 best_rq = cpu_rq(env.best_cpu);
1851 if (env.best_task == NULL) {
1852 ret = migrate_task_to(p, env.best_cpu);
1853 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1854 if (ret != 0)
1855 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1856 return ret;
1857 }
1858
1859 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1860 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1861
1862 if (ret != 0)
1863 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1864 put_task_struct(env.best_task);
1865 return ret;
1866}
1867
1868/* Attempt to migrate a task to a CPU on the preferred node. */
1869static void numa_migrate_preferred(struct task_struct *p)
1870{
1871 unsigned long interval = HZ;
1872
1873 /* This task has no NUMA fault statistics yet */
1874 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1875 return;
1876
1877 /* Periodically retry migrating the task to the preferred node */
1878 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1879 p->numa_migrate_retry = jiffies + interval;
1880
1881 /* Success if task is already running on preferred CPU */
1882 if (task_node(p) == p->numa_preferred_nid)
1883 return;
1884
1885 /* Otherwise, try migrate to a CPU on the preferred node */
1886 task_numa_migrate(p);
1887}
1888
1889/*
1890 * Find out how many nodes on the workload is actively running on. Do this by
1891 * tracking the nodes from which NUMA hinting faults are triggered. This can
1892 * be different from the set of nodes where the workload's memory is currently
1893 * located.
1894 */
1895static void numa_group_count_active_nodes(struct numa_group *numa_group)
1896{
1897 unsigned long faults, max_faults = 0;
1898 int nid, active_nodes = 0;
1899
1900 for_each_online_node(nid) {
1901 faults = group_faults_cpu(numa_group, nid);
1902 if (faults > max_faults)
1903 max_faults = faults;
1904 }
1905
1906 for_each_online_node(nid) {
1907 faults = group_faults_cpu(numa_group, nid);
1908 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1909 active_nodes++;
1910 }
1911
1912 numa_group->max_faults_cpu = max_faults;
1913 numa_group->active_nodes = active_nodes;
1914}
1915
1916/*
1917 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1918 * increments. The more local the fault statistics are, the higher the scan
1919 * period will be for the next scan window. If local/(local+remote) ratio is
1920 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1921 * the scan period will decrease. Aim for 70% local accesses.
1922 */
1923#define NUMA_PERIOD_SLOTS 10
1924#define NUMA_PERIOD_THRESHOLD 7
1925
1926/*
1927 * Increase the scan period (slow down scanning) if the majority of
1928 * our memory is already on our local node, or if the majority of
1929 * the page accesses are shared with other processes.
1930 * Otherwise, decrease the scan period.
1931 */
1932static void update_task_scan_period(struct task_struct *p,
1933 unsigned long shared, unsigned long private)
1934{
1935 unsigned int period_slot;
1936 int lr_ratio, ps_ratio;
1937 int diff;
1938
1939 unsigned long remote = p->numa_faults_locality[0];
1940 unsigned long local = p->numa_faults_locality[1];
1941
1942 /*
1943 * If there were no record hinting faults then either the task is
1944 * completely idle or all activity is areas that are not of interest
1945 * to automatic numa balancing. Related to that, if there were failed
1946 * migration then it implies we are migrating too quickly or the local
1947 * node is overloaded. In either case, scan slower
1948 */
1949 if (local + shared == 0 || p->numa_faults_locality[2]) {
1950 p->numa_scan_period = min(p->numa_scan_period_max,
1951 p->numa_scan_period << 1);
1952
1953 p->mm->numa_next_scan = jiffies +
1954 msecs_to_jiffies(p->numa_scan_period);
1955
1956 return;
1957 }
1958
1959 /*
1960 * Prepare to scale scan period relative to the current period.
1961 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1962 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1963 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1964 */
1965 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1966 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1967 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1968
1969 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1970 /*
1971 * Most memory accesses are local. There is no need to
1972 * do fast NUMA scanning, since memory is already local.
1973 */
1974 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1975 if (!slot)
1976 slot = 1;
1977 diff = slot * period_slot;
1978 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1979 /*
1980 * Most memory accesses are shared with other tasks.
1981 * There is no point in continuing fast NUMA scanning,
1982 * since other tasks may just move the memory elsewhere.
1983 */
1984 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1985 if (!slot)
1986 slot = 1;
1987 diff = slot * period_slot;
1988 } else {
1989 /*
1990 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1991 * yet they are not on the local NUMA node. Speed up
1992 * NUMA scanning to get the memory moved over.
1993 */
1994 int ratio = max(lr_ratio, ps_ratio);
1995 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1996 }
1997
1998 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1999 task_scan_min(p), task_scan_max(p));
2000 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2001}
2002
2003/*
2004 * Get the fraction of time the task has been running since the last
2005 * NUMA placement cycle. The scheduler keeps similar statistics, but
2006 * decays those on a 32ms period, which is orders of magnitude off
2007 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2008 * stats only if the task is so new there are no NUMA statistics yet.
2009 */
2010static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2011{
2012 u64 runtime, delta, now;
2013 /* Use the start of this time slice to avoid calculations. */
2014 now = p->se.exec_start;
2015 runtime = p->se.sum_exec_runtime;
2016
2017 if (p->last_task_numa_placement) {
2018 delta = runtime - p->last_sum_exec_runtime;
2019 *period = now - p->last_task_numa_placement;
2020
2021 /* Avoid time going backwards, prevent potential divide error: */
2022 if (unlikely((s64)*period < 0))
2023 *period = 0;
2024 } else {
2025 delta = p->se.avg.load_sum;
2026 *period = LOAD_AVG_MAX;
2027 }
2028
2029 p->last_sum_exec_runtime = runtime;
2030 p->last_task_numa_placement = now;
2031
2032 return delta;
2033}
2034
2035/*
2036 * Determine the preferred nid for a task in a numa_group. This needs to
2037 * be done in a way that produces consistent results with group_weight,
2038 * otherwise workloads might not converge.
2039 */
2040static int preferred_group_nid(struct task_struct *p, int nid)
2041{
2042 nodemask_t nodes;
2043 int dist;
2044
2045 /* Direct connections between all NUMA nodes. */
2046 if (sched_numa_topology_type == NUMA_DIRECT)
2047 return nid;
2048
2049 /*
2050 * On a system with glueless mesh NUMA topology, group_weight
2051 * scores nodes according to the number of NUMA hinting faults on
2052 * both the node itself, and on nearby nodes.
2053 */
2054 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2055 unsigned long score, max_score = 0;
2056 int node, max_node = nid;
2057
2058 dist = sched_max_numa_distance;
2059
2060 for_each_online_node(node) {
2061 score = group_weight(p, node, dist);
2062 if (score > max_score) {
2063 max_score = score;
2064 max_node = node;
2065 }
2066 }
2067 return max_node;
2068 }
2069
2070 /*
2071 * Finding the preferred nid in a system with NUMA backplane
2072 * interconnect topology is more involved. The goal is to locate
2073 * tasks from numa_groups near each other in the system, and
2074 * untangle workloads from different sides of the system. This requires
2075 * searching down the hierarchy of node groups, recursively searching
2076 * inside the highest scoring group of nodes. The nodemask tricks
2077 * keep the complexity of the search down.
2078 */
2079 nodes = node_online_map;
2080 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2081 unsigned long max_faults = 0;
2082 nodemask_t max_group = NODE_MASK_NONE;
2083 int a, b;
2084
2085 /* Are there nodes at this distance from each other? */
2086 if (!find_numa_distance(dist))
2087 continue;
2088
2089 for_each_node_mask(a, nodes) {
2090 unsigned long faults = 0;
2091 nodemask_t this_group;
2092 nodes_clear(this_group);
2093
2094 /* Sum group's NUMA faults; includes a==b case. */
2095 for_each_node_mask(b, nodes) {
2096 if (node_distance(a, b) < dist) {
2097 faults += group_faults(p, b);
2098 node_set(b, this_group);
2099 node_clear(b, nodes);
2100 }
2101 }
2102
2103 /* Remember the top group. */
2104 if (faults > max_faults) {
2105 max_faults = faults;
2106 max_group = this_group;
2107 /*
2108 * subtle: at the smallest distance there is
2109 * just one node left in each "group", the
2110 * winner is the preferred nid.
2111 */
2112 nid = a;
2113 }
2114 }
2115 /* Next round, evaluate the nodes within max_group. */
2116 if (!max_faults)
2117 break;
2118 nodes = max_group;
2119 }
2120 return nid;
2121}
2122
2123static void task_numa_placement(struct task_struct *p)
2124{
2125 int seq, nid, max_nid = NUMA_NO_NODE;
2126 unsigned long max_faults = 0;
2127 unsigned long fault_types[2] = { 0, 0 };
2128 unsigned long total_faults;
2129 u64 runtime, period;
2130 spinlock_t *group_lock = NULL;
2131 struct numa_group *ng;
2132
2133 /*
2134 * The p->mm->numa_scan_seq field gets updated without
2135 * exclusive access. Use READ_ONCE() here to ensure
2136 * that the field is read in a single access:
2137 */
2138 seq = READ_ONCE(p->mm->numa_scan_seq);
2139 if (p->numa_scan_seq == seq)
2140 return;
2141 p->numa_scan_seq = seq;
2142 p->numa_scan_period_max = task_scan_max(p);
2143
2144 total_faults = p->numa_faults_locality[0] +
2145 p->numa_faults_locality[1];
2146 runtime = numa_get_avg_runtime(p, &period);
2147
2148 /* If the task is part of a group prevent parallel updates to group stats */
2149 ng = deref_curr_numa_group(p);
2150 if (ng) {
2151 group_lock = &ng->lock;
2152 spin_lock_irq(group_lock);
2153 }
2154
2155 /* Find the node with the highest number of faults */
2156 for_each_online_node(nid) {
2157 /* Keep track of the offsets in numa_faults array */
2158 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2159 unsigned long faults = 0, group_faults = 0;
2160 int priv;
2161
2162 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2163 long diff, f_diff, f_weight;
2164
2165 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2166 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2167 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2168 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2169
2170 /* Decay existing window, copy faults since last scan */
2171 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2172 fault_types[priv] += p->numa_faults[membuf_idx];
2173 p->numa_faults[membuf_idx] = 0;
2174
2175 /*
2176 * Normalize the faults_from, so all tasks in a group
2177 * count according to CPU use, instead of by the raw
2178 * number of faults. Tasks with little runtime have
2179 * little over-all impact on throughput, and thus their
2180 * faults are less important.
2181 */
2182 f_weight = div64_u64(runtime << 16, period + 1);
2183 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2184 (total_faults + 1);
2185 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2186 p->numa_faults[cpubuf_idx] = 0;
2187
2188 p->numa_faults[mem_idx] += diff;
2189 p->numa_faults[cpu_idx] += f_diff;
2190 faults += p->numa_faults[mem_idx];
2191 p->total_numa_faults += diff;
2192 if (ng) {
2193 /*
2194 * safe because we can only change our own group
2195 *
2196 * mem_idx represents the offset for a given
2197 * nid and priv in a specific region because it
2198 * is at the beginning of the numa_faults array.
2199 */
2200 ng->faults[mem_idx] += diff;
2201 ng->faults_cpu[mem_idx] += f_diff;
2202 ng->total_faults += diff;
2203 group_faults += ng->faults[mem_idx];
2204 }
2205 }
2206
2207 if (!ng) {
2208 if (faults > max_faults) {
2209 max_faults = faults;
2210 max_nid = nid;
2211 }
2212 } else if (group_faults > max_faults) {
2213 max_faults = group_faults;
2214 max_nid = nid;
2215 }
2216 }
2217
2218 if (ng) {
2219 numa_group_count_active_nodes(ng);
2220 spin_unlock_irq(group_lock);
2221 max_nid = preferred_group_nid(p, max_nid);
2222 }
2223
2224 if (max_faults) {
2225 /* Set the new preferred node */
2226 if (max_nid != p->numa_preferred_nid)
2227 sched_setnuma(p, max_nid);
2228 }
2229
2230 update_task_scan_period(p, fault_types[0], fault_types[1]);
2231}
2232
2233static inline int get_numa_group(struct numa_group *grp)
2234{
2235 return refcount_inc_not_zero(&grp->refcount);
2236}
2237
2238static inline void put_numa_group(struct numa_group *grp)
2239{
2240 if (refcount_dec_and_test(&grp->refcount))
2241 kfree_rcu(grp, rcu);
2242}
2243
2244static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2245 int *priv)
2246{
2247 struct numa_group *grp, *my_grp;
2248 struct task_struct *tsk;
2249 bool join = false;
2250 int cpu = cpupid_to_cpu(cpupid);
2251 int i;
2252
2253 if (unlikely(!deref_curr_numa_group(p))) {
2254 unsigned int size = sizeof(struct numa_group) +
2255 4*nr_node_ids*sizeof(unsigned long);
2256
2257 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2258 if (!grp)
2259 return;
2260
2261 refcount_set(&grp->refcount, 1);
2262 grp->active_nodes = 1;
2263 grp->max_faults_cpu = 0;
2264 spin_lock_init(&grp->lock);
2265 grp->gid = p->pid;
2266 /* Second half of the array tracks nids where faults happen */
2267 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2268 nr_node_ids;
2269
2270 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2271 grp->faults[i] = p->numa_faults[i];
2272
2273 grp->total_faults = p->total_numa_faults;
2274
2275 grp->nr_tasks++;
2276 rcu_assign_pointer(p->numa_group, grp);
2277 }
2278
2279 rcu_read_lock();
2280 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2281
2282 if (!cpupid_match_pid(tsk, cpupid))
2283 goto no_join;
2284
2285 grp = rcu_dereference(tsk->numa_group);
2286 if (!grp)
2287 goto no_join;
2288
2289 my_grp = deref_curr_numa_group(p);
2290 if (grp == my_grp)
2291 goto no_join;
2292
2293 /*
2294 * Only join the other group if its bigger; if we're the bigger group,
2295 * the other task will join us.
2296 */
2297 if (my_grp->nr_tasks > grp->nr_tasks)
2298 goto no_join;
2299
2300 /*
2301 * Tie-break on the grp address.
2302 */
2303 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2304 goto no_join;
2305
2306 /* Always join threads in the same process. */
2307 if (tsk->mm == current->mm)
2308 join = true;
2309
2310 /* Simple filter to avoid false positives due to PID collisions */
2311 if (flags & TNF_SHARED)
2312 join = true;
2313
2314 /* Update priv based on whether false sharing was detected */
2315 *priv = !join;
2316
2317 if (join && !get_numa_group(grp))
2318 goto no_join;
2319
2320 rcu_read_unlock();
2321
2322 if (!join)
2323 return;
2324
2325 BUG_ON(irqs_disabled());
2326 double_lock_irq(&my_grp->lock, &grp->lock);
2327
2328 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2329 my_grp->faults[i] -= p->numa_faults[i];
2330 grp->faults[i] += p->numa_faults[i];
2331 }
2332 my_grp->total_faults -= p->total_numa_faults;
2333 grp->total_faults += p->total_numa_faults;
2334
2335 my_grp->nr_tasks--;
2336 grp->nr_tasks++;
2337
2338 spin_unlock(&my_grp->lock);
2339 spin_unlock_irq(&grp->lock);
2340
2341 rcu_assign_pointer(p->numa_group, grp);
2342
2343 put_numa_group(my_grp);
2344 return;
2345
2346no_join:
2347 rcu_read_unlock();
2348 return;
2349}
2350
2351/*
2352 * Get rid of NUMA staticstics associated with a task (either current or dead).
2353 * If @final is set, the task is dead and has reached refcount zero, so we can
2354 * safely free all relevant data structures. Otherwise, there might be
2355 * concurrent reads from places like load balancing and procfs, and we should
2356 * reset the data back to default state without freeing ->numa_faults.
2357 */
2358void task_numa_free(struct task_struct *p, bool final)
2359{
2360 /* safe: p either is current or is being freed by current */
2361 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2362 unsigned long *numa_faults = p->numa_faults;
2363 unsigned long flags;
2364 int i;
2365
2366 if (!numa_faults)
2367 return;
2368
2369 if (grp) {
2370 spin_lock_irqsave(&grp->lock, flags);
2371 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2372 grp->faults[i] -= p->numa_faults[i];
2373 grp->total_faults -= p->total_numa_faults;
2374
2375 grp->nr_tasks--;
2376 spin_unlock_irqrestore(&grp->lock, flags);
2377 RCU_INIT_POINTER(p->numa_group, NULL);
2378 put_numa_group(grp);
2379 }
2380
2381 if (final) {
2382 p->numa_faults = NULL;
2383 kfree(numa_faults);
2384 } else {
2385 p->total_numa_faults = 0;
2386 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2387 numa_faults[i] = 0;
2388 }
2389}
2390
2391/*
2392 * Got a PROT_NONE fault for a page on @node.
2393 */
2394void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2395{
2396 struct task_struct *p = current;
2397 bool migrated = flags & TNF_MIGRATED;
2398 int cpu_node = task_node(current);
2399 int local = !!(flags & TNF_FAULT_LOCAL);
2400 struct numa_group *ng;
2401 int priv;
2402
2403 if (!static_branch_likely(&sched_numa_balancing))
2404 return;
2405
2406 /* for example, ksmd faulting in a user's mm */
2407 if (!p->mm)
2408 return;
2409
2410 /* Allocate buffer to track faults on a per-node basis */
2411 if (unlikely(!p->numa_faults)) {
2412 int size = sizeof(*p->numa_faults) *
2413 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2414
2415 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2416 if (!p->numa_faults)
2417 return;
2418
2419 p->total_numa_faults = 0;
2420 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2421 }
2422
2423 /*
2424 * First accesses are treated as private, otherwise consider accesses
2425 * to be private if the accessing pid has not changed
2426 */
2427 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2428 priv = 1;
2429 } else {
2430 priv = cpupid_match_pid(p, last_cpupid);
2431 if (!priv && !(flags & TNF_NO_GROUP))
2432 task_numa_group(p, last_cpupid, flags, &priv);
2433 }
2434
2435 /*
2436 * If a workload spans multiple NUMA nodes, a shared fault that
2437 * occurs wholly within the set of nodes that the workload is
2438 * actively using should be counted as local. This allows the
2439 * scan rate to slow down when a workload has settled down.
2440 */
2441 ng = deref_curr_numa_group(p);
2442 if (!priv && !local && ng && ng->active_nodes > 1 &&
2443 numa_is_active_node(cpu_node, ng) &&
2444 numa_is_active_node(mem_node, ng))
2445 local = 1;
2446
2447 /*
2448 * Retry to migrate task to preferred node periodically, in case it
2449 * previously failed, or the scheduler moved us.
2450 */
2451 if (time_after(jiffies, p->numa_migrate_retry)) {
2452 task_numa_placement(p);
2453 numa_migrate_preferred(p);
2454 }
2455
2456 if (migrated)
2457 p->numa_pages_migrated += pages;
2458 if (flags & TNF_MIGRATE_FAIL)
2459 p->numa_faults_locality[2] += pages;
2460
2461 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2462 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2463 p->numa_faults_locality[local] += pages;
2464}
2465
2466static void reset_ptenuma_scan(struct task_struct *p)
2467{
2468 /*
2469 * We only did a read acquisition of the mmap sem, so
2470 * p->mm->numa_scan_seq is written to without exclusive access
2471 * and the update is not guaranteed to be atomic. That's not
2472 * much of an issue though, since this is just used for
2473 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2474 * expensive, to avoid any form of compiler optimizations:
2475 */
2476 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2477 p->mm->numa_scan_offset = 0;
2478}
2479
2480/*
2481 * The expensive part of numa migration is done from task_work context.
2482 * Triggered from task_tick_numa().
2483 */
2484static void task_numa_work(struct callback_head *work)
2485{
2486 unsigned long migrate, next_scan, now = jiffies;
2487 struct task_struct *p = current;
2488 struct mm_struct *mm = p->mm;
2489 u64 runtime = p->se.sum_exec_runtime;
2490 struct vm_area_struct *vma;
2491 unsigned long start, end;
2492 unsigned long nr_pte_updates = 0;
2493 long pages, virtpages;
2494
2495 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2496
2497 work->next = work;
2498 /*
2499 * Who cares about NUMA placement when they're dying.
2500 *
2501 * NOTE: make sure not to dereference p->mm before this check,
2502 * exit_task_work() happens _after_ exit_mm() so we could be called
2503 * without p->mm even though we still had it when we enqueued this
2504 * work.
2505 */
2506 if (p->flags & PF_EXITING)
2507 return;
2508
2509 if (!mm->numa_next_scan) {
2510 mm->numa_next_scan = now +
2511 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2512 }
2513
2514 /*
2515 * Enforce maximal scan/migration frequency..
2516 */
2517 migrate = mm->numa_next_scan;
2518 if (time_before(now, migrate))
2519 return;
2520
2521 if (p->numa_scan_period == 0) {
2522 p->numa_scan_period_max = task_scan_max(p);
2523 p->numa_scan_period = task_scan_start(p);
2524 }
2525
2526 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2527 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2528 return;
2529
2530 /*
2531 * Delay this task enough that another task of this mm will likely win
2532 * the next time around.
2533 */
2534 p->node_stamp += 2 * TICK_NSEC;
2535
2536 start = mm->numa_scan_offset;
2537 pages = sysctl_numa_balancing_scan_size;
2538 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2539 virtpages = pages * 8; /* Scan up to this much virtual space */
2540 if (!pages)
2541 return;
2542
2543
2544 if (!down_read_trylock(&mm->mmap_sem))
2545 return;
2546 vma = find_vma(mm, start);
2547 if (!vma) {
2548 reset_ptenuma_scan(p);
2549 start = 0;
2550 vma = mm->mmap;
2551 }
2552 for (; vma; vma = vma->vm_next) {
2553 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2554 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2555 continue;
2556 }
2557
2558 /*
2559 * Shared library pages mapped by multiple processes are not
2560 * migrated as it is expected they are cache replicated. Avoid
2561 * hinting faults in read-only file-backed mappings or the vdso
2562 * as migrating the pages will be of marginal benefit.
2563 */
2564 if (!vma->vm_mm ||
2565 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2566 continue;
2567
2568 /*
2569 * Skip inaccessible VMAs to avoid any confusion between
2570 * PROT_NONE and NUMA hinting ptes
2571 */
2572 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2573 continue;
2574
2575 do {
2576 start = max(start, vma->vm_start);
2577 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2578 end = min(end, vma->vm_end);
2579 nr_pte_updates = change_prot_numa(vma, start, end);
2580
2581 /*
2582 * Try to scan sysctl_numa_balancing_size worth of
2583 * hpages that have at least one present PTE that
2584 * is not already pte-numa. If the VMA contains
2585 * areas that are unused or already full of prot_numa
2586 * PTEs, scan up to virtpages, to skip through those
2587 * areas faster.
2588 */
2589 if (nr_pte_updates)
2590 pages -= (end - start) >> PAGE_SHIFT;
2591 virtpages -= (end - start) >> PAGE_SHIFT;
2592
2593 start = end;
2594 if (pages <= 0 || virtpages <= 0)
2595 goto out;
2596
2597 cond_resched();
2598 } while (end != vma->vm_end);
2599 }
2600
2601out:
2602 /*
2603 * It is possible to reach the end of the VMA list but the last few
2604 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2605 * would find the !migratable VMA on the next scan but not reset the
2606 * scanner to the start so check it now.
2607 */
2608 if (vma)
2609 mm->numa_scan_offset = start;
2610 else
2611 reset_ptenuma_scan(p);
2612 up_read(&mm->mmap_sem);
2613
2614 /*
2615 * Make sure tasks use at least 32x as much time to run other code
2616 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2617 * Usually update_task_scan_period slows down scanning enough; on an
2618 * overloaded system we need to limit overhead on a per task basis.
2619 */
2620 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2621 u64 diff = p->se.sum_exec_runtime - runtime;
2622 p->node_stamp += 32 * diff;
2623 }
2624}
2625
2626void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2627{
2628 int mm_users = 0;
2629 struct mm_struct *mm = p->mm;
2630
2631 if (mm) {
2632 mm_users = atomic_read(&mm->mm_users);
2633 if (mm_users == 1) {
2634 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2635 mm->numa_scan_seq = 0;
2636 }
2637 }
2638 p->node_stamp = 0;
2639 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2640 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2641 /* Protect against double add, see task_tick_numa and task_numa_work */
2642 p->numa_work.next = &p->numa_work;
2643 p->numa_faults = NULL;
2644 RCU_INIT_POINTER(p->numa_group, NULL);
2645 p->last_task_numa_placement = 0;
2646 p->last_sum_exec_runtime = 0;
2647
2648 init_task_work(&p->numa_work, task_numa_work);
2649
2650 /* New address space, reset the preferred nid */
2651 if (!(clone_flags & CLONE_VM)) {
2652 p->numa_preferred_nid = NUMA_NO_NODE;
2653 return;
2654 }
2655
2656 /*
2657 * New thread, keep existing numa_preferred_nid which should be copied
2658 * already by arch_dup_task_struct but stagger when scans start.
2659 */
2660 if (mm) {
2661 unsigned int delay;
2662
2663 delay = min_t(unsigned int, task_scan_max(current),
2664 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2665 delay += 2 * TICK_NSEC;
2666 p->node_stamp = delay;
2667 }
2668}
2669
2670/*
2671 * Drive the periodic memory faults..
2672 */
2673static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2674{
2675 struct callback_head *work = &curr->numa_work;
2676 u64 period, now;
2677
2678 /*
2679 * We don't care about NUMA placement if we don't have memory.
2680 */
2681 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2682 return;
2683
2684 /*
2685 * Using runtime rather than walltime has the dual advantage that
2686 * we (mostly) drive the selection from busy threads and that the
2687 * task needs to have done some actual work before we bother with
2688 * NUMA placement.
2689 */
2690 now = curr->se.sum_exec_runtime;
2691 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2692
2693 if (now > curr->node_stamp + period) {
2694 if (!curr->node_stamp)
2695 curr->numa_scan_period = task_scan_start(curr);
2696 curr->node_stamp += period;
2697
2698 if (!time_before(jiffies, curr->mm->numa_next_scan))
2699 task_work_add(curr, work, true);
2700 }
2701}
2702
2703static void update_scan_period(struct task_struct *p, int new_cpu)
2704{
2705 int src_nid = cpu_to_node(task_cpu(p));
2706 int dst_nid = cpu_to_node(new_cpu);
2707
2708 if (!static_branch_likely(&sched_numa_balancing))
2709 return;
2710
2711 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2712 return;
2713
2714 if (src_nid == dst_nid)
2715 return;
2716
2717 /*
2718 * Allow resets if faults have been trapped before one scan
2719 * has completed. This is most likely due to a new task that
2720 * is pulled cross-node due to wakeups or load balancing.
2721 */
2722 if (p->numa_scan_seq) {
2723 /*
2724 * Avoid scan adjustments if moving to the preferred
2725 * node or if the task was not previously running on
2726 * the preferred node.
2727 */
2728 if (dst_nid == p->numa_preferred_nid ||
2729 (p->numa_preferred_nid != NUMA_NO_NODE &&
2730 src_nid != p->numa_preferred_nid))
2731 return;
2732 }
2733
2734 p->numa_scan_period = task_scan_start(p);
2735}
2736
2737#else
2738static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2739{
2740}
2741
2742static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2743{
2744}
2745
2746static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2747{
2748}
2749
2750static inline void update_scan_period(struct task_struct *p, int new_cpu)
2751{
2752}
2753
2754#endif /* CONFIG_NUMA_BALANCING */
2755
2756static void
2757account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2758{
2759 update_load_add(&cfs_rq->load, se->load.weight);
2760#ifdef CONFIG_SMP
2761 if (entity_is_task(se)) {
2762 struct rq *rq = rq_of(cfs_rq);
2763
2764 account_numa_enqueue(rq, task_of(se));
2765 list_add(&se->group_node, &rq->cfs_tasks);
2766 }
2767#endif
2768 cfs_rq->nr_running++;
2769}
2770
2771static void
2772account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2773{
2774 update_load_sub(&cfs_rq->load, se->load.weight);
2775#ifdef CONFIG_SMP
2776 if (entity_is_task(se)) {
2777 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2778 list_del_init(&se->group_node);
2779 }
2780#endif
2781 cfs_rq->nr_running--;
2782}
2783
2784/*
2785 * Signed add and clamp on underflow.
2786 *
2787 * Explicitly do a load-store to ensure the intermediate value never hits
2788 * memory. This allows lockless observations without ever seeing the negative
2789 * values.
2790 */
2791#define add_positive(_ptr, _val) do { \
2792 typeof(_ptr) ptr = (_ptr); \
2793 typeof(_val) val = (_val); \
2794 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2795 \
2796 res = var + val; \
2797 \
2798 if (val < 0 && res > var) \
2799 res = 0; \
2800 \
2801 WRITE_ONCE(*ptr, res); \
2802} while (0)
2803
2804/*
2805 * Unsigned subtract and clamp on underflow.
2806 *
2807 * Explicitly do a load-store to ensure the intermediate value never hits
2808 * memory. This allows lockless observations without ever seeing the negative
2809 * values.
2810 */
2811#define sub_positive(_ptr, _val) do { \
2812 typeof(_ptr) ptr = (_ptr); \
2813 typeof(*ptr) val = (_val); \
2814 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2815 res = var - val; \
2816 if (res > var) \
2817 res = 0; \
2818 WRITE_ONCE(*ptr, res); \
2819} while (0)
2820
2821/*
2822 * Remove and clamp on negative, from a local variable.
2823 *
2824 * A variant of sub_positive(), which does not use explicit load-store
2825 * and is thus optimized for local variable updates.
2826 */
2827#define lsub_positive(_ptr, _val) do { \
2828 typeof(_ptr) ptr = (_ptr); \
2829 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
2830} while (0)
2831
2832#ifdef CONFIG_SMP
2833static inline void
2834enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2835{
2836 cfs_rq->runnable_weight += se->runnable_weight;
2837
2838 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2839 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2840}
2841
2842static inline void
2843dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2844{
2845 cfs_rq->runnable_weight -= se->runnable_weight;
2846
2847 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2848 sub_positive(&cfs_rq->avg.runnable_load_sum,
2849 se_runnable(se) * se->avg.runnable_load_sum);
2850}
2851
2852static inline void
2853enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2854{
2855 cfs_rq->avg.load_avg += se->avg.load_avg;
2856 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2857}
2858
2859static inline void
2860dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2861{
2862 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2863 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2864}
2865#else
2866static inline void
2867enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2868static inline void
2869dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2870static inline void
2871enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2872static inline void
2873dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2874#endif
2875
2876static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2877 unsigned long weight, unsigned long runnable)
2878{
2879 if (se->on_rq) {
2880 /* commit outstanding execution time */
2881 if (cfs_rq->curr == se)
2882 update_curr(cfs_rq);
2883 account_entity_dequeue(cfs_rq, se);
2884 dequeue_runnable_load_avg(cfs_rq, se);
2885 }
2886 dequeue_load_avg(cfs_rq, se);
2887
2888 se->runnable_weight = runnable;
2889 update_load_set(&se->load, weight);
2890
2891#ifdef CONFIG_SMP
2892 do {
2893 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2894
2895 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2896 se->avg.runnable_load_avg =
2897 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2898 } while (0);
2899#endif
2900
2901 enqueue_load_avg(cfs_rq, se);
2902 if (se->on_rq) {
2903 account_entity_enqueue(cfs_rq, se);
2904 enqueue_runnable_load_avg(cfs_rq, se);
2905 }
2906}
2907
2908void reweight_task(struct task_struct *p, int prio)
2909{
2910 struct sched_entity *se = &p->se;
2911 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2912 struct load_weight *load = &se->load;
2913 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2914
2915 reweight_entity(cfs_rq, se, weight, weight);
2916 load->inv_weight = sched_prio_to_wmult[prio];
2917}
2918
2919#ifdef CONFIG_FAIR_GROUP_SCHED
2920#ifdef CONFIG_SMP
2921/*
2922 * All this does is approximate the hierarchical proportion which includes that
2923 * global sum we all love to hate.
2924 *
2925 * That is, the weight of a group entity, is the proportional share of the
2926 * group weight based on the group runqueue weights. That is:
2927 *
2928 * tg->weight * grq->load.weight
2929 * ge->load.weight = ----------------------------- (1)
2930 * \Sum grq->load.weight
2931 *
2932 * Now, because computing that sum is prohibitively expensive to compute (been
2933 * there, done that) we approximate it with this average stuff. The average
2934 * moves slower and therefore the approximation is cheaper and more stable.
2935 *
2936 * So instead of the above, we substitute:
2937 *
2938 * grq->load.weight -> grq->avg.load_avg (2)
2939 *
2940 * which yields the following:
2941 *
2942 * tg->weight * grq->avg.load_avg
2943 * ge->load.weight = ------------------------------ (3)
2944 * tg->load_avg
2945 *
2946 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2947 *
2948 * That is shares_avg, and it is right (given the approximation (2)).
2949 *
2950 * The problem with it is that because the average is slow -- it was designed
2951 * to be exactly that of course -- this leads to transients in boundary
2952 * conditions. In specific, the case where the group was idle and we start the
2953 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2954 * yielding bad latency etc..
2955 *
2956 * Now, in that special case (1) reduces to:
2957 *
2958 * tg->weight * grq->load.weight
2959 * ge->load.weight = ----------------------------- = tg->weight (4)
2960 * grp->load.weight
2961 *
2962 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2963 *
2964 * So what we do is modify our approximation (3) to approach (4) in the (near)
2965 * UP case, like:
2966 *
2967 * ge->load.weight =
2968 *
2969 * tg->weight * grq->load.weight
2970 * --------------------------------------------------- (5)
2971 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2972 *
2973 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2974 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2975 *
2976 *
2977 * tg->weight * grq->load.weight
2978 * ge->load.weight = ----------------------------- (6)
2979 * tg_load_avg'
2980 *
2981 * Where:
2982 *
2983 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2984 * max(grq->load.weight, grq->avg.load_avg)
2985 *
2986 * And that is shares_weight and is icky. In the (near) UP case it approaches
2987 * (4) while in the normal case it approaches (3). It consistently
2988 * overestimates the ge->load.weight and therefore:
2989 *
2990 * \Sum ge->load.weight >= tg->weight
2991 *
2992 * hence icky!
2993 */
2994static long calc_group_shares(struct cfs_rq *cfs_rq)
2995{
2996 long tg_weight, tg_shares, load, shares;
2997 struct task_group *tg = cfs_rq->tg;
2998
2999 tg_shares = READ_ONCE(tg->shares);
3000
3001 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3002
3003 tg_weight = atomic_long_read(&tg->load_avg);
3004
3005 /* Ensure tg_weight >= load */
3006 tg_weight -= cfs_rq->tg_load_avg_contrib;
3007 tg_weight += load;
3008
3009 shares = (tg_shares * load);
3010 if (tg_weight)
3011 shares /= tg_weight;
3012
3013 /*
3014 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3015 * of a group with small tg->shares value. It is a floor value which is
3016 * assigned as a minimum load.weight to the sched_entity representing
3017 * the group on a CPU.
3018 *
3019 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3020 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3021 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3022 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3023 * instead of 0.
3024 */
3025 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3026}
3027
3028/*
3029 * This calculates the effective runnable weight for a group entity based on
3030 * the group entity weight calculated above.
3031 *
3032 * Because of the above approximation (2), our group entity weight is
3033 * an load_avg based ratio (3). This means that it includes blocked load and
3034 * does not represent the runnable weight.
3035 *
3036 * Approximate the group entity's runnable weight per ratio from the group
3037 * runqueue:
3038 *
3039 * grq->avg.runnable_load_avg
3040 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
3041 * grq->avg.load_avg
3042 *
3043 * However, analogous to above, since the avg numbers are slow, this leads to
3044 * transients in the from-idle case. Instead we use:
3045 *
3046 * ge->runnable_weight = ge->load.weight *
3047 *
3048 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3049 * ----------------------------------------------------- (8)
3050 * max(grq->avg.load_avg, grq->load.weight)
3051 *
3052 * Where these max() serve both to use the 'instant' values to fix the slow
3053 * from-idle and avoid the /0 on to-idle, similar to (6).
3054 */
3055static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3056{
3057 long runnable, load_avg;
3058
3059 load_avg = max(cfs_rq->avg.load_avg,
3060 scale_load_down(cfs_rq->load.weight));
3061
3062 runnable = max(cfs_rq->avg.runnable_load_avg,
3063 scale_load_down(cfs_rq->runnable_weight));
3064
3065 runnable *= shares;
3066 if (load_avg)
3067 runnable /= load_avg;
3068
3069 return clamp_t(long, runnable, MIN_SHARES, shares);
3070}
3071#endif /* CONFIG_SMP */
3072
3073static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3074
3075/*
3076 * Recomputes the group entity based on the current state of its group
3077 * runqueue.
3078 */
3079static void update_cfs_group(struct sched_entity *se)
3080{
3081 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3082 long shares, runnable;
3083
3084 if (!gcfs_rq)
3085 return;
3086
3087 if (throttled_hierarchy(gcfs_rq))
3088 return;
3089
3090#ifndef CONFIG_SMP
3091 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3092
3093 if (likely(se->load.weight == shares))
3094 return;
3095#else
3096 shares = calc_group_shares(gcfs_rq);
3097 runnable = calc_group_runnable(gcfs_rq, shares);
3098#endif
3099
3100 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3101}
3102
3103#else /* CONFIG_FAIR_GROUP_SCHED */
3104static inline void update_cfs_group(struct sched_entity *se)
3105{
3106}
3107#endif /* CONFIG_FAIR_GROUP_SCHED */
3108
3109static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3110{
3111 struct rq *rq = rq_of(cfs_rq);
3112
3113 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3114 /*
3115 * There are a few boundary cases this might miss but it should
3116 * get called often enough that that should (hopefully) not be
3117 * a real problem.
3118 *
3119 * It will not get called when we go idle, because the idle
3120 * thread is a different class (!fair), nor will the utilization
3121 * number include things like RT tasks.
3122 *
3123 * As is, the util number is not freq-invariant (we'd have to
3124 * implement arch_scale_freq_capacity() for that).
3125 *
3126 * See cpu_util().
3127 */
3128 cpufreq_update_util(rq, flags);
3129 }
3130}
3131
3132#ifdef CONFIG_SMP
3133#ifdef CONFIG_FAIR_GROUP_SCHED
3134/**
3135 * update_tg_load_avg - update the tg's load avg
3136 * @cfs_rq: the cfs_rq whose avg changed
3137 * @force: update regardless of how small the difference
3138 *
3139 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3140 * However, because tg->load_avg is a global value there are performance
3141 * considerations.
3142 *
3143 * In order to avoid having to look at the other cfs_rq's, we use a
3144 * differential update where we store the last value we propagated. This in
3145 * turn allows skipping updates if the differential is 'small'.
3146 *
3147 * Updating tg's load_avg is necessary before update_cfs_share().
3148 */
3149static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3150{
3151 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3152
3153 /*
3154 * No need to update load_avg for root_task_group as it is not used.
3155 */
3156 if (cfs_rq->tg == &root_task_group)
3157 return;
3158
3159 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3160 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3161 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3162 }
3163}
3164
3165/*
3166 * Called within set_task_rq() right before setting a task's CPU. The
3167 * caller only guarantees p->pi_lock is held; no other assumptions,
3168 * including the state of rq->lock, should be made.
3169 */
3170void set_task_rq_fair(struct sched_entity *se,
3171 struct cfs_rq *prev, struct cfs_rq *next)
3172{
3173 u64 p_last_update_time;
3174 u64 n_last_update_time;
3175
3176 if (!sched_feat(ATTACH_AGE_LOAD))
3177 return;
3178
3179 /*
3180 * We are supposed to update the task to "current" time, then its up to
3181 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3182 * getting what current time is, so simply throw away the out-of-date
3183 * time. This will result in the wakee task is less decayed, but giving
3184 * the wakee more load sounds not bad.
3185 */
3186 if (!(se->avg.last_update_time && prev))
3187 return;
3188
3189#ifndef CONFIG_64BIT
3190 {
3191 u64 p_last_update_time_copy;
3192 u64 n_last_update_time_copy;
3193
3194 do {
3195 p_last_update_time_copy = prev->load_last_update_time_copy;
3196 n_last_update_time_copy = next->load_last_update_time_copy;
3197
3198 smp_rmb();
3199
3200 p_last_update_time = prev->avg.last_update_time;
3201 n_last_update_time = next->avg.last_update_time;
3202
3203 } while (p_last_update_time != p_last_update_time_copy ||
3204 n_last_update_time != n_last_update_time_copy);
3205 }
3206#else
3207 p_last_update_time = prev->avg.last_update_time;
3208 n_last_update_time = next->avg.last_update_time;
3209#endif
3210 __update_load_avg_blocked_se(p_last_update_time, se);
3211 se->avg.last_update_time = n_last_update_time;
3212}
3213
3214
3215/*
3216 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3217 * propagate its contribution. The key to this propagation is the invariant
3218 * that for each group:
3219 *
3220 * ge->avg == grq->avg (1)
3221 *
3222 * _IFF_ we look at the pure running and runnable sums. Because they
3223 * represent the very same entity, just at different points in the hierarchy.
3224 *
3225 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3226 * sum over (but still wrong, because the group entity and group rq do not have
3227 * their PELT windows aligned).
3228 *
3229 * However, update_tg_cfs_runnable() is more complex. So we have:
3230 *
3231 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3232 *
3233 * And since, like util, the runnable part should be directly transferable,
3234 * the following would _appear_ to be the straight forward approach:
3235 *
3236 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3237 *
3238 * And per (1) we have:
3239 *
3240 * ge->avg.runnable_avg == grq->avg.runnable_avg
3241 *
3242 * Which gives:
3243 *
3244 * ge->load.weight * grq->avg.load_avg
3245 * ge->avg.load_avg = ----------------------------------- (4)
3246 * grq->load.weight
3247 *
3248 * Except that is wrong!
3249 *
3250 * Because while for entities historical weight is not important and we
3251 * really only care about our future and therefore can consider a pure
3252 * runnable sum, runqueues can NOT do this.
3253 *
3254 * We specifically want runqueues to have a load_avg that includes
3255 * historical weights. Those represent the blocked load, the load we expect
3256 * to (shortly) return to us. This only works by keeping the weights as
3257 * integral part of the sum. We therefore cannot decompose as per (3).
3258 *
3259 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3260 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3261 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3262 * runnable section of these tasks overlap (or not). If they were to perfectly
3263 * align the rq as a whole would be runnable 2/3 of the time. If however we
3264 * always have at least 1 runnable task, the rq as a whole is always runnable.
3265 *
3266 * So we'll have to approximate.. :/
3267 *
3268 * Given the constraint:
3269 *
3270 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3271 *
3272 * We can construct a rule that adds runnable to a rq by assuming minimal
3273 * overlap.
3274 *
3275 * On removal, we'll assume each task is equally runnable; which yields:
3276 *
3277 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3278 *
3279 * XXX: only do this for the part of runnable > running ?
3280 *
3281 */
3282
3283static inline void
3284update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3285{
3286 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3287
3288 /* Nothing to update */
3289 if (!delta)
3290 return;
3291
3292 /*
3293 * The relation between sum and avg is:
3294 *
3295 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3296 *
3297 * however, the PELT windows are not aligned between grq and gse.
3298 */
3299
3300 /* Set new sched_entity's utilization */
3301 se->avg.util_avg = gcfs_rq->avg.util_avg;
3302 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3303
3304 /* Update parent cfs_rq utilization */
3305 add_positive(&cfs_rq->avg.util_avg, delta);
3306 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3307}
3308
3309static inline void
3310update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3311{
3312 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3313 unsigned long runnable_load_avg, load_avg;
3314 u64 runnable_load_sum, load_sum = 0;
3315 s64 delta_sum;
3316
3317 if (!runnable_sum)
3318 return;
3319
3320 gcfs_rq->prop_runnable_sum = 0;
3321
3322 if (runnable_sum >= 0) {
3323 /*
3324 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3325 * the CPU is saturated running == runnable.
3326 */
3327 runnable_sum += se->avg.load_sum;
3328 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3329 } else {
3330 /*
3331 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3332 * assuming all tasks are equally runnable.
3333 */
3334 if (scale_load_down(gcfs_rq->load.weight)) {
3335 load_sum = div_s64(gcfs_rq->avg.load_sum,
3336 scale_load_down(gcfs_rq->load.weight));
3337 }
3338
3339 /* But make sure to not inflate se's runnable */
3340 runnable_sum = min(se->avg.load_sum, load_sum);
3341 }
3342
3343 /*
3344 * runnable_sum can't be lower than running_sum
3345 * Rescale running sum to be in the same range as runnable sum
3346 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3347 * runnable_sum is in [0 : LOAD_AVG_MAX]
3348 */
3349 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3350 runnable_sum = max(runnable_sum, running_sum);
3351
3352 load_sum = (s64)se_weight(se) * runnable_sum;
3353 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3354
3355 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3356 delta_avg = load_avg - se->avg.load_avg;
3357
3358 se->avg.load_sum = runnable_sum;
3359 se->avg.load_avg = load_avg;
3360 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3361 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3362
3363 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3364 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3365 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3366 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3367
3368 se->avg.runnable_load_sum = runnable_sum;
3369 se->avg.runnable_load_avg = runnable_load_avg;
3370
3371 if (se->on_rq) {
3372 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3373 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3374 }
3375}
3376
3377static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3378{
3379 cfs_rq->propagate = 1;
3380 cfs_rq->prop_runnable_sum += runnable_sum;
3381}
3382
3383/* Update task and its cfs_rq load average */
3384static inline int propagate_entity_load_avg(struct sched_entity *se)
3385{
3386 struct cfs_rq *cfs_rq, *gcfs_rq;
3387
3388 if (entity_is_task(se))
3389 return 0;
3390
3391 gcfs_rq = group_cfs_rq(se);
3392 if (!gcfs_rq->propagate)
3393 return 0;
3394
3395 gcfs_rq->propagate = 0;
3396
3397 cfs_rq = cfs_rq_of(se);
3398
3399 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3400
3401 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3402 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3403
3404 trace_pelt_cfs_tp(cfs_rq);
3405 trace_pelt_se_tp(se);
3406
3407 return 1;
3408}
3409
3410/*
3411 * Check if we need to update the load and the utilization of a blocked
3412 * group_entity:
3413 */
3414static inline bool skip_blocked_update(struct sched_entity *se)
3415{
3416 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3417
3418 /*
3419 * If sched_entity still have not zero load or utilization, we have to
3420 * decay it:
3421 */
3422 if (se->avg.load_avg || se->avg.util_avg)
3423 return false;
3424
3425 /*
3426 * If there is a pending propagation, we have to update the load and
3427 * the utilization of the sched_entity:
3428 */
3429 if (gcfs_rq->propagate)
3430 return false;
3431
3432 /*
3433 * Otherwise, the load and the utilization of the sched_entity is
3434 * already zero and there is no pending propagation, so it will be a
3435 * waste of time to try to decay it:
3436 */
3437 return true;
3438}
3439
3440#else /* CONFIG_FAIR_GROUP_SCHED */
3441
3442static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3443
3444static inline int propagate_entity_load_avg(struct sched_entity *se)
3445{
3446 return 0;
3447}
3448
3449static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3450
3451#endif /* CONFIG_FAIR_GROUP_SCHED */
3452
3453/**
3454 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3455 * @now: current time, as per cfs_rq_clock_pelt()
3456 * @cfs_rq: cfs_rq to update
3457 *
3458 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3459 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3460 * post_init_entity_util_avg().
3461 *
3462 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3463 *
3464 * Returns true if the load decayed or we removed load.
3465 *
3466 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3467 * call update_tg_load_avg() when this function returns true.
3468 */
3469static inline int
3470update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3471{
3472 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3473 struct sched_avg *sa = &cfs_rq->avg;
3474 int decayed = 0;
3475
3476 if (cfs_rq->removed.nr) {
3477 unsigned long r;
3478 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3479
3480 raw_spin_lock(&cfs_rq->removed.lock);
3481 swap(cfs_rq->removed.util_avg, removed_util);
3482 swap(cfs_rq->removed.load_avg, removed_load);
3483 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3484 cfs_rq->removed.nr = 0;
3485 raw_spin_unlock(&cfs_rq->removed.lock);
3486
3487 r = removed_load;
3488 sub_positive(&sa->load_avg, r);
3489 sub_positive(&sa->load_sum, r * divider);
3490
3491 r = removed_util;
3492 sub_positive(&sa->util_avg, r);
3493 sub_positive(&sa->util_sum, r * divider);
3494
3495 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3496
3497 decayed = 1;
3498 }
3499
3500 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3501
3502#ifndef CONFIG_64BIT
3503 smp_wmb();
3504 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3505#endif
3506
3507 if (decayed)
3508 cfs_rq_util_change(cfs_rq, 0);
3509
3510 return decayed;
3511}
3512
3513/**
3514 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3515 * @cfs_rq: cfs_rq to attach to
3516 * @se: sched_entity to attach
3517 * @flags: migration hints
3518 *
3519 * Must call update_cfs_rq_load_avg() before this, since we rely on
3520 * cfs_rq->avg.last_update_time being current.
3521 */
3522static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3523{
3524 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3525
3526 /*
3527 * When we attach the @se to the @cfs_rq, we must align the decay
3528 * window because without that, really weird and wonderful things can
3529 * happen.
3530 *
3531 * XXX illustrate
3532 */
3533 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3534 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3535
3536 /*
3537 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3538 * period_contrib. This isn't strictly correct, but since we're
3539 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3540 * _sum a little.
3541 */
3542 se->avg.util_sum = se->avg.util_avg * divider;
3543
3544 se->avg.load_sum = divider;
3545 if (se_weight(se)) {
3546 se->avg.load_sum =
3547 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3548 }
3549
3550 se->avg.runnable_load_sum = se->avg.load_sum;
3551
3552 enqueue_load_avg(cfs_rq, se);
3553 cfs_rq->avg.util_avg += se->avg.util_avg;
3554 cfs_rq->avg.util_sum += se->avg.util_sum;
3555
3556 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3557
3558 cfs_rq_util_change(cfs_rq, flags);
3559
3560 trace_pelt_cfs_tp(cfs_rq);
3561}
3562
3563/**
3564 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3565 * @cfs_rq: cfs_rq to detach from
3566 * @se: sched_entity to detach
3567 *
3568 * Must call update_cfs_rq_load_avg() before this, since we rely on
3569 * cfs_rq->avg.last_update_time being current.
3570 */
3571static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3572{
3573 dequeue_load_avg(cfs_rq, se);
3574 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3575 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3576
3577 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3578
3579 cfs_rq_util_change(cfs_rq, 0);
3580
3581 trace_pelt_cfs_tp(cfs_rq);
3582}
3583
3584/*
3585 * Optional action to be done while updating the load average
3586 */
3587#define UPDATE_TG 0x1
3588#define SKIP_AGE_LOAD 0x2
3589#define DO_ATTACH 0x4
3590
3591/* Update task and its cfs_rq load average */
3592static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3593{
3594 u64 now = cfs_rq_clock_pelt(cfs_rq);
3595 int decayed;
3596
3597 /*
3598 * Track task load average for carrying it to new CPU after migrated, and
3599 * track group sched_entity load average for task_h_load calc in migration
3600 */
3601 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3602 __update_load_avg_se(now, cfs_rq, se);
3603
3604 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3605 decayed |= propagate_entity_load_avg(se);
3606
3607 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3608
3609 /*
3610 * DO_ATTACH means we're here from enqueue_entity().
3611 * !last_update_time means we've passed through
3612 * migrate_task_rq_fair() indicating we migrated.
3613 *
3614 * IOW we're enqueueing a task on a new CPU.
3615 */
3616 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3617 update_tg_load_avg(cfs_rq, 0);
3618
3619 } else if (decayed && (flags & UPDATE_TG))
3620 update_tg_load_avg(cfs_rq, 0);
3621}
3622
3623#ifndef CONFIG_64BIT
3624static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3625{
3626 u64 last_update_time_copy;
3627 u64 last_update_time;
3628
3629 do {
3630 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3631 smp_rmb();
3632 last_update_time = cfs_rq->avg.last_update_time;
3633 } while (last_update_time != last_update_time_copy);
3634
3635 return last_update_time;
3636}
3637#else
3638static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3639{
3640 return cfs_rq->avg.last_update_time;
3641}
3642#endif
3643
3644/*
3645 * Synchronize entity load avg of dequeued entity without locking
3646 * the previous rq.
3647 */
3648static void sync_entity_load_avg(struct sched_entity *se)
3649{
3650 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3651 u64 last_update_time;
3652
3653 last_update_time = cfs_rq_last_update_time(cfs_rq);
3654 __update_load_avg_blocked_se(last_update_time, se);
3655}
3656
3657/*
3658 * Task first catches up with cfs_rq, and then subtract
3659 * itself from the cfs_rq (task must be off the queue now).
3660 */
3661static void remove_entity_load_avg(struct sched_entity *se)
3662{
3663 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3664 unsigned long flags;
3665
3666 /*
3667 * tasks cannot exit without having gone through wake_up_new_task() ->
3668 * post_init_entity_util_avg() which will have added things to the
3669 * cfs_rq, so we can remove unconditionally.
3670 */
3671
3672 sync_entity_load_avg(se);
3673
3674 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3675 ++cfs_rq->removed.nr;
3676 cfs_rq->removed.util_avg += se->avg.util_avg;
3677 cfs_rq->removed.load_avg += se->avg.load_avg;
3678 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3679 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3680}
3681
3682static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3683{
3684 return cfs_rq->avg.runnable_load_avg;
3685}
3686
3687static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3688{
3689 return cfs_rq->avg.load_avg;
3690}
3691
3692static inline unsigned long task_util(struct task_struct *p)
3693{
3694 return READ_ONCE(p->se.avg.util_avg);
3695}
3696
3697static inline unsigned long _task_util_est(struct task_struct *p)
3698{
3699 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3700
3701 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3702}
3703
3704static inline unsigned long task_util_est(struct task_struct *p)
3705{
3706 return max(task_util(p), _task_util_est(p));
3707}
3708
3709static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3710 struct task_struct *p)
3711{
3712 unsigned int enqueued;
3713
3714 if (!sched_feat(UTIL_EST))
3715 return;
3716
3717 /* Update root cfs_rq's estimated utilization */
3718 enqueued = cfs_rq->avg.util_est.enqueued;
3719 enqueued += _task_util_est(p);
3720 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3721}
3722
3723/*
3724 * Check if a (signed) value is within a specified (unsigned) margin,
3725 * based on the observation that:
3726 *
3727 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3728 *
3729 * NOTE: this only works when value + maring < INT_MAX.
3730 */
3731static inline bool within_margin(int value, int margin)
3732{
3733 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3734}
3735
3736static void
3737util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3738{
3739 long last_ewma_diff;
3740 struct util_est ue;
3741 int cpu;
3742
3743 if (!sched_feat(UTIL_EST))
3744 return;
3745
3746 /* Update root cfs_rq's estimated utilization */
3747 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3748 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3749 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3750
3751 /*
3752 * Skip update of task's estimated utilization when the task has not
3753 * yet completed an activation, e.g. being migrated.
3754 */
3755 if (!task_sleep)
3756 return;
3757
3758 /*
3759 * If the PELT values haven't changed since enqueue time,
3760 * skip the util_est update.
3761 */
3762 ue = p->se.avg.util_est;
3763 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3764 return;
3765
3766 /*
3767 * Skip update of task's estimated utilization when its EWMA is
3768 * already ~1% close to its last activation value.
3769 */
3770 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3771 last_ewma_diff = ue.enqueued - ue.ewma;
3772 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3773 return;
3774
3775 /*
3776 * To avoid overestimation of actual task utilization, skip updates if
3777 * we cannot grant there is idle time in this CPU.
3778 */
3779 cpu = cpu_of(rq_of(cfs_rq));
3780 if (task_util(p) > capacity_orig_of(cpu))
3781 return;
3782
3783 /*
3784 * Update Task's estimated utilization
3785 *
3786 * When *p completes an activation we can consolidate another sample
3787 * of the task size. This is done by storing the current PELT value
3788 * as ue.enqueued and by using this value to update the Exponential
3789 * Weighted Moving Average (EWMA):
3790 *
3791 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3792 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3793 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3794 * = w * ( last_ewma_diff ) + ewma(t-1)
3795 * = w * (last_ewma_diff + ewma(t-1) / w)
3796 *
3797 * Where 'w' is the weight of new samples, which is configured to be
3798 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3799 */
3800 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3801 ue.ewma += last_ewma_diff;
3802 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3803 WRITE_ONCE(p->se.avg.util_est, ue);
3804}
3805
3806static inline int task_fits_capacity(struct task_struct *p, long capacity)
3807{
3808 return fits_capacity(task_util_est(p), capacity);
3809}
3810
3811static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3812{
3813 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3814 return;
3815
3816 if (!p) {
3817 rq->misfit_task_load = 0;
3818 return;
3819 }
3820
3821 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3822 rq->misfit_task_load = 0;
3823 return;
3824 }
3825
3826 rq->misfit_task_load = task_h_load(p);
3827}
3828
3829#else /* CONFIG_SMP */
3830
3831#define UPDATE_TG 0x0
3832#define SKIP_AGE_LOAD 0x0
3833#define DO_ATTACH 0x0
3834
3835static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3836{
3837 cfs_rq_util_change(cfs_rq, 0);
3838}
3839
3840static inline void remove_entity_load_avg(struct sched_entity *se) {}
3841
3842static inline void
3843attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3844static inline void
3845detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3846
3847static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3848{
3849 return 0;
3850}
3851
3852static inline void
3853util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3854
3855static inline void
3856util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3857 bool task_sleep) {}
3858static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3859
3860#endif /* CONFIG_SMP */
3861
3862static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3863{
3864#ifdef CONFIG_SCHED_DEBUG
3865 s64 d = se->vruntime - cfs_rq->min_vruntime;
3866
3867 if (d < 0)
3868 d = -d;
3869
3870 if (d > 3*sysctl_sched_latency)
3871 schedstat_inc(cfs_rq->nr_spread_over);
3872#endif
3873}
3874
3875static void
3876place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3877{
3878 u64 vruntime = cfs_rq->min_vruntime;
3879
3880 /*
3881 * The 'current' period is already promised to the current tasks,
3882 * however the extra weight of the new task will slow them down a
3883 * little, place the new task so that it fits in the slot that
3884 * stays open at the end.
3885 */
3886 if (initial && sched_feat(START_DEBIT))
3887 vruntime += sched_vslice(cfs_rq, se);
3888
3889 /* sleeps up to a single latency don't count. */
3890 if (!initial) {
3891 unsigned long thresh = sysctl_sched_latency;
3892
3893 /*
3894 * Halve their sleep time's effect, to allow
3895 * for a gentler effect of sleepers:
3896 */
3897 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3898 thresh >>= 1;
3899
3900 vruntime -= thresh;
3901 }
3902
3903 /* ensure we never gain time by being placed backwards. */
3904 se->vruntime = max_vruntime(se->vruntime, vruntime);
3905}
3906
3907static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3908
3909static inline void check_schedstat_required(void)
3910{
3911#ifdef CONFIG_SCHEDSTATS
3912 if (schedstat_enabled())
3913 return;
3914
3915 /* Force schedstat enabled if a dependent tracepoint is active */
3916 if (trace_sched_stat_wait_enabled() ||
3917 trace_sched_stat_sleep_enabled() ||
3918 trace_sched_stat_iowait_enabled() ||
3919 trace_sched_stat_blocked_enabled() ||
3920 trace_sched_stat_runtime_enabled()) {
3921 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3922 "stat_blocked and stat_runtime require the "
3923 "kernel parameter schedstats=enable or "
3924 "kernel.sched_schedstats=1\n");
3925 }
3926#endif
3927}
3928
3929
3930/*
3931 * MIGRATION
3932 *
3933 * dequeue
3934 * update_curr()
3935 * update_min_vruntime()
3936 * vruntime -= min_vruntime
3937 *
3938 * enqueue
3939 * update_curr()
3940 * update_min_vruntime()
3941 * vruntime += min_vruntime
3942 *
3943 * this way the vruntime transition between RQs is done when both
3944 * min_vruntime are up-to-date.
3945 *
3946 * WAKEUP (remote)
3947 *
3948 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3949 * vruntime -= min_vruntime
3950 *
3951 * enqueue
3952 * update_curr()
3953 * update_min_vruntime()
3954 * vruntime += min_vruntime
3955 *
3956 * this way we don't have the most up-to-date min_vruntime on the originating
3957 * CPU and an up-to-date min_vruntime on the destination CPU.
3958 */
3959
3960static void
3961enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3962{
3963 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3964 bool curr = cfs_rq->curr == se;
3965
3966 /*
3967 * If we're the current task, we must renormalise before calling
3968 * update_curr().
3969 */
3970 if (renorm && curr)
3971 se->vruntime += cfs_rq->min_vruntime;
3972
3973 update_curr(cfs_rq);
3974
3975 /*
3976 * Otherwise, renormalise after, such that we're placed at the current
3977 * moment in time, instead of some random moment in the past. Being
3978 * placed in the past could significantly boost this task to the
3979 * fairness detriment of existing tasks.
3980 */
3981 if (renorm && !curr)
3982 se->vruntime += cfs_rq->min_vruntime;
3983
3984 /*
3985 * When enqueuing a sched_entity, we must:
3986 * - Update loads to have both entity and cfs_rq synced with now.
3987 * - Add its load to cfs_rq->runnable_avg
3988 * - For group_entity, update its weight to reflect the new share of
3989 * its group cfs_rq
3990 * - Add its new weight to cfs_rq->load.weight
3991 */
3992 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3993 update_cfs_group(se);
3994 enqueue_runnable_load_avg(cfs_rq, se);
3995 account_entity_enqueue(cfs_rq, se);
3996
3997 if (flags & ENQUEUE_WAKEUP)
3998 place_entity(cfs_rq, se, 0);
3999
4000 check_schedstat_required();
4001 update_stats_enqueue(cfs_rq, se, flags);
4002 check_spread(cfs_rq, se);
4003 if (!curr)
4004 __enqueue_entity(cfs_rq, se);
4005 se->on_rq = 1;
4006
4007 if (cfs_rq->nr_running == 1) {
4008 list_add_leaf_cfs_rq(cfs_rq);
4009 check_enqueue_throttle(cfs_rq);
4010 }
4011}
4012
4013static void __clear_buddies_last(struct sched_entity *se)
4014{
4015 for_each_sched_entity(se) {
4016 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4017 if (cfs_rq->last != se)
4018 break;
4019
4020 cfs_rq->last = NULL;
4021 }
4022}
4023
4024static void __clear_buddies_next(struct sched_entity *se)
4025{
4026 for_each_sched_entity(se) {
4027 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4028 if (cfs_rq->next != se)
4029 break;
4030
4031 cfs_rq->next = NULL;
4032 }
4033}
4034
4035static void __clear_buddies_skip(struct sched_entity *se)
4036{
4037 for_each_sched_entity(se) {
4038 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4039 if (cfs_rq->skip != se)
4040 break;
4041
4042 cfs_rq->skip = NULL;
4043 }
4044}
4045
4046static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4047{
4048 if (cfs_rq->last == se)
4049 __clear_buddies_last(se);
4050
4051 if (cfs_rq->next == se)
4052 __clear_buddies_next(se);
4053
4054 if (cfs_rq->skip == se)
4055 __clear_buddies_skip(se);
4056}
4057
4058static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4059
4060static void
4061dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4062{
4063 /*
4064 * Update run-time statistics of the 'current'.
4065 */
4066 update_curr(cfs_rq);
4067
4068 /*
4069 * When dequeuing a sched_entity, we must:
4070 * - Update loads to have both entity and cfs_rq synced with now.
4071 * - Subtract its load from the cfs_rq->runnable_avg.
4072 * - Subtract its previous weight from cfs_rq->load.weight.
4073 * - For group entity, update its weight to reflect the new share
4074 * of its group cfs_rq.
4075 */
4076 update_load_avg(cfs_rq, se, UPDATE_TG);
4077 dequeue_runnable_load_avg(cfs_rq, se);
4078
4079 update_stats_dequeue(cfs_rq, se, flags);
4080
4081 clear_buddies(cfs_rq, se);
4082
4083 if (se != cfs_rq->curr)
4084 __dequeue_entity(cfs_rq, se);
4085 se->on_rq = 0;
4086 account_entity_dequeue(cfs_rq, se);
4087
4088 /*
4089 * Normalize after update_curr(); which will also have moved
4090 * min_vruntime if @se is the one holding it back. But before doing
4091 * update_min_vruntime() again, which will discount @se's position and
4092 * can move min_vruntime forward still more.
4093 */
4094 if (!(flags & DEQUEUE_SLEEP))
4095 se->vruntime -= cfs_rq->min_vruntime;
4096
4097 /* return excess runtime on last dequeue */
4098 return_cfs_rq_runtime(cfs_rq);
4099
4100 update_cfs_group(se);
4101
4102 /*
4103 * Now advance min_vruntime if @se was the entity holding it back,
4104 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4105 * put back on, and if we advance min_vruntime, we'll be placed back
4106 * further than we started -- ie. we'll be penalized.
4107 */
4108 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4109 update_min_vruntime(cfs_rq);
4110}
4111
4112/*
4113 * Preempt the current task with a newly woken task if needed:
4114 */
4115static void
4116check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4117{
4118 unsigned long ideal_runtime, delta_exec;
4119 struct sched_entity *se;
4120 s64 delta;
4121
4122 ideal_runtime = sched_slice(cfs_rq, curr);
4123 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4124 if (delta_exec > ideal_runtime) {
4125 resched_curr(rq_of(cfs_rq));
4126 /*
4127 * The current task ran long enough, ensure it doesn't get
4128 * re-elected due to buddy favours.
4129 */
4130 clear_buddies(cfs_rq, curr);
4131 return;
4132 }
4133
4134 /*
4135 * Ensure that a task that missed wakeup preemption by a
4136 * narrow margin doesn't have to wait for a full slice.
4137 * This also mitigates buddy induced latencies under load.
4138 */
4139 if (delta_exec < sysctl_sched_min_granularity)
4140 return;
4141
4142 se = __pick_first_entity(cfs_rq);
4143 delta = curr->vruntime - se->vruntime;
4144
4145 if (delta < 0)
4146 return;
4147
4148 if (delta > ideal_runtime)
4149 resched_curr(rq_of(cfs_rq));
4150}
4151
4152static void
4153set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4154{
4155 /* 'current' is not kept within the tree. */
4156 if (se->on_rq) {
4157 /*
4158 * Any task has to be enqueued before it get to execute on
4159 * a CPU. So account for the time it spent waiting on the
4160 * runqueue.
4161 */
4162 update_stats_wait_end(cfs_rq, se);
4163 __dequeue_entity(cfs_rq, se);
4164 update_load_avg(cfs_rq, se, UPDATE_TG);
4165 }
4166
4167 update_stats_curr_start(cfs_rq, se);
4168 cfs_rq->curr = se;
4169
4170 /*
4171 * Track our maximum slice length, if the CPU's load is at
4172 * least twice that of our own weight (i.e. dont track it
4173 * when there are only lesser-weight tasks around):
4174 */
4175 if (schedstat_enabled() &&
4176 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4177 schedstat_set(se->statistics.slice_max,
4178 max((u64)schedstat_val(se->statistics.slice_max),
4179 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4180 }
4181
4182 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4183}
4184
4185static int
4186wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4187
4188/*
4189 * Pick the next process, keeping these things in mind, in this order:
4190 * 1) keep things fair between processes/task groups
4191 * 2) pick the "next" process, since someone really wants that to run
4192 * 3) pick the "last" process, for cache locality
4193 * 4) do not run the "skip" process, if something else is available
4194 */
4195static struct sched_entity *
4196pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4197{
4198 struct sched_entity *left = __pick_first_entity(cfs_rq);
4199 struct sched_entity *se;
4200
4201 /*
4202 * If curr is set we have to see if its left of the leftmost entity
4203 * still in the tree, provided there was anything in the tree at all.
4204 */
4205 if (!left || (curr && entity_before(curr, left)))
4206 left = curr;
4207
4208 se = left; /* ideally we run the leftmost entity */
4209
4210 /*
4211 * Avoid running the skip buddy, if running something else can
4212 * be done without getting too unfair.
4213 */
4214 if (cfs_rq->skip == se) {
4215 struct sched_entity *second;
4216
4217 if (se == curr) {
4218 second = __pick_first_entity(cfs_rq);
4219 } else {
4220 second = __pick_next_entity(se);
4221 if (!second || (curr && entity_before(curr, second)))
4222 second = curr;
4223 }
4224
4225 if (second && wakeup_preempt_entity(second, left) < 1)
4226 se = second;
4227 }
4228
4229 /*
4230 * Prefer last buddy, try to return the CPU to a preempted task.
4231 */
4232 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4233 se = cfs_rq->last;
4234
4235 /*
4236 * Someone really wants this to run. If it's not unfair, run it.
4237 */
4238 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4239 se = cfs_rq->next;
4240
4241 clear_buddies(cfs_rq, se);
4242
4243 return se;
4244}
4245
4246static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4247
4248static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4249{
4250 /*
4251 * If still on the runqueue then deactivate_task()
4252 * was not called and update_curr() has to be done:
4253 */
4254 if (prev->on_rq)
4255 update_curr(cfs_rq);
4256
4257 /* throttle cfs_rqs exceeding runtime */
4258 check_cfs_rq_runtime(cfs_rq);
4259
4260 check_spread(cfs_rq, prev);
4261
4262 if (prev->on_rq) {
4263 update_stats_wait_start(cfs_rq, prev);
4264 /* Put 'current' back into the tree. */
4265 __enqueue_entity(cfs_rq, prev);
4266 /* in !on_rq case, update occurred at dequeue */
4267 update_load_avg(cfs_rq, prev, 0);
4268 }
4269 cfs_rq->curr = NULL;
4270}
4271
4272static void
4273entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4274{
4275 /*
4276 * Update run-time statistics of the 'current'.
4277 */
4278 update_curr(cfs_rq);
4279
4280 /*
4281 * Ensure that runnable average is periodically updated.
4282 */
4283 update_load_avg(cfs_rq, curr, UPDATE_TG);
4284 update_cfs_group(curr);
4285
4286#ifdef CONFIG_SCHED_HRTICK
4287 /*
4288 * queued ticks are scheduled to match the slice, so don't bother
4289 * validating it and just reschedule.
4290 */
4291 if (queued) {
4292 resched_curr(rq_of(cfs_rq));
4293 return;
4294 }
4295 /*
4296 * don't let the period tick interfere with the hrtick preemption
4297 */
4298 if (!sched_feat(DOUBLE_TICK) &&
4299 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4300 return;
4301#endif
4302
4303 if (cfs_rq->nr_running > 1)
4304 check_preempt_tick(cfs_rq, curr);
4305}
4306
4307
4308/**************************************************
4309 * CFS bandwidth control machinery
4310 */
4311
4312#ifdef CONFIG_CFS_BANDWIDTH
4313
4314#ifdef CONFIG_JUMP_LABEL
4315static struct static_key __cfs_bandwidth_used;
4316
4317static inline bool cfs_bandwidth_used(void)
4318{
4319 return static_key_false(&__cfs_bandwidth_used);
4320}
4321
4322void cfs_bandwidth_usage_inc(void)
4323{
4324 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4325}
4326
4327void cfs_bandwidth_usage_dec(void)
4328{
4329 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4330}
4331#else /* CONFIG_JUMP_LABEL */
4332static bool cfs_bandwidth_used(void)
4333{
4334 return true;
4335}
4336
4337void cfs_bandwidth_usage_inc(void) {}
4338void cfs_bandwidth_usage_dec(void) {}
4339#endif /* CONFIG_JUMP_LABEL */
4340
4341/*
4342 * default period for cfs group bandwidth.
4343 * default: 0.1s, units: nanoseconds
4344 */
4345static inline u64 default_cfs_period(void)
4346{
4347 return 100000000ULL;
4348}
4349
4350static inline u64 sched_cfs_bandwidth_slice(void)
4351{
4352 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4353}
4354
4355/*
4356 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4357 * directly instead of rq->clock to avoid adding additional synchronization
4358 * around rq->lock.
4359 *
4360 * requires cfs_b->lock
4361 */
4362void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4363{
4364 if (cfs_b->quota != RUNTIME_INF)
4365 cfs_b->runtime = cfs_b->quota;
4366}
4367
4368static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4369{
4370 return &tg->cfs_bandwidth;
4371}
4372
4373/* returns 0 on failure to allocate runtime */
4374static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4375{
4376 struct task_group *tg = cfs_rq->tg;
4377 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4378 u64 amount = 0, min_amount;
4379
4380 /* note: this is a positive sum as runtime_remaining <= 0 */
4381 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4382
4383 raw_spin_lock(&cfs_b->lock);
4384 if (cfs_b->quota == RUNTIME_INF)
4385 amount = min_amount;
4386 else {
4387 start_cfs_bandwidth(cfs_b);
4388
4389 if (cfs_b->runtime > 0) {
4390 amount = min(cfs_b->runtime, min_amount);
4391 cfs_b->runtime -= amount;
4392 cfs_b->idle = 0;
4393 }
4394 }
4395 raw_spin_unlock(&cfs_b->lock);
4396
4397 cfs_rq->runtime_remaining += amount;
4398
4399 return cfs_rq->runtime_remaining > 0;
4400}
4401
4402static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4403{
4404 /* dock delta_exec before expiring quota (as it could span periods) */
4405 cfs_rq->runtime_remaining -= delta_exec;
4406
4407 if (likely(cfs_rq->runtime_remaining > 0))
4408 return;
4409
4410 if (cfs_rq->throttled)
4411 return;
4412 /*
4413 * if we're unable to extend our runtime we resched so that the active
4414 * hierarchy can be throttled
4415 */
4416 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4417 resched_curr(rq_of(cfs_rq));
4418}
4419
4420static __always_inline
4421void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4422{
4423 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4424 return;
4425
4426 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4427}
4428
4429static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4430{
4431 return cfs_bandwidth_used() && cfs_rq->throttled;
4432}
4433
4434/* check whether cfs_rq, or any parent, is throttled */
4435static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4436{
4437 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4438}
4439
4440/*
4441 * Ensure that neither of the group entities corresponding to src_cpu or
4442 * dest_cpu are members of a throttled hierarchy when performing group
4443 * load-balance operations.
4444 */
4445static inline int throttled_lb_pair(struct task_group *tg,
4446 int src_cpu, int dest_cpu)
4447{
4448 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4449
4450 src_cfs_rq = tg->cfs_rq[src_cpu];
4451 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4452
4453 return throttled_hierarchy(src_cfs_rq) ||
4454 throttled_hierarchy(dest_cfs_rq);
4455}
4456
4457static int tg_unthrottle_up(struct task_group *tg, void *data)
4458{
4459 struct rq *rq = data;
4460 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4461
4462 cfs_rq->throttle_count--;
4463 if (!cfs_rq->throttle_count) {
4464 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4465 cfs_rq->throttled_clock_task;
4466
4467 /* Add cfs_rq with already running entity in the list */
4468 if (cfs_rq->nr_running >= 1)
4469 list_add_leaf_cfs_rq(cfs_rq);
4470 }
4471
4472 return 0;
4473}
4474
4475static int tg_throttle_down(struct task_group *tg, void *data)
4476{
4477 struct rq *rq = data;
4478 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4479
4480 /* group is entering throttled state, stop time */
4481 if (!cfs_rq->throttle_count) {
4482 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4483 list_del_leaf_cfs_rq(cfs_rq);
4484 }
4485 cfs_rq->throttle_count++;
4486
4487 return 0;
4488}
4489
4490static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4491{
4492 struct rq *rq = rq_of(cfs_rq);
4493 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4494 struct sched_entity *se;
4495 long task_delta, idle_task_delta, dequeue = 1;
4496 bool empty;
4497
4498 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4499
4500 /* freeze hierarchy runnable averages while throttled */
4501 rcu_read_lock();
4502 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4503 rcu_read_unlock();
4504
4505 task_delta = cfs_rq->h_nr_running;
4506 idle_task_delta = cfs_rq->idle_h_nr_running;
4507 for_each_sched_entity(se) {
4508 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4509 /* throttled entity or throttle-on-deactivate */
4510 if (!se->on_rq)
4511 break;
4512
4513 if (dequeue)
4514 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4515 qcfs_rq->h_nr_running -= task_delta;
4516 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4517
4518 if (qcfs_rq->load.weight)
4519 dequeue = 0;
4520 }
4521
4522 if (!se)
4523 sub_nr_running(rq, task_delta);
4524
4525 cfs_rq->throttled = 1;
4526 cfs_rq->throttled_clock = rq_clock(rq);
4527 raw_spin_lock(&cfs_b->lock);
4528 empty = list_empty(&cfs_b->throttled_cfs_rq);
4529
4530 /*
4531 * Add to the _head_ of the list, so that an already-started
4532 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4533 * not running add to the tail so that later runqueues don't get starved.
4534 */
4535 if (cfs_b->distribute_running)
4536 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4537 else
4538 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4539
4540 /*
4541 * If we're the first throttled task, make sure the bandwidth
4542 * timer is running.
4543 */
4544 if (empty)
4545 start_cfs_bandwidth(cfs_b);
4546
4547 raw_spin_unlock(&cfs_b->lock);
4548}
4549
4550void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4551{
4552 struct rq *rq = rq_of(cfs_rq);
4553 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4554 struct sched_entity *se;
4555 int enqueue = 1;
4556 long task_delta, idle_task_delta;
4557
4558 se = cfs_rq->tg->se[cpu_of(rq)];
4559
4560 cfs_rq->throttled = 0;
4561
4562 update_rq_clock(rq);
4563
4564 raw_spin_lock(&cfs_b->lock);
4565 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4566 list_del_rcu(&cfs_rq->throttled_list);
4567 raw_spin_unlock(&cfs_b->lock);
4568
4569 /* update hierarchical throttle state */
4570 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4571
4572 if (!cfs_rq->load.weight)
4573 return;
4574
4575 task_delta = cfs_rq->h_nr_running;
4576 idle_task_delta = cfs_rq->idle_h_nr_running;
4577 for_each_sched_entity(se) {
4578 if (se->on_rq)
4579 enqueue = 0;
4580
4581 cfs_rq = cfs_rq_of(se);
4582 if (enqueue)
4583 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4584 cfs_rq->h_nr_running += task_delta;
4585 cfs_rq->idle_h_nr_running += idle_task_delta;
4586
4587 if (cfs_rq_throttled(cfs_rq))
4588 break;
4589 }
4590
4591 assert_list_leaf_cfs_rq(rq);
4592
4593 if (!se)
4594 add_nr_running(rq, task_delta);
4595
4596 /* Determine whether we need to wake up potentially idle CPU: */
4597 if (rq->curr == rq->idle && rq->cfs.nr_running)
4598 resched_curr(rq);
4599}
4600
4601static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4602{
4603 struct cfs_rq *cfs_rq;
4604 u64 runtime;
4605 u64 starting_runtime = remaining;
4606
4607 rcu_read_lock();
4608 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4609 throttled_list) {
4610 struct rq *rq = rq_of(cfs_rq);
4611 struct rq_flags rf;
4612
4613 rq_lock_irqsave(rq, &rf);
4614 if (!cfs_rq_throttled(cfs_rq))
4615 goto next;
4616
4617 /* By the above check, this should never be true */
4618 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4619
4620 runtime = -cfs_rq->runtime_remaining + 1;
4621 if (runtime > remaining)
4622 runtime = remaining;
4623 remaining -= runtime;
4624
4625 cfs_rq->runtime_remaining += runtime;
4626
4627 /* we check whether we're throttled above */
4628 if (cfs_rq->runtime_remaining > 0)
4629 unthrottle_cfs_rq(cfs_rq);
4630
4631next:
4632 rq_unlock_irqrestore(rq, &rf);
4633
4634 if (!remaining)
4635 break;
4636 }
4637 rcu_read_unlock();
4638
4639 return starting_runtime - remaining;
4640}
4641
4642/*
4643 * Responsible for refilling a task_group's bandwidth and unthrottling its
4644 * cfs_rqs as appropriate. If there has been no activity within the last
4645 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4646 * used to track this state.
4647 */
4648static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4649{
4650 u64 runtime;
4651 int throttled;
4652
4653 /* no need to continue the timer with no bandwidth constraint */
4654 if (cfs_b->quota == RUNTIME_INF)
4655 goto out_deactivate;
4656
4657 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4658 cfs_b->nr_periods += overrun;
4659
4660 /*
4661 * idle depends on !throttled (for the case of a large deficit), and if
4662 * we're going inactive then everything else can be deferred
4663 */
4664 if (cfs_b->idle && !throttled)
4665 goto out_deactivate;
4666
4667 __refill_cfs_bandwidth_runtime(cfs_b);
4668
4669 if (!throttled) {
4670 /* mark as potentially idle for the upcoming period */
4671 cfs_b->idle = 1;
4672 return 0;
4673 }
4674
4675 /* account preceding periods in which throttling occurred */
4676 cfs_b->nr_throttled += overrun;
4677
4678 /*
4679 * This check is repeated as we are holding onto the new bandwidth while
4680 * we unthrottle. This can potentially race with an unthrottled group
4681 * trying to acquire new bandwidth from the global pool. This can result
4682 * in us over-using our runtime if it is all used during this loop, but
4683 * only by limited amounts in that extreme case.
4684 */
4685 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4686 runtime = cfs_b->runtime;
4687 cfs_b->distribute_running = 1;
4688 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4689 /* we can't nest cfs_b->lock while distributing bandwidth */
4690 runtime = distribute_cfs_runtime(cfs_b, runtime);
4691 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4692
4693 cfs_b->distribute_running = 0;
4694 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4695
4696 lsub_positive(&cfs_b->runtime, runtime);
4697 }
4698
4699 /*
4700 * While we are ensured activity in the period following an
4701 * unthrottle, this also covers the case in which the new bandwidth is
4702 * insufficient to cover the existing bandwidth deficit. (Forcing the
4703 * timer to remain active while there are any throttled entities.)
4704 */
4705 cfs_b->idle = 0;
4706
4707 return 0;
4708
4709out_deactivate:
4710 return 1;
4711}
4712
4713/* a cfs_rq won't donate quota below this amount */
4714static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4715/* minimum remaining period time to redistribute slack quota */
4716static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4717/* how long we wait to gather additional slack before distributing */
4718static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4719
4720/*
4721 * Are we near the end of the current quota period?
4722 *
4723 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4724 * hrtimer base being cleared by hrtimer_start. In the case of
4725 * migrate_hrtimers, base is never cleared, so we are fine.
4726 */
4727static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4728{
4729 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4730 u64 remaining;
4731
4732 /* if the call-back is running a quota refresh is already occurring */
4733 if (hrtimer_callback_running(refresh_timer))
4734 return 1;
4735
4736 /* is a quota refresh about to occur? */
4737 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4738 if (remaining < min_expire)
4739 return 1;
4740
4741 return 0;
4742}
4743
4744static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4745{
4746 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4747
4748 /* if there's a quota refresh soon don't bother with slack */
4749 if (runtime_refresh_within(cfs_b, min_left))
4750 return;
4751
4752 /* don't push forwards an existing deferred unthrottle */
4753 if (cfs_b->slack_started)
4754 return;
4755 cfs_b->slack_started = true;
4756
4757 hrtimer_start(&cfs_b->slack_timer,
4758 ns_to_ktime(cfs_bandwidth_slack_period),
4759 HRTIMER_MODE_REL);
4760}
4761
4762/* we know any runtime found here is valid as update_curr() precedes return */
4763static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4764{
4765 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4766 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4767
4768 if (slack_runtime <= 0)
4769 return;
4770
4771 raw_spin_lock(&cfs_b->lock);
4772 if (cfs_b->quota != RUNTIME_INF) {
4773 cfs_b->runtime += slack_runtime;
4774
4775 /* we are under rq->lock, defer unthrottling using a timer */
4776 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4777 !list_empty(&cfs_b->throttled_cfs_rq))
4778 start_cfs_slack_bandwidth(cfs_b);
4779 }
4780 raw_spin_unlock(&cfs_b->lock);
4781
4782 /* even if it's not valid for return we don't want to try again */
4783 cfs_rq->runtime_remaining -= slack_runtime;
4784}
4785
4786static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4787{
4788 if (!cfs_bandwidth_used())
4789 return;
4790
4791 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4792 return;
4793
4794 __return_cfs_rq_runtime(cfs_rq);
4795}
4796
4797/*
4798 * This is done with a timer (instead of inline with bandwidth return) since
4799 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4800 */
4801static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4802{
4803 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4804 unsigned long flags;
4805
4806 /* confirm we're still not at a refresh boundary */
4807 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4808 cfs_b->slack_started = false;
4809 if (cfs_b->distribute_running) {
4810 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4811 return;
4812 }
4813
4814 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4815 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4816 return;
4817 }
4818
4819 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4820 runtime = cfs_b->runtime;
4821
4822 if (runtime)
4823 cfs_b->distribute_running = 1;
4824
4825 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4826
4827 if (!runtime)
4828 return;
4829
4830 runtime = distribute_cfs_runtime(cfs_b, runtime);
4831
4832 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4833 lsub_positive(&cfs_b->runtime, runtime);
4834 cfs_b->distribute_running = 0;
4835 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4836}
4837
4838/*
4839 * When a group wakes up we want to make sure that its quota is not already
4840 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4841 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4842 */
4843static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4844{
4845 if (!cfs_bandwidth_used())
4846 return;
4847
4848 /* an active group must be handled by the update_curr()->put() path */
4849 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4850 return;
4851
4852 /* ensure the group is not already throttled */
4853 if (cfs_rq_throttled(cfs_rq))
4854 return;
4855
4856 /* update runtime allocation */
4857 account_cfs_rq_runtime(cfs_rq, 0);
4858 if (cfs_rq->runtime_remaining <= 0)
4859 throttle_cfs_rq(cfs_rq);
4860}
4861
4862static void sync_throttle(struct task_group *tg, int cpu)
4863{
4864 struct cfs_rq *pcfs_rq, *cfs_rq;
4865
4866 if (!cfs_bandwidth_used())
4867 return;
4868
4869 if (!tg->parent)
4870 return;
4871
4872 cfs_rq = tg->cfs_rq[cpu];
4873 pcfs_rq = tg->parent->cfs_rq[cpu];
4874
4875 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4876 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4877}
4878
4879/* conditionally throttle active cfs_rq's from put_prev_entity() */
4880static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4881{
4882 if (!cfs_bandwidth_used())
4883 return false;
4884
4885 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4886 return false;
4887
4888 /*
4889 * it's possible for a throttled entity to be forced into a running
4890 * state (e.g. set_curr_task), in this case we're finished.
4891 */
4892 if (cfs_rq_throttled(cfs_rq))
4893 return true;
4894
4895 throttle_cfs_rq(cfs_rq);
4896 return true;
4897}
4898
4899static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4900{
4901 struct cfs_bandwidth *cfs_b =
4902 container_of(timer, struct cfs_bandwidth, slack_timer);
4903
4904 do_sched_cfs_slack_timer(cfs_b);
4905
4906 return HRTIMER_NORESTART;
4907}
4908
4909extern const u64 max_cfs_quota_period;
4910
4911static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4912{
4913 struct cfs_bandwidth *cfs_b =
4914 container_of(timer, struct cfs_bandwidth, period_timer);
4915 unsigned long flags;
4916 int overrun;
4917 int idle = 0;
4918 int count = 0;
4919
4920 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4921 for (;;) {
4922 overrun = hrtimer_forward_now(timer, cfs_b->period);
4923 if (!overrun)
4924 break;
4925
4926 if (++count > 3) {
4927 u64 new, old = ktime_to_ns(cfs_b->period);
4928
4929 /*
4930 * Grow period by a factor of 2 to avoid losing precision.
4931 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4932 * to fail.
4933 */
4934 new = old * 2;
4935 if (new < max_cfs_quota_period) {
4936 cfs_b->period = ns_to_ktime(new);
4937 cfs_b->quota *= 2;
4938
4939 pr_warn_ratelimited(
4940 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4941 smp_processor_id(),
4942 div_u64(new, NSEC_PER_USEC),
4943 div_u64(cfs_b->quota, NSEC_PER_USEC));
4944 } else {
4945 pr_warn_ratelimited(
4946 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4947 smp_processor_id(),
4948 div_u64(old, NSEC_PER_USEC),
4949 div_u64(cfs_b->quota, NSEC_PER_USEC));
4950 }
4951
4952 /* reset count so we don't come right back in here */
4953 count = 0;
4954 }
4955
4956 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
4957 }
4958 if (idle)
4959 cfs_b->period_active = 0;
4960 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4961
4962 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4963}
4964
4965void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4966{
4967 raw_spin_lock_init(&cfs_b->lock);
4968 cfs_b->runtime = 0;
4969 cfs_b->quota = RUNTIME_INF;
4970 cfs_b->period = ns_to_ktime(default_cfs_period());
4971
4972 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4973 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4974 cfs_b->period_timer.function = sched_cfs_period_timer;
4975 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4976 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4977 cfs_b->distribute_running = 0;
4978 cfs_b->slack_started = false;
4979}
4980
4981static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4982{
4983 cfs_rq->runtime_enabled = 0;
4984 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4985}
4986
4987void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4988{
4989 lockdep_assert_held(&cfs_b->lock);
4990
4991 if (cfs_b->period_active)
4992 return;
4993
4994 cfs_b->period_active = 1;
4995 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4996 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4997}
4998
4999static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5000{
5001 /* init_cfs_bandwidth() was not called */
5002 if (!cfs_b->throttled_cfs_rq.next)
5003 return;
5004
5005 hrtimer_cancel(&cfs_b->period_timer);
5006 hrtimer_cancel(&cfs_b->slack_timer);
5007}
5008
5009/*
5010 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5011 *
5012 * The race is harmless, since modifying bandwidth settings of unhooked group
5013 * bits doesn't do much.
5014 */
5015
5016/* cpu online calback */
5017static void __maybe_unused update_runtime_enabled(struct rq *rq)
5018{
5019 struct task_group *tg;
5020
5021 lockdep_assert_held(&rq->lock);
5022
5023 rcu_read_lock();
5024 list_for_each_entry_rcu(tg, &task_groups, list) {
5025 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5026 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5027
5028 raw_spin_lock(&cfs_b->lock);
5029 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5030 raw_spin_unlock(&cfs_b->lock);
5031 }
5032 rcu_read_unlock();
5033}
5034
5035/* cpu offline callback */
5036static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5037{
5038 struct task_group *tg;
5039
5040 lockdep_assert_held(&rq->lock);
5041
5042 rcu_read_lock();
5043 list_for_each_entry_rcu(tg, &task_groups, list) {
5044 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5045
5046 if (!cfs_rq->runtime_enabled)
5047 continue;
5048
5049 /*
5050 * clock_task is not advancing so we just need to make sure
5051 * there's some valid quota amount
5052 */
5053 cfs_rq->runtime_remaining = 1;
5054 /*
5055 * Offline rq is schedulable till CPU is completely disabled
5056 * in take_cpu_down(), so we prevent new cfs throttling here.
5057 */
5058 cfs_rq->runtime_enabled = 0;
5059
5060 if (cfs_rq_throttled(cfs_rq))
5061 unthrottle_cfs_rq(cfs_rq);
5062 }
5063 rcu_read_unlock();
5064}
5065
5066#else /* CONFIG_CFS_BANDWIDTH */
5067
5068static inline bool cfs_bandwidth_used(void)
5069{
5070 return false;
5071}
5072
5073static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5074static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5075static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5076static inline void sync_throttle(struct task_group *tg, int cpu) {}
5077static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5078
5079static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5080{
5081 return 0;
5082}
5083
5084static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5085{
5086 return 0;
5087}
5088
5089static inline int throttled_lb_pair(struct task_group *tg,
5090 int src_cpu, int dest_cpu)
5091{
5092 return 0;
5093}
5094
5095void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5096
5097#ifdef CONFIG_FAIR_GROUP_SCHED
5098static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5099#endif
5100
5101static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5102{
5103 return NULL;
5104}
5105static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5106static inline void update_runtime_enabled(struct rq *rq) {}
5107static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5108
5109#endif /* CONFIG_CFS_BANDWIDTH */
5110
5111/**************************************************
5112 * CFS operations on tasks:
5113 */
5114
5115#ifdef CONFIG_SCHED_HRTICK
5116static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5117{
5118 struct sched_entity *se = &p->se;
5119 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5120
5121 SCHED_WARN_ON(task_rq(p) != rq);
5122
5123 if (rq->cfs.h_nr_running > 1) {
5124 u64 slice = sched_slice(cfs_rq, se);
5125 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5126 s64 delta = slice - ran;
5127
5128 if (delta < 0) {
5129 if (rq->curr == p)
5130 resched_curr(rq);
5131 return;
5132 }
5133 hrtick_start(rq, delta);
5134 }
5135}
5136
5137/*
5138 * called from enqueue/dequeue and updates the hrtick when the
5139 * current task is from our class and nr_running is low enough
5140 * to matter.
5141 */
5142static void hrtick_update(struct rq *rq)
5143{
5144 struct task_struct *curr = rq->curr;
5145
5146 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5147 return;
5148
5149 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5150 hrtick_start_fair(rq, curr);
5151}
5152#else /* !CONFIG_SCHED_HRTICK */
5153static inline void
5154hrtick_start_fair(struct rq *rq, struct task_struct *p)
5155{
5156}
5157
5158static inline void hrtick_update(struct rq *rq)
5159{
5160}
5161#endif
5162
5163#ifdef CONFIG_SMP
5164static inline unsigned long cpu_util(int cpu);
5165
5166static inline bool cpu_overutilized(int cpu)
5167{
5168 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5169}
5170
5171static inline void update_overutilized_status(struct rq *rq)
5172{
5173 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5174 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5175 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5176 }
5177}
5178#else
5179static inline void update_overutilized_status(struct rq *rq) { }
5180#endif
5181
5182/*
5183 * The enqueue_task method is called before nr_running is
5184 * increased. Here we update the fair scheduling stats and
5185 * then put the task into the rbtree:
5186 */
5187static void
5188enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5189{
5190 struct cfs_rq *cfs_rq;
5191 struct sched_entity *se = &p->se;
5192 int idle_h_nr_running = task_has_idle_policy(p);
5193
5194 /*
5195 * The code below (indirectly) updates schedutil which looks at
5196 * the cfs_rq utilization to select a frequency.
5197 * Let's add the task's estimated utilization to the cfs_rq's
5198 * estimated utilization, before we update schedutil.
5199 */
5200 util_est_enqueue(&rq->cfs, p);
5201
5202 /*
5203 * If in_iowait is set, the code below may not trigger any cpufreq
5204 * utilization updates, so do it here explicitly with the IOWAIT flag
5205 * passed.
5206 */
5207 if (p->in_iowait)
5208 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5209
5210 for_each_sched_entity(se) {
5211 if (se->on_rq)
5212 break;
5213 cfs_rq = cfs_rq_of(se);
5214 enqueue_entity(cfs_rq, se, flags);
5215
5216 /*
5217 * end evaluation on encountering a throttled cfs_rq
5218 *
5219 * note: in the case of encountering a throttled cfs_rq we will
5220 * post the final h_nr_running increment below.
5221 */
5222 if (cfs_rq_throttled(cfs_rq))
5223 break;
5224 cfs_rq->h_nr_running++;
5225 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5226
5227 flags = ENQUEUE_WAKEUP;
5228 }
5229
5230 for_each_sched_entity(se) {
5231 cfs_rq = cfs_rq_of(se);
5232 cfs_rq->h_nr_running++;
5233 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5234
5235 if (cfs_rq_throttled(cfs_rq))
5236 break;
5237
5238 update_load_avg(cfs_rq, se, UPDATE_TG);
5239 update_cfs_group(se);
5240 }
5241
5242 if (!se) {
5243 add_nr_running(rq, 1);
5244 /*
5245 * Since new tasks are assigned an initial util_avg equal to
5246 * half of the spare capacity of their CPU, tiny tasks have the
5247 * ability to cross the overutilized threshold, which will
5248 * result in the load balancer ruining all the task placement
5249 * done by EAS. As a way to mitigate that effect, do not account
5250 * for the first enqueue operation of new tasks during the
5251 * overutilized flag detection.
5252 *
5253 * A better way of solving this problem would be to wait for
5254 * the PELT signals of tasks to converge before taking them
5255 * into account, but that is not straightforward to implement,
5256 * and the following generally works well enough in practice.
5257 */
5258 if (flags & ENQUEUE_WAKEUP)
5259 update_overutilized_status(rq);
5260
5261 }
5262
5263 if (cfs_bandwidth_used()) {
5264 /*
5265 * When bandwidth control is enabled; the cfs_rq_throttled()
5266 * breaks in the above iteration can result in incomplete
5267 * leaf list maintenance, resulting in triggering the assertion
5268 * below.
5269 */
5270 for_each_sched_entity(se) {
5271 cfs_rq = cfs_rq_of(se);
5272
5273 if (list_add_leaf_cfs_rq(cfs_rq))
5274 break;
5275 }
5276 }
5277
5278 assert_list_leaf_cfs_rq(rq);
5279
5280 hrtick_update(rq);
5281}
5282
5283static void set_next_buddy(struct sched_entity *se);
5284
5285/*
5286 * The dequeue_task method is called before nr_running is
5287 * decreased. We remove the task from the rbtree and
5288 * update the fair scheduling stats:
5289 */
5290static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5291{
5292 struct cfs_rq *cfs_rq;
5293 struct sched_entity *se = &p->se;
5294 int task_sleep = flags & DEQUEUE_SLEEP;
5295 int idle_h_nr_running = task_has_idle_policy(p);
5296
5297 for_each_sched_entity(se) {
5298 cfs_rq = cfs_rq_of(se);
5299 dequeue_entity(cfs_rq, se, flags);
5300
5301 /*
5302 * end evaluation on encountering a throttled cfs_rq
5303 *
5304 * note: in the case of encountering a throttled cfs_rq we will
5305 * post the final h_nr_running decrement below.
5306 */
5307 if (cfs_rq_throttled(cfs_rq))
5308 break;
5309 cfs_rq->h_nr_running--;
5310 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5311
5312 /* Don't dequeue parent if it has other entities besides us */
5313 if (cfs_rq->load.weight) {
5314 /* Avoid re-evaluating load for this entity: */
5315 se = parent_entity(se);
5316 /*
5317 * Bias pick_next to pick a task from this cfs_rq, as
5318 * p is sleeping when it is within its sched_slice.
5319 */
5320 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5321 set_next_buddy(se);
5322 break;
5323 }
5324 flags |= DEQUEUE_SLEEP;
5325 }
5326
5327 for_each_sched_entity(se) {
5328 cfs_rq = cfs_rq_of(se);
5329 cfs_rq->h_nr_running--;
5330 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5331
5332 if (cfs_rq_throttled(cfs_rq))
5333 break;
5334
5335 update_load_avg(cfs_rq, se, UPDATE_TG);
5336 update_cfs_group(se);
5337 }
5338
5339 if (!se)
5340 sub_nr_running(rq, 1);
5341
5342 util_est_dequeue(&rq->cfs, p, task_sleep);
5343 hrtick_update(rq);
5344}
5345
5346#ifdef CONFIG_SMP
5347
5348/* Working cpumask for: load_balance, load_balance_newidle. */
5349DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5350DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5351
5352#ifdef CONFIG_NO_HZ_COMMON
5353
5354static struct {
5355 cpumask_var_t idle_cpus_mask;
5356 atomic_t nr_cpus;
5357 int has_blocked; /* Idle CPUS has blocked load */
5358 unsigned long next_balance; /* in jiffy units */
5359 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5360} nohz ____cacheline_aligned;
5361
5362#endif /* CONFIG_NO_HZ_COMMON */
5363
5364/* CPU only has SCHED_IDLE tasks enqueued */
5365static int sched_idle_cpu(int cpu)
5366{
5367 struct rq *rq = cpu_rq(cpu);
5368
5369 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5370 rq->nr_running);
5371}
5372
5373static unsigned long cpu_runnable_load(struct rq *rq)
5374{
5375 return cfs_rq_runnable_load_avg(&rq->cfs);
5376}
5377
5378static unsigned long capacity_of(int cpu)
5379{
5380 return cpu_rq(cpu)->cpu_capacity;
5381}
5382
5383static unsigned long cpu_avg_load_per_task(int cpu)
5384{
5385 struct rq *rq = cpu_rq(cpu);
5386 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5387 unsigned long load_avg = cpu_runnable_load(rq);
5388
5389 if (nr_running)
5390 return load_avg / nr_running;
5391
5392 return 0;
5393}
5394
5395static void record_wakee(struct task_struct *p)
5396{
5397 /*
5398 * Only decay a single time; tasks that have less then 1 wakeup per
5399 * jiffy will not have built up many flips.
5400 */
5401 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5402 current->wakee_flips >>= 1;
5403 current->wakee_flip_decay_ts = jiffies;
5404 }
5405
5406 if (current->last_wakee != p) {
5407 current->last_wakee = p;
5408 current->wakee_flips++;
5409 }
5410}
5411
5412/*
5413 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5414 *
5415 * A waker of many should wake a different task than the one last awakened
5416 * at a frequency roughly N times higher than one of its wakees.
5417 *
5418 * In order to determine whether we should let the load spread vs consolidating
5419 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5420 * partner, and a factor of lls_size higher frequency in the other.
5421 *
5422 * With both conditions met, we can be relatively sure that the relationship is
5423 * non-monogamous, with partner count exceeding socket size.
5424 *
5425 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5426 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5427 * socket size.
5428 */
5429static int wake_wide(struct task_struct *p)
5430{
5431 unsigned int master = current->wakee_flips;
5432 unsigned int slave = p->wakee_flips;
5433 int factor = this_cpu_read(sd_llc_size);
5434
5435 if (master < slave)
5436 swap(master, slave);
5437 if (slave < factor || master < slave * factor)
5438 return 0;
5439 return 1;
5440}
5441
5442/*
5443 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5444 * soonest. For the purpose of speed we only consider the waking and previous
5445 * CPU.
5446 *
5447 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5448 * cache-affine and is (or will be) idle.
5449 *
5450 * wake_affine_weight() - considers the weight to reflect the average
5451 * scheduling latency of the CPUs. This seems to work
5452 * for the overloaded case.
5453 */
5454static int
5455wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5456{
5457 /*
5458 * If this_cpu is idle, it implies the wakeup is from interrupt
5459 * context. Only allow the move if cache is shared. Otherwise an
5460 * interrupt intensive workload could force all tasks onto one
5461 * node depending on the IO topology or IRQ affinity settings.
5462 *
5463 * If the prev_cpu is idle and cache affine then avoid a migration.
5464 * There is no guarantee that the cache hot data from an interrupt
5465 * is more important than cache hot data on the prev_cpu and from
5466 * a cpufreq perspective, it's better to have higher utilisation
5467 * on one CPU.
5468 */
5469 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5470 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5471
5472 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5473 return this_cpu;
5474
5475 return nr_cpumask_bits;
5476}
5477
5478static int
5479wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5480 int this_cpu, int prev_cpu, int sync)
5481{
5482 s64 this_eff_load, prev_eff_load;
5483 unsigned long task_load;
5484
5485 this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5486
5487 if (sync) {
5488 unsigned long current_load = task_h_load(current);
5489
5490 if (current_load > this_eff_load)
5491 return this_cpu;
5492
5493 this_eff_load -= current_load;
5494 }
5495
5496 task_load = task_h_load(p);
5497
5498 this_eff_load += task_load;
5499 if (sched_feat(WA_BIAS))
5500 this_eff_load *= 100;
5501 this_eff_load *= capacity_of(prev_cpu);
5502
5503 prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5504 prev_eff_load -= task_load;
5505 if (sched_feat(WA_BIAS))
5506 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5507 prev_eff_load *= capacity_of(this_cpu);
5508
5509 /*
5510 * If sync, adjust the weight of prev_eff_load such that if
5511 * prev_eff == this_eff that select_idle_sibling() will consider
5512 * stacking the wakee on top of the waker if no other CPU is
5513 * idle.
5514 */
5515 if (sync)
5516 prev_eff_load += 1;
5517
5518 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5519}
5520
5521static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5522 int this_cpu, int prev_cpu, int sync)
5523{
5524 int target = nr_cpumask_bits;
5525
5526 if (sched_feat(WA_IDLE))
5527 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5528
5529 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5530 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5531
5532 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5533 if (target == nr_cpumask_bits)
5534 return prev_cpu;
5535
5536 schedstat_inc(sd->ttwu_move_affine);
5537 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5538 return target;
5539}
5540
5541static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5542
5543static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5544{
5545 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5546}
5547
5548/*
5549 * find_idlest_group finds and returns the least busy CPU group within the
5550 * domain.
5551 *
5552 * Assumes p is allowed on at least one CPU in sd.
5553 */
5554static struct sched_group *
5555find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5556 int this_cpu, int sd_flag)
5557{
5558 struct sched_group *idlest = NULL, *group = sd->groups;
5559 struct sched_group *most_spare_sg = NULL;
5560 unsigned long min_runnable_load = ULONG_MAX;
5561 unsigned long this_runnable_load = ULONG_MAX;
5562 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5563 unsigned long most_spare = 0, this_spare = 0;
5564 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5565 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5566 (sd->imbalance_pct-100) / 100;
5567
5568 do {
5569 unsigned long load, avg_load, runnable_load;
5570 unsigned long spare_cap, max_spare_cap;
5571 int local_group;
5572 int i;
5573
5574 /* Skip over this group if it has no CPUs allowed */
5575 if (!cpumask_intersects(sched_group_span(group),
5576 p->cpus_ptr))
5577 continue;
5578
5579 local_group = cpumask_test_cpu(this_cpu,
5580 sched_group_span(group));
5581
5582 /*
5583 * Tally up the load of all CPUs in the group and find
5584 * the group containing the CPU with most spare capacity.
5585 */
5586 avg_load = 0;
5587 runnable_load = 0;
5588 max_spare_cap = 0;
5589
5590 for_each_cpu(i, sched_group_span(group)) {
5591 load = cpu_runnable_load(cpu_rq(i));
5592 runnable_load += load;
5593
5594 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5595
5596 spare_cap = capacity_spare_without(i, p);
5597
5598 if (spare_cap > max_spare_cap)
5599 max_spare_cap = spare_cap;
5600 }
5601
5602 /* Adjust by relative CPU capacity of the group */
5603 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5604 group->sgc->capacity;
5605 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5606 group->sgc->capacity;
5607
5608 if (local_group) {
5609 this_runnable_load = runnable_load;
5610 this_avg_load = avg_load;
5611 this_spare = max_spare_cap;
5612 } else {
5613 if (min_runnable_load > (runnable_load + imbalance)) {
5614 /*
5615 * The runnable load is significantly smaller
5616 * so we can pick this new CPU:
5617 */
5618 min_runnable_load = runnable_load;
5619 min_avg_load = avg_load;
5620 idlest = group;
5621 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5622 (100*min_avg_load > imbalance_scale*avg_load)) {
5623 /*
5624 * The runnable loads are close so take the
5625 * blocked load into account through avg_load:
5626 */
5627 min_avg_load = avg_load;
5628 idlest = group;
5629 }
5630
5631 if (most_spare < max_spare_cap) {
5632 most_spare = max_spare_cap;
5633 most_spare_sg = group;
5634 }
5635 }
5636 } while (group = group->next, group != sd->groups);
5637
5638 /*
5639 * The cross-over point between using spare capacity or least load
5640 * is too conservative for high utilization tasks on partially
5641 * utilized systems if we require spare_capacity > task_util(p),
5642 * so we allow for some task stuffing by using
5643 * spare_capacity > task_util(p)/2.
5644 *
5645 * Spare capacity can't be used for fork because the utilization has
5646 * not been set yet, we must first select a rq to compute the initial
5647 * utilization.
5648 */
5649 if (sd_flag & SD_BALANCE_FORK)
5650 goto skip_spare;
5651
5652 if (this_spare > task_util(p) / 2 &&
5653 imbalance_scale*this_spare > 100*most_spare)
5654 return NULL;
5655
5656 if (most_spare > task_util(p) / 2)
5657 return most_spare_sg;
5658
5659skip_spare:
5660 if (!idlest)
5661 return NULL;
5662
5663 /*
5664 * When comparing groups across NUMA domains, it's possible for the
5665 * local domain to be very lightly loaded relative to the remote
5666 * domains but "imbalance" skews the comparison making remote CPUs
5667 * look much more favourable. When considering cross-domain, add
5668 * imbalance to the runnable load on the remote node and consider
5669 * staying local.
5670 */
5671 if ((sd->flags & SD_NUMA) &&
5672 min_runnable_load + imbalance >= this_runnable_load)
5673 return NULL;
5674
5675 if (min_runnable_load > (this_runnable_load + imbalance))
5676 return NULL;
5677
5678 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5679 (100*this_avg_load < imbalance_scale*min_avg_load))
5680 return NULL;
5681
5682 return idlest;
5683}
5684
5685/*
5686 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5687 */
5688static int
5689find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5690{
5691 unsigned long load, min_load = ULONG_MAX;
5692 unsigned int min_exit_latency = UINT_MAX;
5693 u64 latest_idle_timestamp = 0;
5694 int least_loaded_cpu = this_cpu;
5695 int shallowest_idle_cpu = -1, si_cpu = -1;
5696 int i;
5697
5698 /* Check if we have any choice: */
5699 if (group->group_weight == 1)
5700 return cpumask_first(sched_group_span(group));
5701
5702 /* Traverse only the allowed CPUs */
5703 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5704 if (available_idle_cpu(i)) {
5705 struct rq *rq = cpu_rq(i);
5706 struct cpuidle_state *idle = idle_get_state(rq);
5707 if (idle && idle->exit_latency < min_exit_latency) {
5708 /*
5709 * We give priority to a CPU whose idle state
5710 * has the smallest exit latency irrespective
5711 * of any idle timestamp.
5712 */
5713 min_exit_latency = idle->exit_latency;
5714 latest_idle_timestamp = rq->idle_stamp;
5715 shallowest_idle_cpu = i;
5716 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5717 rq->idle_stamp > latest_idle_timestamp) {
5718 /*
5719 * If equal or no active idle state, then
5720 * the most recently idled CPU might have
5721 * a warmer cache.
5722 */
5723 latest_idle_timestamp = rq->idle_stamp;
5724 shallowest_idle_cpu = i;
5725 }
5726 } else if (shallowest_idle_cpu == -1 && si_cpu == -1) {
5727 if (sched_idle_cpu(i)) {
5728 si_cpu = i;
5729 continue;
5730 }
5731
5732 load = cpu_runnable_load(cpu_rq(i));
5733 if (load < min_load) {
5734 min_load = load;
5735 least_loaded_cpu = i;
5736 }
5737 }
5738 }
5739
5740 if (shallowest_idle_cpu != -1)
5741 return shallowest_idle_cpu;
5742 if (si_cpu != -1)
5743 return si_cpu;
5744 return least_loaded_cpu;
5745}
5746
5747static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5748 int cpu, int prev_cpu, int sd_flag)
5749{
5750 int new_cpu = cpu;
5751
5752 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5753 return prev_cpu;
5754
5755 /*
5756 * We need task's util for capacity_spare_without, sync it up to
5757 * prev_cpu's last_update_time.
5758 */
5759 if (!(sd_flag & SD_BALANCE_FORK))
5760 sync_entity_load_avg(&p->se);
5761
5762 while (sd) {
5763 struct sched_group *group;
5764 struct sched_domain *tmp;
5765 int weight;
5766
5767 if (!(sd->flags & sd_flag)) {
5768 sd = sd->child;
5769 continue;
5770 }
5771
5772 group = find_idlest_group(sd, p, cpu, sd_flag);
5773 if (!group) {
5774 sd = sd->child;
5775 continue;
5776 }
5777
5778 new_cpu = find_idlest_group_cpu(group, p, cpu);
5779 if (new_cpu == cpu) {
5780 /* Now try balancing at a lower domain level of 'cpu': */
5781 sd = sd->child;
5782 continue;
5783 }
5784
5785 /* Now try balancing at a lower domain level of 'new_cpu': */
5786 cpu = new_cpu;
5787 weight = sd->span_weight;
5788 sd = NULL;
5789 for_each_domain(cpu, tmp) {
5790 if (weight <= tmp->span_weight)
5791 break;
5792 if (tmp->flags & sd_flag)
5793 sd = tmp;
5794 }
5795 }
5796
5797 return new_cpu;
5798}
5799
5800#ifdef CONFIG_SCHED_SMT
5801DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5802EXPORT_SYMBOL_GPL(sched_smt_present);
5803
5804static inline void set_idle_cores(int cpu, int val)
5805{
5806 struct sched_domain_shared *sds;
5807
5808 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5809 if (sds)
5810 WRITE_ONCE(sds->has_idle_cores, val);
5811}
5812
5813static inline bool test_idle_cores(int cpu, bool def)
5814{
5815 struct sched_domain_shared *sds;
5816
5817 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5818 if (sds)
5819 return READ_ONCE(sds->has_idle_cores);
5820
5821 return def;
5822}
5823
5824/*
5825 * Scans the local SMT mask to see if the entire core is idle, and records this
5826 * information in sd_llc_shared->has_idle_cores.
5827 *
5828 * Since SMT siblings share all cache levels, inspecting this limited remote
5829 * state should be fairly cheap.
5830 */
5831void __update_idle_core(struct rq *rq)
5832{
5833 int core = cpu_of(rq);
5834 int cpu;
5835
5836 rcu_read_lock();
5837 if (test_idle_cores(core, true))
5838 goto unlock;
5839
5840 for_each_cpu(cpu, cpu_smt_mask(core)) {
5841 if (cpu == core)
5842 continue;
5843
5844 if (!available_idle_cpu(cpu))
5845 goto unlock;
5846 }
5847
5848 set_idle_cores(core, 1);
5849unlock:
5850 rcu_read_unlock();
5851}
5852
5853/*
5854 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5855 * there are no idle cores left in the system; tracked through
5856 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5857 */
5858static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5859{
5860 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5861 int core, cpu;
5862
5863 if (!static_branch_likely(&sched_smt_present))
5864 return -1;
5865
5866 if (!test_idle_cores(target, false))
5867 return -1;
5868
5869 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5870
5871 for_each_cpu_wrap(core, cpus, target) {
5872 bool idle = true;
5873
5874 for_each_cpu(cpu, cpu_smt_mask(core)) {
5875 __cpumask_clear_cpu(cpu, cpus);
5876 if (!available_idle_cpu(cpu))
5877 idle = false;
5878 }
5879
5880 if (idle)
5881 return core;
5882 }
5883
5884 /*
5885 * Failed to find an idle core; stop looking for one.
5886 */
5887 set_idle_cores(target, 0);
5888
5889 return -1;
5890}
5891
5892/*
5893 * Scan the local SMT mask for idle CPUs.
5894 */
5895static int select_idle_smt(struct task_struct *p, int target)
5896{
5897 int cpu, si_cpu = -1;
5898
5899 if (!static_branch_likely(&sched_smt_present))
5900 return -1;
5901
5902 for_each_cpu(cpu, cpu_smt_mask(target)) {
5903 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5904 continue;
5905 if (available_idle_cpu(cpu))
5906 return cpu;
5907 if (si_cpu == -1 && sched_idle_cpu(cpu))
5908 si_cpu = cpu;
5909 }
5910
5911 return si_cpu;
5912}
5913
5914#else /* CONFIG_SCHED_SMT */
5915
5916static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5917{
5918 return -1;
5919}
5920
5921static inline int select_idle_smt(struct task_struct *p, int target)
5922{
5923 return -1;
5924}
5925
5926#endif /* CONFIG_SCHED_SMT */
5927
5928/*
5929 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5930 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5931 * average idle time for this rq (as found in rq->avg_idle).
5932 */
5933static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5934{
5935 struct sched_domain *this_sd;
5936 u64 avg_cost, avg_idle;
5937 u64 time, cost;
5938 s64 delta;
5939 int this = smp_processor_id();
5940 int cpu, nr = INT_MAX, si_cpu = -1;
5941
5942 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5943 if (!this_sd)
5944 return -1;
5945
5946 /*
5947 * Due to large variance we need a large fuzz factor; hackbench in
5948 * particularly is sensitive here.
5949 */
5950 avg_idle = this_rq()->avg_idle / 512;
5951 avg_cost = this_sd->avg_scan_cost + 1;
5952
5953 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5954 return -1;
5955
5956 if (sched_feat(SIS_PROP)) {
5957 u64 span_avg = sd->span_weight * avg_idle;
5958 if (span_avg > 4*avg_cost)
5959 nr = div_u64(span_avg, avg_cost);
5960 else
5961 nr = 4;
5962 }
5963
5964 time = cpu_clock(this);
5965
5966 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
5967 if (!--nr)
5968 return si_cpu;
5969 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5970 continue;
5971 if (available_idle_cpu(cpu))
5972 break;
5973 if (si_cpu == -1 && sched_idle_cpu(cpu))
5974 si_cpu = cpu;
5975 }
5976
5977 time = cpu_clock(this) - time;
5978 cost = this_sd->avg_scan_cost;
5979 delta = (s64)(time - cost) / 8;
5980 this_sd->avg_scan_cost += delta;
5981
5982 return cpu;
5983}
5984
5985/*
5986 * Try and locate an idle core/thread in the LLC cache domain.
5987 */
5988static int select_idle_sibling(struct task_struct *p, int prev, int target)
5989{
5990 struct sched_domain *sd;
5991 int i, recent_used_cpu;
5992
5993 if (available_idle_cpu(target) || sched_idle_cpu(target))
5994 return target;
5995
5996 /*
5997 * If the previous CPU is cache affine and idle, don't be stupid:
5998 */
5999 if (prev != target && cpus_share_cache(prev, target) &&
6000 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
6001 return prev;
6002
6003 /* Check a recently used CPU as a potential idle candidate: */
6004 recent_used_cpu = p->recent_used_cpu;
6005 if (recent_used_cpu != prev &&
6006 recent_used_cpu != target &&
6007 cpus_share_cache(recent_used_cpu, target) &&
6008 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6009 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6010 /*
6011 * Replace recent_used_cpu with prev as it is a potential
6012 * candidate for the next wake:
6013 */
6014 p->recent_used_cpu = prev;
6015 return recent_used_cpu;
6016 }
6017
6018 sd = rcu_dereference(per_cpu(sd_llc, target));
6019 if (!sd)
6020 return target;
6021
6022 i = select_idle_core(p, sd, target);
6023 if ((unsigned)i < nr_cpumask_bits)
6024 return i;
6025
6026 i = select_idle_cpu(p, sd, target);
6027 if ((unsigned)i < nr_cpumask_bits)
6028 return i;
6029
6030 i = select_idle_smt(p, target);
6031 if ((unsigned)i < nr_cpumask_bits)
6032 return i;
6033
6034 return target;
6035}
6036
6037/**
6038 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6039 * @cpu: the CPU to get the utilization of
6040 *
6041 * The unit of the return value must be the one of capacity so we can compare
6042 * the utilization with the capacity of the CPU that is available for CFS task
6043 * (ie cpu_capacity).
6044 *
6045 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6046 * recent utilization of currently non-runnable tasks on a CPU. It represents
6047 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6048 * capacity_orig is the cpu_capacity available at the highest frequency
6049 * (arch_scale_freq_capacity()).
6050 * The utilization of a CPU converges towards a sum equal to or less than the
6051 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6052 * the running time on this CPU scaled by capacity_curr.
6053 *
6054 * The estimated utilization of a CPU is defined to be the maximum between its
6055 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6056 * currently RUNNABLE on that CPU.
6057 * This allows to properly represent the expected utilization of a CPU which
6058 * has just got a big task running since a long sleep period. At the same time
6059 * however it preserves the benefits of the "blocked utilization" in
6060 * describing the potential for other tasks waking up on the same CPU.
6061 *
6062 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6063 * higher than capacity_orig because of unfortunate rounding in
6064 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6065 * the average stabilizes with the new running time. We need to check that the
6066 * utilization stays within the range of [0..capacity_orig] and cap it if
6067 * necessary. Without utilization capping, a group could be seen as overloaded
6068 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6069 * available capacity. We allow utilization to overshoot capacity_curr (but not
6070 * capacity_orig) as it useful for predicting the capacity required after task
6071 * migrations (scheduler-driven DVFS).
6072 *
6073 * Return: the (estimated) utilization for the specified CPU
6074 */
6075static inline unsigned long cpu_util(int cpu)
6076{
6077 struct cfs_rq *cfs_rq;
6078 unsigned int util;
6079
6080 cfs_rq = &cpu_rq(cpu)->cfs;
6081 util = READ_ONCE(cfs_rq->avg.util_avg);
6082
6083 if (sched_feat(UTIL_EST))
6084 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6085
6086 return min_t(unsigned long, util, capacity_orig_of(cpu));
6087}
6088
6089/*
6090 * cpu_util_without: compute cpu utilization without any contributions from *p
6091 * @cpu: the CPU which utilization is requested
6092 * @p: the task which utilization should be discounted
6093 *
6094 * The utilization of a CPU is defined by the utilization of tasks currently
6095 * enqueued on that CPU as well as tasks which are currently sleeping after an
6096 * execution on that CPU.
6097 *
6098 * This method returns the utilization of the specified CPU by discounting the
6099 * utilization of the specified task, whenever the task is currently
6100 * contributing to the CPU utilization.
6101 */
6102static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6103{
6104 struct cfs_rq *cfs_rq;
6105 unsigned int util;
6106
6107 /* Task has no contribution or is new */
6108 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6109 return cpu_util(cpu);
6110
6111 cfs_rq = &cpu_rq(cpu)->cfs;
6112 util = READ_ONCE(cfs_rq->avg.util_avg);
6113
6114 /* Discount task's util from CPU's util */
6115 lsub_positive(&util, task_util(p));
6116
6117 /*
6118 * Covered cases:
6119 *
6120 * a) if *p is the only task sleeping on this CPU, then:
6121 * cpu_util (== task_util) > util_est (== 0)
6122 * and thus we return:
6123 * cpu_util_without = (cpu_util - task_util) = 0
6124 *
6125 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6126 * IDLE, then:
6127 * cpu_util >= task_util
6128 * cpu_util > util_est (== 0)
6129 * and thus we discount *p's blocked utilization to return:
6130 * cpu_util_without = (cpu_util - task_util) >= 0
6131 *
6132 * c) if other tasks are RUNNABLE on that CPU and
6133 * util_est > cpu_util
6134 * then we use util_est since it returns a more restrictive
6135 * estimation of the spare capacity on that CPU, by just
6136 * considering the expected utilization of tasks already
6137 * runnable on that CPU.
6138 *
6139 * Cases a) and b) are covered by the above code, while case c) is
6140 * covered by the following code when estimated utilization is
6141 * enabled.
6142 */
6143 if (sched_feat(UTIL_EST)) {
6144 unsigned int estimated =
6145 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6146
6147 /*
6148 * Despite the following checks we still have a small window
6149 * for a possible race, when an execl's select_task_rq_fair()
6150 * races with LB's detach_task():
6151 *
6152 * detach_task()
6153 * p->on_rq = TASK_ON_RQ_MIGRATING;
6154 * ---------------------------------- A
6155 * deactivate_task() \
6156 * dequeue_task() + RaceTime
6157 * util_est_dequeue() /
6158 * ---------------------------------- B
6159 *
6160 * The additional check on "current == p" it's required to
6161 * properly fix the execl regression and it helps in further
6162 * reducing the chances for the above race.
6163 */
6164 if (unlikely(task_on_rq_queued(p) || current == p))
6165 lsub_positive(&estimated, _task_util_est(p));
6166
6167 util = max(util, estimated);
6168 }
6169
6170 /*
6171 * Utilization (estimated) can exceed the CPU capacity, thus let's
6172 * clamp to the maximum CPU capacity to ensure consistency with
6173 * the cpu_util call.
6174 */
6175 return min_t(unsigned long, util, capacity_orig_of(cpu));
6176}
6177
6178/*
6179 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6180 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6181 *
6182 * In that case WAKE_AFFINE doesn't make sense and we'll let
6183 * BALANCE_WAKE sort things out.
6184 */
6185static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6186{
6187 long min_cap, max_cap;
6188
6189 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6190 return 0;
6191
6192 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6193 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6194
6195 /* Minimum capacity is close to max, no need to abort wake_affine */
6196 if (max_cap - min_cap < max_cap >> 3)
6197 return 0;
6198
6199 /* Bring task utilization in sync with prev_cpu */
6200 sync_entity_load_avg(&p->se);
6201
6202 return !task_fits_capacity(p, min_cap);
6203}
6204
6205/*
6206 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6207 * to @dst_cpu.
6208 */
6209static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6210{
6211 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6212 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6213
6214 /*
6215 * If @p migrates from @cpu to another, remove its contribution. Or,
6216 * if @p migrates from another CPU to @cpu, add its contribution. In
6217 * the other cases, @cpu is not impacted by the migration, so the
6218 * util_avg should already be correct.
6219 */
6220 if (task_cpu(p) == cpu && dst_cpu != cpu)
6221 sub_positive(&util, task_util(p));
6222 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6223 util += task_util(p);
6224
6225 if (sched_feat(UTIL_EST)) {
6226 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6227
6228 /*
6229 * During wake-up, the task isn't enqueued yet and doesn't
6230 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6231 * so just add it (if needed) to "simulate" what will be
6232 * cpu_util() after the task has been enqueued.
6233 */
6234 if (dst_cpu == cpu)
6235 util_est += _task_util_est(p);
6236
6237 util = max(util, util_est);
6238 }
6239
6240 return min(util, capacity_orig_of(cpu));
6241}
6242
6243/*
6244 * compute_energy(): Estimates the energy that @pd would consume if @p was
6245 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6246 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6247 * to compute what would be the energy if we decided to actually migrate that
6248 * task.
6249 */
6250static long
6251compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6252{
6253 struct cpumask *pd_mask = perf_domain_span(pd);
6254 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6255 unsigned long max_util = 0, sum_util = 0;
6256 int cpu;
6257
6258 /*
6259 * The capacity state of CPUs of the current rd can be driven by CPUs
6260 * of another rd if they belong to the same pd. So, account for the
6261 * utilization of these CPUs too by masking pd with cpu_online_mask
6262 * instead of the rd span.
6263 *
6264 * If an entire pd is outside of the current rd, it will not appear in
6265 * its pd list and will not be accounted by compute_energy().
6266 */
6267 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6268 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6269 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6270
6271 /*
6272 * Busy time computation: utilization clamping is not
6273 * required since the ratio (sum_util / cpu_capacity)
6274 * is already enough to scale the EM reported power
6275 * consumption at the (eventually clamped) cpu_capacity.
6276 */
6277 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6278 ENERGY_UTIL, NULL);
6279
6280 /*
6281 * Performance domain frequency: utilization clamping
6282 * must be considered since it affects the selection
6283 * of the performance domain frequency.
6284 * NOTE: in case RT tasks are running, by default the
6285 * FREQUENCY_UTIL's utilization can be max OPP.
6286 */
6287 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6288 FREQUENCY_UTIL, tsk);
6289 max_util = max(max_util, cpu_util);
6290 }
6291
6292 return em_pd_energy(pd->em_pd, max_util, sum_util);
6293}
6294
6295/*
6296 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6297 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6298 * spare capacity in each performance domain and uses it as a potential
6299 * candidate to execute the task. Then, it uses the Energy Model to figure
6300 * out which of the CPU candidates is the most energy-efficient.
6301 *
6302 * The rationale for this heuristic is as follows. In a performance domain,
6303 * all the most energy efficient CPU candidates (according to the Energy
6304 * Model) are those for which we'll request a low frequency. When there are
6305 * several CPUs for which the frequency request will be the same, we don't
6306 * have enough data to break the tie between them, because the Energy Model
6307 * only includes active power costs. With this model, if we assume that
6308 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6309 * the maximum spare capacity in a performance domain is guaranteed to be among
6310 * the best candidates of the performance domain.
6311 *
6312 * In practice, it could be preferable from an energy standpoint to pack
6313 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6314 * but that could also hurt our chances to go cluster idle, and we have no
6315 * ways to tell with the current Energy Model if this is actually a good
6316 * idea or not. So, find_energy_efficient_cpu() basically favors
6317 * cluster-packing, and spreading inside a cluster. That should at least be
6318 * a good thing for latency, and this is consistent with the idea that most
6319 * of the energy savings of EAS come from the asymmetry of the system, and
6320 * not so much from breaking the tie between identical CPUs. That's also the
6321 * reason why EAS is enabled in the topology code only for systems where
6322 * SD_ASYM_CPUCAPACITY is set.
6323 *
6324 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6325 * they don't have any useful utilization data yet and it's not possible to
6326 * forecast their impact on energy consumption. Consequently, they will be
6327 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6328 * to be energy-inefficient in some use-cases. The alternative would be to
6329 * bias new tasks towards specific types of CPUs first, or to try to infer
6330 * their util_avg from the parent task, but those heuristics could hurt
6331 * other use-cases too. So, until someone finds a better way to solve this,
6332 * let's keep things simple by re-using the existing slow path.
6333 */
6334static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6335{
6336 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6337 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6338 unsigned long cpu_cap, util, base_energy = 0;
6339 int cpu, best_energy_cpu = prev_cpu;
6340 struct sched_domain *sd;
6341 struct perf_domain *pd;
6342
6343 rcu_read_lock();
6344 pd = rcu_dereference(rd->pd);
6345 if (!pd || READ_ONCE(rd->overutilized))
6346 goto fail;
6347
6348 /*
6349 * Energy-aware wake-up happens on the lowest sched_domain starting
6350 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6351 */
6352 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6353 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6354 sd = sd->parent;
6355 if (!sd)
6356 goto fail;
6357
6358 sync_entity_load_avg(&p->se);
6359 if (!task_util_est(p))
6360 goto unlock;
6361
6362 for (; pd; pd = pd->next) {
6363 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6364 unsigned long base_energy_pd;
6365 int max_spare_cap_cpu = -1;
6366
6367 /* Compute the 'base' energy of the pd, without @p */
6368 base_energy_pd = compute_energy(p, -1, pd);
6369 base_energy += base_energy_pd;
6370
6371 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6372 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6373 continue;
6374
6375 /* Skip CPUs that will be overutilized. */
6376 util = cpu_util_next(cpu, p, cpu);
6377 cpu_cap = capacity_of(cpu);
6378 if (!fits_capacity(util, cpu_cap))
6379 continue;
6380
6381 /* Always use prev_cpu as a candidate. */
6382 if (cpu == prev_cpu) {
6383 prev_delta = compute_energy(p, prev_cpu, pd);
6384 prev_delta -= base_energy_pd;
6385 best_delta = min(best_delta, prev_delta);
6386 }
6387
6388 /*
6389 * Find the CPU with the maximum spare capacity in
6390 * the performance domain
6391 */
6392 spare_cap = cpu_cap - util;
6393 if (spare_cap > max_spare_cap) {
6394 max_spare_cap = spare_cap;
6395 max_spare_cap_cpu = cpu;
6396 }
6397 }
6398
6399 /* Evaluate the energy impact of using this CPU. */
6400 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6401 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6402 cur_delta -= base_energy_pd;
6403 if (cur_delta < best_delta) {
6404 best_delta = cur_delta;
6405 best_energy_cpu = max_spare_cap_cpu;
6406 }
6407 }
6408 }
6409unlock:
6410 rcu_read_unlock();
6411
6412 /*
6413 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6414 * least 6% of the energy used by prev_cpu.
6415 */
6416 if (prev_delta == ULONG_MAX)
6417 return best_energy_cpu;
6418
6419 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6420 return best_energy_cpu;
6421
6422 return prev_cpu;
6423
6424fail:
6425 rcu_read_unlock();
6426
6427 return -1;
6428}
6429
6430/*
6431 * select_task_rq_fair: Select target runqueue for the waking task in domains
6432 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6433 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6434 *
6435 * Balances load by selecting the idlest CPU in the idlest group, or under
6436 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6437 *
6438 * Returns the target CPU number.
6439 *
6440 * preempt must be disabled.
6441 */
6442static int
6443select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6444{
6445 struct sched_domain *tmp, *sd = NULL;
6446 int cpu = smp_processor_id();
6447 int new_cpu = prev_cpu;
6448 int want_affine = 0;
6449 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6450
6451 if (sd_flag & SD_BALANCE_WAKE) {
6452 record_wakee(p);
6453
6454 if (sched_energy_enabled()) {
6455 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6456 if (new_cpu >= 0)
6457 return new_cpu;
6458 new_cpu = prev_cpu;
6459 }
6460
6461 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6462 cpumask_test_cpu(cpu, p->cpus_ptr);
6463 }
6464
6465 rcu_read_lock();
6466 for_each_domain(cpu, tmp) {
6467 if (!(tmp->flags & SD_LOAD_BALANCE))
6468 break;
6469
6470 /*
6471 * If both 'cpu' and 'prev_cpu' are part of this domain,
6472 * cpu is a valid SD_WAKE_AFFINE target.
6473 */
6474 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6475 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6476 if (cpu != prev_cpu)
6477 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6478
6479 sd = NULL; /* Prefer wake_affine over balance flags */
6480 break;
6481 }
6482
6483 if (tmp->flags & sd_flag)
6484 sd = tmp;
6485 else if (!want_affine)
6486 break;
6487 }
6488
6489 if (unlikely(sd)) {
6490 /* Slow path */
6491 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6492 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6493 /* Fast path */
6494
6495 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6496
6497 if (want_affine)
6498 current->recent_used_cpu = cpu;
6499 }
6500 rcu_read_unlock();
6501
6502 return new_cpu;
6503}
6504
6505static void detach_entity_cfs_rq(struct sched_entity *se);
6506
6507/*
6508 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6509 * cfs_rq_of(p) references at time of call are still valid and identify the
6510 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6511 */
6512static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6513{
6514 /*
6515 * As blocked tasks retain absolute vruntime the migration needs to
6516 * deal with this by subtracting the old and adding the new
6517 * min_vruntime -- the latter is done by enqueue_entity() when placing
6518 * the task on the new runqueue.
6519 */
6520 if (p->state == TASK_WAKING) {
6521 struct sched_entity *se = &p->se;
6522 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6523 u64 min_vruntime;
6524
6525#ifndef CONFIG_64BIT
6526 u64 min_vruntime_copy;
6527
6528 do {
6529 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6530 smp_rmb();
6531 min_vruntime = cfs_rq->min_vruntime;
6532 } while (min_vruntime != min_vruntime_copy);
6533#else
6534 min_vruntime = cfs_rq->min_vruntime;
6535#endif
6536
6537 se->vruntime -= min_vruntime;
6538 }
6539
6540 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6541 /*
6542 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6543 * rq->lock and can modify state directly.
6544 */
6545 lockdep_assert_held(&task_rq(p)->lock);
6546 detach_entity_cfs_rq(&p->se);
6547
6548 } else {
6549 /*
6550 * We are supposed to update the task to "current" time, then
6551 * its up to date and ready to go to new CPU/cfs_rq. But we
6552 * have difficulty in getting what current time is, so simply
6553 * throw away the out-of-date time. This will result in the
6554 * wakee task is less decayed, but giving the wakee more load
6555 * sounds not bad.
6556 */
6557 remove_entity_load_avg(&p->se);
6558 }
6559
6560 /* Tell new CPU we are migrated */
6561 p->se.avg.last_update_time = 0;
6562
6563 /* We have migrated, no longer consider this task hot */
6564 p->se.exec_start = 0;
6565
6566 update_scan_period(p, new_cpu);
6567}
6568
6569static void task_dead_fair(struct task_struct *p)
6570{
6571 remove_entity_load_avg(&p->se);
6572}
6573
6574static int
6575balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6576{
6577 if (rq->nr_running)
6578 return 1;
6579
6580 return newidle_balance(rq, rf) != 0;
6581}
6582#endif /* CONFIG_SMP */
6583
6584static unsigned long wakeup_gran(struct sched_entity *se)
6585{
6586 unsigned long gran = sysctl_sched_wakeup_granularity;
6587
6588 /*
6589 * Since its curr running now, convert the gran from real-time
6590 * to virtual-time in his units.
6591 *
6592 * By using 'se' instead of 'curr' we penalize light tasks, so
6593 * they get preempted easier. That is, if 'se' < 'curr' then
6594 * the resulting gran will be larger, therefore penalizing the
6595 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6596 * be smaller, again penalizing the lighter task.
6597 *
6598 * This is especially important for buddies when the leftmost
6599 * task is higher priority than the buddy.
6600 */
6601 return calc_delta_fair(gran, se);
6602}
6603
6604/*
6605 * Should 'se' preempt 'curr'.
6606 *
6607 * |s1
6608 * |s2
6609 * |s3
6610 * g
6611 * |<--->|c
6612 *
6613 * w(c, s1) = -1
6614 * w(c, s2) = 0
6615 * w(c, s3) = 1
6616 *
6617 */
6618static int
6619wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6620{
6621 s64 gran, vdiff = curr->vruntime - se->vruntime;
6622
6623 if (vdiff <= 0)
6624 return -1;
6625
6626 gran = wakeup_gran(se);
6627 if (vdiff > gran)
6628 return 1;
6629
6630 return 0;
6631}
6632
6633static void set_last_buddy(struct sched_entity *se)
6634{
6635 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6636 return;
6637
6638 for_each_sched_entity(se) {
6639 if (SCHED_WARN_ON(!se->on_rq))
6640 return;
6641 cfs_rq_of(se)->last = se;
6642 }
6643}
6644
6645static void set_next_buddy(struct sched_entity *se)
6646{
6647 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6648 return;
6649
6650 for_each_sched_entity(se) {
6651 if (SCHED_WARN_ON(!se->on_rq))
6652 return;
6653 cfs_rq_of(se)->next = se;
6654 }
6655}
6656
6657static void set_skip_buddy(struct sched_entity *se)
6658{
6659 for_each_sched_entity(se)
6660 cfs_rq_of(se)->skip = se;
6661}
6662
6663/*
6664 * Preempt the current task with a newly woken task if needed:
6665 */
6666static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6667{
6668 struct task_struct *curr = rq->curr;
6669 struct sched_entity *se = &curr->se, *pse = &p->se;
6670 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6671 int scale = cfs_rq->nr_running >= sched_nr_latency;
6672 int next_buddy_marked = 0;
6673
6674 if (unlikely(se == pse))
6675 return;
6676
6677 /*
6678 * This is possible from callers such as attach_tasks(), in which we
6679 * unconditionally check_prempt_curr() after an enqueue (which may have
6680 * lead to a throttle). This both saves work and prevents false
6681 * next-buddy nomination below.
6682 */
6683 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6684 return;
6685
6686 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6687 set_next_buddy(pse);
6688 next_buddy_marked = 1;
6689 }
6690
6691 /*
6692 * We can come here with TIF_NEED_RESCHED already set from new task
6693 * wake up path.
6694 *
6695 * Note: this also catches the edge-case of curr being in a throttled
6696 * group (e.g. via set_curr_task), since update_curr() (in the
6697 * enqueue of curr) will have resulted in resched being set. This
6698 * prevents us from potentially nominating it as a false LAST_BUDDY
6699 * below.
6700 */
6701 if (test_tsk_need_resched(curr))
6702 return;
6703
6704 /* Idle tasks are by definition preempted by non-idle tasks. */
6705 if (unlikely(task_has_idle_policy(curr)) &&
6706 likely(!task_has_idle_policy(p)))
6707 goto preempt;
6708
6709 /*
6710 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6711 * is driven by the tick):
6712 */
6713 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6714 return;
6715
6716 find_matching_se(&se, &pse);
6717 update_curr(cfs_rq_of(se));
6718 BUG_ON(!pse);
6719 if (wakeup_preempt_entity(se, pse) == 1) {
6720 /*
6721 * Bias pick_next to pick the sched entity that is
6722 * triggering this preemption.
6723 */
6724 if (!next_buddy_marked)
6725 set_next_buddy(pse);
6726 goto preempt;
6727 }
6728
6729 return;
6730
6731preempt:
6732 resched_curr(rq);
6733 /*
6734 * Only set the backward buddy when the current task is still
6735 * on the rq. This can happen when a wakeup gets interleaved
6736 * with schedule on the ->pre_schedule() or idle_balance()
6737 * point, either of which can * drop the rq lock.
6738 *
6739 * Also, during early boot the idle thread is in the fair class,
6740 * for obvious reasons its a bad idea to schedule back to it.
6741 */
6742 if (unlikely(!se->on_rq || curr == rq->idle))
6743 return;
6744
6745 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6746 set_last_buddy(se);
6747}
6748
6749static struct task_struct *
6750pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6751{
6752 struct cfs_rq *cfs_rq = &rq->cfs;
6753 struct sched_entity *se;
6754 struct task_struct *p;
6755 int new_tasks;
6756
6757again:
6758 if (!sched_fair_runnable(rq))
6759 goto idle;
6760
6761#ifdef CONFIG_FAIR_GROUP_SCHED
6762 if (!prev || prev->sched_class != &fair_sched_class)
6763 goto simple;
6764
6765 /*
6766 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6767 * likely that a next task is from the same cgroup as the current.
6768 *
6769 * Therefore attempt to avoid putting and setting the entire cgroup
6770 * hierarchy, only change the part that actually changes.
6771 */
6772
6773 do {
6774 struct sched_entity *curr = cfs_rq->curr;
6775
6776 /*
6777 * Since we got here without doing put_prev_entity() we also
6778 * have to consider cfs_rq->curr. If it is still a runnable
6779 * entity, update_curr() will update its vruntime, otherwise
6780 * forget we've ever seen it.
6781 */
6782 if (curr) {
6783 if (curr->on_rq)
6784 update_curr(cfs_rq);
6785 else
6786 curr = NULL;
6787
6788 /*
6789 * This call to check_cfs_rq_runtime() will do the
6790 * throttle and dequeue its entity in the parent(s).
6791 * Therefore the nr_running test will indeed
6792 * be correct.
6793 */
6794 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6795 cfs_rq = &rq->cfs;
6796
6797 if (!cfs_rq->nr_running)
6798 goto idle;
6799
6800 goto simple;
6801 }
6802 }
6803
6804 se = pick_next_entity(cfs_rq, curr);
6805 cfs_rq = group_cfs_rq(se);
6806 } while (cfs_rq);
6807
6808 p = task_of(se);
6809
6810 /*
6811 * Since we haven't yet done put_prev_entity and if the selected task
6812 * is a different task than we started out with, try and touch the
6813 * least amount of cfs_rqs.
6814 */
6815 if (prev != p) {
6816 struct sched_entity *pse = &prev->se;
6817
6818 while (!(cfs_rq = is_same_group(se, pse))) {
6819 int se_depth = se->depth;
6820 int pse_depth = pse->depth;
6821
6822 if (se_depth <= pse_depth) {
6823 put_prev_entity(cfs_rq_of(pse), pse);
6824 pse = parent_entity(pse);
6825 }
6826 if (se_depth >= pse_depth) {
6827 set_next_entity(cfs_rq_of(se), se);
6828 se = parent_entity(se);
6829 }
6830 }
6831
6832 put_prev_entity(cfs_rq, pse);
6833 set_next_entity(cfs_rq, se);
6834 }
6835
6836 goto done;
6837simple:
6838#endif
6839 if (prev)
6840 put_prev_task(rq, prev);
6841
6842 do {
6843 se = pick_next_entity(cfs_rq, NULL);
6844 set_next_entity(cfs_rq, se);
6845 cfs_rq = group_cfs_rq(se);
6846 } while (cfs_rq);
6847
6848 p = task_of(se);
6849
6850done: __maybe_unused;
6851#ifdef CONFIG_SMP
6852 /*
6853 * Move the next running task to the front of
6854 * the list, so our cfs_tasks list becomes MRU
6855 * one.
6856 */
6857 list_move(&p->se.group_node, &rq->cfs_tasks);
6858#endif
6859
6860 if (hrtick_enabled(rq))
6861 hrtick_start_fair(rq, p);
6862
6863 update_misfit_status(p, rq);
6864
6865 return p;
6866
6867idle:
6868 if (!rf)
6869 return NULL;
6870
6871 new_tasks = newidle_balance(rq, rf);
6872
6873 /*
6874 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
6875 * possible for any higher priority task to appear. In that case we
6876 * must re-start the pick_next_entity() loop.
6877 */
6878 if (new_tasks < 0)
6879 return RETRY_TASK;
6880
6881 if (new_tasks > 0)
6882 goto again;
6883
6884 /*
6885 * rq is about to be idle, check if we need to update the
6886 * lost_idle_time of clock_pelt
6887 */
6888 update_idle_rq_clock_pelt(rq);
6889
6890 return NULL;
6891}
6892
6893/*
6894 * Account for a descheduled task:
6895 */
6896static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6897{
6898 struct sched_entity *se = &prev->se;
6899 struct cfs_rq *cfs_rq;
6900
6901 for_each_sched_entity(se) {
6902 cfs_rq = cfs_rq_of(se);
6903 put_prev_entity(cfs_rq, se);
6904 }
6905}
6906
6907/*
6908 * sched_yield() is very simple
6909 *
6910 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6911 */
6912static void yield_task_fair(struct rq *rq)
6913{
6914 struct task_struct *curr = rq->curr;
6915 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6916 struct sched_entity *se = &curr->se;
6917
6918 /*
6919 * Are we the only task in the tree?
6920 */
6921 if (unlikely(rq->nr_running == 1))
6922 return;
6923
6924 clear_buddies(cfs_rq, se);
6925
6926 if (curr->policy != SCHED_BATCH) {
6927 update_rq_clock(rq);
6928 /*
6929 * Update run-time statistics of the 'current'.
6930 */
6931 update_curr(cfs_rq);
6932 /*
6933 * Tell update_rq_clock() that we've just updated,
6934 * so we don't do microscopic update in schedule()
6935 * and double the fastpath cost.
6936 */
6937 rq_clock_skip_update(rq);
6938 }
6939
6940 set_skip_buddy(se);
6941}
6942
6943static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6944{
6945 struct sched_entity *se = &p->se;
6946
6947 /* throttled hierarchies are not runnable */
6948 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6949 return false;
6950
6951 /* Tell the scheduler that we'd really like pse to run next. */
6952 set_next_buddy(se);
6953
6954 yield_task_fair(rq);
6955
6956 return true;
6957}
6958
6959#ifdef CONFIG_SMP
6960/**************************************************
6961 * Fair scheduling class load-balancing methods.
6962 *
6963 * BASICS
6964 *
6965 * The purpose of load-balancing is to achieve the same basic fairness the
6966 * per-CPU scheduler provides, namely provide a proportional amount of compute
6967 * time to each task. This is expressed in the following equation:
6968 *
6969 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6970 *
6971 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6972 * W_i,0 is defined as:
6973 *
6974 * W_i,0 = \Sum_j w_i,j (2)
6975 *
6976 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6977 * is derived from the nice value as per sched_prio_to_weight[].
6978 *
6979 * The weight average is an exponential decay average of the instantaneous
6980 * weight:
6981 *
6982 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6983 *
6984 * C_i is the compute capacity of CPU i, typically it is the
6985 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6986 * can also include other factors [XXX].
6987 *
6988 * To achieve this balance we define a measure of imbalance which follows
6989 * directly from (1):
6990 *
6991 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6992 *
6993 * We them move tasks around to minimize the imbalance. In the continuous
6994 * function space it is obvious this converges, in the discrete case we get
6995 * a few fun cases generally called infeasible weight scenarios.
6996 *
6997 * [XXX expand on:
6998 * - infeasible weights;
6999 * - local vs global optima in the discrete case. ]
7000 *
7001 *
7002 * SCHED DOMAINS
7003 *
7004 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7005 * for all i,j solution, we create a tree of CPUs that follows the hardware
7006 * topology where each level pairs two lower groups (or better). This results
7007 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7008 * tree to only the first of the previous level and we decrease the frequency
7009 * of load-balance at each level inv. proportional to the number of CPUs in
7010 * the groups.
7011 *
7012 * This yields:
7013 *
7014 * log_2 n 1 n
7015 * \Sum { --- * --- * 2^i } = O(n) (5)
7016 * i = 0 2^i 2^i
7017 * `- size of each group
7018 * | | `- number of CPUs doing load-balance
7019 * | `- freq
7020 * `- sum over all levels
7021 *
7022 * Coupled with a limit on how many tasks we can migrate every balance pass,
7023 * this makes (5) the runtime complexity of the balancer.
7024 *
7025 * An important property here is that each CPU is still (indirectly) connected
7026 * to every other CPU in at most O(log n) steps:
7027 *
7028 * The adjacency matrix of the resulting graph is given by:
7029 *
7030 * log_2 n
7031 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7032 * k = 0
7033 *
7034 * And you'll find that:
7035 *
7036 * A^(log_2 n)_i,j != 0 for all i,j (7)
7037 *
7038 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7039 * The task movement gives a factor of O(m), giving a convergence complexity
7040 * of:
7041 *
7042 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7043 *
7044 *
7045 * WORK CONSERVING
7046 *
7047 * In order to avoid CPUs going idle while there's still work to do, new idle
7048 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7049 * tree itself instead of relying on other CPUs to bring it work.
7050 *
7051 * This adds some complexity to both (5) and (8) but it reduces the total idle
7052 * time.
7053 *
7054 * [XXX more?]
7055 *
7056 *
7057 * CGROUPS
7058 *
7059 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7060 *
7061 * s_k,i
7062 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7063 * S_k
7064 *
7065 * Where
7066 *
7067 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7068 *
7069 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7070 *
7071 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7072 * property.
7073 *
7074 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7075 * rewrite all of this once again.]
7076 */
7077
7078static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7079
7080enum fbq_type { regular, remote, all };
7081
7082enum group_type {
7083 group_other = 0,
7084 group_misfit_task,
7085 group_imbalanced,
7086 group_overloaded,
7087};
7088
7089#define LBF_ALL_PINNED 0x01
7090#define LBF_NEED_BREAK 0x02
7091#define LBF_DST_PINNED 0x04
7092#define LBF_SOME_PINNED 0x08
7093#define LBF_NOHZ_STATS 0x10
7094#define LBF_NOHZ_AGAIN 0x20
7095
7096struct lb_env {
7097 struct sched_domain *sd;
7098
7099 struct rq *src_rq;
7100 int src_cpu;
7101
7102 int dst_cpu;
7103 struct rq *dst_rq;
7104
7105 struct cpumask *dst_grpmask;
7106 int new_dst_cpu;
7107 enum cpu_idle_type idle;
7108 long imbalance;
7109 /* The set of CPUs under consideration for load-balancing */
7110 struct cpumask *cpus;
7111
7112 unsigned int flags;
7113
7114 unsigned int loop;
7115 unsigned int loop_break;
7116 unsigned int loop_max;
7117
7118 enum fbq_type fbq_type;
7119 enum group_type src_grp_type;
7120 struct list_head tasks;
7121};
7122
7123/*
7124 * Is this task likely cache-hot:
7125 */
7126static int task_hot(struct task_struct *p, struct lb_env *env)
7127{
7128 s64 delta;
7129
7130 lockdep_assert_held(&env->src_rq->lock);
7131
7132 if (p->sched_class != &fair_sched_class)
7133 return 0;
7134
7135 if (unlikely(task_has_idle_policy(p)))
7136 return 0;
7137
7138 /*
7139 * Buddy candidates are cache hot:
7140 */
7141 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7142 (&p->se == cfs_rq_of(&p->se)->next ||
7143 &p->se == cfs_rq_of(&p->se)->last))
7144 return 1;
7145
7146 if (sysctl_sched_migration_cost == -1)
7147 return 1;
7148 if (sysctl_sched_migration_cost == 0)
7149 return 0;
7150
7151 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7152
7153 return delta < (s64)sysctl_sched_migration_cost;
7154}
7155
7156#ifdef CONFIG_NUMA_BALANCING
7157/*
7158 * Returns 1, if task migration degrades locality
7159 * Returns 0, if task migration improves locality i.e migration preferred.
7160 * Returns -1, if task migration is not affected by locality.
7161 */
7162static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7163{
7164 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7165 unsigned long src_weight, dst_weight;
7166 int src_nid, dst_nid, dist;
7167
7168 if (!static_branch_likely(&sched_numa_balancing))
7169 return -1;
7170
7171 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7172 return -1;
7173
7174 src_nid = cpu_to_node(env->src_cpu);
7175 dst_nid = cpu_to_node(env->dst_cpu);
7176
7177 if (src_nid == dst_nid)
7178 return -1;
7179
7180 /* Migrating away from the preferred node is always bad. */
7181 if (src_nid == p->numa_preferred_nid) {
7182 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7183 return 1;
7184 else
7185 return -1;
7186 }
7187
7188 /* Encourage migration to the preferred node. */
7189 if (dst_nid == p->numa_preferred_nid)
7190 return 0;
7191
7192 /* Leaving a core idle is often worse than degrading locality. */
7193 if (env->idle == CPU_IDLE)
7194 return -1;
7195
7196 dist = node_distance(src_nid, dst_nid);
7197 if (numa_group) {
7198 src_weight = group_weight(p, src_nid, dist);
7199 dst_weight = group_weight(p, dst_nid, dist);
7200 } else {
7201 src_weight = task_weight(p, src_nid, dist);
7202 dst_weight = task_weight(p, dst_nid, dist);
7203 }
7204
7205 return dst_weight < src_weight;
7206}
7207
7208#else
7209static inline int migrate_degrades_locality(struct task_struct *p,
7210 struct lb_env *env)
7211{
7212 return -1;
7213}
7214#endif
7215
7216/*
7217 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7218 */
7219static
7220int can_migrate_task(struct task_struct *p, struct lb_env *env)
7221{
7222 int tsk_cache_hot;
7223
7224 lockdep_assert_held(&env->src_rq->lock);
7225
7226 /*
7227 * We do not migrate tasks that are:
7228 * 1) throttled_lb_pair, or
7229 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7230 * 3) running (obviously), or
7231 * 4) are cache-hot on their current CPU.
7232 */
7233 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7234 return 0;
7235
7236 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7237 int cpu;
7238
7239 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7240
7241 env->flags |= LBF_SOME_PINNED;
7242
7243 /*
7244 * Remember if this task can be migrated to any other CPU in
7245 * our sched_group. We may want to revisit it if we couldn't
7246 * meet load balance goals by pulling other tasks on src_cpu.
7247 *
7248 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7249 * already computed one in current iteration.
7250 */
7251 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7252 return 0;
7253
7254 /* Prevent to re-select dst_cpu via env's CPUs: */
7255 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7256 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7257 env->flags |= LBF_DST_PINNED;
7258 env->new_dst_cpu = cpu;
7259 break;
7260 }
7261 }
7262
7263 return 0;
7264 }
7265
7266 /* Record that we found atleast one task that could run on dst_cpu */
7267 env->flags &= ~LBF_ALL_PINNED;
7268
7269 if (task_running(env->src_rq, p)) {
7270 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7271 return 0;
7272 }
7273
7274 /*
7275 * Aggressive migration if:
7276 * 1) destination numa is preferred
7277 * 2) task is cache cold, or
7278 * 3) too many balance attempts have failed.
7279 */
7280 tsk_cache_hot = migrate_degrades_locality(p, env);
7281 if (tsk_cache_hot == -1)
7282 tsk_cache_hot = task_hot(p, env);
7283
7284 if (tsk_cache_hot <= 0 ||
7285 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7286 if (tsk_cache_hot == 1) {
7287 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7288 schedstat_inc(p->se.statistics.nr_forced_migrations);
7289 }
7290 return 1;
7291 }
7292
7293 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7294 return 0;
7295}
7296
7297/*
7298 * detach_task() -- detach the task for the migration specified in env
7299 */
7300static void detach_task(struct task_struct *p, struct lb_env *env)
7301{
7302 lockdep_assert_held(&env->src_rq->lock);
7303
7304 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7305 set_task_cpu(p, env->dst_cpu);
7306}
7307
7308/*
7309 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7310 * part of active balancing operations within "domain".
7311 *
7312 * Returns a task if successful and NULL otherwise.
7313 */
7314static struct task_struct *detach_one_task(struct lb_env *env)
7315{
7316 struct task_struct *p;
7317
7318 lockdep_assert_held(&env->src_rq->lock);
7319
7320 list_for_each_entry_reverse(p,
7321 &env->src_rq->cfs_tasks, se.group_node) {
7322 if (!can_migrate_task(p, env))
7323 continue;
7324
7325 detach_task(p, env);
7326
7327 /*
7328 * Right now, this is only the second place where
7329 * lb_gained[env->idle] is updated (other is detach_tasks)
7330 * so we can safely collect stats here rather than
7331 * inside detach_tasks().
7332 */
7333 schedstat_inc(env->sd->lb_gained[env->idle]);
7334 return p;
7335 }
7336 return NULL;
7337}
7338
7339static const unsigned int sched_nr_migrate_break = 32;
7340
7341/*
7342 * detach_tasks() -- tries to detach up to imbalance runnable load from
7343 * busiest_rq, as part of a balancing operation within domain "sd".
7344 *
7345 * Returns number of detached tasks if successful and 0 otherwise.
7346 */
7347static int detach_tasks(struct lb_env *env)
7348{
7349 struct list_head *tasks = &env->src_rq->cfs_tasks;
7350 struct task_struct *p;
7351 unsigned long load;
7352 int detached = 0;
7353
7354 lockdep_assert_held(&env->src_rq->lock);
7355
7356 if (env->imbalance <= 0)
7357 return 0;
7358
7359 while (!list_empty(tasks)) {
7360 /*
7361 * We don't want to steal all, otherwise we may be treated likewise,
7362 * which could at worst lead to a livelock crash.
7363 */
7364 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7365 break;
7366
7367 p = list_last_entry(tasks, struct task_struct, se.group_node);
7368
7369 env->loop++;
7370 /* We've more or less seen every task there is, call it quits */
7371 if (env->loop > env->loop_max)
7372 break;
7373
7374 /* take a breather every nr_migrate tasks */
7375 if (env->loop > env->loop_break) {
7376 env->loop_break += sched_nr_migrate_break;
7377 env->flags |= LBF_NEED_BREAK;
7378 break;
7379 }
7380
7381 if (!can_migrate_task(p, env))
7382 goto next;
7383
7384 load = task_h_load(p);
7385
7386 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7387 goto next;
7388
7389 if ((load / 2) > env->imbalance)
7390 goto next;
7391
7392 detach_task(p, env);
7393 list_add(&p->se.group_node, &env->tasks);
7394
7395 detached++;
7396 env->imbalance -= load;
7397
7398#ifdef CONFIG_PREEMPTION
7399 /*
7400 * NEWIDLE balancing is a source of latency, so preemptible
7401 * kernels will stop after the first task is detached to minimize
7402 * the critical section.
7403 */
7404 if (env->idle == CPU_NEWLY_IDLE)
7405 break;
7406#endif
7407
7408 /*
7409 * We only want to steal up to the prescribed amount of
7410 * runnable load.
7411 */
7412 if (env->imbalance <= 0)
7413 break;
7414
7415 continue;
7416next:
7417 list_move(&p->se.group_node, tasks);
7418 }
7419
7420 /*
7421 * Right now, this is one of only two places we collect this stat
7422 * so we can safely collect detach_one_task() stats here rather
7423 * than inside detach_one_task().
7424 */
7425 schedstat_add(env->sd->lb_gained[env->idle], detached);
7426
7427 return detached;
7428}
7429
7430/*
7431 * attach_task() -- attach the task detached by detach_task() to its new rq.
7432 */
7433static void attach_task(struct rq *rq, struct task_struct *p)
7434{
7435 lockdep_assert_held(&rq->lock);
7436
7437 BUG_ON(task_rq(p) != rq);
7438 activate_task(rq, p, ENQUEUE_NOCLOCK);
7439 check_preempt_curr(rq, p, 0);
7440}
7441
7442/*
7443 * attach_one_task() -- attaches the task returned from detach_one_task() to
7444 * its new rq.
7445 */
7446static void attach_one_task(struct rq *rq, struct task_struct *p)
7447{
7448 struct rq_flags rf;
7449
7450 rq_lock(rq, &rf);
7451 update_rq_clock(rq);
7452 attach_task(rq, p);
7453 rq_unlock(rq, &rf);
7454}
7455
7456/*
7457 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7458 * new rq.
7459 */
7460static void attach_tasks(struct lb_env *env)
7461{
7462 struct list_head *tasks = &env->tasks;
7463 struct task_struct *p;
7464 struct rq_flags rf;
7465
7466 rq_lock(env->dst_rq, &rf);
7467 update_rq_clock(env->dst_rq);
7468
7469 while (!list_empty(tasks)) {
7470 p = list_first_entry(tasks, struct task_struct, se.group_node);
7471 list_del_init(&p->se.group_node);
7472
7473 attach_task(env->dst_rq, p);
7474 }
7475
7476 rq_unlock(env->dst_rq, &rf);
7477}
7478
7479#ifdef CONFIG_NO_HZ_COMMON
7480static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7481{
7482 if (cfs_rq->avg.load_avg)
7483 return true;
7484
7485 if (cfs_rq->avg.util_avg)
7486 return true;
7487
7488 return false;
7489}
7490
7491static inline bool others_have_blocked(struct rq *rq)
7492{
7493 if (READ_ONCE(rq->avg_rt.util_avg))
7494 return true;
7495
7496 if (READ_ONCE(rq->avg_dl.util_avg))
7497 return true;
7498
7499#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7500 if (READ_ONCE(rq->avg_irq.util_avg))
7501 return true;
7502#endif
7503
7504 return false;
7505}
7506
7507static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7508{
7509 rq->last_blocked_load_update_tick = jiffies;
7510
7511 if (!has_blocked)
7512 rq->has_blocked_load = 0;
7513}
7514#else
7515static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7516static inline bool others_have_blocked(struct rq *rq) { return false; }
7517static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7518#endif
7519
7520#ifdef CONFIG_FAIR_GROUP_SCHED
7521
7522static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7523{
7524 if (cfs_rq->load.weight)
7525 return false;
7526
7527 if (cfs_rq->avg.load_sum)
7528 return false;
7529
7530 if (cfs_rq->avg.util_sum)
7531 return false;
7532
7533 if (cfs_rq->avg.runnable_load_sum)
7534 return false;
7535
7536 return true;
7537}
7538
7539static void update_blocked_averages(int cpu)
7540{
7541 struct rq *rq = cpu_rq(cpu);
7542 struct cfs_rq *cfs_rq, *pos;
7543 const struct sched_class *curr_class;
7544 struct rq_flags rf;
7545 bool done = true;
7546
7547 rq_lock_irqsave(rq, &rf);
7548 update_rq_clock(rq);
7549
7550 /*
7551 * update_cfs_rq_load_avg() can call cpufreq_update_util(). Make sure
7552 * that RT, DL and IRQ signals have been updated before updating CFS.
7553 */
7554 curr_class = rq->curr->sched_class;
7555 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7556 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7557 update_irq_load_avg(rq, 0);
7558
7559 /* Don't need periodic decay once load/util_avg are null */
7560 if (others_have_blocked(rq))
7561 done = false;
7562
7563 /*
7564 * Iterates the task_group tree in a bottom up fashion, see
7565 * list_add_leaf_cfs_rq() for details.
7566 */
7567 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7568 struct sched_entity *se;
7569
7570 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7571 update_tg_load_avg(cfs_rq, 0);
7572
7573 /* Propagate pending load changes to the parent, if any: */
7574 se = cfs_rq->tg->se[cpu];
7575 if (se && !skip_blocked_update(se))
7576 update_load_avg(cfs_rq_of(se), se, 0);
7577
7578 /*
7579 * There can be a lot of idle CPU cgroups. Don't let fully
7580 * decayed cfs_rqs linger on the list.
7581 */
7582 if (cfs_rq_is_decayed(cfs_rq))
7583 list_del_leaf_cfs_rq(cfs_rq);
7584
7585 /* Don't need periodic decay once load/util_avg are null */
7586 if (cfs_rq_has_blocked(cfs_rq))
7587 done = false;
7588 }
7589
7590 update_blocked_load_status(rq, !done);
7591 rq_unlock_irqrestore(rq, &rf);
7592}
7593
7594/*
7595 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7596 * This needs to be done in a top-down fashion because the load of a child
7597 * group is a fraction of its parents load.
7598 */
7599static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7600{
7601 struct rq *rq = rq_of(cfs_rq);
7602 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7603 unsigned long now = jiffies;
7604 unsigned long load;
7605
7606 if (cfs_rq->last_h_load_update == now)
7607 return;
7608
7609 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7610 for_each_sched_entity(se) {
7611 cfs_rq = cfs_rq_of(se);
7612 WRITE_ONCE(cfs_rq->h_load_next, se);
7613 if (cfs_rq->last_h_load_update == now)
7614 break;
7615 }
7616
7617 if (!se) {
7618 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7619 cfs_rq->last_h_load_update = now;
7620 }
7621
7622 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7623 load = cfs_rq->h_load;
7624 load = div64_ul(load * se->avg.load_avg,
7625 cfs_rq_load_avg(cfs_rq) + 1);
7626 cfs_rq = group_cfs_rq(se);
7627 cfs_rq->h_load = load;
7628 cfs_rq->last_h_load_update = now;
7629 }
7630}
7631
7632static unsigned long task_h_load(struct task_struct *p)
7633{
7634 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7635
7636 update_cfs_rq_h_load(cfs_rq);
7637 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7638 cfs_rq_load_avg(cfs_rq) + 1);
7639}
7640#else
7641static inline void update_blocked_averages(int cpu)
7642{
7643 struct rq *rq = cpu_rq(cpu);
7644 struct cfs_rq *cfs_rq = &rq->cfs;
7645 const struct sched_class *curr_class;
7646 struct rq_flags rf;
7647
7648 rq_lock_irqsave(rq, &rf);
7649 update_rq_clock(rq);
7650
7651 /*
7652 * update_cfs_rq_load_avg() can call cpufreq_update_util(). Make sure
7653 * that RT, DL and IRQ signals have been updated before updating CFS.
7654 */
7655 curr_class = rq->curr->sched_class;
7656 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7657 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7658 update_irq_load_avg(rq, 0);
7659
7660 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7661
7662 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7663 rq_unlock_irqrestore(rq, &rf);
7664}
7665
7666static unsigned long task_h_load(struct task_struct *p)
7667{
7668 return p->se.avg.load_avg;
7669}
7670#endif
7671
7672/********** Helpers for find_busiest_group ************************/
7673
7674/*
7675 * sg_lb_stats - stats of a sched_group required for load_balancing
7676 */
7677struct sg_lb_stats {
7678 unsigned long avg_load; /*Avg load across the CPUs of the group */
7679 unsigned long group_load; /* Total load over the CPUs of the group */
7680 unsigned long load_per_task;
7681 unsigned long group_capacity;
7682 unsigned long group_util; /* Total utilization of the group */
7683 unsigned int sum_nr_running; /* Nr tasks running in the group */
7684 unsigned int idle_cpus;
7685 unsigned int group_weight;
7686 enum group_type group_type;
7687 int group_no_capacity;
7688 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7689#ifdef CONFIG_NUMA_BALANCING
7690 unsigned int nr_numa_running;
7691 unsigned int nr_preferred_running;
7692#endif
7693};
7694
7695/*
7696 * sd_lb_stats - Structure to store the statistics of a sched_domain
7697 * during load balancing.
7698 */
7699struct sd_lb_stats {
7700 struct sched_group *busiest; /* Busiest group in this sd */
7701 struct sched_group *local; /* Local group in this sd */
7702 unsigned long total_running;
7703 unsigned long total_load; /* Total load of all groups in sd */
7704 unsigned long total_capacity; /* Total capacity of all groups in sd */
7705 unsigned long avg_load; /* Average load across all groups in sd */
7706
7707 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7708 struct sg_lb_stats local_stat; /* Statistics of the local group */
7709};
7710
7711static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7712{
7713 /*
7714 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7715 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7716 * We must however clear busiest_stat::avg_load because
7717 * update_sd_pick_busiest() reads this before assignment.
7718 */
7719 *sds = (struct sd_lb_stats){
7720 .busiest = NULL,
7721 .local = NULL,
7722 .total_running = 0UL,
7723 .total_load = 0UL,
7724 .total_capacity = 0UL,
7725 .busiest_stat = {
7726 .avg_load = 0UL,
7727 .sum_nr_running = 0,
7728 .group_type = group_other,
7729 },
7730 };
7731}
7732
7733static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7734{
7735 struct rq *rq = cpu_rq(cpu);
7736 unsigned long max = arch_scale_cpu_capacity(cpu);
7737 unsigned long used, free;
7738 unsigned long irq;
7739
7740 irq = cpu_util_irq(rq);
7741
7742 if (unlikely(irq >= max))
7743 return 1;
7744
7745 used = READ_ONCE(rq->avg_rt.util_avg);
7746 used += READ_ONCE(rq->avg_dl.util_avg);
7747
7748 if (unlikely(used >= max))
7749 return 1;
7750
7751 free = max - used;
7752
7753 return scale_irq_capacity(free, irq, max);
7754}
7755
7756static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7757{
7758 unsigned long capacity = scale_rt_capacity(sd, cpu);
7759 struct sched_group *sdg = sd->groups;
7760
7761 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7762
7763 if (!capacity)
7764 capacity = 1;
7765
7766 cpu_rq(cpu)->cpu_capacity = capacity;
7767 sdg->sgc->capacity = capacity;
7768 sdg->sgc->min_capacity = capacity;
7769 sdg->sgc->max_capacity = capacity;
7770}
7771
7772void update_group_capacity(struct sched_domain *sd, int cpu)
7773{
7774 struct sched_domain *child = sd->child;
7775 struct sched_group *group, *sdg = sd->groups;
7776 unsigned long capacity, min_capacity, max_capacity;
7777 unsigned long interval;
7778
7779 interval = msecs_to_jiffies(sd->balance_interval);
7780 interval = clamp(interval, 1UL, max_load_balance_interval);
7781 sdg->sgc->next_update = jiffies + interval;
7782
7783 if (!child) {
7784 update_cpu_capacity(sd, cpu);
7785 return;
7786 }
7787
7788 capacity = 0;
7789 min_capacity = ULONG_MAX;
7790 max_capacity = 0;
7791
7792 if (child->flags & SD_OVERLAP) {
7793 /*
7794 * SD_OVERLAP domains cannot assume that child groups
7795 * span the current group.
7796 */
7797
7798 for_each_cpu(cpu, sched_group_span(sdg)) {
7799 struct sched_group_capacity *sgc;
7800 struct rq *rq = cpu_rq(cpu);
7801
7802 /*
7803 * build_sched_domains() -> init_sched_groups_capacity()
7804 * gets here before we've attached the domains to the
7805 * runqueues.
7806 *
7807 * Use capacity_of(), which is set irrespective of domains
7808 * in update_cpu_capacity().
7809 *
7810 * This avoids capacity from being 0 and
7811 * causing divide-by-zero issues on boot.
7812 */
7813 if (unlikely(!rq->sd)) {
7814 capacity += capacity_of(cpu);
7815 } else {
7816 sgc = rq->sd->groups->sgc;
7817 capacity += sgc->capacity;
7818 }
7819
7820 min_capacity = min(capacity, min_capacity);
7821 max_capacity = max(capacity, max_capacity);
7822 }
7823 } else {
7824 /*
7825 * !SD_OVERLAP domains can assume that child groups
7826 * span the current group.
7827 */
7828
7829 group = child->groups;
7830 do {
7831 struct sched_group_capacity *sgc = group->sgc;
7832
7833 capacity += sgc->capacity;
7834 min_capacity = min(sgc->min_capacity, min_capacity);
7835 max_capacity = max(sgc->max_capacity, max_capacity);
7836 group = group->next;
7837 } while (group != child->groups);
7838 }
7839
7840 sdg->sgc->capacity = capacity;
7841 sdg->sgc->min_capacity = min_capacity;
7842 sdg->sgc->max_capacity = max_capacity;
7843}
7844
7845/*
7846 * Check whether the capacity of the rq has been noticeably reduced by side
7847 * activity. The imbalance_pct is used for the threshold.
7848 * Return true is the capacity is reduced
7849 */
7850static inline int
7851check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7852{
7853 return ((rq->cpu_capacity * sd->imbalance_pct) <
7854 (rq->cpu_capacity_orig * 100));
7855}
7856
7857/*
7858 * Check whether a rq has a misfit task and if it looks like we can actually
7859 * help that task: we can migrate the task to a CPU of higher capacity, or
7860 * the task's current CPU is heavily pressured.
7861 */
7862static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7863{
7864 return rq->misfit_task_load &&
7865 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7866 check_cpu_capacity(rq, sd));
7867}
7868
7869/*
7870 * Group imbalance indicates (and tries to solve) the problem where balancing
7871 * groups is inadequate due to ->cpus_ptr constraints.
7872 *
7873 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7874 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7875 * Something like:
7876 *
7877 * { 0 1 2 3 } { 4 5 6 7 }
7878 * * * * *
7879 *
7880 * If we were to balance group-wise we'd place two tasks in the first group and
7881 * two tasks in the second group. Clearly this is undesired as it will overload
7882 * cpu 3 and leave one of the CPUs in the second group unused.
7883 *
7884 * The current solution to this issue is detecting the skew in the first group
7885 * by noticing the lower domain failed to reach balance and had difficulty
7886 * moving tasks due to affinity constraints.
7887 *
7888 * When this is so detected; this group becomes a candidate for busiest; see
7889 * update_sd_pick_busiest(). And calculate_imbalance() and
7890 * find_busiest_group() avoid some of the usual balance conditions to allow it
7891 * to create an effective group imbalance.
7892 *
7893 * This is a somewhat tricky proposition since the next run might not find the
7894 * group imbalance and decide the groups need to be balanced again. A most
7895 * subtle and fragile situation.
7896 */
7897
7898static inline int sg_imbalanced(struct sched_group *group)
7899{
7900 return group->sgc->imbalance;
7901}
7902
7903/*
7904 * group_has_capacity returns true if the group has spare capacity that could
7905 * be used by some tasks.
7906 * We consider that a group has spare capacity if the * number of task is
7907 * smaller than the number of CPUs or if the utilization is lower than the
7908 * available capacity for CFS tasks.
7909 * For the latter, we use a threshold to stabilize the state, to take into
7910 * account the variance of the tasks' load and to return true if the available
7911 * capacity in meaningful for the load balancer.
7912 * As an example, an available capacity of 1% can appear but it doesn't make
7913 * any benefit for the load balance.
7914 */
7915static inline bool
7916group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7917{
7918 if (sgs->sum_nr_running < sgs->group_weight)
7919 return true;
7920
7921 if ((sgs->group_capacity * 100) >
7922 (sgs->group_util * env->sd->imbalance_pct))
7923 return true;
7924
7925 return false;
7926}
7927
7928/*
7929 * group_is_overloaded returns true if the group has more tasks than it can
7930 * handle.
7931 * group_is_overloaded is not equals to !group_has_capacity because a group
7932 * with the exact right number of tasks, has no more spare capacity but is not
7933 * overloaded so both group_has_capacity and group_is_overloaded return
7934 * false.
7935 */
7936static inline bool
7937group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7938{
7939 if (sgs->sum_nr_running <= sgs->group_weight)
7940 return false;
7941
7942 if ((sgs->group_capacity * 100) <
7943 (sgs->group_util * env->sd->imbalance_pct))
7944 return true;
7945
7946 return false;
7947}
7948
7949/*
7950 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
7951 * per-CPU capacity than sched_group ref.
7952 */
7953static inline bool
7954group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7955{
7956 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
7957}
7958
7959/*
7960 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
7961 * per-CPU capacity_orig than sched_group ref.
7962 */
7963static inline bool
7964group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7965{
7966 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
7967}
7968
7969static inline enum
7970group_type group_classify(struct sched_group *group,
7971 struct sg_lb_stats *sgs)
7972{
7973 if (sgs->group_no_capacity)
7974 return group_overloaded;
7975
7976 if (sg_imbalanced(group))
7977 return group_imbalanced;
7978
7979 if (sgs->group_misfit_task_load)
7980 return group_misfit_task;
7981
7982 return group_other;
7983}
7984
7985static bool update_nohz_stats(struct rq *rq, bool force)
7986{
7987#ifdef CONFIG_NO_HZ_COMMON
7988 unsigned int cpu = rq->cpu;
7989
7990 if (!rq->has_blocked_load)
7991 return false;
7992
7993 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7994 return false;
7995
7996 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7997 return true;
7998
7999 update_blocked_averages(cpu);
8000
8001 return rq->has_blocked_load;
8002#else
8003 return false;
8004#endif
8005}
8006
8007/**
8008 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8009 * @env: The load balancing environment.
8010 * @group: sched_group whose statistics are to be updated.
8011 * @sgs: variable to hold the statistics for this group.
8012 * @sg_status: Holds flag indicating the status of the sched_group
8013 */
8014static inline void update_sg_lb_stats(struct lb_env *env,
8015 struct sched_group *group,
8016 struct sg_lb_stats *sgs,
8017 int *sg_status)
8018{
8019 int i, nr_running;
8020
8021 memset(sgs, 0, sizeof(*sgs));
8022
8023 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8024 struct rq *rq = cpu_rq(i);
8025
8026 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8027 env->flags |= LBF_NOHZ_AGAIN;
8028
8029 sgs->group_load += cpu_runnable_load(rq);
8030 sgs->group_util += cpu_util(i);
8031 sgs->sum_nr_running += rq->cfs.h_nr_running;
8032
8033 nr_running = rq->nr_running;
8034 if (nr_running > 1)
8035 *sg_status |= SG_OVERLOAD;
8036
8037 if (cpu_overutilized(i))
8038 *sg_status |= SG_OVERUTILIZED;
8039
8040#ifdef CONFIG_NUMA_BALANCING
8041 sgs->nr_numa_running += rq->nr_numa_running;
8042 sgs->nr_preferred_running += rq->nr_preferred_running;
8043#endif
8044 /*
8045 * No need to call idle_cpu() if nr_running is not 0
8046 */
8047 if (!nr_running && idle_cpu(i))
8048 sgs->idle_cpus++;
8049
8050 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8051 sgs->group_misfit_task_load < rq->misfit_task_load) {
8052 sgs->group_misfit_task_load = rq->misfit_task_load;
8053 *sg_status |= SG_OVERLOAD;
8054 }
8055 }
8056
8057 /* Adjust by relative CPU capacity of the group */
8058 sgs->group_capacity = group->sgc->capacity;
8059 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8060
8061 if (sgs->sum_nr_running)
8062 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8063
8064 sgs->group_weight = group->group_weight;
8065
8066 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8067 sgs->group_type = group_classify(group, sgs);
8068}
8069
8070/**
8071 * update_sd_pick_busiest - return 1 on busiest group
8072 * @env: The load balancing environment.
8073 * @sds: sched_domain statistics
8074 * @sg: sched_group candidate to be checked for being the busiest
8075 * @sgs: sched_group statistics
8076 *
8077 * Determine if @sg is a busier group than the previously selected
8078 * busiest group.
8079 *
8080 * Return: %true if @sg is a busier group than the previously selected
8081 * busiest group. %false otherwise.
8082 */
8083static bool update_sd_pick_busiest(struct lb_env *env,
8084 struct sd_lb_stats *sds,
8085 struct sched_group *sg,
8086 struct sg_lb_stats *sgs)
8087{
8088 struct sg_lb_stats *busiest = &sds->busiest_stat;
8089
8090 /*
8091 * Don't try to pull misfit tasks we can't help.
8092 * We can use max_capacity here as reduction in capacity on some
8093 * CPUs in the group should either be possible to resolve
8094 * internally or be covered by avg_load imbalance (eventually).
8095 */
8096 if (sgs->group_type == group_misfit_task &&
8097 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8098 !group_has_capacity(env, &sds->local_stat)))
8099 return false;
8100
8101 if (sgs->group_type > busiest->group_type)
8102 return true;
8103
8104 if (sgs->group_type < busiest->group_type)
8105 return false;
8106
8107 if (sgs->avg_load <= busiest->avg_load)
8108 return false;
8109
8110 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8111 goto asym_packing;
8112
8113 /*
8114 * Candidate sg has no more than one task per CPU and
8115 * has higher per-CPU capacity. Migrating tasks to less
8116 * capable CPUs may harm throughput. Maximize throughput,
8117 * power/energy consequences are not considered.
8118 */
8119 if (sgs->sum_nr_running <= sgs->group_weight &&
8120 group_smaller_min_cpu_capacity(sds->local, sg))
8121 return false;
8122
8123 /*
8124 * If we have more than one misfit sg go with the biggest misfit.
8125 */
8126 if (sgs->group_type == group_misfit_task &&
8127 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8128 return false;
8129
8130asym_packing:
8131 /* This is the busiest node in its class. */
8132 if (!(env->sd->flags & SD_ASYM_PACKING))
8133 return true;
8134
8135 /* No ASYM_PACKING if target CPU is already busy */
8136 if (env->idle == CPU_NOT_IDLE)
8137 return true;
8138 /*
8139 * ASYM_PACKING needs to move all the work to the highest
8140 * prority CPUs in the group, therefore mark all groups
8141 * of lower priority than ourself as busy.
8142 */
8143 if (sgs->sum_nr_running &&
8144 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8145 if (!sds->busiest)
8146 return true;
8147
8148 /* Prefer to move from lowest priority CPU's work */
8149 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8150 sg->asym_prefer_cpu))
8151 return true;
8152 }
8153
8154 return false;
8155}
8156
8157#ifdef CONFIG_NUMA_BALANCING
8158static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8159{
8160 if (sgs->sum_nr_running > sgs->nr_numa_running)
8161 return regular;
8162 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8163 return remote;
8164 return all;
8165}
8166
8167static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8168{
8169 if (rq->nr_running > rq->nr_numa_running)
8170 return regular;
8171 if (rq->nr_running > rq->nr_preferred_running)
8172 return remote;
8173 return all;
8174}
8175#else
8176static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8177{
8178 return all;
8179}
8180
8181static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8182{
8183 return regular;
8184}
8185#endif /* CONFIG_NUMA_BALANCING */
8186
8187/**
8188 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8189 * @env: The load balancing environment.
8190 * @sds: variable to hold the statistics for this sched_domain.
8191 */
8192static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8193{
8194 struct sched_domain *child = env->sd->child;
8195 struct sched_group *sg = env->sd->groups;
8196 struct sg_lb_stats *local = &sds->local_stat;
8197 struct sg_lb_stats tmp_sgs;
8198 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8199 int sg_status = 0;
8200
8201#ifdef CONFIG_NO_HZ_COMMON
8202 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8203 env->flags |= LBF_NOHZ_STATS;
8204#endif
8205
8206 do {
8207 struct sg_lb_stats *sgs = &tmp_sgs;
8208 int local_group;
8209
8210 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8211 if (local_group) {
8212 sds->local = sg;
8213 sgs = local;
8214
8215 if (env->idle != CPU_NEWLY_IDLE ||
8216 time_after_eq(jiffies, sg->sgc->next_update))
8217 update_group_capacity(env->sd, env->dst_cpu);
8218 }
8219
8220 update_sg_lb_stats(env, sg, sgs, &sg_status);
8221
8222 if (local_group)
8223 goto next_group;
8224
8225 /*
8226 * In case the child domain prefers tasks go to siblings
8227 * first, lower the sg capacity so that we'll try
8228 * and move all the excess tasks away. We lower the capacity
8229 * of a group only if the local group has the capacity to fit
8230 * these excess tasks. The extra check prevents the case where
8231 * you always pull from the heaviest group when it is already
8232 * under-utilized (possible with a large weight task outweighs
8233 * the tasks on the system).
8234 */
8235 if (prefer_sibling && sds->local &&
8236 group_has_capacity(env, local) &&
8237 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8238 sgs->group_no_capacity = 1;
8239 sgs->group_type = group_classify(sg, sgs);
8240 }
8241
8242 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8243 sds->busiest = sg;
8244 sds->busiest_stat = *sgs;
8245 }
8246
8247next_group:
8248 /* Now, start updating sd_lb_stats */
8249 sds->total_running += sgs->sum_nr_running;
8250 sds->total_load += sgs->group_load;
8251 sds->total_capacity += sgs->group_capacity;
8252
8253 sg = sg->next;
8254 } while (sg != env->sd->groups);
8255
8256#ifdef CONFIG_NO_HZ_COMMON
8257 if ((env->flags & LBF_NOHZ_AGAIN) &&
8258 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8259
8260 WRITE_ONCE(nohz.next_blocked,
8261 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8262 }
8263#endif
8264
8265 if (env->sd->flags & SD_NUMA)
8266 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8267
8268 if (!env->sd->parent) {
8269 struct root_domain *rd = env->dst_rq->rd;
8270
8271 /* update overload indicator if we are at root domain */
8272 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8273
8274 /* Update over-utilization (tipping point, U >= 0) indicator */
8275 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8276 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8277 } else if (sg_status & SG_OVERUTILIZED) {
8278 struct root_domain *rd = env->dst_rq->rd;
8279
8280 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8281 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8282 }
8283}
8284
8285/**
8286 * check_asym_packing - Check to see if the group is packed into the
8287 * sched domain.
8288 *
8289 * This is primarily intended to used at the sibling level. Some
8290 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8291 * case of POWER7, it can move to lower SMT modes only when higher
8292 * threads are idle. When in lower SMT modes, the threads will
8293 * perform better since they share less core resources. Hence when we
8294 * have idle threads, we want them to be the higher ones.
8295 *
8296 * This packing function is run on idle threads. It checks to see if
8297 * the busiest CPU in this domain (core in the P7 case) has a higher
8298 * CPU number than the packing function is being run on. Here we are
8299 * assuming lower CPU number will be equivalent to lower a SMT thread
8300 * number.
8301 *
8302 * Return: 1 when packing is required and a task should be moved to
8303 * this CPU. The amount of the imbalance is returned in env->imbalance.
8304 *
8305 * @env: The load balancing environment.
8306 * @sds: Statistics of the sched_domain which is to be packed
8307 */
8308static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8309{
8310 int busiest_cpu;
8311
8312 if (!(env->sd->flags & SD_ASYM_PACKING))
8313 return 0;
8314
8315 if (env->idle == CPU_NOT_IDLE)
8316 return 0;
8317
8318 if (!sds->busiest)
8319 return 0;
8320
8321 busiest_cpu = sds->busiest->asym_prefer_cpu;
8322 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8323 return 0;
8324
8325 env->imbalance = sds->busiest_stat.group_load;
8326
8327 return 1;
8328}
8329
8330/**
8331 * fix_small_imbalance - Calculate the minor imbalance that exists
8332 * amongst the groups of a sched_domain, during
8333 * load balancing.
8334 * @env: The load balancing environment.
8335 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8336 */
8337static inline
8338void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8339{
8340 unsigned long tmp, capa_now = 0, capa_move = 0;
8341 unsigned int imbn = 2;
8342 unsigned long scaled_busy_load_per_task;
8343 struct sg_lb_stats *local, *busiest;
8344
8345 local = &sds->local_stat;
8346 busiest = &sds->busiest_stat;
8347
8348 if (!local->sum_nr_running)
8349 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8350 else if (busiest->load_per_task > local->load_per_task)
8351 imbn = 1;
8352
8353 scaled_busy_load_per_task =
8354 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8355 busiest->group_capacity;
8356
8357 if (busiest->avg_load + scaled_busy_load_per_task >=
8358 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8359 env->imbalance = busiest->load_per_task;
8360 return;
8361 }
8362
8363 /*
8364 * OK, we don't have enough imbalance to justify moving tasks,
8365 * however we may be able to increase total CPU capacity used by
8366 * moving them.
8367 */
8368
8369 capa_now += busiest->group_capacity *
8370 min(busiest->load_per_task, busiest->avg_load);
8371 capa_now += local->group_capacity *
8372 min(local->load_per_task, local->avg_load);
8373 capa_now /= SCHED_CAPACITY_SCALE;
8374
8375 /* Amount of load we'd subtract */
8376 if (busiest->avg_load > scaled_busy_load_per_task) {
8377 capa_move += busiest->group_capacity *
8378 min(busiest->load_per_task,
8379 busiest->avg_load - scaled_busy_load_per_task);
8380 }
8381
8382 /* Amount of load we'd add */
8383 if (busiest->avg_load * busiest->group_capacity <
8384 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8385 tmp = (busiest->avg_load * busiest->group_capacity) /
8386 local->group_capacity;
8387 } else {
8388 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8389 local->group_capacity;
8390 }
8391 capa_move += local->group_capacity *
8392 min(local->load_per_task, local->avg_load + tmp);
8393 capa_move /= SCHED_CAPACITY_SCALE;
8394
8395 /* Move if we gain throughput */
8396 if (capa_move > capa_now)
8397 env->imbalance = busiest->load_per_task;
8398}
8399
8400/**
8401 * calculate_imbalance - Calculate the amount of imbalance present within the
8402 * groups of a given sched_domain during load balance.
8403 * @env: load balance environment
8404 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8405 */
8406static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8407{
8408 unsigned long max_pull, load_above_capacity = ~0UL;
8409 struct sg_lb_stats *local, *busiest;
8410
8411 local = &sds->local_stat;
8412 busiest = &sds->busiest_stat;
8413
8414 if (busiest->group_type == group_imbalanced) {
8415 /*
8416 * In the group_imb case we cannot rely on group-wide averages
8417 * to ensure CPU-load equilibrium, look at wider averages. XXX
8418 */
8419 busiest->load_per_task =
8420 min(busiest->load_per_task, sds->avg_load);
8421 }
8422
8423 /*
8424 * Avg load of busiest sg can be less and avg load of local sg can
8425 * be greater than avg load across all sgs of sd because avg load
8426 * factors in sg capacity and sgs with smaller group_type are
8427 * skipped when updating the busiest sg:
8428 */
8429 if (busiest->group_type != group_misfit_task &&
8430 (busiest->avg_load <= sds->avg_load ||
8431 local->avg_load >= sds->avg_load)) {
8432 env->imbalance = 0;
8433 return fix_small_imbalance(env, sds);
8434 }
8435
8436 /*
8437 * If there aren't any idle CPUs, avoid creating some.
8438 */
8439 if (busiest->group_type == group_overloaded &&
8440 local->group_type == group_overloaded) {
8441 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8442 if (load_above_capacity > busiest->group_capacity) {
8443 load_above_capacity -= busiest->group_capacity;
8444 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8445 load_above_capacity /= busiest->group_capacity;
8446 } else
8447 load_above_capacity = ~0UL;
8448 }
8449
8450 /*
8451 * We're trying to get all the CPUs to the average_load, so we don't
8452 * want to push ourselves above the average load, nor do we wish to
8453 * reduce the max loaded CPU below the average load. At the same time,
8454 * we also don't want to reduce the group load below the group
8455 * capacity. Thus we look for the minimum possible imbalance.
8456 */
8457 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8458
8459 /* How much load to actually move to equalise the imbalance */
8460 env->imbalance = min(
8461 max_pull * busiest->group_capacity,
8462 (sds->avg_load - local->avg_load) * local->group_capacity
8463 ) / SCHED_CAPACITY_SCALE;
8464
8465 /* Boost imbalance to allow misfit task to be balanced. */
8466 if (busiest->group_type == group_misfit_task) {
8467 env->imbalance = max_t(long, env->imbalance,
8468 busiest->group_misfit_task_load);
8469 }
8470
8471 /*
8472 * if *imbalance is less than the average load per runnable task
8473 * there is no guarantee that any tasks will be moved so we'll have
8474 * a think about bumping its value to force at least one task to be
8475 * moved
8476 */
8477 if (env->imbalance < busiest->load_per_task)
8478 return fix_small_imbalance(env, sds);
8479}
8480
8481/******* find_busiest_group() helpers end here *********************/
8482
8483/**
8484 * find_busiest_group - Returns the busiest group within the sched_domain
8485 * if there is an imbalance.
8486 *
8487 * Also calculates the amount of runnable load which should be moved
8488 * to restore balance.
8489 *
8490 * @env: The load balancing environment.
8491 *
8492 * Return: - The busiest group if imbalance exists.
8493 */
8494static struct sched_group *find_busiest_group(struct lb_env *env)
8495{
8496 struct sg_lb_stats *local, *busiest;
8497 struct sd_lb_stats sds;
8498
8499 init_sd_lb_stats(&sds);
8500
8501 /*
8502 * Compute the various statistics relavent for load balancing at
8503 * this level.
8504 */
8505 update_sd_lb_stats(env, &sds);
8506
8507 if (sched_energy_enabled()) {
8508 struct root_domain *rd = env->dst_rq->rd;
8509
8510 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8511 goto out_balanced;
8512 }
8513
8514 local = &sds.local_stat;
8515 busiest = &sds.busiest_stat;
8516
8517 /* ASYM feature bypasses nice load balance check */
8518 if (check_asym_packing(env, &sds))
8519 return sds.busiest;
8520
8521 /* There is no busy sibling group to pull tasks from */
8522 if (!sds.busiest || busiest->sum_nr_running == 0)
8523 goto out_balanced;
8524
8525 /* XXX broken for overlapping NUMA groups */
8526 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8527 / sds.total_capacity;
8528
8529 /*
8530 * If the busiest group is imbalanced the below checks don't
8531 * work because they assume all things are equal, which typically
8532 * isn't true due to cpus_ptr constraints and the like.
8533 */
8534 if (busiest->group_type == group_imbalanced)
8535 goto force_balance;
8536
8537 /*
8538 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8539 * capacities from resulting in underutilization due to avg_load.
8540 */
8541 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8542 busiest->group_no_capacity)
8543 goto force_balance;
8544
8545 /* Misfit tasks should be dealt with regardless of the avg load */
8546 if (busiest->group_type == group_misfit_task)
8547 goto force_balance;
8548
8549 /*
8550 * If the local group is busier than the selected busiest group
8551 * don't try and pull any tasks.
8552 */
8553 if (local->avg_load >= busiest->avg_load)
8554 goto out_balanced;
8555
8556 /*
8557 * Don't pull any tasks if this group is already above the domain
8558 * average load.
8559 */
8560 if (local->avg_load >= sds.avg_load)
8561 goto out_balanced;
8562
8563 if (env->idle == CPU_IDLE) {
8564 /*
8565 * This CPU is idle. If the busiest group is not overloaded
8566 * and there is no imbalance between this and busiest group
8567 * wrt idle CPUs, it is balanced. The imbalance becomes
8568 * significant if the diff is greater than 1 otherwise we
8569 * might end up to just move the imbalance on another group
8570 */
8571 if ((busiest->group_type != group_overloaded) &&
8572 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8573 goto out_balanced;
8574 } else {
8575 /*
8576 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8577 * imbalance_pct to be conservative.
8578 */
8579 if (100 * busiest->avg_load <=
8580 env->sd->imbalance_pct * local->avg_load)
8581 goto out_balanced;
8582 }
8583
8584force_balance:
8585 /* Looks like there is an imbalance. Compute it */
8586 env->src_grp_type = busiest->group_type;
8587 calculate_imbalance(env, &sds);
8588 return env->imbalance ? sds.busiest : NULL;
8589
8590out_balanced:
8591 env->imbalance = 0;
8592 return NULL;
8593}
8594
8595/*
8596 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8597 */
8598static struct rq *find_busiest_queue(struct lb_env *env,
8599 struct sched_group *group)
8600{
8601 struct rq *busiest = NULL, *rq;
8602 unsigned long busiest_load = 0, busiest_capacity = 1;
8603 int i;
8604
8605 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8606 unsigned long capacity, load;
8607 enum fbq_type rt;
8608
8609 rq = cpu_rq(i);
8610 rt = fbq_classify_rq(rq);
8611
8612 /*
8613 * We classify groups/runqueues into three groups:
8614 * - regular: there are !numa tasks
8615 * - remote: there are numa tasks that run on the 'wrong' node
8616 * - all: there is no distinction
8617 *
8618 * In order to avoid migrating ideally placed numa tasks,
8619 * ignore those when there's better options.
8620 *
8621 * If we ignore the actual busiest queue to migrate another
8622 * task, the next balance pass can still reduce the busiest
8623 * queue by moving tasks around inside the node.
8624 *
8625 * If we cannot move enough load due to this classification
8626 * the next pass will adjust the group classification and
8627 * allow migration of more tasks.
8628 *
8629 * Both cases only affect the total convergence complexity.
8630 */
8631 if (rt > env->fbq_type)
8632 continue;
8633
8634 /*
8635 * For ASYM_CPUCAPACITY domains with misfit tasks we simply
8636 * seek the "biggest" misfit task.
8637 */
8638 if (env->src_grp_type == group_misfit_task) {
8639 if (rq->misfit_task_load > busiest_load) {
8640 busiest_load = rq->misfit_task_load;
8641 busiest = rq;
8642 }
8643
8644 continue;
8645 }
8646
8647 capacity = capacity_of(i);
8648
8649 /*
8650 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
8651 * eventually lead to active_balancing high->low capacity.
8652 * Higher per-CPU capacity is considered better than balancing
8653 * average load.
8654 */
8655 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8656 capacity_of(env->dst_cpu) < capacity &&
8657 rq->nr_running == 1)
8658 continue;
8659
8660 load = cpu_runnable_load(rq);
8661
8662 /*
8663 * When comparing with imbalance, use cpu_runnable_load()
8664 * which is not scaled with the CPU capacity.
8665 */
8666
8667 if (rq->nr_running == 1 && load > env->imbalance &&
8668 !check_cpu_capacity(rq, env->sd))
8669 continue;
8670
8671 /*
8672 * For the load comparisons with the other CPU's, consider
8673 * the cpu_runnable_load() scaled with the CPU capacity, so
8674 * that the load can be moved away from the CPU that is
8675 * potentially running at a lower capacity.
8676 *
8677 * Thus we're looking for max(load_i / capacity_i), crosswise
8678 * multiplication to rid ourselves of the division works out
8679 * to: load_i * capacity_j > load_j * capacity_i; where j is
8680 * our previous maximum.
8681 */
8682 if (load * busiest_capacity > busiest_load * capacity) {
8683 busiest_load = load;
8684 busiest_capacity = capacity;
8685 busiest = rq;
8686 }
8687 }
8688
8689 return busiest;
8690}
8691
8692/*
8693 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8694 * so long as it is large enough.
8695 */
8696#define MAX_PINNED_INTERVAL 512
8697
8698static inline bool
8699asym_active_balance(struct lb_env *env)
8700{
8701 /*
8702 * ASYM_PACKING needs to force migrate tasks from busy but
8703 * lower priority CPUs in order to pack all tasks in the
8704 * highest priority CPUs.
8705 */
8706 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8707 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8708}
8709
8710static inline bool
8711voluntary_active_balance(struct lb_env *env)
8712{
8713 struct sched_domain *sd = env->sd;
8714
8715 if (asym_active_balance(env))
8716 return 1;
8717
8718 /*
8719 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8720 * It's worth migrating the task if the src_cpu's capacity is reduced
8721 * because of other sched_class or IRQs if more capacity stays
8722 * available on dst_cpu.
8723 */
8724 if ((env->idle != CPU_NOT_IDLE) &&
8725 (env->src_rq->cfs.h_nr_running == 1)) {
8726 if ((check_cpu_capacity(env->src_rq, sd)) &&
8727 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8728 return 1;
8729 }
8730
8731 if (env->src_grp_type == group_misfit_task)
8732 return 1;
8733
8734 return 0;
8735}
8736
8737static int need_active_balance(struct lb_env *env)
8738{
8739 struct sched_domain *sd = env->sd;
8740
8741 if (voluntary_active_balance(env))
8742 return 1;
8743
8744 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8745}
8746
8747static int active_load_balance_cpu_stop(void *data);
8748
8749static int should_we_balance(struct lb_env *env)
8750{
8751 struct sched_group *sg = env->sd->groups;
8752 int cpu, balance_cpu = -1;
8753
8754 /*
8755 * Ensure the balancing environment is consistent; can happen
8756 * when the softirq triggers 'during' hotplug.
8757 */
8758 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8759 return 0;
8760
8761 /*
8762 * In the newly idle case, we will allow all the CPUs
8763 * to do the newly idle load balance.
8764 */
8765 if (env->idle == CPU_NEWLY_IDLE)
8766 return 1;
8767
8768 /* Try to find first idle CPU */
8769 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8770 if (!idle_cpu(cpu))
8771 continue;
8772
8773 balance_cpu = cpu;
8774 break;
8775 }
8776
8777 if (balance_cpu == -1)
8778 balance_cpu = group_balance_cpu(sg);
8779
8780 /*
8781 * First idle CPU or the first CPU(busiest) in this sched group
8782 * is eligible for doing load balancing at this and above domains.
8783 */
8784 return balance_cpu == env->dst_cpu;
8785}
8786
8787/*
8788 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8789 * tasks if there is an imbalance.
8790 */
8791static int load_balance(int this_cpu, struct rq *this_rq,
8792 struct sched_domain *sd, enum cpu_idle_type idle,
8793 int *continue_balancing)
8794{
8795 int ld_moved, cur_ld_moved, active_balance = 0;
8796 struct sched_domain *sd_parent = sd->parent;
8797 struct sched_group *group;
8798 struct rq *busiest;
8799 struct rq_flags rf;
8800 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8801
8802 struct lb_env env = {
8803 .sd = sd,
8804 .dst_cpu = this_cpu,
8805 .dst_rq = this_rq,
8806 .dst_grpmask = sched_group_span(sd->groups),
8807 .idle = idle,
8808 .loop_break = sched_nr_migrate_break,
8809 .cpus = cpus,
8810 .fbq_type = all,
8811 .tasks = LIST_HEAD_INIT(env.tasks),
8812 };
8813
8814 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8815
8816 schedstat_inc(sd->lb_count[idle]);
8817
8818redo:
8819 if (!should_we_balance(&env)) {
8820 *continue_balancing = 0;
8821 goto out_balanced;
8822 }
8823
8824 group = find_busiest_group(&env);
8825 if (!group) {
8826 schedstat_inc(sd->lb_nobusyg[idle]);
8827 goto out_balanced;
8828 }
8829
8830 busiest = find_busiest_queue(&env, group);
8831 if (!busiest) {
8832 schedstat_inc(sd->lb_nobusyq[idle]);
8833 goto out_balanced;
8834 }
8835
8836 BUG_ON(busiest == env.dst_rq);
8837
8838 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8839
8840 env.src_cpu = busiest->cpu;
8841 env.src_rq = busiest;
8842
8843 ld_moved = 0;
8844 if (busiest->nr_running > 1) {
8845 /*
8846 * Attempt to move tasks. If find_busiest_group has found
8847 * an imbalance but busiest->nr_running <= 1, the group is
8848 * still unbalanced. ld_moved simply stays zero, so it is
8849 * correctly treated as an imbalance.
8850 */
8851 env.flags |= LBF_ALL_PINNED;
8852 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8853
8854more_balance:
8855 rq_lock_irqsave(busiest, &rf);
8856 update_rq_clock(busiest);
8857
8858 /*
8859 * cur_ld_moved - load moved in current iteration
8860 * ld_moved - cumulative load moved across iterations
8861 */
8862 cur_ld_moved = detach_tasks(&env);
8863
8864 /*
8865 * We've detached some tasks from busiest_rq. Every
8866 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8867 * unlock busiest->lock, and we are able to be sure
8868 * that nobody can manipulate the tasks in parallel.
8869 * See task_rq_lock() family for the details.
8870 */
8871
8872 rq_unlock(busiest, &rf);
8873
8874 if (cur_ld_moved) {
8875 attach_tasks(&env);
8876 ld_moved += cur_ld_moved;
8877 }
8878
8879 local_irq_restore(rf.flags);
8880
8881 if (env.flags & LBF_NEED_BREAK) {
8882 env.flags &= ~LBF_NEED_BREAK;
8883 goto more_balance;
8884 }
8885
8886 /*
8887 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8888 * us and move them to an alternate dst_cpu in our sched_group
8889 * where they can run. The upper limit on how many times we
8890 * iterate on same src_cpu is dependent on number of CPUs in our
8891 * sched_group.
8892 *
8893 * This changes load balance semantics a bit on who can move
8894 * load to a given_cpu. In addition to the given_cpu itself
8895 * (or a ilb_cpu acting on its behalf where given_cpu is
8896 * nohz-idle), we now have balance_cpu in a position to move
8897 * load to given_cpu. In rare situations, this may cause
8898 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8899 * _independently_ and at _same_ time to move some load to
8900 * given_cpu) causing exceess load to be moved to given_cpu.
8901 * This however should not happen so much in practice and
8902 * moreover subsequent load balance cycles should correct the
8903 * excess load moved.
8904 */
8905 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8906
8907 /* Prevent to re-select dst_cpu via env's CPUs */
8908 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8909
8910 env.dst_rq = cpu_rq(env.new_dst_cpu);
8911 env.dst_cpu = env.new_dst_cpu;
8912 env.flags &= ~LBF_DST_PINNED;
8913 env.loop = 0;
8914 env.loop_break = sched_nr_migrate_break;
8915
8916 /*
8917 * Go back to "more_balance" rather than "redo" since we
8918 * need to continue with same src_cpu.
8919 */
8920 goto more_balance;
8921 }
8922
8923 /*
8924 * We failed to reach balance because of affinity.
8925 */
8926 if (sd_parent) {
8927 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8928
8929 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8930 *group_imbalance = 1;
8931 }
8932
8933 /* All tasks on this runqueue were pinned by CPU affinity */
8934 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8935 __cpumask_clear_cpu(cpu_of(busiest), cpus);
8936 /*
8937 * Attempting to continue load balancing at the current
8938 * sched_domain level only makes sense if there are
8939 * active CPUs remaining as possible busiest CPUs to
8940 * pull load from which are not contained within the
8941 * destination group that is receiving any migrated
8942 * load.
8943 */
8944 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8945 env.loop = 0;
8946 env.loop_break = sched_nr_migrate_break;
8947 goto redo;
8948 }
8949 goto out_all_pinned;
8950 }
8951 }
8952
8953 if (!ld_moved) {
8954 schedstat_inc(sd->lb_failed[idle]);
8955 /*
8956 * Increment the failure counter only on periodic balance.
8957 * We do not want newidle balance, which can be very
8958 * frequent, pollute the failure counter causing
8959 * excessive cache_hot migrations and active balances.
8960 */
8961 if (idle != CPU_NEWLY_IDLE)
8962 sd->nr_balance_failed++;
8963
8964 if (need_active_balance(&env)) {
8965 unsigned long flags;
8966
8967 raw_spin_lock_irqsave(&busiest->lock, flags);
8968
8969 /*
8970 * Don't kick the active_load_balance_cpu_stop,
8971 * if the curr task on busiest CPU can't be
8972 * moved to this_cpu:
8973 */
8974 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
8975 raw_spin_unlock_irqrestore(&busiest->lock,
8976 flags);
8977 env.flags |= LBF_ALL_PINNED;
8978 goto out_one_pinned;
8979 }
8980
8981 /*
8982 * ->active_balance synchronizes accesses to
8983 * ->active_balance_work. Once set, it's cleared
8984 * only after active load balance is finished.
8985 */
8986 if (!busiest->active_balance) {
8987 busiest->active_balance = 1;
8988 busiest->push_cpu = this_cpu;
8989 active_balance = 1;
8990 }
8991 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8992
8993 if (active_balance) {
8994 stop_one_cpu_nowait(cpu_of(busiest),
8995 active_load_balance_cpu_stop, busiest,
8996 &busiest->active_balance_work);
8997 }
8998
8999 /* We've kicked active balancing, force task migration. */
9000 sd->nr_balance_failed = sd->cache_nice_tries+1;
9001 }
9002 } else
9003 sd->nr_balance_failed = 0;
9004
9005 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9006 /* We were unbalanced, so reset the balancing interval */
9007 sd->balance_interval = sd->min_interval;
9008 } else {
9009 /*
9010 * If we've begun active balancing, start to back off. This
9011 * case may not be covered by the all_pinned logic if there
9012 * is only 1 task on the busy runqueue (because we don't call
9013 * detach_tasks).
9014 */
9015 if (sd->balance_interval < sd->max_interval)
9016 sd->balance_interval *= 2;
9017 }
9018
9019 goto out;
9020
9021out_balanced:
9022 /*
9023 * We reach balance although we may have faced some affinity
9024 * constraints. Clear the imbalance flag only if other tasks got
9025 * a chance to move and fix the imbalance.
9026 */
9027 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9028 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9029
9030 if (*group_imbalance)
9031 *group_imbalance = 0;
9032 }
9033
9034out_all_pinned:
9035 /*
9036 * We reach balance because all tasks are pinned at this level so
9037 * we can't migrate them. Let the imbalance flag set so parent level
9038 * can try to migrate them.
9039 */
9040 schedstat_inc(sd->lb_balanced[idle]);
9041
9042 sd->nr_balance_failed = 0;
9043
9044out_one_pinned:
9045 ld_moved = 0;
9046
9047 /*
9048 * newidle_balance() disregards balance intervals, so we could
9049 * repeatedly reach this code, which would lead to balance_interval
9050 * skyrocketting in a short amount of time. Skip the balance_interval
9051 * increase logic to avoid that.
9052 */
9053 if (env.idle == CPU_NEWLY_IDLE)
9054 goto out;
9055
9056 /* tune up the balancing interval */
9057 if ((env.flags & LBF_ALL_PINNED &&
9058 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9059 sd->balance_interval < sd->max_interval)
9060 sd->balance_interval *= 2;
9061out:
9062 return ld_moved;
9063}
9064
9065static inline unsigned long
9066get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9067{
9068 unsigned long interval = sd->balance_interval;
9069
9070 if (cpu_busy)
9071 interval *= sd->busy_factor;
9072
9073 /* scale ms to jiffies */
9074 interval = msecs_to_jiffies(interval);
9075 interval = clamp(interval, 1UL, max_load_balance_interval);
9076
9077 return interval;
9078}
9079
9080static inline void
9081update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9082{
9083 unsigned long interval, next;
9084
9085 /* used by idle balance, so cpu_busy = 0 */
9086 interval = get_sd_balance_interval(sd, 0);
9087 next = sd->last_balance + interval;
9088
9089 if (time_after(*next_balance, next))
9090 *next_balance = next;
9091}
9092
9093/*
9094 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9095 * running tasks off the busiest CPU onto idle CPUs. It requires at
9096 * least 1 task to be running on each physical CPU where possible, and
9097 * avoids physical / logical imbalances.
9098 */
9099static int active_load_balance_cpu_stop(void *data)
9100{
9101 struct rq *busiest_rq = data;
9102 int busiest_cpu = cpu_of(busiest_rq);
9103 int target_cpu = busiest_rq->push_cpu;
9104 struct rq *target_rq = cpu_rq(target_cpu);
9105 struct sched_domain *sd;
9106 struct task_struct *p = NULL;
9107 struct rq_flags rf;
9108
9109 rq_lock_irq(busiest_rq, &rf);
9110 /*
9111 * Between queueing the stop-work and running it is a hole in which
9112 * CPUs can become inactive. We should not move tasks from or to
9113 * inactive CPUs.
9114 */
9115 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9116 goto out_unlock;
9117
9118 /* Make sure the requested CPU hasn't gone down in the meantime: */
9119 if (unlikely(busiest_cpu != smp_processor_id() ||
9120 !busiest_rq->active_balance))
9121 goto out_unlock;
9122
9123 /* Is there any task to move? */
9124 if (busiest_rq->nr_running <= 1)
9125 goto out_unlock;
9126
9127 /*
9128 * This condition is "impossible", if it occurs
9129 * we need to fix it. Originally reported by
9130 * Bjorn Helgaas on a 128-CPU setup.
9131 */
9132 BUG_ON(busiest_rq == target_rq);
9133
9134 /* Search for an sd spanning us and the target CPU. */
9135 rcu_read_lock();
9136 for_each_domain(target_cpu, sd) {
9137 if ((sd->flags & SD_LOAD_BALANCE) &&
9138 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9139 break;
9140 }
9141
9142 if (likely(sd)) {
9143 struct lb_env env = {
9144 .sd = sd,
9145 .dst_cpu = target_cpu,
9146 .dst_rq = target_rq,
9147 .src_cpu = busiest_rq->cpu,
9148 .src_rq = busiest_rq,
9149 .idle = CPU_IDLE,
9150 /*
9151 * can_migrate_task() doesn't need to compute new_dst_cpu
9152 * for active balancing. Since we have CPU_IDLE, but no
9153 * @dst_grpmask we need to make that test go away with lying
9154 * about DST_PINNED.
9155 */
9156 .flags = LBF_DST_PINNED,
9157 };
9158
9159 schedstat_inc(sd->alb_count);
9160 update_rq_clock(busiest_rq);
9161
9162 p = detach_one_task(&env);
9163 if (p) {
9164 schedstat_inc(sd->alb_pushed);
9165 /* Active balancing done, reset the failure counter. */
9166 sd->nr_balance_failed = 0;
9167 } else {
9168 schedstat_inc(sd->alb_failed);
9169 }
9170 }
9171 rcu_read_unlock();
9172out_unlock:
9173 busiest_rq->active_balance = 0;
9174 rq_unlock(busiest_rq, &rf);
9175
9176 if (p)
9177 attach_one_task(target_rq, p);
9178
9179 local_irq_enable();
9180
9181 return 0;
9182}
9183
9184static DEFINE_SPINLOCK(balancing);
9185
9186/*
9187 * Scale the max load_balance interval with the number of CPUs in the system.
9188 * This trades load-balance latency on larger machines for less cross talk.
9189 */
9190void update_max_interval(void)
9191{
9192 max_load_balance_interval = HZ*num_online_cpus()/10;
9193}
9194
9195/*
9196 * It checks each scheduling domain to see if it is due to be balanced,
9197 * and initiates a balancing operation if so.
9198 *
9199 * Balancing parameters are set up in init_sched_domains.
9200 */
9201static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9202{
9203 int continue_balancing = 1;
9204 int cpu = rq->cpu;
9205 unsigned long interval;
9206 struct sched_domain *sd;
9207 /* Earliest time when we have to do rebalance again */
9208 unsigned long next_balance = jiffies + 60*HZ;
9209 int update_next_balance = 0;
9210 int need_serialize, need_decay = 0;
9211 u64 max_cost = 0;
9212
9213 rcu_read_lock();
9214 for_each_domain(cpu, sd) {
9215 /*
9216 * Decay the newidle max times here because this is a regular
9217 * visit to all the domains. Decay ~1% per second.
9218 */
9219 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9220 sd->max_newidle_lb_cost =
9221 (sd->max_newidle_lb_cost * 253) / 256;
9222 sd->next_decay_max_lb_cost = jiffies + HZ;
9223 need_decay = 1;
9224 }
9225 max_cost += sd->max_newidle_lb_cost;
9226
9227 if (!(sd->flags & SD_LOAD_BALANCE))
9228 continue;
9229
9230 /*
9231 * Stop the load balance at this level. There is another
9232 * CPU in our sched group which is doing load balancing more
9233 * actively.
9234 */
9235 if (!continue_balancing) {
9236 if (need_decay)
9237 continue;
9238 break;
9239 }
9240
9241 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9242
9243 need_serialize = sd->flags & SD_SERIALIZE;
9244 if (need_serialize) {
9245 if (!spin_trylock(&balancing))
9246 goto out;
9247 }
9248
9249 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9250 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9251 /*
9252 * The LBF_DST_PINNED logic could have changed
9253 * env->dst_cpu, so we can't know our idle
9254 * state even if we migrated tasks. Update it.
9255 */
9256 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9257 }
9258 sd->last_balance = jiffies;
9259 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9260 }
9261 if (need_serialize)
9262 spin_unlock(&balancing);
9263out:
9264 if (time_after(next_balance, sd->last_balance + interval)) {
9265 next_balance = sd->last_balance + interval;
9266 update_next_balance = 1;
9267 }
9268 }
9269 if (need_decay) {
9270 /*
9271 * Ensure the rq-wide value also decays but keep it at a
9272 * reasonable floor to avoid funnies with rq->avg_idle.
9273 */
9274 rq->max_idle_balance_cost =
9275 max((u64)sysctl_sched_migration_cost, max_cost);
9276 }
9277 rcu_read_unlock();
9278
9279 /*
9280 * next_balance will be updated only when there is a need.
9281 * When the cpu is attached to null domain for ex, it will not be
9282 * updated.
9283 */
9284 if (likely(update_next_balance)) {
9285 rq->next_balance = next_balance;
9286
9287#ifdef CONFIG_NO_HZ_COMMON
9288 /*
9289 * If this CPU has been elected to perform the nohz idle
9290 * balance. Other idle CPUs have already rebalanced with
9291 * nohz_idle_balance() and nohz.next_balance has been
9292 * updated accordingly. This CPU is now running the idle load
9293 * balance for itself and we need to update the
9294 * nohz.next_balance accordingly.
9295 */
9296 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9297 nohz.next_balance = rq->next_balance;
9298#endif
9299 }
9300}
9301
9302static inline int on_null_domain(struct rq *rq)
9303{
9304 return unlikely(!rcu_dereference_sched(rq->sd));
9305}
9306
9307#ifdef CONFIG_NO_HZ_COMMON
9308/*
9309 * idle load balancing details
9310 * - When one of the busy CPUs notice that there may be an idle rebalancing
9311 * needed, they will kick the idle load balancer, which then does idle
9312 * load balancing for all the idle CPUs.
9313 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9314 * anywhere yet.
9315 */
9316
9317static inline int find_new_ilb(void)
9318{
9319 int ilb;
9320
9321 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9322 housekeeping_cpumask(HK_FLAG_MISC)) {
9323 if (idle_cpu(ilb))
9324 return ilb;
9325 }
9326
9327 return nr_cpu_ids;
9328}
9329
9330/*
9331 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9332 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9333 */
9334static void kick_ilb(unsigned int flags)
9335{
9336 int ilb_cpu;
9337
9338 nohz.next_balance++;
9339
9340 ilb_cpu = find_new_ilb();
9341
9342 if (ilb_cpu >= nr_cpu_ids)
9343 return;
9344
9345 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9346 if (flags & NOHZ_KICK_MASK)
9347 return;
9348
9349 /*
9350 * Use smp_send_reschedule() instead of resched_cpu().
9351 * This way we generate a sched IPI on the target CPU which
9352 * is idle. And the softirq performing nohz idle load balance
9353 * will be run before returning from the IPI.
9354 */
9355 smp_send_reschedule(ilb_cpu);
9356}
9357
9358/*
9359 * Current decision point for kicking the idle load balancer in the presence
9360 * of idle CPUs in the system.
9361 */
9362static void nohz_balancer_kick(struct rq *rq)
9363{
9364 unsigned long now = jiffies;
9365 struct sched_domain_shared *sds;
9366 struct sched_domain *sd;
9367 int nr_busy, i, cpu = rq->cpu;
9368 unsigned int flags = 0;
9369
9370 if (unlikely(rq->idle_balance))
9371 return;
9372
9373 /*
9374 * We may be recently in ticked or tickless idle mode. At the first
9375 * busy tick after returning from idle, we will update the busy stats.
9376 */
9377 nohz_balance_exit_idle(rq);
9378
9379 /*
9380 * None are in tickless mode and hence no need for NOHZ idle load
9381 * balancing.
9382 */
9383 if (likely(!atomic_read(&nohz.nr_cpus)))
9384 return;
9385
9386 if (READ_ONCE(nohz.has_blocked) &&
9387 time_after(now, READ_ONCE(nohz.next_blocked)))
9388 flags = NOHZ_STATS_KICK;
9389
9390 if (time_before(now, nohz.next_balance))
9391 goto out;
9392
9393 if (rq->nr_running >= 2) {
9394 flags = NOHZ_KICK_MASK;
9395 goto out;
9396 }
9397
9398 rcu_read_lock();
9399
9400 sd = rcu_dereference(rq->sd);
9401 if (sd) {
9402 /*
9403 * If there's a CFS task and the current CPU has reduced
9404 * capacity; kick the ILB to see if there's a better CPU to run
9405 * on.
9406 */
9407 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9408 flags = NOHZ_KICK_MASK;
9409 goto unlock;
9410 }
9411 }
9412
9413 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9414 if (sd) {
9415 /*
9416 * When ASYM_PACKING; see if there's a more preferred CPU
9417 * currently idle; in which case, kick the ILB to move tasks
9418 * around.
9419 */
9420 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9421 if (sched_asym_prefer(i, cpu)) {
9422 flags = NOHZ_KICK_MASK;
9423 goto unlock;
9424 }
9425 }
9426 }
9427
9428 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9429 if (sd) {
9430 /*
9431 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
9432 * to run the misfit task on.
9433 */
9434 if (check_misfit_status(rq, sd)) {
9435 flags = NOHZ_KICK_MASK;
9436 goto unlock;
9437 }
9438
9439 /*
9440 * For asymmetric systems, we do not want to nicely balance
9441 * cache use, instead we want to embrace asymmetry and only
9442 * ensure tasks have enough CPU capacity.
9443 *
9444 * Skip the LLC logic because it's not relevant in that case.
9445 */
9446 goto unlock;
9447 }
9448
9449 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9450 if (sds) {
9451 /*
9452 * If there is an imbalance between LLC domains (IOW we could
9453 * increase the overall cache use), we need some less-loaded LLC
9454 * domain to pull some load. Likewise, we may need to spread
9455 * load within the current LLC domain (e.g. packed SMT cores but
9456 * other CPUs are idle). We can't really know from here how busy
9457 * the others are - so just get a nohz balance going if it looks
9458 * like this LLC domain has tasks we could move.
9459 */
9460 nr_busy = atomic_read(&sds->nr_busy_cpus);
9461 if (nr_busy > 1) {
9462 flags = NOHZ_KICK_MASK;
9463 goto unlock;
9464 }
9465 }
9466unlock:
9467 rcu_read_unlock();
9468out:
9469 if (flags)
9470 kick_ilb(flags);
9471}
9472
9473static void set_cpu_sd_state_busy(int cpu)
9474{
9475 struct sched_domain *sd;
9476
9477 rcu_read_lock();
9478 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9479
9480 if (!sd || !sd->nohz_idle)
9481 goto unlock;
9482 sd->nohz_idle = 0;
9483
9484 atomic_inc(&sd->shared->nr_busy_cpus);
9485unlock:
9486 rcu_read_unlock();
9487}
9488
9489void nohz_balance_exit_idle(struct rq *rq)
9490{
9491 SCHED_WARN_ON(rq != this_rq());
9492
9493 if (likely(!rq->nohz_tick_stopped))
9494 return;
9495
9496 rq->nohz_tick_stopped = 0;
9497 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9498 atomic_dec(&nohz.nr_cpus);
9499
9500 set_cpu_sd_state_busy(rq->cpu);
9501}
9502
9503static void set_cpu_sd_state_idle(int cpu)
9504{
9505 struct sched_domain *sd;
9506
9507 rcu_read_lock();
9508 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9509
9510 if (!sd || sd->nohz_idle)
9511 goto unlock;
9512 sd->nohz_idle = 1;
9513
9514 atomic_dec(&sd->shared->nr_busy_cpus);
9515unlock:
9516 rcu_read_unlock();
9517}
9518
9519/*
9520 * This routine will record that the CPU is going idle with tick stopped.
9521 * This info will be used in performing idle load balancing in the future.
9522 */
9523void nohz_balance_enter_idle(int cpu)
9524{
9525 struct rq *rq = cpu_rq(cpu);
9526
9527 SCHED_WARN_ON(cpu != smp_processor_id());
9528
9529 /* If this CPU is going down, then nothing needs to be done: */
9530 if (!cpu_active(cpu))
9531 return;
9532
9533 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9534 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9535 return;
9536
9537 /*
9538 * Can be set safely without rq->lock held
9539 * If a clear happens, it will have evaluated last additions because
9540 * rq->lock is held during the check and the clear
9541 */
9542 rq->has_blocked_load = 1;
9543
9544 /*
9545 * The tick is still stopped but load could have been added in the
9546 * meantime. We set the nohz.has_blocked flag to trig a check of the
9547 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9548 * of nohz.has_blocked can only happen after checking the new load
9549 */
9550 if (rq->nohz_tick_stopped)
9551 goto out;
9552
9553 /* If we're a completely isolated CPU, we don't play: */
9554 if (on_null_domain(rq))
9555 return;
9556
9557 rq->nohz_tick_stopped = 1;
9558
9559 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9560 atomic_inc(&nohz.nr_cpus);
9561
9562 /*
9563 * Ensures that if nohz_idle_balance() fails to observe our
9564 * @idle_cpus_mask store, it must observe the @has_blocked
9565 * store.
9566 */
9567 smp_mb__after_atomic();
9568
9569 set_cpu_sd_state_idle(cpu);
9570
9571out:
9572 /*
9573 * Each time a cpu enter idle, we assume that it has blocked load and
9574 * enable the periodic update of the load of idle cpus
9575 */
9576 WRITE_ONCE(nohz.has_blocked, 1);
9577}
9578
9579/*
9580 * Internal function that runs load balance for all idle cpus. The load balance
9581 * can be a simple update of blocked load or a complete load balance with
9582 * tasks movement depending of flags.
9583 * The function returns false if the loop has stopped before running
9584 * through all idle CPUs.
9585 */
9586static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9587 enum cpu_idle_type idle)
9588{
9589 /* Earliest time when we have to do rebalance again */
9590 unsigned long now = jiffies;
9591 unsigned long next_balance = now + 60*HZ;
9592 bool has_blocked_load = false;
9593 int update_next_balance = 0;
9594 int this_cpu = this_rq->cpu;
9595 int balance_cpu;
9596 int ret = false;
9597 struct rq *rq;
9598
9599 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9600
9601 /*
9602 * We assume there will be no idle load after this update and clear
9603 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9604 * set the has_blocked flag and trig another update of idle load.
9605 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9606 * setting the flag, we are sure to not clear the state and not
9607 * check the load of an idle cpu.
9608 */
9609 WRITE_ONCE(nohz.has_blocked, 0);
9610
9611 /*
9612 * Ensures that if we miss the CPU, we must see the has_blocked
9613 * store from nohz_balance_enter_idle().
9614 */
9615 smp_mb();
9616
9617 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9618 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9619 continue;
9620
9621 /*
9622 * If this CPU gets work to do, stop the load balancing
9623 * work being done for other CPUs. Next load
9624 * balancing owner will pick it up.
9625 */
9626 if (need_resched()) {
9627 has_blocked_load = true;
9628 goto abort;
9629 }
9630
9631 rq = cpu_rq(balance_cpu);
9632
9633 has_blocked_load |= update_nohz_stats(rq, true);
9634
9635 /*
9636 * If time for next balance is due,
9637 * do the balance.
9638 */
9639 if (time_after_eq(jiffies, rq->next_balance)) {
9640 struct rq_flags rf;
9641
9642 rq_lock_irqsave(rq, &rf);
9643 update_rq_clock(rq);
9644 rq_unlock_irqrestore(rq, &rf);
9645
9646 if (flags & NOHZ_BALANCE_KICK)
9647 rebalance_domains(rq, CPU_IDLE);
9648 }
9649
9650 if (time_after(next_balance, rq->next_balance)) {
9651 next_balance = rq->next_balance;
9652 update_next_balance = 1;
9653 }
9654 }
9655
9656 /* Newly idle CPU doesn't need an update */
9657 if (idle != CPU_NEWLY_IDLE) {
9658 update_blocked_averages(this_cpu);
9659 has_blocked_load |= this_rq->has_blocked_load;
9660 }
9661
9662 if (flags & NOHZ_BALANCE_KICK)
9663 rebalance_domains(this_rq, CPU_IDLE);
9664
9665 WRITE_ONCE(nohz.next_blocked,
9666 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9667
9668 /* The full idle balance loop has been done */
9669 ret = true;
9670
9671abort:
9672 /* There is still blocked load, enable periodic update */
9673 if (has_blocked_load)
9674 WRITE_ONCE(nohz.has_blocked, 1);
9675
9676 /*
9677 * next_balance will be updated only when there is a need.
9678 * When the CPU is attached to null domain for ex, it will not be
9679 * updated.
9680 */
9681 if (likely(update_next_balance))
9682 nohz.next_balance = next_balance;
9683
9684 return ret;
9685}
9686
9687/*
9688 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9689 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9690 */
9691static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9692{
9693 int this_cpu = this_rq->cpu;
9694 unsigned int flags;
9695
9696 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9697 return false;
9698
9699 if (idle != CPU_IDLE) {
9700 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9701 return false;
9702 }
9703
9704 /* could be _relaxed() */
9705 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9706 if (!(flags & NOHZ_KICK_MASK))
9707 return false;
9708
9709 _nohz_idle_balance(this_rq, flags, idle);
9710
9711 return true;
9712}
9713
9714static void nohz_newidle_balance(struct rq *this_rq)
9715{
9716 int this_cpu = this_rq->cpu;
9717
9718 /*
9719 * This CPU doesn't want to be disturbed by scheduler
9720 * housekeeping
9721 */
9722 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9723 return;
9724
9725 /* Will wake up very soon. No time for doing anything else*/
9726 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9727 return;
9728
9729 /* Don't need to update blocked load of idle CPUs*/
9730 if (!READ_ONCE(nohz.has_blocked) ||
9731 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9732 return;
9733
9734 raw_spin_unlock(&this_rq->lock);
9735 /*
9736 * This CPU is going to be idle and blocked load of idle CPUs
9737 * need to be updated. Run the ilb locally as it is a good
9738 * candidate for ilb instead of waking up another idle CPU.
9739 * Kick an normal ilb if we failed to do the update.
9740 */
9741 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9742 kick_ilb(NOHZ_STATS_KICK);
9743 raw_spin_lock(&this_rq->lock);
9744}
9745
9746#else /* !CONFIG_NO_HZ_COMMON */
9747static inline void nohz_balancer_kick(struct rq *rq) { }
9748
9749static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9750{
9751 return false;
9752}
9753
9754static inline void nohz_newidle_balance(struct rq *this_rq) { }
9755#endif /* CONFIG_NO_HZ_COMMON */
9756
9757/*
9758 * idle_balance is called by schedule() if this_cpu is about to become
9759 * idle. Attempts to pull tasks from other CPUs.
9760 */
9761int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
9762{
9763 unsigned long next_balance = jiffies + HZ;
9764 int this_cpu = this_rq->cpu;
9765 struct sched_domain *sd;
9766 int pulled_task = 0;
9767 u64 curr_cost = 0;
9768
9769 update_misfit_status(NULL, this_rq);
9770 /*
9771 * We must set idle_stamp _before_ calling idle_balance(), such that we
9772 * measure the duration of idle_balance() as idle time.
9773 */
9774 this_rq->idle_stamp = rq_clock(this_rq);
9775
9776 /*
9777 * Do not pull tasks towards !active CPUs...
9778 */
9779 if (!cpu_active(this_cpu))
9780 return 0;
9781
9782 /*
9783 * This is OK, because current is on_cpu, which avoids it being picked
9784 * for load-balance and preemption/IRQs are still disabled avoiding
9785 * further scheduler activity on it and we're being very careful to
9786 * re-start the picking loop.
9787 */
9788 rq_unpin_lock(this_rq, rf);
9789
9790 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9791 !READ_ONCE(this_rq->rd->overload)) {
9792
9793 rcu_read_lock();
9794 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9795 if (sd)
9796 update_next_balance(sd, &next_balance);
9797 rcu_read_unlock();
9798
9799 nohz_newidle_balance(this_rq);
9800
9801 goto out;
9802 }
9803
9804 raw_spin_unlock(&this_rq->lock);
9805
9806 update_blocked_averages(this_cpu);
9807 rcu_read_lock();
9808 for_each_domain(this_cpu, sd) {
9809 int continue_balancing = 1;
9810 u64 t0, domain_cost;
9811
9812 if (!(sd->flags & SD_LOAD_BALANCE))
9813 continue;
9814
9815 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9816 update_next_balance(sd, &next_balance);
9817 break;
9818 }
9819
9820 if (sd->flags & SD_BALANCE_NEWIDLE) {
9821 t0 = sched_clock_cpu(this_cpu);
9822
9823 pulled_task = load_balance(this_cpu, this_rq,
9824 sd, CPU_NEWLY_IDLE,
9825 &continue_balancing);
9826
9827 domain_cost = sched_clock_cpu(this_cpu) - t0;
9828 if (domain_cost > sd->max_newidle_lb_cost)
9829 sd->max_newidle_lb_cost = domain_cost;
9830
9831 curr_cost += domain_cost;
9832 }
9833
9834 update_next_balance(sd, &next_balance);
9835
9836 /*
9837 * Stop searching for tasks to pull if there are
9838 * now runnable tasks on this rq.
9839 */
9840 if (pulled_task || this_rq->nr_running > 0)
9841 break;
9842 }
9843 rcu_read_unlock();
9844
9845 raw_spin_lock(&this_rq->lock);
9846
9847 if (curr_cost > this_rq->max_idle_balance_cost)
9848 this_rq->max_idle_balance_cost = curr_cost;
9849
9850out:
9851 /*
9852 * While browsing the domains, we released the rq lock, a task could
9853 * have been enqueued in the meantime. Since we're not going idle,
9854 * pretend we pulled a task.
9855 */
9856 if (this_rq->cfs.h_nr_running && !pulled_task)
9857 pulled_task = 1;
9858
9859 /* Move the next balance forward */
9860 if (time_after(this_rq->next_balance, next_balance))
9861 this_rq->next_balance = next_balance;
9862
9863 /* Is there a task of a high priority class? */
9864 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9865 pulled_task = -1;
9866
9867 if (pulled_task)
9868 this_rq->idle_stamp = 0;
9869
9870 rq_repin_lock(this_rq, rf);
9871
9872 return pulled_task;
9873}
9874
9875/*
9876 * run_rebalance_domains is triggered when needed from the scheduler tick.
9877 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9878 */
9879static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9880{
9881 struct rq *this_rq = this_rq();
9882 enum cpu_idle_type idle = this_rq->idle_balance ?
9883 CPU_IDLE : CPU_NOT_IDLE;
9884
9885 /*
9886 * If this CPU has a pending nohz_balance_kick, then do the
9887 * balancing on behalf of the other idle CPUs whose ticks are
9888 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9889 * give the idle CPUs a chance to load balance. Else we may
9890 * load balance only within the local sched_domain hierarchy
9891 * and abort nohz_idle_balance altogether if we pull some load.
9892 */
9893 if (nohz_idle_balance(this_rq, idle))
9894 return;
9895
9896 /* normal load balance */
9897 update_blocked_averages(this_rq->cpu);
9898 rebalance_domains(this_rq, idle);
9899}
9900
9901/*
9902 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9903 */
9904void trigger_load_balance(struct rq *rq)
9905{
9906 /* Don't need to rebalance while attached to NULL domain */
9907 if (unlikely(on_null_domain(rq)))
9908 return;
9909
9910 if (time_after_eq(jiffies, rq->next_balance))
9911 raise_softirq(SCHED_SOFTIRQ);
9912
9913 nohz_balancer_kick(rq);
9914}
9915
9916static void rq_online_fair(struct rq *rq)
9917{
9918 update_sysctl();
9919
9920 update_runtime_enabled(rq);
9921}
9922
9923static void rq_offline_fair(struct rq *rq)
9924{
9925 update_sysctl();
9926
9927 /* Ensure any throttled groups are reachable by pick_next_task */
9928 unthrottle_offline_cfs_rqs(rq);
9929}
9930
9931#endif /* CONFIG_SMP */
9932
9933/*
9934 * scheduler tick hitting a task of our scheduling class.
9935 *
9936 * NOTE: This function can be called remotely by the tick offload that
9937 * goes along full dynticks. Therefore no local assumption can be made
9938 * and everything must be accessed through the @rq and @curr passed in
9939 * parameters.
9940 */
9941static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9942{
9943 struct cfs_rq *cfs_rq;
9944 struct sched_entity *se = &curr->se;
9945
9946 for_each_sched_entity(se) {
9947 cfs_rq = cfs_rq_of(se);
9948 entity_tick(cfs_rq, se, queued);
9949 }
9950
9951 if (static_branch_unlikely(&sched_numa_balancing))
9952 task_tick_numa(rq, curr);
9953
9954 update_misfit_status(curr, rq);
9955 update_overutilized_status(task_rq(curr));
9956}
9957
9958/*
9959 * called on fork with the child task as argument from the parent's context
9960 * - child not yet on the tasklist
9961 * - preemption disabled
9962 */
9963static void task_fork_fair(struct task_struct *p)
9964{
9965 struct cfs_rq *cfs_rq;
9966 struct sched_entity *se = &p->se, *curr;
9967 struct rq *rq = this_rq();
9968 struct rq_flags rf;
9969
9970 rq_lock(rq, &rf);
9971 update_rq_clock(rq);
9972
9973 cfs_rq = task_cfs_rq(current);
9974 curr = cfs_rq->curr;
9975 if (curr) {
9976 update_curr(cfs_rq);
9977 se->vruntime = curr->vruntime;
9978 }
9979 place_entity(cfs_rq, se, 1);
9980
9981 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9982 /*
9983 * Upon rescheduling, sched_class::put_prev_task() will place
9984 * 'current' within the tree based on its new key value.
9985 */
9986 swap(curr->vruntime, se->vruntime);
9987 resched_curr(rq);
9988 }
9989
9990 se->vruntime -= cfs_rq->min_vruntime;
9991 rq_unlock(rq, &rf);
9992}
9993
9994/*
9995 * Priority of the task has changed. Check to see if we preempt
9996 * the current task.
9997 */
9998static void
9999prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10000{
10001 if (!task_on_rq_queued(p))
10002 return;
10003
10004 /*
10005 * Reschedule if we are currently running on this runqueue and
10006 * our priority decreased, or if we are not currently running on
10007 * this runqueue and our priority is higher than the current's
10008 */
10009 if (rq->curr == p) {
10010 if (p->prio > oldprio)
10011 resched_curr(rq);
10012 } else
10013 check_preempt_curr(rq, p, 0);
10014}
10015
10016static inline bool vruntime_normalized(struct task_struct *p)
10017{
10018 struct sched_entity *se = &p->se;
10019
10020 /*
10021 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10022 * the dequeue_entity(.flags=0) will already have normalized the
10023 * vruntime.
10024 */
10025 if (p->on_rq)
10026 return true;
10027
10028 /*
10029 * When !on_rq, vruntime of the task has usually NOT been normalized.
10030 * But there are some cases where it has already been normalized:
10031 *
10032 * - A forked child which is waiting for being woken up by
10033 * wake_up_new_task().
10034 * - A task which has been woken up by try_to_wake_up() and
10035 * waiting for actually being woken up by sched_ttwu_pending().
10036 */
10037 if (!se->sum_exec_runtime ||
10038 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10039 return true;
10040
10041 return false;
10042}
10043
10044#ifdef CONFIG_FAIR_GROUP_SCHED
10045/*
10046 * Propagate the changes of the sched_entity across the tg tree to make it
10047 * visible to the root
10048 */
10049static void propagate_entity_cfs_rq(struct sched_entity *se)
10050{
10051 struct cfs_rq *cfs_rq;
10052
10053 /* Start to propagate at parent */
10054 se = se->parent;
10055
10056 for_each_sched_entity(se) {
10057 cfs_rq = cfs_rq_of(se);
10058
10059 if (cfs_rq_throttled(cfs_rq))
10060 break;
10061
10062 update_load_avg(cfs_rq, se, UPDATE_TG);
10063 }
10064}
10065#else
10066static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10067#endif
10068
10069static void detach_entity_cfs_rq(struct sched_entity *se)
10070{
10071 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10072
10073 /* Catch up with the cfs_rq and remove our load when we leave */
10074 update_load_avg(cfs_rq, se, 0);
10075 detach_entity_load_avg(cfs_rq, se);
10076 update_tg_load_avg(cfs_rq, false);
10077 propagate_entity_cfs_rq(se);
10078}
10079
10080static void attach_entity_cfs_rq(struct sched_entity *se)
10081{
10082 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10083
10084#ifdef CONFIG_FAIR_GROUP_SCHED
10085 /*
10086 * Since the real-depth could have been changed (only FAIR
10087 * class maintain depth value), reset depth properly.
10088 */
10089 se->depth = se->parent ? se->parent->depth + 1 : 0;
10090#endif
10091
10092 /* Synchronize entity with its cfs_rq */
10093 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10094 attach_entity_load_avg(cfs_rq, se, 0);
10095 update_tg_load_avg(cfs_rq, false);
10096 propagate_entity_cfs_rq(se);
10097}
10098
10099static void detach_task_cfs_rq(struct task_struct *p)
10100{
10101 struct sched_entity *se = &p->se;
10102 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10103
10104 if (!vruntime_normalized(p)) {
10105 /*
10106 * Fix up our vruntime so that the current sleep doesn't
10107 * cause 'unlimited' sleep bonus.
10108 */
10109 place_entity(cfs_rq, se, 0);
10110 se->vruntime -= cfs_rq->min_vruntime;
10111 }
10112
10113 detach_entity_cfs_rq(se);
10114}
10115
10116static void attach_task_cfs_rq(struct task_struct *p)
10117{
10118 struct sched_entity *se = &p->se;
10119 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10120
10121 attach_entity_cfs_rq(se);
10122
10123 if (!vruntime_normalized(p))
10124 se->vruntime += cfs_rq->min_vruntime;
10125}
10126
10127static void switched_from_fair(struct rq *rq, struct task_struct *p)
10128{
10129 detach_task_cfs_rq(p);
10130}
10131
10132static void switched_to_fair(struct rq *rq, struct task_struct *p)
10133{
10134 attach_task_cfs_rq(p);
10135
10136 if (task_on_rq_queued(p)) {
10137 /*
10138 * We were most likely switched from sched_rt, so
10139 * kick off the schedule if running, otherwise just see
10140 * if we can still preempt the current task.
10141 */
10142 if (rq->curr == p)
10143 resched_curr(rq);
10144 else
10145 check_preempt_curr(rq, p, 0);
10146 }
10147}
10148
10149/* Account for a task changing its policy or group.
10150 *
10151 * This routine is mostly called to set cfs_rq->curr field when a task
10152 * migrates between groups/classes.
10153 */
10154static void set_next_task_fair(struct rq *rq, struct task_struct *p)
10155{
10156 struct sched_entity *se = &p->se;
10157
10158#ifdef CONFIG_SMP
10159 if (task_on_rq_queued(p)) {
10160 /*
10161 * Move the next running task to the front of the list, so our
10162 * cfs_tasks list becomes MRU one.
10163 */
10164 list_move(&se->group_node, &rq->cfs_tasks);
10165 }
10166#endif
10167
10168 for_each_sched_entity(se) {
10169 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10170
10171 set_next_entity(cfs_rq, se);
10172 /* ensure bandwidth has been allocated on our new cfs_rq */
10173 account_cfs_rq_runtime(cfs_rq, 0);
10174 }
10175}
10176
10177void init_cfs_rq(struct cfs_rq *cfs_rq)
10178{
10179 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10180 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10181#ifndef CONFIG_64BIT
10182 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10183#endif
10184#ifdef CONFIG_SMP
10185 raw_spin_lock_init(&cfs_rq->removed.lock);
10186#endif
10187}
10188
10189#ifdef CONFIG_FAIR_GROUP_SCHED
10190static void task_set_group_fair(struct task_struct *p)
10191{
10192 struct sched_entity *se = &p->se;
10193
10194 set_task_rq(p, task_cpu(p));
10195 se->depth = se->parent ? se->parent->depth + 1 : 0;
10196}
10197
10198static void task_move_group_fair(struct task_struct *p)
10199{
10200 detach_task_cfs_rq(p);
10201 set_task_rq(p, task_cpu(p));
10202
10203#ifdef CONFIG_SMP
10204 /* Tell se's cfs_rq has been changed -- migrated */
10205 p->se.avg.last_update_time = 0;
10206#endif
10207 attach_task_cfs_rq(p);
10208}
10209
10210static void task_change_group_fair(struct task_struct *p, int type)
10211{
10212 switch (type) {
10213 case TASK_SET_GROUP:
10214 task_set_group_fair(p);
10215 break;
10216
10217 case TASK_MOVE_GROUP:
10218 task_move_group_fair(p);
10219 break;
10220 }
10221}
10222
10223void free_fair_sched_group(struct task_group *tg)
10224{
10225 int i;
10226
10227 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10228
10229 for_each_possible_cpu(i) {
10230 if (tg->cfs_rq)
10231 kfree(tg->cfs_rq[i]);
10232 if (tg->se)
10233 kfree(tg->se[i]);
10234 }
10235
10236 kfree(tg->cfs_rq);
10237 kfree(tg->se);
10238}
10239
10240int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10241{
10242 struct sched_entity *se;
10243 struct cfs_rq *cfs_rq;
10244 int i;
10245
10246 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10247 if (!tg->cfs_rq)
10248 goto err;
10249 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10250 if (!tg->se)
10251 goto err;
10252
10253 tg->shares = NICE_0_LOAD;
10254
10255 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10256
10257 for_each_possible_cpu(i) {
10258 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10259 GFP_KERNEL, cpu_to_node(i));
10260 if (!cfs_rq)
10261 goto err;
10262
10263 se = kzalloc_node(sizeof(struct sched_entity),
10264 GFP_KERNEL, cpu_to_node(i));
10265 if (!se)
10266 goto err_free_rq;
10267
10268 init_cfs_rq(cfs_rq);
10269 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10270 init_entity_runnable_average(se);
10271 }
10272
10273 return 1;
10274
10275err_free_rq:
10276 kfree(cfs_rq);
10277err:
10278 return 0;
10279}
10280
10281void online_fair_sched_group(struct task_group *tg)
10282{
10283 struct sched_entity *se;
10284 struct rq_flags rf;
10285 struct rq *rq;
10286 int i;
10287
10288 for_each_possible_cpu(i) {
10289 rq = cpu_rq(i);
10290 se = tg->se[i];
10291 rq_lock_irq(rq, &rf);
10292 update_rq_clock(rq);
10293 attach_entity_cfs_rq(se);
10294 sync_throttle(tg, i);
10295 rq_unlock_irq(rq, &rf);
10296 }
10297}
10298
10299void unregister_fair_sched_group(struct task_group *tg)
10300{
10301 unsigned long flags;
10302 struct rq *rq;
10303 int cpu;
10304
10305 for_each_possible_cpu(cpu) {
10306 if (tg->se[cpu])
10307 remove_entity_load_avg(tg->se[cpu]);
10308
10309 /*
10310 * Only empty task groups can be destroyed; so we can speculatively
10311 * check on_list without danger of it being re-added.
10312 */
10313 if (!tg->cfs_rq[cpu]->on_list)
10314 continue;
10315
10316 rq = cpu_rq(cpu);
10317
10318 raw_spin_lock_irqsave(&rq->lock, flags);
10319 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10320 raw_spin_unlock_irqrestore(&rq->lock, flags);
10321 }
10322}
10323
10324void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10325 struct sched_entity *se, int cpu,
10326 struct sched_entity *parent)
10327{
10328 struct rq *rq = cpu_rq(cpu);
10329
10330 cfs_rq->tg = tg;
10331 cfs_rq->rq = rq;
10332 init_cfs_rq_runtime(cfs_rq);
10333
10334 tg->cfs_rq[cpu] = cfs_rq;
10335 tg->se[cpu] = se;
10336
10337 /* se could be NULL for root_task_group */
10338 if (!se)
10339 return;
10340
10341 if (!parent) {
10342 se->cfs_rq = &rq->cfs;
10343 se->depth = 0;
10344 } else {
10345 se->cfs_rq = parent->my_q;
10346 se->depth = parent->depth + 1;
10347 }
10348
10349 se->my_q = cfs_rq;
10350 /* guarantee group entities always have weight */
10351 update_load_set(&se->load, NICE_0_LOAD);
10352 se->parent = parent;
10353}
10354
10355static DEFINE_MUTEX(shares_mutex);
10356
10357int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10358{
10359 int i;
10360
10361 /*
10362 * We can't change the weight of the root cgroup.
10363 */
10364 if (!tg->se[0])
10365 return -EINVAL;
10366
10367 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10368
10369 mutex_lock(&shares_mutex);
10370 if (tg->shares == shares)
10371 goto done;
10372
10373 tg->shares = shares;
10374 for_each_possible_cpu(i) {
10375 struct rq *rq = cpu_rq(i);
10376 struct sched_entity *se = tg->se[i];
10377 struct rq_flags rf;
10378
10379 /* Propagate contribution to hierarchy */
10380 rq_lock_irqsave(rq, &rf);
10381 update_rq_clock(rq);
10382 for_each_sched_entity(se) {
10383 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10384 update_cfs_group(se);
10385 }
10386 rq_unlock_irqrestore(rq, &rf);
10387 }
10388
10389done:
10390 mutex_unlock(&shares_mutex);
10391 return 0;
10392}
10393#else /* CONFIG_FAIR_GROUP_SCHED */
10394
10395void free_fair_sched_group(struct task_group *tg) { }
10396
10397int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10398{
10399 return 1;
10400}
10401
10402void online_fair_sched_group(struct task_group *tg) { }
10403
10404void unregister_fair_sched_group(struct task_group *tg) { }
10405
10406#endif /* CONFIG_FAIR_GROUP_SCHED */
10407
10408
10409static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10410{
10411 struct sched_entity *se = &task->se;
10412 unsigned int rr_interval = 0;
10413
10414 /*
10415 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10416 * idle runqueue:
10417 */
10418 if (rq->cfs.load.weight)
10419 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10420
10421 return rr_interval;
10422}
10423
10424/*
10425 * All the scheduling class methods:
10426 */
10427const struct sched_class fair_sched_class = {
10428 .next = &idle_sched_class,
10429 .enqueue_task = enqueue_task_fair,
10430 .dequeue_task = dequeue_task_fair,
10431 .yield_task = yield_task_fair,
10432 .yield_to_task = yield_to_task_fair,
10433
10434 .check_preempt_curr = check_preempt_wakeup,
10435
10436 .pick_next_task = pick_next_task_fair,
10437 .put_prev_task = put_prev_task_fair,
10438 .set_next_task = set_next_task_fair,
10439
10440#ifdef CONFIG_SMP
10441 .balance = balance_fair,
10442 .select_task_rq = select_task_rq_fair,
10443 .migrate_task_rq = migrate_task_rq_fair,
10444
10445 .rq_online = rq_online_fair,
10446 .rq_offline = rq_offline_fair,
10447
10448 .task_dead = task_dead_fair,
10449 .set_cpus_allowed = set_cpus_allowed_common,
10450#endif
10451
10452 .task_tick = task_tick_fair,
10453 .task_fork = task_fork_fair,
10454
10455 .prio_changed = prio_changed_fair,
10456 .switched_from = switched_from_fair,
10457 .switched_to = switched_to_fair,
10458
10459 .get_rr_interval = get_rr_interval_fair,
10460
10461 .update_curr = update_curr_fair,
10462
10463#ifdef CONFIG_FAIR_GROUP_SCHED
10464 .task_change_group = task_change_group_fair,
10465#endif
10466
10467#ifdef CONFIG_UCLAMP_TASK
10468 .uclamp_enabled = 1,
10469#endif
10470};
10471
10472#ifdef CONFIG_SCHED_DEBUG
10473void print_cfs_stats(struct seq_file *m, int cpu)
10474{
10475 struct cfs_rq *cfs_rq, *pos;
10476
10477 rcu_read_lock();
10478 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10479 print_cfs_rq(m, cpu, cfs_rq);
10480 rcu_read_unlock();
10481}
10482
10483#ifdef CONFIG_NUMA_BALANCING
10484void show_numa_stats(struct task_struct *p, struct seq_file *m)
10485{
10486 int node;
10487 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10488 struct numa_group *ng;
10489
10490 rcu_read_lock();
10491 ng = rcu_dereference(p->numa_group);
10492 for_each_online_node(node) {
10493 if (p->numa_faults) {
10494 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10495 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10496 }
10497 if (ng) {
10498 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10499 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10500 }
10501 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10502 }
10503 rcu_read_unlock();
10504}
10505#endif /* CONFIG_NUMA_BALANCING */
10506#endif /* CONFIG_SCHED_DEBUG */
10507
10508__init void init_sched_fair_class(void)
10509{
10510#ifdef CONFIG_SMP
10511 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10512
10513#ifdef CONFIG_NO_HZ_COMMON
10514 nohz.next_balance = jiffies;
10515 nohz.next_blocked = jiffies;
10516 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10517#endif
10518#endif /* SMP */
10519
10520}
10521
10522/*
10523 * Helper functions to facilitate extracting info from tracepoints.
10524 */
10525
10526const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10527{
10528#ifdef CONFIG_SMP
10529 return cfs_rq ? &cfs_rq->avg : NULL;
10530#else
10531 return NULL;
10532#endif
10533}
10534EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10535
10536char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10537{
10538 if (!cfs_rq) {
10539 if (str)
10540 strlcpy(str, "(null)", len);
10541 else
10542 return NULL;
10543 }
10544
10545 cfs_rq_tg_path(cfs_rq, str, len);
10546 return str;
10547}
10548EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10549
10550int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10551{
10552 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10553}
10554EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10555
10556const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10557{
10558#ifdef CONFIG_SMP
10559 return rq ? &rq->avg_rt : NULL;
10560#else
10561 return NULL;
10562#endif
10563}
10564EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10565
10566const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10567{
10568#ifdef CONFIG_SMP
10569 return rq ? &rq->avg_dl : NULL;
10570#else
10571 return NULL;
10572#endif
10573}
10574EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10575
10576const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10577{
10578#if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10579 return rq ? &rq->avg_irq : NULL;
10580#else
10581 return NULL;
10582#endif
10583}
10584EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10585
10586int sched_trace_rq_cpu(struct rq *rq)
10587{
10588 return rq ? cpu_of(rq) : -1;
10589}
10590EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10591
10592const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10593{
10594#ifdef CONFIG_SMP
10595 return rd ? rd->span : NULL;
10596#else
10597 return NULL;
10598#endif
10599}
10600EXPORT_SYMBOL_GPL(sched_trace_rd_span);
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 __always_inline
420void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, 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
779static void
780account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
781{
782 update_load_add(&cfs_rq->load, se->load.weight);
783 if (!parent_entity(se))
784 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
785#ifdef CONFIG_SMP
786 if (entity_is_task(se))
787 list_add(&se->group_node, &rq_of(cfs_rq)->cfs_tasks);
788#endif
789 cfs_rq->nr_running++;
790}
791
792static void
793account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
794{
795 update_load_sub(&cfs_rq->load, se->load.weight);
796 if (!parent_entity(se))
797 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
798 if (entity_is_task(se))
799 list_del_init(&se->group_node);
800 cfs_rq->nr_running--;
801}
802
803#ifdef CONFIG_FAIR_GROUP_SCHED
804/* we need this in update_cfs_load and load-balance functions below */
805static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
806# ifdef CONFIG_SMP
807static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
808 int global_update)
809{
810 struct task_group *tg = cfs_rq->tg;
811 long load_avg;
812
813 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
814 load_avg -= cfs_rq->load_contribution;
815
816 if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
817 atomic_add(load_avg, &tg->load_weight);
818 cfs_rq->load_contribution += load_avg;
819 }
820}
821
822static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
823{
824 u64 period = sysctl_sched_shares_window;
825 u64 now, delta;
826 unsigned long load = cfs_rq->load.weight;
827
828 if (cfs_rq->tg == &root_task_group || throttled_hierarchy(cfs_rq))
829 return;
830
831 now = rq_of(cfs_rq)->clock_task;
832 delta = now - cfs_rq->load_stamp;
833
834 /* truncate load history at 4 idle periods */
835 if (cfs_rq->load_stamp > cfs_rq->load_last &&
836 now - cfs_rq->load_last > 4 * period) {
837 cfs_rq->load_period = 0;
838 cfs_rq->load_avg = 0;
839 delta = period - 1;
840 }
841
842 cfs_rq->load_stamp = now;
843 cfs_rq->load_unacc_exec_time = 0;
844 cfs_rq->load_period += delta;
845 if (load) {
846 cfs_rq->load_last = now;
847 cfs_rq->load_avg += delta * load;
848 }
849
850 /* consider updating load contribution on each fold or truncate */
851 if (global_update || cfs_rq->load_period > period
852 || !cfs_rq->load_period)
853 update_cfs_rq_load_contribution(cfs_rq, global_update);
854
855 while (cfs_rq->load_period > period) {
856 /*
857 * Inline assembly required to prevent the compiler
858 * optimising this loop into a divmod call.
859 * See __iter_div_u64_rem() for another example of this.
860 */
861 asm("" : "+rm" (cfs_rq->load_period));
862 cfs_rq->load_period /= 2;
863 cfs_rq->load_avg /= 2;
864 }
865
866 if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
867 list_del_leaf_cfs_rq(cfs_rq);
868}
869
870static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
871{
872 long tg_weight;
873
874 /*
875 * Use this CPU's actual weight instead of the last load_contribution
876 * to gain a more accurate current total weight. See
877 * update_cfs_rq_load_contribution().
878 */
879 tg_weight = atomic_read(&tg->load_weight);
880 tg_weight -= cfs_rq->load_contribution;
881 tg_weight += cfs_rq->load.weight;
882
883 return tg_weight;
884}
885
886static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
887{
888 long tg_weight, load, shares;
889
890 tg_weight = calc_tg_weight(tg, cfs_rq);
891 load = cfs_rq->load.weight;
892
893 shares = (tg->shares * load);
894 if (tg_weight)
895 shares /= tg_weight;
896
897 if (shares < MIN_SHARES)
898 shares = MIN_SHARES;
899 if (shares > tg->shares)
900 shares = tg->shares;
901
902 return shares;
903}
904
905static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
906{
907 if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
908 update_cfs_load(cfs_rq, 0);
909 update_cfs_shares(cfs_rq);
910 }
911}
912# else /* CONFIG_SMP */
913static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
914{
915}
916
917static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
918{
919 return tg->shares;
920}
921
922static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
923{
924}
925# endif /* CONFIG_SMP */
926static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
927 unsigned long weight)
928{
929 if (se->on_rq) {
930 /* commit outstanding execution time */
931 if (cfs_rq->curr == se)
932 update_curr(cfs_rq);
933 account_entity_dequeue(cfs_rq, se);
934 }
935
936 update_load_set(&se->load, weight);
937
938 if (se->on_rq)
939 account_entity_enqueue(cfs_rq, se);
940}
941
942static void update_cfs_shares(struct cfs_rq *cfs_rq)
943{
944 struct task_group *tg;
945 struct sched_entity *se;
946 long shares;
947
948 tg = cfs_rq->tg;
949 se = tg->se[cpu_of(rq_of(cfs_rq))];
950 if (!se || throttled_hierarchy(cfs_rq))
951 return;
952#ifndef CONFIG_SMP
953 if (likely(se->load.weight == tg->shares))
954 return;
955#endif
956 shares = calc_cfs_shares(cfs_rq, tg);
957
958 reweight_entity(cfs_rq_of(se), se, shares);
959}
960#else /* CONFIG_FAIR_GROUP_SCHED */
961static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
962{
963}
964
965static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
966{
967}
968
969static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
970{
971}
972#endif /* CONFIG_FAIR_GROUP_SCHED */
973
974static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
975{
976#ifdef CONFIG_SCHEDSTATS
977 struct task_struct *tsk = NULL;
978
979 if (entity_is_task(se))
980 tsk = task_of(se);
981
982 if (se->statistics.sleep_start) {
983 u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
984
985 if ((s64)delta < 0)
986 delta = 0;
987
988 if (unlikely(delta > se->statistics.sleep_max))
989 se->statistics.sleep_max = delta;
990
991 se->statistics.sleep_start = 0;
992 se->statistics.sum_sleep_runtime += delta;
993
994 if (tsk) {
995 account_scheduler_latency(tsk, delta >> 10, 1);
996 trace_sched_stat_sleep(tsk, delta);
997 }
998 }
999 if (se->statistics.block_start) {
1000 u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
1001
1002 if ((s64)delta < 0)
1003 delta = 0;
1004
1005 if (unlikely(delta > se->statistics.block_max))
1006 se->statistics.block_max = delta;
1007
1008 se->statistics.block_start = 0;
1009 se->statistics.sum_sleep_runtime += delta;
1010
1011 if (tsk) {
1012 if (tsk->in_iowait) {
1013 se->statistics.iowait_sum += delta;
1014 se->statistics.iowait_count++;
1015 trace_sched_stat_iowait(tsk, delta);
1016 }
1017
1018 trace_sched_stat_blocked(tsk, delta);
1019
1020 /*
1021 * Blocking time is in units of nanosecs, so shift by
1022 * 20 to get a milliseconds-range estimation of the
1023 * amount of time that the task spent sleeping:
1024 */
1025 if (unlikely(prof_on == SLEEP_PROFILING)) {
1026 profile_hits(SLEEP_PROFILING,
1027 (void *)get_wchan(tsk),
1028 delta >> 20);
1029 }
1030 account_scheduler_latency(tsk, delta >> 10, 0);
1031 }
1032 }
1033#endif
1034}
1035
1036static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
1037{
1038#ifdef CONFIG_SCHED_DEBUG
1039 s64 d = se->vruntime - cfs_rq->min_vruntime;
1040
1041 if (d < 0)
1042 d = -d;
1043
1044 if (d > 3*sysctl_sched_latency)
1045 schedstat_inc(cfs_rq, nr_spread_over);
1046#endif
1047}
1048
1049static void
1050place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
1051{
1052 u64 vruntime = cfs_rq->min_vruntime;
1053
1054 /*
1055 * The 'current' period is already promised to the current tasks,
1056 * however the extra weight of the new task will slow them down a
1057 * little, place the new task so that it fits in the slot that
1058 * stays open at the end.
1059 */
1060 if (initial && sched_feat(START_DEBIT))
1061 vruntime += sched_vslice(cfs_rq, se);
1062
1063 /* sleeps up to a single latency don't count. */
1064 if (!initial) {
1065 unsigned long thresh = sysctl_sched_latency;
1066
1067 /*
1068 * Halve their sleep time's effect, to allow
1069 * for a gentler effect of sleepers:
1070 */
1071 if (sched_feat(GENTLE_FAIR_SLEEPERS))
1072 thresh >>= 1;
1073
1074 vruntime -= thresh;
1075 }
1076
1077 /* ensure we never gain time by being placed backwards. */
1078 vruntime = max_vruntime(se->vruntime, vruntime);
1079
1080 se->vruntime = vruntime;
1081}
1082
1083static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
1084
1085static void
1086enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1087{
1088 /*
1089 * Update the normalized vruntime before updating min_vruntime
1090 * through callig update_curr().
1091 */
1092 if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
1093 se->vruntime += cfs_rq->min_vruntime;
1094
1095 /*
1096 * Update run-time statistics of the 'current'.
1097 */
1098 update_curr(cfs_rq);
1099 update_cfs_load(cfs_rq, 0);
1100 account_entity_enqueue(cfs_rq, se);
1101 update_cfs_shares(cfs_rq);
1102
1103 if (flags & ENQUEUE_WAKEUP) {
1104 place_entity(cfs_rq, se, 0);
1105 enqueue_sleeper(cfs_rq, se);
1106 }
1107
1108 update_stats_enqueue(cfs_rq, se);
1109 check_spread(cfs_rq, se);
1110 if (se != cfs_rq->curr)
1111 __enqueue_entity(cfs_rq, se);
1112 se->on_rq = 1;
1113
1114 if (cfs_rq->nr_running == 1) {
1115 list_add_leaf_cfs_rq(cfs_rq);
1116 check_enqueue_throttle(cfs_rq);
1117 }
1118}
1119
1120static void __clear_buddies_last(struct sched_entity *se)
1121{
1122 for_each_sched_entity(se) {
1123 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1124 if (cfs_rq->last == se)
1125 cfs_rq->last = NULL;
1126 else
1127 break;
1128 }
1129}
1130
1131static void __clear_buddies_next(struct sched_entity *se)
1132{
1133 for_each_sched_entity(se) {
1134 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1135 if (cfs_rq->next == se)
1136 cfs_rq->next = NULL;
1137 else
1138 break;
1139 }
1140}
1141
1142static void __clear_buddies_skip(struct sched_entity *se)
1143{
1144 for_each_sched_entity(se) {
1145 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1146 if (cfs_rq->skip == se)
1147 cfs_rq->skip = NULL;
1148 else
1149 break;
1150 }
1151}
1152
1153static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
1154{
1155 if (cfs_rq->last == se)
1156 __clear_buddies_last(se);
1157
1158 if (cfs_rq->next == se)
1159 __clear_buddies_next(se);
1160
1161 if (cfs_rq->skip == se)
1162 __clear_buddies_skip(se);
1163}
1164
1165static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1166
1167static void
1168dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1169{
1170 /*
1171 * Update run-time statistics of the 'current'.
1172 */
1173 update_curr(cfs_rq);
1174
1175 update_stats_dequeue(cfs_rq, se);
1176 if (flags & DEQUEUE_SLEEP) {
1177#ifdef CONFIG_SCHEDSTATS
1178 if (entity_is_task(se)) {
1179 struct task_struct *tsk = task_of(se);
1180
1181 if (tsk->state & TASK_INTERRUPTIBLE)
1182 se->statistics.sleep_start = rq_of(cfs_rq)->clock;
1183 if (tsk->state & TASK_UNINTERRUPTIBLE)
1184 se->statistics.block_start = rq_of(cfs_rq)->clock;
1185 }
1186#endif
1187 }
1188
1189 clear_buddies(cfs_rq, se);
1190
1191 if (se != cfs_rq->curr)
1192 __dequeue_entity(cfs_rq, se);
1193 se->on_rq = 0;
1194 update_cfs_load(cfs_rq, 0);
1195 account_entity_dequeue(cfs_rq, se);
1196
1197 /*
1198 * Normalize the entity after updating the min_vruntime because the
1199 * update can refer to the ->curr item and we need to reflect this
1200 * movement in our normalized position.
1201 */
1202 if (!(flags & DEQUEUE_SLEEP))
1203 se->vruntime -= cfs_rq->min_vruntime;
1204
1205 /* return excess runtime on last dequeue */
1206 return_cfs_rq_runtime(cfs_rq);
1207
1208 update_min_vruntime(cfs_rq);
1209 update_cfs_shares(cfs_rq);
1210}
1211
1212/*
1213 * Preempt the current task with a newly woken task if needed:
1214 */
1215static void
1216check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
1217{
1218 unsigned long ideal_runtime, delta_exec;
1219 struct sched_entity *se;
1220 s64 delta;
1221
1222 ideal_runtime = sched_slice(cfs_rq, curr);
1223 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
1224 if (delta_exec > ideal_runtime) {
1225 resched_task(rq_of(cfs_rq)->curr);
1226 /*
1227 * The current task ran long enough, ensure it doesn't get
1228 * re-elected due to buddy favours.
1229 */
1230 clear_buddies(cfs_rq, curr);
1231 return;
1232 }
1233
1234 /*
1235 * Ensure that a task that missed wakeup preemption by a
1236 * narrow margin doesn't have to wait for a full slice.
1237 * This also mitigates buddy induced latencies under load.
1238 */
1239 if (delta_exec < sysctl_sched_min_granularity)
1240 return;
1241
1242 se = __pick_first_entity(cfs_rq);
1243 delta = curr->vruntime - se->vruntime;
1244
1245 if (delta < 0)
1246 return;
1247
1248 if (delta > ideal_runtime)
1249 resched_task(rq_of(cfs_rq)->curr);
1250}
1251
1252static void
1253set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
1254{
1255 /* 'current' is not kept within the tree. */
1256 if (se->on_rq) {
1257 /*
1258 * Any task has to be enqueued before it get to execute on
1259 * a CPU. So account for the time it spent waiting on the
1260 * runqueue.
1261 */
1262 update_stats_wait_end(cfs_rq, se);
1263 __dequeue_entity(cfs_rq, se);
1264 }
1265
1266 update_stats_curr_start(cfs_rq, se);
1267 cfs_rq->curr = se;
1268#ifdef CONFIG_SCHEDSTATS
1269 /*
1270 * Track our maximum slice length, if the CPU's load is at
1271 * least twice that of our own weight (i.e. dont track it
1272 * when there are only lesser-weight tasks around):
1273 */
1274 if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
1275 se->statistics.slice_max = max(se->statistics.slice_max,
1276 se->sum_exec_runtime - se->prev_sum_exec_runtime);
1277 }
1278#endif
1279 se->prev_sum_exec_runtime = se->sum_exec_runtime;
1280}
1281
1282static int
1283wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
1284
1285/*
1286 * Pick the next process, keeping these things in mind, in this order:
1287 * 1) keep things fair between processes/task groups
1288 * 2) pick the "next" process, since someone really wants that to run
1289 * 3) pick the "last" process, for cache locality
1290 * 4) do not run the "skip" process, if something else is available
1291 */
1292static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
1293{
1294 struct sched_entity *se = __pick_first_entity(cfs_rq);
1295 struct sched_entity *left = se;
1296
1297 /*
1298 * Avoid running the skip buddy, if running something else can
1299 * be done without getting too unfair.
1300 */
1301 if (cfs_rq->skip == se) {
1302 struct sched_entity *second = __pick_next_entity(se);
1303 if (second && wakeup_preempt_entity(second, left) < 1)
1304 se = second;
1305 }
1306
1307 /*
1308 * Prefer last buddy, try to return the CPU to a preempted task.
1309 */
1310 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
1311 se = cfs_rq->last;
1312
1313 /*
1314 * Someone really wants this to run. If it's not unfair, run it.
1315 */
1316 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
1317 se = cfs_rq->next;
1318
1319 clear_buddies(cfs_rq, se);
1320
1321 return se;
1322}
1323
1324static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
1325
1326static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
1327{
1328 /*
1329 * If still on the runqueue then deactivate_task()
1330 * was not called and update_curr() has to be done:
1331 */
1332 if (prev->on_rq)
1333 update_curr(cfs_rq);
1334
1335 /* throttle cfs_rqs exceeding runtime */
1336 check_cfs_rq_runtime(cfs_rq);
1337
1338 check_spread(cfs_rq, prev);
1339 if (prev->on_rq) {
1340 update_stats_wait_start(cfs_rq, prev);
1341 /* Put 'current' back into the tree. */
1342 __enqueue_entity(cfs_rq, prev);
1343 }
1344 cfs_rq->curr = NULL;
1345}
1346
1347static void
1348entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
1349{
1350 /*
1351 * Update run-time statistics of the 'current'.
1352 */
1353 update_curr(cfs_rq);
1354
1355 /*
1356 * Update share accounting for long-running entities.
1357 */
1358 update_entity_shares_tick(cfs_rq);
1359
1360#ifdef CONFIG_SCHED_HRTICK
1361 /*
1362 * queued ticks are scheduled to match the slice, so don't bother
1363 * validating it and just reschedule.
1364 */
1365 if (queued) {
1366 resched_task(rq_of(cfs_rq)->curr);
1367 return;
1368 }
1369 /*
1370 * don't let the period tick interfere with the hrtick preemption
1371 */
1372 if (!sched_feat(DOUBLE_TICK) &&
1373 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
1374 return;
1375#endif
1376
1377 if (cfs_rq->nr_running > 1)
1378 check_preempt_tick(cfs_rq, curr);
1379}
1380
1381
1382/**************************************************
1383 * CFS bandwidth control machinery
1384 */
1385
1386#ifdef CONFIG_CFS_BANDWIDTH
1387
1388#ifdef HAVE_JUMP_LABEL
1389static struct static_key __cfs_bandwidth_used;
1390
1391static inline bool cfs_bandwidth_used(void)
1392{
1393 return static_key_false(&__cfs_bandwidth_used);
1394}
1395
1396void account_cfs_bandwidth_used(int enabled, int was_enabled)
1397{
1398 /* only need to count groups transitioning between enabled/!enabled */
1399 if (enabled && !was_enabled)
1400 static_key_slow_inc(&__cfs_bandwidth_used);
1401 else if (!enabled && was_enabled)
1402 static_key_slow_dec(&__cfs_bandwidth_used);
1403}
1404#else /* HAVE_JUMP_LABEL */
1405static bool cfs_bandwidth_used(void)
1406{
1407 return true;
1408}
1409
1410void account_cfs_bandwidth_used(int enabled, int was_enabled) {}
1411#endif /* HAVE_JUMP_LABEL */
1412
1413/*
1414 * default period for cfs group bandwidth.
1415 * default: 0.1s, units: nanoseconds
1416 */
1417static inline u64 default_cfs_period(void)
1418{
1419 return 100000000ULL;
1420}
1421
1422static inline u64 sched_cfs_bandwidth_slice(void)
1423{
1424 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
1425}
1426
1427/*
1428 * Replenish runtime according to assigned quota and update expiration time.
1429 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
1430 * additional synchronization around rq->lock.
1431 *
1432 * requires cfs_b->lock
1433 */
1434void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
1435{
1436 u64 now;
1437
1438 if (cfs_b->quota == RUNTIME_INF)
1439 return;
1440
1441 now = sched_clock_cpu(smp_processor_id());
1442 cfs_b->runtime = cfs_b->quota;
1443 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
1444}
1445
1446static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
1447{
1448 return &tg->cfs_bandwidth;
1449}
1450
1451/* returns 0 on failure to allocate runtime */
1452static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1453{
1454 struct task_group *tg = cfs_rq->tg;
1455 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
1456 u64 amount = 0, min_amount, expires;
1457
1458 /* note: this is a positive sum as runtime_remaining <= 0 */
1459 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
1460
1461 raw_spin_lock(&cfs_b->lock);
1462 if (cfs_b->quota == RUNTIME_INF)
1463 amount = min_amount;
1464 else {
1465 /*
1466 * If the bandwidth pool has become inactive, then at least one
1467 * period must have elapsed since the last consumption.
1468 * Refresh the global state and ensure bandwidth timer becomes
1469 * active.
1470 */
1471 if (!cfs_b->timer_active) {
1472 __refill_cfs_bandwidth_runtime(cfs_b);
1473 __start_cfs_bandwidth(cfs_b);
1474 }
1475
1476 if (cfs_b->runtime > 0) {
1477 amount = min(cfs_b->runtime, min_amount);
1478 cfs_b->runtime -= amount;
1479 cfs_b->idle = 0;
1480 }
1481 }
1482 expires = cfs_b->runtime_expires;
1483 raw_spin_unlock(&cfs_b->lock);
1484
1485 cfs_rq->runtime_remaining += amount;
1486 /*
1487 * we may have advanced our local expiration to account for allowed
1488 * spread between our sched_clock and the one on which runtime was
1489 * issued.
1490 */
1491 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
1492 cfs_rq->runtime_expires = expires;
1493
1494 return cfs_rq->runtime_remaining > 0;
1495}
1496
1497/*
1498 * Note: This depends on the synchronization provided by sched_clock and the
1499 * fact that rq->clock snapshots this value.
1500 */
1501static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1502{
1503 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1504 struct rq *rq = rq_of(cfs_rq);
1505
1506 /* if the deadline is ahead of our clock, nothing to do */
1507 if (likely((s64)(rq->clock - cfs_rq->runtime_expires) < 0))
1508 return;
1509
1510 if (cfs_rq->runtime_remaining < 0)
1511 return;
1512
1513 /*
1514 * If the local deadline has passed we have to consider the
1515 * possibility that our sched_clock is 'fast' and the global deadline
1516 * has not truly expired.
1517 *
1518 * Fortunately we can check determine whether this the case by checking
1519 * whether the global deadline has advanced.
1520 */
1521
1522 if ((s64)(cfs_rq->runtime_expires - cfs_b->runtime_expires) >= 0) {
1523 /* extend local deadline, drift is bounded above by 2 ticks */
1524 cfs_rq->runtime_expires += TICK_NSEC;
1525 } else {
1526 /* global deadline is ahead, expiration has passed */
1527 cfs_rq->runtime_remaining = 0;
1528 }
1529}
1530
1531static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq,
1532 unsigned long delta_exec)
1533{
1534 /* dock delta_exec before expiring quota (as it could span periods) */
1535 cfs_rq->runtime_remaining -= delta_exec;
1536 expire_cfs_rq_runtime(cfs_rq);
1537
1538 if (likely(cfs_rq->runtime_remaining > 0))
1539 return;
1540
1541 /*
1542 * if we're unable to extend our runtime we resched so that the active
1543 * hierarchy can be throttled
1544 */
1545 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
1546 resched_task(rq_of(cfs_rq)->curr);
1547}
1548
1549static __always_inline
1550void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec)
1551{
1552 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
1553 return;
1554
1555 __account_cfs_rq_runtime(cfs_rq, delta_exec);
1556}
1557
1558static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
1559{
1560 return cfs_bandwidth_used() && cfs_rq->throttled;
1561}
1562
1563/* check whether cfs_rq, or any parent, is throttled */
1564static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
1565{
1566 return cfs_bandwidth_used() && cfs_rq->throttle_count;
1567}
1568
1569/*
1570 * Ensure that neither of the group entities corresponding to src_cpu or
1571 * dest_cpu are members of a throttled hierarchy when performing group
1572 * load-balance operations.
1573 */
1574static inline int throttled_lb_pair(struct task_group *tg,
1575 int src_cpu, int dest_cpu)
1576{
1577 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
1578
1579 src_cfs_rq = tg->cfs_rq[src_cpu];
1580 dest_cfs_rq = tg->cfs_rq[dest_cpu];
1581
1582 return throttled_hierarchy(src_cfs_rq) ||
1583 throttled_hierarchy(dest_cfs_rq);
1584}
1585
1586/* updated child weight may affect parent so we have to do this bottom up */
1587static int tg_unthrottle_up(struct task_group *tg, void *data)
1588{
1589 struct rq *rq = data;
1590 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1591
1592 cfs_rq->throttle_count--;
1593#ifdef CONFIG_SMP
1594 if (!cfs_rq->throttle_count) {
1595 u64 delta = rq->clock_task - cfs_rq->load_stamp;
1596
1597 /* leaving throttled state, advance shares averaging windows */
1598 cfs_rq->load_stamp += delta;
1599 cfs_rq->load_last += delta;
1600
1601 /* update entity weight now that we are on_rq again */
1602 update_cfs_shares(cfs_rq);
1603 }
1604#endif
1605
1606 return 0;
1607}
1608
1609static int tg_throttle_down(struct task_group *tg, void *data)
1610{
1611 struct rq *rq = data;
1612 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
1613
1614 /* group is entering throttled state, record last load */
1615 if (!cfs_rq->throttle_count)
1616 update_cfs_load(cfs_rq, 0);
1617 cfs_rq->throttle_count++;
1618
1619 return 0;
1620}
1621
1622static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
1623{
1624 struct rq *rq = rq_of(cfs_rq);
1625 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1626 struct sched_entity *se;
1627 long task_delta, dequeue = 1;
1628
1629 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1630
1631 /* account load preceding throttle */
1632 rcu_read_lock();
1633 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
1634 rcu_read_unlock();
1635
1636 task_delta = cfs_rq->h_nr_running;
1637 for_each_sched_entity(se) {
1638 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
1639 /* throttled entity or throttle-on-deactivate */
1640 if (!se->on_rq)
1641 break;
1642
1643 if (dequeue)
1644 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
1645 qcfs_rq->h_nr_running -= task_delta;
1646
1647 if (qcfs_rq->load.weight)
1648 dequeue = 0;
1649 }
1650
1651 if (!se)
1652 rq->nr_running -= task_delta;
1653
1654 cfs_rq->throttled = 1;
1655 cfs_rq->throttled_timestamp = rq->clock;
1656 raw_spin_lock(&cfs_b->lock);
1657 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
1658 raw_spin_unlock(&cfs_b->lock);
1659}
1660
1661void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
1662{
1663 struct rq *rq = rq_of(cfs_rq);
1664 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1665 struct sched_entity *se;
1666 int enqueue = 1;
1667 long task_delta;
1668
1669 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
1670
1671 cfs_rq->throttled = 0;
1672 raw_spin_lock(&cfs_b->lock);
1673 cfs_b->throttled_time += rq->clock - cfs_rq->throttled_timestamp;
1674 list_del_rcu(&cfs_rq->throttled_list);
1675 raw_spin_unlock(&cfs_b->lock);
1676 cfs_rq->throttled_timestamp = 0;
1677
1678 update_rq_clock(rq);
1679 /* update hierarchical throttle state */
1680 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
1681
1682 if (!cfs_rq->load.weight)
1683 return;
1684
1685 task_delta = cfs_rq->h_nr_running;
1686 for_each_sched_entity(se) {
1687 if (se->on_rq)
1688 enqueue = 0;
1689
1690 cfs_rq = cfs_rq_of(se);
1691 if (enqueue)
1692 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
1693 cfs_rq->h_nr_running += task_delta;
1694
1695 if (cfs_rq_throttled(cfs_rq))
1696 break;
1697 }
1698
1699 if (!se)
1700 rq->nr_running += task_delta;
1701
1702 /* determine whether we need to wake up potentially idle cpu */
1703 if (rq->curr == rq->idle && rq->cfs.nr_running)
1704 resched_task(rq->curr);
1705}
1706
1707static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
1708 u64 remaining, u64 expires)
1709{
1710 struct cfs_rq *cfs_rq;
1711 u64 runtime = remaining;
1712
1713 rcu_read_lock();
1714 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
1715 throttled_list) {
1716 struct rq *rq = rq_of(cfs_rq);
1717
1718 raw_spin_lock(&rq->lock);
1719 if (!cfs_rq_throttled(cfs_rq))
1720 goto next;
1721
1722 runtime = -cfs_rq->runtime_remaining + 1;
1723 if (runtime > remaining)
1724 runtime = remaining;
1725 remaining -= runtime;
1726
1727 cfs_rq->runtime_remaining += runtime;
1728 cfs_rq->runtime_expires = expires;
1729
1730 /* we check whether we're throttled above */
1731 if (cfs_rq->runtime_remaining > 0)
1732 unthrottle_cfs_rq(cfs_rq);
1733
1734next:
1735 raw_spin_unlock(&rq->lock);
1736
1737 if (!remaining)
1738 break;
1739 }
1740 rcu_read_unlock();
1741
1742 return remaining;
1743}
1744
1745/*
1746 * Responsible for refilling a task_group's bandwidth and unthrottling its
1747 * cfs_rqs as appropriate. If there has been no activity within the last
1748 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
1749 * used to track this state.
1750 */
1751static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
1752{
1753 u64 runtime, runtime_expires;
1754 int idle = 1, throttled;
1755
1756 raw_spin_lock(&cfs_b->lock);
1757 /* no need to continue the timer with no bandwidth constraint */
1758 if (cfs_b->quota == RUNTIME_INF)
1759 goto out_unlock;
1760
1761 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1762 /* idle depends on !throttled (for the case of a large deficit) */
1763 idle = cfs_b->idle && !throttled;
1764 cfs_b->nr_periods += overrun;
1765
1766 /* if we're going inactive then everything else can be deferred */
1767 if (idle)
1768 goto out_unlock;
1769
1770 __refill_cfs_bandwidth_runtime(cfs_b);
1771
1772 if (!throttled) {
1773 /* mark as potentially idle for the upcoming period */
1774 cfs_b->idle = 1;
1775 goto out_unlock;
1776 }
1777
1778 /* account preceding periods in which throttling occurred */
1779 cfs_b->nr_throttled += overrun;
1780
1781 /*
1782 * There are throttled entities so we must first use the new bandwidth
1783 * to unthrottle them before making it generally available. This
1784 * ensures that all existing debts will be paid before a new cfs_rq is
1785 * allowed to run.
1786 */
1787 runtime = cfs_b->runtime;
1788 runtime_expires = cfs_b->runtime_expires;
1789 cfs_b->runtime = 0;
1790
1791 /*
1792 * This check is repeated as we are holding onto the new bandwidth
1793 * while we unthrottle. This can potentially race with an unthrottled
1794 * group trying to acquire new bandwidth from the global pool.
1795 */
1796 while (throttled && runtime > 0) {
1797 raw_spin_unlock(&cfs_b->lock);
1798 /* we can't nest cfs_b->lock while distributing bandwidth */
1799 runtime = distribute_cfs_runtime(cfs_b, runtime,
1800 runtime_expires);
1801 raw_spin_lock(&cfs_b->lock);
1802
1803 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
1804 }
1805
1806 /* return (any) remaining runtime */
1807 cfs_b->runtime = runtime;
1808 /*
1809 * While we are ensured activity in the period following an
1810 * unthrottle, this also covers the case in which the new bandwidth is
1811 * insufficient to cover the existing bandwidth deficit. (Forcing the
1812 * timer to remain active while there are any throttled entities.)
1813 */
1814 cfs_b->idle = 0;
1815out_unlock:
1816 if (idle)
1817 cfs_b->timer_active = 0;
1818 raw_spin_unlock(&cfs_b->lock);
1819
1820 return idle;
1821}
1822
1823/* a cfs_rq won't donate quota below this amount */
1824static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
1825/* minimum remaining period time to redistribute slack quota */
1826static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
1827/* how long we wait to gather additional slack before distributing */
1828static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
1829
1830/* are we near the end of the current quota period? */
1831static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
1832{
1833 struct hrtimer *refresh_timer = &cfs_b->period_timer;
1834 u64 remaining;
1835
1836 /* if the call-back is running a quota refresh is already occurring */
1837 if (hrtimer_callback_running(refresh_timer))
1838 return 1;
1839
1840 /* is a quota refresh about to occur? */
1841 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
1842 if (remaining < min_expire)
1843 return 1;
1844
1845 return 0;
1846}
1847
1848static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
1849{
1850 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
1851
1852 /* if there's a quota refresh soon don't bother with slack */
1853 if (runtime_refresh_within(cfs_b, min_left))
1854 return;
1855
1856 start_bandwidth_timer(&cfs_b->slack_timer,
1857 ns_to_ktime(cfs_bandwidth_slack_period));
1858}
1859
1860/* we know any runtime found here is valid as update_curr() precedes return */
1861static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1862{
1863 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
1864 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
1865
1866 if (slack_runtime <= 0)
1867 return;
1868
1869 raw_spin_lock(&cfs_b->lock);
1870 if (cfs_b->quota != RUNTIME_INF &&
1871 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
1872 cfs_b->runtime += slack_runtime;
1873
1874 /* we are under rq->lock, defer unthrottling using a timer */
1875 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
1876 !list_empty(&cfs_b->throttled_cfs_rq))
1877 start_cfs_slack_bandwidth(cfs_b);
1878 }
1879 raw_spin_unlock(&cfs_b->lock);
1880
1881 /* even if it's not valid for return we don't want to try again */
1882 cfs_rq->runtime_remaining -= slack_runtime;
1883}
1884
1885static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1886{
1887 if (!cfs_bandwidth_used())
1888 return;
1889
1890 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
1891 return;
1892
1893 __return_cfs_rq_runtime(cfs_rq);
1894}
1895
1896/*
1897 * This is done with a timer (instead of inline with bandwidth return) since
1898 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
1899 */
1900static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
1901{
1902 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
1903 u64 expires;
1904
1905 /* confirm we're still not at a refresh boundary */
1906 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration))
1907 return;
1908
1909 raw_spin_lock(&cfs_b->lock);
1910 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice) {
1911 runtime = cfs_b->runtime;
1912 cfs_b->runtime = 0;
1913 }
1914 expires = cfs_b->runtime_expires;
1915 raw_spin_unlock(&cfs_b->lock);
1916
1917 if (!runtime)
1918 return;
1919
1920 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
1921
1922 raw_spin_lock(&cfs_b->lock);
1923 if (expires == cfs_b->runtime_expires)
1924 cfs_b->runtime = runtime;
1925 raw_spin_unlock(&cfs_b->lock);
1926}
1927
1928/*
1929 * When a group wakes up we want to make sure that its quota is not already
1930 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
1931 * runtime as update_curr() throttling can not not trigger until it's on-rq.
1932 */
1933static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
1934{
1935 if (!cfs_bandwidth_used())
1936 return;
1937
1938 /* an active group must be handled by the update_curr()->put() path */
1939 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
1940 return;
1941
1942 /* ensure the group is not already throttled */
1943 if (cfs_rq_throttled(cfs_rq))
1944 return;
1945
1946 /* update runtime allocation */
1947 account_cfs_rq_runtime(cfs_rq, 0);
1948 if (cfs_rq->runtime_remaining <= 0)
1949 throttle_cfs_rq(cfs_rq);
1950}
1951
1952/* conditionally throttle active cfs_rq's from put_prev_entity() */
1953static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
1954{
1955 if (!cfs_bandwidth_used())
1956 return;
1957
1958 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
1959 return;
1960
1961 /*
1962 * it's possible for a throttled entity to be forced into a running
1963 * state (e.g. set_curr_task), in this case we're finished.
1964 */
1965 if (cfs_rq_throttled(cfs_rq))
1966 return;
1967
1968 throttle_cfs_rq(cfs_rq);
1969}
1970
1971static inline u64 default_cfs_period(void);
1972static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
1973static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
1974
1975static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
1976{
1977 struct cfs_bandwidth *cfs_b =
1978 container_of(timer, struct cfs_bandwidth, slack_timer);
1979 do_sched_cfs_slack_timer(cfs_b);
1980
1981 return HRTIMER_NORESTART;
1982}
1983
1984static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
1985{
1986 struct cfs_bandwidth *cfs_b =
1987 container_of(timer, struct cfs_bandwidth, period_timer);
1988 ktime_t now;
1989 int overrun;
1990 int idle = 0;
1991
1992 for (;;) {
1993 now = hrtimer_cb_get_time(timer);
1994 overrun = hrtimer_forward(timer, now, cfs_b->period);
1995
1996 if (!overrun)
1997 break;
1998
1999 idle = do_sched_cfs_period_timer(cfs_b, overrun);
2000 }
2001
2002 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
2003}
2004
2005void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2006{
2007 raw_spin_lock_init(&cfs_b->lock);
2008 cfs_b->runtime = 0;
2009 cfs_b->quota = RUNTIME_INF;
2010 cfs_b->period = ns_to_ktime(default_cfs_period());
2011
2012 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
2013 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2014 cfs_b->period_timer.function = sched_cfs_period_timer;
2015 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2016 cfs_b->slack_timer.function = sched_cfs_slack_timer;
2017}
2018
2019static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
2020{
2021 cfs_rq->runtime_enabled = 0;
2022 INIT_LIST_HEAD(&cfs_rq->throttled_list);
2023}
2024
2025/* requires cfs_b->lock, may release to reprogram timer */
2026void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2027{
2028 /*
2029 * The timer may be active because we're trying to set a new bandwidth
2030 * period or because we're racing with the tear-down path
2031 * (timer_active==0 becomes visible before the hrtimer call-back
2032 * terminates). In either case we ensure that it's re-programmed
2033 */
2034 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
2035 raw_spin_unlock(&cfs_b->lock);
2036 /* ensure cfs_b->lock is available while we wait */
2037 hrtimer_cancel(&cfs_b->period_timer);
2038
2039 raw_spin_lock(&cfs_b->lock);
2040 /* if someone else restarted the timer then we're done */
2041 if (cfs_b->timer_active)
2042 return;
2043 }
2044
2045 cfs_b->timer_active = 1;
2046 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
2047}
2048
2049static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
2050{
2051 hrtimer_cancel(&cfs_b->period_timer);
2052 hrtimer_cancel(&cfs_b->slack_timer);
2053}
2054
2055void unthrottle_offline_cfs_rqs(struct rq *rq)
2056{
2057 struct cfs_rq *cfs_rq;
2058
2059 for_each_leaf_cfs_rq(rq, cfs_rq) {
2060 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
2061
2062 if (!cfs_rq->runtime_enabled)
2063 continue;
2064
2065 /*
2066 * clock_task is not advancing so we just need to make sure
2067 * there's some valid quota amount
2068 */
2069 cfs_rq->runtime_remaining = cfs_b->quota;
2070 if (cfs_rq_throttled(cfs_rq))
2071 unthrottle_cfs_rq(cfs_rq);
2072 }
2073}
2074
2075#else /* CONFIG_CFS_BANDWIDTH */
2076static __always_inline
2077void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, unsigned long delta_exec) {}
2078static void check_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2079static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
2080static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2081
2082static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
2083{
2084 return 0;
2085}
2086
2087static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
2088{
2089 return 0;
2090}
2091
2092static inline int throttled_lb_pair(struct task_group *tg,
2093 int src_cpu, int dest_cpu)
2094{
2095 return 0;
2096}
2097
2098void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2099
2100#ifdef CONFIG_FAIR_GROUP_SCHED
2101static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
2102#endif
2103
2104static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
2105{
2106 return NULL;
2107}
2108static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
2109void unthrottle_offline_cfs_rqs(struct rq *rq) {}
2110
2111#endif /* CONFIG_CFS_BANDWIDTH */
2112
2113/**************************************************
2114 * CFS operations on tasks:
2115 */
2116
2117#ifdef CONFIG_SCHED_HRTICK
2118static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
2119{
2120 struct sched_entity *se = &p->se;
2121 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2122
2123 WARN_ON(task_rq(p) != rq);
2124
2125 if (cfs_rq->nr_running > 1) {
2126 u64 slice = sched_slice(cfs_rq, se);
2127 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
2128 s64 delta = slice - ran;
2129
2130 if (delta < 0) {
2131 if (rq->curr == p)
2132 resched_task(p);
2133 return;
2134 }
2135
2136 /*
2137 * Don't schedule slices shorter than 10000ns, that just
2138 * doesn't make sense. Rely on vruntime for fairness.
2139 */
2140 if (rq->curr != p)
2141 delta = max_t(s64, 10000LL, delta);
2142
2143 hrtick_start(rq, delta);
2144 }
2145}
2146
2147/*
2148 * called from enqueue/dequeue and updates the hrtick when the
2149 * current task is from our class and nr_running is low enough
2150 * to matter.
2151 */
2152static void hrtick_update(struct rq *rq)
2153{
2154 struct task_struct *curr = rq->curr;
2155
2156 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
2157 return;
2158
2159 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
2160 hrtick_start_fair(rq, curr);
2161}
2162#else /* !CONFIG_SCHED_HRTICK */
2163static inline void
2164hrtick_start_fair(struct rq *rq, struct task_struct *p)
2165{
2166}
2167
2168static inline void hrtick_update(struct rq *rq)
2169{
2170}
2171#endif
2172
2173/*
2174 * The enqueue_task method is called before nr_running is
2175 * increased. Here we update the fair scheduling stats and
2176 * then put the task into the rbtree:
2177 */
2178static void
2179enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2180{
2181 struct cfs_rq *cfs_rq;
2182 struct sched_entity *se = &p->se;
2183
2184 for_each_sched_entity(se) {
2185 if (se->on_rq)
2186 break;
2187 cfs_rq = cfs_rq_of(se);
2188 enqueue_entity(cfs_rq, se, flags);
2189
2190 /*
2191 * end evaluation on encountering a throttled cfs_rq
2192 *
2193 * note: in the case of encountering a throttled cfs_rq we will
2194 * post the final h_nr_running increment below.
2195 */
2196 if (cfs_rq_throttled(cfs_rq))
2197 break;
2198 cfs_rq->h_nr_running++;
2199
2200 flags = ENQUEUE_WAKEUP;
2201 }
2202
2203 for_each_sched_entity(se) {
2204 cfs_rq = cfs_rq_of(se);
2205 cfs_rq->h_nr_running++;
2206
2207 if (cfs_rq_throttled(cfs_rq))
2208 break;
2209
2210 update_cfs_load(cfs_rq, 0);
2211 update_cfs_shares(cfs_rq);
2212 }
2213
2214 if (!se)
2215 inc_nr_running(rq);
2216 hrtick_update(rq);
2217}
2218
2219static void set_next_buddy(struct sched_entity *se);
2220
2221/*
2222 * The dequeue_task method is called before nr_running is
2223 * decreased. We remove the task from the rbtree and
2224 * update the fair scheduling stats:
2225 */
2226static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
2227{
2228 struct cfs_rq *cfs_rq;
2229 struct sched_entity *se = &p->se;
2230 int task_sleep = flags & DEQUEUE_SLEEP;
2231
2232 for_each_sched_entity(se) {
2233 cfs_rq = cfs_rq_of(se);
2234 dequeue_entity(cfs_rq, se, flags);
2235
2236 /*
2237 * end evaluation on encountering a throttled cfs_rq
2238 *
2239 * note: in the case of encountering a throttled cfs_rq we will
2240 * post the final h_nr_running decrement below.
2241 */
2242 if (cfs_rq_throttled(cfs_rq))
2243 break;
2244 cfs_rq->h_nr_running--;
2245
2246 /* Don't dequeue parent if it has other entities besides us */
2247 if (cfs_rq->load.weight) {
2248 /*
2249 * Bias pick_next to pick a task from this cfs_rq, as
2250 * p is sleeping when it is within its sched_slice.
2251 */
2252 if (task_sleep && parent_entity(se))
2253 set_next_buddy(parent_entity(se));
2254
2255 /* avoid re-evaluating load for this entity */
2256 se = parent_entity(se);
2257 break;
2258 }
2259 flags |= DEQUEUE_SLEEP;
2260 }
2261
2262 for_each_sched_entity(se) {
2263 cfs_rq = cfs_rq_of(se);
2264 cfs_rq->h_nr_running--;
2265
2266 if (cfs_rq_throttled(cfs_rq))
2267 break;
2268
2269 update_cfs_load(cfs_rq, 0);
2270 update_cfs_shares(cfs_rq);
2271 }
2272
2273 if (!se)
2274 dec_nr_running(rq);
2275 hrtick_update(rq);
2276}
2277
2278#ifdef CONFIG_SMP
2279/* Used instead of source_load when we know the type == 0 */
2280static unsigned long weighted_cpuload(const int cpu)
2281{
2282 return cpu_rq(cpu)->load.weight;
2283}
2284
2285/*
2286 * Return a low guess at the load of a migration-source cpu weighted
2287 * according to the scheduling class and "nice" value.
2288 *
2289 * We want to under-estimate the load of migration sources, to
2290 * balance conservatively.
2291 */
2292static unsigned long source_load(int cpu, int type)
2293{
2294 struct rq *rq = cpu_rq(cpu);
2295 unsigned long total = weighted_cpuload(cpu);
2296
2297 if (type == 0 || !sched_feat(LB_BIAS))
2298 return total;
2299
2300 return min(rq->cpu_load[type-1], total);
2301}
2302
2303/*
2304 * Return a high guess at the load of a migration-target cpu weighted
2305 * according to the scheduling class and "nice" value.
2306 */
2307static unsigned long target_load(int cpu, int type)
2308{
2309 struct rq *rq = cpu_rq(cpu);
2310 unsigned long total = weighted_cpuload(cpu);
2311
2312 if (type == 0 || !sched_feat(LB_BIAS))
2313 return total;
2314
2315 return max(rq->cpu_load[type-1], total);
2316}
2317
2318static unsigned long power_of(int cpu)
2319{
2320 return cpu_rq(cpu)->cpu_power;
2321}
2322
2323static unsigned long cpu_avg_load_per_task(int cpu)
2324{
2325 struct rq *rq = cpu_rq(cpu);
2326 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
2327
2328 if (nr_running)
2329 return rq->load.weight / nr_running;
2330
2331 return 0;
2332}
2333
2334
2335static void task_waking_fair(struct task_struct *p)
2336{
2337 struct sched_entity *se = &p->se;
2338 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2339 u64 min_vruntime;
2340
2341#ifndef CONFIG_64BIT
2342 u64 min_vruntime_copy;
2343
2344 do {
2345 min_vruntime_copy = cfs_rq->min_vruntime_copy;
2346 smp_rmb();
2347 min_vruntime = cfs_rq->min_vruntime;
2348 } while (min_vruntime != min_vruntime_copy);
2349#else
2350 min_vruntime = cfs_rq->min_vruntime;
2351#endif
2352
2353 se->vruntime -= min_vruntime;
2354}
2355
2356#ifdef CONFIG_FAIR_GROUP_SCHED
2357/*
2358 * effective_load() calculates the load change as seen from the root_task_group
2359 *
2360 * Adding load to a group doesn't make a group heavier, but can cause movement
2361 * of group shares between cpus. Assuming the shares were perfectly aligned one
2362 * can calculate the shift in shares.
2363 *
2364 * Calculate the effective load difference if @wl is added (subtracted) to @tg
2365 * on this @cpu and results in a total addition (subtraction) of @wg to the
2366 * total group weight.
2367 *
2368 * Given a runqueue weight distribution (rw_i) we can compute a shares
2369 * distribution (s_i) using:
2370 *
2371 * s_i = rw_i / \Sum rw_j (1)
2372 *
2373 * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
2374 * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
2375 * shares distribution (s_i):
2376 *
2377 * rw_i = { 2, 4, 1, 0 }
2378 * s_i = { 2/7, 4/7, 1/7, 0 }
2379 *
2380 * As per wake_affine() we're interested in the load of two CPUs (the CPU the
2381 * task used to run on and the CPU the waker is running on), we need to
2382 * compute the effect of waking a task on either CPU and, in case of a sync
2383 * wakeup, compute the effect of the current task going to sleep.
2384 *
2385 * So for a change of @wl to the local @cpu with an overall group weight change
2386 * of @wl we can compute the new shares distribution (s'_i) using:
2387 *
2388 * s'_i = (rw_i + @wl) / (@wg + \Sum rw_j) (2)
2389 *
2390 * Suppose we're interested in CPUs 0 and 1, and want to compute the load
2391 * differences in waking a task to CPU 0. The additional task changes the
2392 * weight and shares distributions like:
2393 *
2394 * rw'_i = { 3, 4, 1, 0 }
2395 * s'_i = { 3/8, 4/8, 1/8, 0 }
2396 *
2397 * We can then compute the difference in effective weight by using:
2398 *
2399 * dw_i = S * (s'_i - s_i) (3)
2400 *
2401 * Where 'S' is the group weight as seen by its parent.
2402 *
2403 * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
2404 * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
2405 * 4/7) times the weight of the group.
2406 */
2407static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
2408{
2409 struct sched_entity *se = tg->se[cpu];
2410
2411 if (!tg->parent) /* the trivial, non-cgroup case */
2412 return wl;
2413
2414 for_each_sched_entity(se) {
2415 long w, W;
2416
2417 tg = se->my_q->tg;
2418
2419 /*
2420 * W = @wg + \Sum rw_j
2421 */
2422 W = wg + calc_tg_weight(tg, se->my_q);
2423
2424 /*
2425 * w = rw_i + @wl
2426 */
2427 w = se->my_q->load.weight + wl;
2428
2429 /*
2430 * wl = S * s'_i; see (2)
2431 */
2432 if (W > 0 && w < W)
2433 wl = (w * tg->shares) / W;
2434 else
2435 wl = tg->shares;
2436
2437 /*
2438 * Per the above, wl is the new se->load.weight value; since
2439 * those are clipped to [MIN_SHARES, ...) do so now. See
2440 * calc_cfs_shares().
2441 */
2442 if (wl < MIN_SHARES)
2443 wl = MIN_SHARES;
2444
2445 /*
2446 * wl = dw_i = S * (s'_i - s_i); see (3)
2447 */
2448 wl -= se->load.weight;
2449
2450 /*
2451 * Recursively apply this logic to all parent groups to compute
2452 * the final effective load change on the root group. Since
2453 * only the @tg group gets extra weight, all parent groups can
2454 * only redistribute existing shares. @wl is the shift in shares
2455 * resulting from this level per the above.
2456 */
2457 wg = 0;
2458 }
2459
2460 return wl;
2461}
2462#else
2463
2464static inline unsigned long effective_load(struct task_group *tg, int cpu,
2465 unsigned long wl, unsigned long wg)
2466{
2467 return wl;
2468}
2469
2470#endif
2471
2472static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
2473{
2474 s64 this_load, load;
2475 int idx, this_cpu, prev_cpu;
2476 unsigned long tl_per_task;
2477 struct task_group *tg;
2478 unsigned long weight;
2479 int balanced;
2480
2481 idx = sd->wake_idx;
2482 this_cpu = smp_processor_id();
2483 prev_cpu = task_cpu(p);
2484 load = source_load(prev_cpu, idx);
2485 this_load = target_load(this_cpu, idx);
2486
2487 /*
2488 * If sync wakeup then subtract the (maximum possible)
2489 * effect of the currently running task from the load
2490 * of the current CPU:
2491 */
2492 if (sync) {
2493 tg = task_group(current);
2494 weight = current->se.load.weight;
2495
2496 this_load += effective_load(tg, this_cpu, -weight, -weight);
2497 load += effective_load(tg, prev_cpu, 0, -weight);
2498 }
2499
2500 tg = task_group(p);
2501 weight = p->se.load.weight;
2502
2503 /*
2504 * In low-load situations, where prev_cpu is idle and this_cpu is idle
2505 * due to the sync cause above having dropped this_load to 0, we'll
2506 * always have an imbalance, but there's really nothing you can do
2507 * about that, so that's good too.
2508 *
2509 * Otherwise check if either cpus are near enough in load to allow this
2510 * task to be woken on this_cpu.
2511 */
2512 if (this_load > 0) {
2513 s64 this_eff_load, prev_eff_load;
2514
2515 this_eff_load = 100;
2516 this_eff_load *= power_of(prev_cpu);
2517 this_eff_load *= this_load +
2518 effective_load(tg, this_cpu, weight, weight);
2519
2520 prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
2521 prev_eff_load *= power_of(this_cpu);
2522 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
2523
2524 balanced = this_eff_load <= prev_eff_load;
2525 } else
2526 balanced = true;
2527
2528 /*
2529 * If the currently running task will sleep within
2530 * a reasonable amount of time then attract this newly
2531 * woken task:
2532 */
2533 if (sync && balanced)
2534 return 1;
2535
2536 schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
2537 tl_per_task = cpu_avg_load_per_task(this_cpu);
2538
2539 if (balanced ||
2540 (this_load <= load &&
2541 this_load + target_load(prev_cpu, idx) <= tl_per_task)) {
2542 /*
2543 * This domain has SD_WAKE_AFFINE and
2544 * p is cache cold in this domain, and
2545 * there is no bad imbalance.
2546 */
2547 schedstat_inc(sd, ttwu_move_affine);
2548 schedstat_inc(p, se.statistics.nr_wakeups_affine);
2549
2550 return 1;
2551 }
2552 return 0;
2553}
2554
2555/*
2556 * find_idlest_group finds and returns the least busy CPU group within the
2557 * domain.
2558 */
2559static struct sched_group *
2560find_idlest_group(struct sched_domain *sd, struct task_struct *p,
2561 int this_cpu, int load_idx)
2562{
2563 struct sched_group *idlest = NULL, *group = sd->groups;
2564 unsigned long min_load = ULONG_MAX, this_load = 0;
2565 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2566
2567 do {
2568 unsigned long load, avg_load;
2569 int local_group;
2570 int i;
2571
2572 /* Skip over this group if it has no CPUs allowed */
2573 if (!cpumask_intersects(sched_group_cpus(group),
2574 tsk_cpus_allowed(p)))
2575 continue;
2576
2577 local_group = cpumask_test_cpu(this_cpu,
2578 sched_group_cpus(group));
2579
2580 /* Tally up the load of all CPUs in the group */
2581 avg_load = 0;
2582
2583 for_each_cpu(i, sched_group_cpus(group)) {
2584 /* Bias balancing toward cpus of our domain */
2585 if (local_group)
2586 load = source_load(i, load_idx);
2587 else
2588 load = target_load(i, load_idx);
2589
2590 avg_load += load;
2591 }
2592
2593 /* Adjust by relative CPU power of the group */
2594 avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
2595
2596 if (local_group) {
2597 this_load = avg_load;
2598 } else if (avg_load < min_load) {
2599 min_load = avg_load;
2600 idlest = group;
2601 }
2602 } while (group = group->next, group != sd->groups);
2603
2604 if (!idlest || 100*this_load < imbalance*min_load)
2605 return NULL;
2606 return idlest;
2607}
2608
2609/*
2610 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2611 */
2612static int
2613find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2614{
2615 unsigned long load, min_load = ULONG_MAX;
2616 int idlest = -1;
2617 int i;
2618
2619 /* Traverse only the allowed CPUs */
2620 for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
2621 load = weighted_cpuload(i);
2622
2623 if (load < min_load || (load == min_load && i == this_cpu)) {
2624 min_load = load;
2625 idlest = i;
2626 }
2627 }
2628
2629 return idlest;
2630}
2631
2632/*
2633 * Try and locate an idle CPU in the sched_domain.
2634 */
2635static int select_idle_sibling(struct task_struct *p, int target)
2636{
2637 int cpu = smp_processor_id();
2638 int prev_cpu = task_cpu(p);
2639 struct sched_domain *sd;
2640 struct sched_group *sg;
2641 int i;
2642
2643 /*
2644 * If the task is going to be woken-up on this cpu and if it is
2645 * already idle, then it is the right target.
2646 */
2647 if (target == cpu && idle_cpu(cpu))
2648 return cpu;
2649
2650 /*
2651 * If the task is going to be woken-up on the cpu where it previously
2652 * ran and if it is currently idle, then it the right target.
2653 */
2654 if (target == prev_cpu && idle_cpu(prev_cpu))
2655 return prev_cpu;
2656
2657 /*
2658 * Otherwise, iterate the domains and find an elegible idle cpu.
2659 */
2660 sd = rcu_dereference(per_cpu(sd_llc, target));
2661 for_each_lower_domain(sd) {
2662 sg = sd->groups;
2663 do {
2664 if (!cpumask_intersects(sched_group_cpus(sg),
2665 tsk_cpus_allowed(p)))
2666 goto next;
2667
2668 for_each_cpu(i, sched_group_cpus(sg)) {
2669 if (!idle_cpu(i))
2670 goto next;
2671 }
2672
2673 target = cpumask_first_and(sched_group_cpus(sg),
2674 tsk_cpus_allowed(p));
2675 goto done;
2676next:
2677 sg = sg->next;
2678 } while (sg != sd->groups);
2679 }
2680done:
2681 return target;
2682}
2683
2684/*
2685 * sched_balance_self: balance the current task (running on cpu) in domains
2686 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2687 * SD_BALANCE_EXEC.
2688 *
2689 * Balance, ie. select the least loaded group.
2690 *
2691 * Returns the target CPU number, or the same CPU if no balancing is needed.
2692 *
2693 * preempt must be disabled.
2694 */
2695static int
2696select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
2697{
2698 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
2699 int cpu = smp_processor_id();
2700 int prev_cpu = task_cpu(p);
2701 int new_cpu = cpu;
2702 int want_affine = 0;
2703 int want_sd = 1;
2704 int sync = wake_flags & WF_SYNC;
2705
2706 if (p->nr_cpus_allowed == 1)
2707 return prev_cpu;
2708
2709 if (sd_flag & SD_BALANCE_WAKE) {
2710 if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
2711 want_affine = 1;
2712 new_cpu = prev_cpu;
2713 }
2714
2715 rcu_read_lock();
2716 for_each_domain(cpu, tmp) {
2717 if (!(tmp->flags & SD_LOAD_BALANCE))
2718 continue;
2719
2720 /*
2721 * If power savings logic is enabled for a domain, see if we
2722 * are not overloaded, if so, don't balance wider.
2723 */
2724 if (tmp->flags & (SD_PREFER_LOCAL)) {
2725 unsigned long power = 0;
2726 unsigned long nr_running = 0;
2727 unsigned long capacity;
2728 int i;
2729
2730 for_each_cpu(i, sched_domain_span(tmp)) {
2731 power += power_of(i);
2732 nr_running += cpu_rq(i)->cfs.nr_running;
2733 }
2734
2735 capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
2736
2737 if (nr_running < capacity)
2738 want_sd = 0;
2739 }
2740
2741 /*
2742 * If both cpu and prev_cpu are part of this domain,
2743 * cpu is a valid SD_WAKE_AFFINE target.
2744 */
2745 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
2746 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
2747 affine_sd = tmp;
2748 want_affine = 0;
2749 }
2750
2751 if (!want_sd && !want_affine)
2752 break;
2753
2754 if (!(tmp->flags & sd_flag))
2755 continue;
2756
2757 if (want_sd)
2758 sd = tmp;
2759 }
2760
2761 if (affine_sd) {
2762 if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
2763 prev_cpu = cpu;
2764
2765 new_cpu = select_idle_sibling(p, prev_cpu);
2766 goto unlock;
2767 }
2768
2769 while (sd) {
2770 int load_idx = sd->forkexec_idx;
2771 struct sched_group *group;
2772 int weight;
2773
2774 if (!(sd->flags & sd_flag)) {
2775 sd = sd->child;
2776 continue;
2777 }
2778
2779 if (sd_flag & SD_BALANCE_WAKE)
2780 load_idx = sd->wake_idx;
2781
2782 group = find_idlest_group(sd, p, cpu, load_idx);
2783 if (!group) {
2784 sd = sd->child;
2785 continue;
2786 }
2787
2788 new_cpu = find_idlest_cpu(group, p, cpu);
2789 if (new_cpu == -1 || new_cpu == cpu) {
2790 /* Now try balancing at a lower domain level of cpu */
2791 sd = sd->child;
2792 continue;
2793 }
2794
2795 /* Now try balancing at a lower domain level of new_cpu */
2796 cpu = new_cpu;
2797 weight = sd->span_weight;
2798 sd = NULL;
2799 for_each_domain(cpu, tmp) {
2800 if (weight <= tmp->span_weight)
2801 break;
2802 if (tmp->flags & sd_flag)
2803 sd = tmp;
2804 }
2805 /* while loop will break here if sd == NULL */
2806 }
2807unlock:
2808 rcu_read_unlock();
2809
2810 return new_cpu;
2811}
2812#endif /* CONFIG_SMP */
2813
2814static unsigned long
2815wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
2816{
2817 unsigned long gran = sysctl_sched_wakeup_granularity;
2818
2819 /*
2820 * Since its curr running now, convert the gran from real-time
2821 * to virtual-time in his units.
2822 *
2823 * By using 'se' instead of 'curr' we penalize light tasks, so
2824 * they get preempted easier. That is, if 'se' < 'curr' then
2825 * the resulting gran will be larger, therefore penalizing the
2826 * lighter, if otoh 'se' > 'curr' then the resulting gran will
2827 * be smaller, again penalizing the lighter task.
2828 *
2829 * This is especially important for buddies when the leftmost
2830 * task is higher priority than the buddy.
2831 */
2832 return calc_delta_fair(gran, se);
2833}
2834
2835/*
2836 * Should 'se' preempt 'curr'.
2837 *
2838 * |s1
2839 * |s2
2840 * |s3
2841 * g
2842 * |<--->|c
2843 *
2844 * w(c, s1) = -1
2845 * w(c, s2) = 0
2846 * w(c, s3) = 1
2847 *
2848 */
2849static int
2850wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
2851{
2852 s64 gran, vdiff = curr->vruntime - se->vruntime;
2853
2854 if (vdiff <= 0)
2855 return -1;
2856
2857 gran = wakeup_gran(curr, se);
2858 if (vdiff > gran)
2859 return 1;
2860
2861 return 0;
2862}
2863
2864static void set_last_buddy(struct sched_entity *se)
2865{
2866 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2867 return;
2868
2869 for_each_sched_entity(se)
2870 cfs_rq_of(se)->last = se;
2871}
2872
2873static void set_next_buddy(struct sched_entity *se)
2874{
2875 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
2876 return;
2877
2878 for_each_sched_entity(se)
2879 cfs_rq_of(se)->next = se;
2880}
2881
2882static void set_skip_buddy(struct sched_entity *se)
2883{
2884 for_each_sched_entity(se)
2885 cfs_rq_of(se)->skip = se;
2886}
2887
2888/*
2889 * Preempt the current task with a newly woken task if needed:
2890 */
2891static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2892{
2893 struct task_struct *curr = rq->curr;
2894 struct sched_entity *se = &curr->se, *pse = &p->se;
2895 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
2896 int scale = cfs_rq->nr_running >= sched_nr_latency;
2897 int next_buddy_marked = 0;
2898
2899 if (unlikely(se == pse))
2900 return;
2901
2902 /*
2903 * This is possible from callers such as move_task(), in which we
2904 * unconditionally check_prempt_curr() after an enqueue (which may have
2905 * lead to a throttle). This both saves work and prevents false
2906 * next-buddy nomination below.
2907 */
2908 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
2909 return;
2910
2911 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
2912 set_next_buddy(pse);
2913 next_buddy_marked = 1;
2914 }
2915
2916 /*
2917 * We can come here with TIF_NEED_RESCHED already set from new task
2918 * wake up path.
2919 *
2920 * Note: this also catches the edge-case of curr being in a throttled
2921 * group (e.g. via set_curr_task), since update_curr() (in the
2922 * enqueue of curr) will have resulted in resched being set. This
2923 * prevents us from potentially nominating it as a false LAST_BUDDY
2924 * below.
2925 */
2926 if (test_tsk_need_resched(curr))
2927 return;
2928
2929 /* Idle tasks are by definition preempted by non-idle tasks. */
2930 if (unlikely(curr->policy == SCHED_IDLE) &&
2931 likely(p->policy != SCHED_IDLE))
2932 goto preempt;
2933
2934 /*
2935 * Batch and idle tasks do not preempt non-idle tasks (their preemption
2936 * is driven by the tick):
2937 */
2938 if (unlikely(p->policy != SCHED_NORMAL))
2939 return;
2940
2941 find_matching_se(&se, &pse);
2942 update_curr(cfs_rq_of(se));
2943 BUG_ON(!pse);
2944 if (wakeup_preempt_entity(se, pse) == 1) {
2945 /*
2946 * Bias pick_next to pick the sched entity that is
2947 * triggering this preemption.
2948 */
2949 if (!next_buddy_marked)
2950 set_next_buddy(pse);
2951 goto preempt;
2952 }
2953
2954 return;
2955
2956preempt:
2957 resched_task(curr);
2958 /*
2959 * Only set the backward buddy when the current task is still
2960 * on the rq. This can happen when a wakeup gets interleaved
2961 * with schedule on the ->pre_schedule() or idle_balance()
2962 * point, either of which can * drop the rq lock.
2963 *
2964 * Also, during early boot the idle thread is in the fair class,
2965 * for obvious reasons its a bad idea to schedule back to it.
2966 */
2967 if (unlikely(!se->on_rq || curr == rq->idle))
2968 return;
2969
2970 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
2971 set_last_buddy(se);
2972}
2973
2974static struct task_struct *pick_next_task_fair(struct rq *rq)
2975{
2976 struct task_struct *p;
2977 struct cfs_rq *cfs_rq = &rq->cfs;
2978 struct sched_entity *se;
2979
2980 if (!cfs_rq->nr_running)
2981 return NULL;
2982
2983 do {
2984 se = pick_next_entity(cfs_rq);
2985 set_next_entity(cfs_rq, se);
2986 cfs_rq = group_cfs_rq(se);
2987 } while (cfs_rq);
2988
2989 p = task_of(se);
2990 if (hrtick_enabled(rq))
2991 hrtick_start_fair(rq, p);
2992
2993 return p;
2994}
2995
2996/*
2997 * Account for a descheduled task:
2998 */
2999static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
3000{
3001 struct sched_entity *se = &prev->se;
3002 struct cfs_rq *cfs_rq;
3003
3004 for_each_sched_entity(se) {
3005 cfs_rq = cfs_rq_of(se);
3006 put_prev_entity(cfs_rq, se);
3007 }
3008}
3009
3010/*
3011 * sched_yield() is very simple
3012 *
3013 * The magic of dealing with the ->skip buddy is in pick_next_entity.
3014 */
3015static void yield_task_fair(struct rq *rq)
3016{
3017 struct task_struct *curr = rq->curr;
3018 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
3019 struct sched_entity *se = &curr->se;
3020
3021 /*
3022 * Are we the only task in the tree?
3023 */
3024 if (unlikely(rq->nr_running == 1))
3025 return;
3026
3027 clear_buddies(cfs_rq, se);
3028
3029 if (curr->policy != SCHED_BATCH) {
3030 update_rq_clock(rq);
3031 /*
3032 * Update run-time statistics of the 'current'.
3033 */
3034 update_curr(cfs_rq);
3035 /*
3036 * Tell update_rq_clock() that we've just updated,
3037 * so we don't do microscopic update in schedule()
3038 * and double the fastpath cost.
3039 */
3040 rq->skip_clock_update = 1;
3041 }
3042
3043 set_skip_buddy(se);
3044}
3045
3046static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
3047{
3048 struct sched_entity *se = &p->se;
3049
3050 /* throttled hierarchies are not runnable */
3051 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
3052 return false;
3053
3054 /* Tell the scheduler that we'd really like pse to run next. */
3055 set_next_buddy(se);
3056
3057 yield_task_fair(rq);
3058
3059 return true;
3060}
3061
3062#ifdef CONFIG_SMP
3063/**************************************************
3064 * Fair scheduling class load-balancing methods:
3065 */
3066
3067static unsigned long __read_mostly max_load_balance_interval = HZ/10;
3068
3069#define LBF_ALL_PINNED 0x01
3070#define LBF_NEED_BREAK 0x02
3071
3072struct lb_env {
3073 struct sched_domain *sd;
3074
3075 int src_cpu;
3076 struct rq *src_rq;
3077
3078 int dst_cpu;
3079 struct rq *dst_rq;
3080
3081 enum cpu_idle_type idle;
3082 long imbalance;
3083 unsigned int flags;
3084
3085 unsigned int loop;
3086 unsigned int loop_break;
3087 unsigned int loop_max;
3088};
3089
3090/*
3091 * move_task - move a task from one runqueue to another runqueue.
3092 * Both runqueues must be locked.
3093 */
3094static void move_task(struct task_struct *p, struct lb_env *env)
3095{
3096 deactivate_task(env->src_rq, p, 0);
3097 set_task_cpu(p, env->dst_cpu);
3098 activate_task(env->dst_rq, p, 0);
3099 check_preempt_curr(env->dst_rq, p, 0);
3100}
3101
3102/*
3103 * Is this task likely cache-hot:
3104 */
3105static int
3106task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
3107{
3108 s64 delta;
3109
3110 if (p->sched_class != &fair_sched_class)
3111 return 0;
3112
3113 if (unlikely(p->policy == SCHED_IDLE))
3114 return 0;
3115
3116 /*
3117 * Buddy candidates are cache hot:
3118 */
3119 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
3120 (&p->se == cfs_rq_of(&p->se)->next ||
3121 &p->se == cfs_rq_of(&p->se)->last))
3122 return 1;
3123
3124 if (sysctl_sched_migration_cost == -1)
3125 return 1;
3126 if (sysctl_sched_migration_cost == 0)
3127 return 0;
3128
3129 delta = now - p->se.exec_start;
3130
3131 return delta < (s64)sysctl_sched_migration_cost;
3132}
3133
3134/*
3135 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3136 */
3137static
3138int can_migrate_task(struct task_struct *p, struct lb_env *env)
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(env->dst_cpu, tsk_cpus_allowed(p))) {
3148 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
3149 return 0;
3150 }
3151 env->flags &= ~LBF_ALL_PINNED;
3152
3153 if (task_running(env->src_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, env->src_rq->clock_task, env->sd);
3165 if (!tsk_cache_hot ||
3166 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
3167#ifdef CONFIG_SCHEDSTATS
3168 if (tsk_cache_hot) {
3169 schedstat_inc(env->sd, lb_hot_gained[env->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 move_one_task(struct lb_env *env)
3191{
3192 struct task_struct *p, *n;
3193
3194 list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
3195 if (throttled_lb_pair(task_group(p), env->src_rq->cpu, env->dst_cpu))
3196 continue;
3197
3198 if (!can_migrate_task(p, env))
3199 continue;
3200
3201 move_task(p, env);
3202 /*
3203 * Right now, this is only the second place move_task()
3204 * is called, so we can safely collect move_task()
3205 * stats here rather than inside move_task().
3206 */
3207 schedstat_inc(env->sd, lb_gained[env->idle]);
3208 return 1;
3209 }
3210 return 0;
3211}
3212
3213static unsigned long task_h_load(struct task_struct *p);
3214
3215static const unsigned int sched_nr_migrate_break = 32;
3216
3217/*
3218 * move_tasks tries to move up to imbalance weighted load from busiest to
3219 * this_rq, as part of a balancing operation within domain "sd".
3220 * Returns 1 if successful and 0 otherwise.
3221 *
3222 * Called with both runqueues locked.
3223 */
3224static int move_tasks(struct lb_env *env)
3225{
3226 struct list_head *tasks = &env->src_rq->cfs_tasks;
3227 struct task_struct *p;
3228 unsigned long load;
3229 int pulled = 0;
3230
3231 if (env->imbalance <= 0)
3232 return 0;
3233
3234 while (!list_empty(tasks)) {
3235 p = list_first_entry(tasks, struct task_struct, se.group_node);
3236
3237 env->loop++;
3238 /* We've more or less seen every task there is, call it quits */
3239 if (env->loop > env->loop_max)
3240 break;
3241
3242 /* take a breather every nr_migrate tasks */
3243 if (env->loop > env->loop_break) {
3244 env->loop_break += sched_nr_migrate_break;
3245 env->flags |= LBF_NEED_BREAK;
3246 break;
3247 }
3248
3249 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
3250 goto next;
3251
3252 load = task_h_load(p);
3253
3254 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
3255 goto next;
3256
3257 if ((load / 2) > env->imbalance)
3258 goto next;
3259
3260 if (!can_migrate_task(p, env))
3261 goto next;
3262
3263 move_task(p, env);
3264 pulled++;
3265 env->imbalance -= load;
3266
3267#ifdef CONFIG_PREEMPT
3268 /*
3269 * NEWIDLE balancing is a source of latency, so preemptible
3270 * kernels will stop after the first task is pulled to minimize
3271 * the critical section.
3272 */
3273 if (env->idle == CPU_NEWLY_IDLE)
3274 break;
3275#endif
3276
3277 /*
3278 * We only want to steal up to the prescribed amount of
3279 * weighted load.
3280 */
3281 if (env->imbalance <= 0)
3282 break;
3283
3284 continue;
3285next:
3286 list_move_tail(&p->se.group_node, tasks);
3287 }
3288
3289 /*
3290 * Right now, this is one of only two places move_task() is called,
3291 * so we can safely collect move_task() stats here rather than
3292 * inside move_task().
3293 */
3294 schedstat_add(env->sd, lb_gained[env->idle], pulled);
3295
3296 return pulled;
3297}
3298
3299#ifdef CONFIG_FAIR_GROUP_SCHED
3300/*
3301 * update tg->load_weight by folding this cpu's load_avg
3302 */
3303static int update_shares_cpu(struct task_group *tg, int cpu)
3304{
3305 struct cfs_rq *cfs_rq;
3306 unsigned long flags;
3307 struct rq *rq;
3308
3309 if (!tg->se[cpu])
3310 return 0;
3311
3312 rq = cpu_rq(cpu);
3313 cfs_rq = tg->cfs_rq[cpu];
3314
3315 raw_spin_lock_irqsave(&rq->lock, flags);
3316
3317 update_rq_clock(rq);
3318 update_cfs_load(cfs_rq, 1);
3319
3320 /*
3321 * We need to update shares after updating tg->load_weight in
3322 * order to adjust the weight of groups with long running tasks.
3323 */
3324 update_cfs_shares(cfs_rq);
3325
3326 raw_spin_unlock_irqrestore(&rq->lock, flags);
3327
3328 return 0;
3329}
3330
3331static void update_shares(int cpu)
3332{
3333 struct cfs_rq *cfs_rq;
3334 struct rq *rq = cpu_rq(cpu);
3335
3336 rcu_read_lock();
3337 /*
3338 * Iterates the task_group tree in a bottom up fashion, see
3339 * list_add_leaf_cfs_rq() for details.
3340 */
3341 for_each_leaf_cfs_rq(rq, cfs_rq) {
3342 /* throttled entities do not contribute to load */
3343 if (throttled_hierarchy(cfs_rq))
3344 continue;
3345
3346 update_shares_cpu(cfs_rq->tg, cpu);
3347 }
3348 rcu_read_unlock();
3349}
3350
3351/*
3352 * Compute the cpu's hierarchical load factor for each task group.
3353 * This needs to be done in a top-down fashion because the load of a child
3354 * group is a fraction of its parents load.
3355 */
3356static int tg_load_down(struct task_group *tg, void *data)
3357{
3358 unsigned long load;
3359 long cpu = (long)data;
3360
3361 if (!tg->parent) {
3362 load = cpu_rq(cpu)->load.weight;
3363 } else {
3364 load = tg->parent->cfs_rq[cpu]->h_load;
3365 load *= tg->se[cpu]->load.weight;
3366 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
3367 }
3368
3369 tg->cfs_rq[cpu]->h_load = load;
3370
3371 return 0;
3372}
3373
3374static void update_h_load(long cpu)
3375{
3376 rcu_read_lock();
3377 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
3378 rcu_read_unlock();
3379}
3380
3381static unsigned long task_h_load(struct task_struct *p)
3382{
3383 struct cfs_rq *cfs_rq = task_cfs_rq(p);
3384 unsigned long load;
3385
3386 load = p->se.load.weight;
3387 load = div_u64(load * cfs_rq->h_load, cfs_rq->load.weight + 1);
3388
3389 return load;
3390}
3391#else
3392static inline void update_shares(int cpu)
3393{
3394}
3395
3396static inline void update_h_load(long cpu)
3397{
3398}
3399
3400static unsigned long task_h_load(struct task_struct *p)
3401{
3402 return p->se.load.weight;
3403}
3404#endif
3405
3406/********** Helpers for find_busiest_group ************************/
3407/*
3408 * sd_lb_stats - Structure to store the statistics of a sched_domain
3409 * during load balancing.
3410 */
3411struct sd_lb_stats {
3412 struct sched_group *busiest; /* Busiest group in this sd */
3413 struct sched_group *this; /* Local group in this sd */
3414 unsigned long total_load; /* Total load of all groups in sd */
3415 unsigned long total_pwr; /* Total power of all groups in sd */
3416 unsigned long avg_load; /* Average load across all groups in sd */
3417
3418 /** Statistics of this group */
3419 unsigned long this_load;
3420 unsigned long this_load_per_task;
3421 unsigned long this_nr_running;
3422 unsigned long this_has_capacity;
3423 unsigned int this_idle_cpus;
3424
3425 /* Statistics of the busiest group */
3426 unsigned int busiest_idle_cpus;
3427 unsigned long max_load;
3428 unsigned long busiest_load_per_task;
3429 unsigned long busiest_nr_running;
3430 unsigned long busiest_group_capacity;
3431 unsigned long busiest_has_capacity;
3432 unsigned int busiest_group_weight;
3433
3434 int group_imb; /* Is there imbalance in this sd */
3435};
3436
3437/*
3438 * sg_lb_stats - stats of a sched_group required for load_balancing
3439 */
3440struct sg_lb_stats {
3441 unsigned long avg_load; /*Avg load across the CPUs of the group */
3442 unsigned long group_load; /* Total load over the CPUs of the group */
3443 unsigned long sum_nr_running; /* Nr tasks running in the group */
3444 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3445 unsigned long group_capacity;
3446 unsigned long idle_cpus;
3447 unsigned long group_weight;
3448 int group_imb; /* Is there an imbalance in the group ? */
3449 int group_has_capacity; /* Is there extra capacity in the group? */
3450};
3451
3452/**
3453 * get_sd_load_idx - Obtain the load index for a given sched domain.
3454 * @sd: The sched_domain whose load_idx is to be obtained.
3455 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3456 */
3457static inline int get_sd_load_idx(struct sched_domain *sd,
3458 enum cpu_idle_type idle)
3459{
3460 int load_idx;
3461
3462 switch (idle) {
3463 case CPU_NOT_IDLE:
3464 load_idx = sd->busy_idx;
3465 break;
3466
3467 case CPU_NEWLY_IDLE:
3468 load_idx = sd->newidle_idx;
3469 break;
3470 default:
3471 load_idx = sd->idle_idx;
3472 break;
3473 }
3474
3475 return load_idx;
3476}
3477
3478unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3479{
3480 return SCHED_POWER_SCALE;
3481}
3482
3483unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3484{
3485 return default_scale_freq_power(sd, cpu);
3486}
3487
3488unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3489{
3490 unsigned long weight = sd->span_weight;
3491 unsigned long smt_gain = sd->smt_gain;
3492
3493 smt_gain /= weight;
3494
3495 return smt_gain;
3496}
3497
3498unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3499{
3500 return default_scale_smt_power(sd, cpu);
3501}
3502
3503unsigned long scale_rt_power(int cpu)
3504{
3505 struct rq *rq = cpu_rq(cpu);
3506 u64 total, available, age_stamp, avg;
3507
3508 /*
3509 * Since we're reading these variables without serialization make sure
3510 * we read them once before doing sanity checks on them.
3511 */
3512 age_stamp = ACCESS_ONCE(rq->age_stamp);
3513 avg = ACCESS_ONCE(rq->rt_avg);
3514
3515 total = sched_avg_period() + (rq->clock - age_stamp);
3516
3517 if (unlikely(total < avg)) {
3518 /* Ensures that power won't end up being negative */
3519 available = 0;
3520 } else {
3521 available = total - avg;
3522 }
3523
3524 if (unlikely((s64)total < SCHED_POWER_SCALE))
3525 total = SCHED_POWER_SCALE;
3526
3527 total >>= SCHED_POWER_SHIFT;
3528
3529 return div_u64(available, total);
3530}
3531
3532static void update_cpu_power(struct sched_domain *sd, int cpu)
3533{
3534 unsigned long weight = sd->span_weight;
3535 unsigned long power = SCHED_POWER_SCALE;
3536 struct sched_group *sdg = sd->groups;
3537
3538 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3539 if (sched_feat(ARCH_POWER))
3540 power *= arch_scale_smt_power(sd, cpu);
3541 else
3542 power *= default_scale_smt_power(sd, cpu);
3543
3544 power >>= SCHED_POWER_SHIFT;
3545 }
3546
3547 sdg->sgp->power_orig = power;
3548
3549 if (sched_feat(ARCH_POWER))
3550 power *= arch_scale_freq_power(sd, cpu);
3551 else
3552 power *= default_scale_freq_power(sd, cpu);
3553
3554 power >>= SCHED_POWER_SHIFT;
3555
3556 power *= scale_rt_power(cpu);
3557 power >>= SCHED_POWER_SHIFT;
3558
3559 if (!power)
3560 power = 1;
3561
3562 cpu_rq(cpu)->cpu_power = power;
3563 sdg->sgp->power = power;
3564}
3565
3566void update_group_power(struct sched_domain *sd, int cpu)
3567{
3568 struct sched_domain *child = sd->child;
3569 struct sched_group *group, *sdg = sd->groups;
3570 unsigned long power;
3571 unsigned long interval;
3572
3573 interval = msecs_to_jiffies(sd->balance_interval);
3574 interval = clamp(interval, 1UL, max_load_balance_interval);
3575 sdg->sgp->next_update = jiffies + interval;
3576
3577 if (!child) {
3578 update_cpu_power(sd, cpu);
3579 return;
3580 }
3581
3582 power = 0;
3583
3584 if (child->flags & SD_OVERLAP) {
3585 /*
3586 * SD_OVERLAP domains cannot assume that child groups
3587 * span the current group.
3588 */
3589
3590 for_each_cpu(cpu, sched_group_cpus(sdg))
3591 power += power_of(cpu);
3592 } else {
3593 /*
3594 * !SD_OVERLAP domains can assume that child groups
3595 * span the current group.
3596 */
3597
3598 group = child->groups;
3599 do {
3600 power += group->sgp->power;
3601 group = group->next;
3602 } while (group != child->groups);
3603 }
3604
3605 sdg->sgp->power_orig = sdg->sgp->power = power;
3606}
3607
3608/*
3609 * Try and fix up capacity for tiny siblings, this is needed when
3610 * things like SD_ASYM_PACKING need f_b_g to select another sibling
3611 * which on its own isn't powerful enough.
3612 *
3613 * See update_sd_pick_busiest() and check_asym_packing().
3614 */
3615static inline int
3616fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
3617{
3618 /*
3619 * Only siblings can have significantly less than SCHED_POWER_SCALE
3620 */
3621 if (!(sd->flags & SD_SHARE_CPUPOWER))
3622 return 0;
3623
3624 /*
3625 * If ~90% of the cpu_power is still there, we're good.
3626 */
3627 if (group->sgp->power * 32 > group->sgp->power_orig * 29)
3628 return 1;
3629
3630 return 0;
3631}
3632
3633/**
3634 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3635 * @env: The load balancing environment.
3636 * @group: sched_group whose statistics are to be updated.
3637 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3638 * @local_group: Does group contain this_cpu.
3639 * @cpus: Set of cpus considered for load balancing.
3640 * @balance: Should we balance.
3641 * @sgs: variable to hold the statistics for this group.
3642 */
3643static inline void update_sg_lb_stats(struct lb_env *env,
3644 struct sched_group *group, int load_idx,
3645 int local_group, const struct cpumask *cpus,
3646 int *balance, struct sg_lb_stats *sgs)
3647{
3648 unsigned long nr_running, max_nr_running, min_nr_running;
3649 unsigned long load, max_cpu_load, min_cpu_load;
3650 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3651 unsigned long avg_load_per_task = 0;
3652 int i;
3653
3654 if (local_group)
3655 balance_cpu = group_balance_cpu(group);
3656
3657 /* Tally up the load of all CPUs in the group */
3658 max_cpu_load = 0;
3659 min_cpu_load = ~0UL;
3660 max_nr_running = 0;
3661 min_nr_running = ~0UL;
3662
3663 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3664 struct rq *rq = cpu_rq(i);
3665
3666 nr_running = rq->nr_running;
3667
3668 /* Bias balancing toward cpus of our domain */
3669 if (local_group) {
3670 if (idle_cpu(i) && !first_idle_cpu &&
3671 cpumask_test_cpu(i, sched_group_mask(group))) {
3672 first_idle_cpu = 1;
3673 balance_cpu = i;
3674 }
3675
3676 load = target_load(i, load_idx);
3677 } else {
3678 load = source_load(i, load_idx);
3679 if (load > max_cpu_load)
3680 max_cpu_load = load;
3681 if (min_cpu_load > load)
3682 min_cpu_load = load;
3683
3684 if (nr_running > max_nr_running)
3685 max_nr_running = nr_running;
3686 if (min_nr_running > nr_running)
3687 min_nr_running = nr_running;
3688 }
3689
3690 sgs->group_load += load;
3691 sgs->sum_nr_running += nr_running;
3692 sgs->sum_weighted_load += weighted_cpuload(i);
3693 if (idle_cpu(i))
3694 sgs->idle_cpus++;
3695 }
3696
3697 /*
3698 * First idle cpu or the first cpu(busiest) in this sched group
3699 * is eligible for doing load balancing at this and above
3700 * domains. In the newly idle case, we will allow all the cpu's
3701 * to do the newly idle load balance.
3702 */
3703 if (local_group) {
3704 if (env->idle != CPU_NEWLY_IDLE) {
3705 if (balance_cpu != env->dst_cpu) {
3706 *balance = 0;
3707 return;
3708 }
3709 update_group_power(env->sd, env->dst_cpu);
3710 } else if (time_after_eq(jiffies, group->sgp->next_update))
3711 update_group_power(env->sd, env->dst_cpu);
3712 }
3713
3714 /* Adjust by relative CPU power of the group */
3715 sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
3716
3717 /*
3718 * Consider the group unbalanced when the imbalance is larger
3719 * than the average weight of a task.
3720 *
3721 * APZ: with cgroup the avg task weight can vary wildly and
3722 * might not be a suitable number - should we keep a
3723 * normalized nr_running number somewhere that negates
3724 * the hierarchy?
3725 */
3726 if (sgs->sum_nr_running)
3727 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3728
3729 if ((max_cpu_load - min_cpu_load) >= avg_load_per_task &&
3730 (max_nr_running - min_nr_running) > 1)
3731 sgs->group_imb = 1;
3732
3733 sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
3734 SCHED_POWER_SCALE);
3735 if (!sgs->group_capacity)
3736 sgs->group_capacity = fix_small_capacity(env->sd, group);
3737 sgs->group_weight = group->group_weight;
3738
3739 if (sgs->group_capacity > sgs->sum_nr_running)
3740 sgs->group_has_capacity = 1;
3741}
3742
3743/**
3744 * update_sd_pick_busiest - return 1 on busiest group
3745 * @env: The load balancing environment.
3746 * @sds: sched_domain statistics
3747 * @sg: sched_group candidate to be checked for being the busiest
3748 * @sgs: sched_group statistics
3749 *
3750 * Determine if @sg is a busier group than the previously selected
3751 * busiest group.
3752 */
3753static bool update_sd_pick_busiest(struct lb_env *env,
3754 struct sd_lb_stats *sds,
3755 struct sched_group *sg,
3756 struct sg_lb_stats *sgs)
3757{
3758 if (sgs->avg_load <= sds->max_load)
3759 return false;
3760
3761 if (sgs->sum_nr_running > sgs->group_capacity)
3762 return true;
3763
3764 if (sgs->group_imb)
3765 return true;
3766
3767 /*
3768 * ASYM_PACKING needs to move all the work to the lowest
3769 * numbered CPUs in the group, therefore mark all groups
3770 * higher than ourself as busy.
3771 */
3772 if ((env->sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
3773 env->dst_cpu < group_first_cpu(sg)) {
3774 if (!sds->busiest)
3775 return true;
3776
3777 if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
3778 return true;
3779 }
3780
3781 return false;
3782}
3783
3784/**
3785 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
3786 * @env: The load balancing environment.
3787 * @cpus: Set of cpus considered for load balancing.
3788 * @balance: Should we balance.
3789 * @sds: variable to hold the statistics for this sched_domain.
3790 */
3791static inline void update_sd_lb_stats(struct lb_env *env,
3792 const struct cpumask *cpus,
3793 int *balance, struct sd_lb_stats *sds)
3794{
3795 struct sched_domain *child = env->sd->child;
3796 struct sched_group *sg = env->sd->groups;
3797 struct sg_lb_stats sgs;
3798 int load_idx, prefer_sibling = 0;
3799
3800 if (child && child->flags & SD_PREFER_SIBLING)
3801 prefer_sibling = 1;
3802
3803 load_idx = get_sd_load_idx(env->sd, env->idle);
3804
3805 do {
3806 int local_group;
3807
3808 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
3809 memset(&sgs, 0, sizeof(sgs));
3810 update_sg_lb_stats(env, sg, load_idx, local_group,
3811 cpus, balance, &sgs);
3812
3813 if (local_group && !(*balance))
3814 return;
3815
3816 sds->total_load += sgs.group_load;
3817 sds->total_pwr += sg->sgp->power;
3818
3819 /*
3820 * In case the child domain prefers tasks go to siblings
3821 * first, lower the sg capacity to one so that we'll try
3822 * and move all the excess tasks away. We lower the capacity
3823 * of a group only if the local group has the capacity to fit
3824 * these excess tasks, i.e. nr_running < group_capacity. The
3825 * extra check prevents the case where you always pull from the
3826 * heaviest group when it is already under-utilized (possible
3827 * with a large weight task outweighs the tasks on the system).
3828 */
3829 if (prefer_sibling && !local_group && sds->this_has_capacity)
3830 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3831
3832 if (local_group) {
3833 sds->this_load = sgs.avg_load;
3834 sds->this = sg;
3835 sds->this_nr_running = sgs.sum_nr_running;
3836 sds->this_load_per_task = sgs.sum_weighted_load;
3837 sds->this_has_capacity = sgs.group_has_capacity;
3838 sds->this_idle_cpus = sgs.idle_cpus;
3839 } else if (update_sd_pick_busiest(env, sds, sg, &sgs)) {
3840 sds->max_load = sgs.avg_load;
3841 sds->busiest = sg;
3842 sds->busiest_nr_running = sgs.sum_nr_running;
3843 sds->busiest_idle_cpus = sgs.idle_cpus;
3844 sds->busiest_group_capacity = sgs.group_capacity;
3845 sds->busiest_load_per_task = sgs.sum_weighted_load;
3846 sds->busiest_has_capacity = sgs.group_has_capacity;
3847 sds->busiest_group_weight = sgs.group_weight;
3848 sds->group_imb = sgs.group_imb;
3849 }
3850
3851 sg = sg->next;
3852 } while (sg != env->sd->groups);
3853}
3854
3855/**
3856 * check_asym_packing - Check to see if the group is packed into the
3857 * sched doman.
3858 *
3859 * This is primarily intended to used at the sibling level. Some
3860 * cores like POWER7 prefer to use lower numbered SMT threads. In the
3861 * case of POWER7, it can move to lower SMT modes only when higher
3862 * threads are idle. When in lower SMT modes, the threads will
3863 * perform better since they share less core resources. Hence when we
3864 * have idle threads, we want them to be the higher ones.
3865 *
3866 * This packing function is run on idle threads. It checks to see if
3867 * the busiest CPU in this domain (core in the P7 case) has a higher
3868 * CPU number than the packing function is being run on. Here we are
3869 * assuming lower CPU number will be equivalent to lower a SMT thread
3870 * number.
3871 *
3872 * Returns 1 when packing is required and a task should be moved to
3873 * this CPU. The amount of the imbalance is returned in *imbalance.
3874 *
3875 * @env: The load balancing environment.
3876 * @sds: Statistics of the sched_domain which is to be packed
3877 */
3878static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
3879{
3880 int busiest_cpu;
3881
3882 if (!(env->sd->flags & SD_ASYM_PACKING))
3883 return 0;
3884
3885 if (!sds->busiest)
3886 return 0;
3887
3888 busiest_cpu = group_first_cpu(sds->busiest);
3889 if (env->dst_cpu > busiest_cpu)
3890 return 0;
3891
3892 env->imbalance = DIV_ROUND_CLOSEST(
3893 sds->max_load * sds->busiest->sgp->power, SCHED_POWER_SCALE);
3894
3895 return 1;
3896}
3897
3898/**
3899 * fix_small_imbalance - Calculate the minor imbalance that exists
3900 * amongst the groups of a sched_domain, during
3901 * load balancing.
3902 * @env: The load balancing environment.
3903 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3904 */
3905static inline
3906void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
3907{
3908 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3909 unsigned int imbn = 2;
3910 unsigned long scaled_busy_load_per_task;
3911
3912 if (sds->this_nr_running) {
3913 sds->this_load_per_task /= sds->this_nr_running;
3914 if (sds->busiest_load_per_task >
3915 sds->this_load_per_task)
3916 imbn = 1;
3917 } else {
3918 sds->this_load_per_task =
3919 cpu_avg_load_per_task(env->dst_cpu);
3920 }
3921
3922 scaled_busy_load_per_task = sds->busiest_load_per_task
3923 * SCHED_POWER_SCALE;
3924 scaled_busy_load_per_task /= sds->busiest->sgp->power;
3925
3926 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3927 (scaled_busy_load_per_task * imbn)) {
3928 env->imbalance = sds->busiest_load_per_task;
3929 return;
3930 }
3931
3932 /*
3933 * OK, we don't have enough imbalance to justify moving tasks,
3934 * however we may be able to increase total CPU power used by
3935 * moving them.
3936 */
3937
3938 pwr_now += sds->busiest->sgp->power *
3939 min(sds->busiest_load_per_task, sds->max_load);
3940 pwr_now += sds->this->sgp->power *
3941 min(sds->this_load_per_task, sds->this_load);
3942 pwr_now /= SCHED_POWER_SCALE;
3943
3944 /* Amount of load we'd subtract */
3945 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3946 sds->busiest->sgp->power;
3947 if (sds->max_load > tmp)
3948 pwr_move += sds->busiest->sgp->power *
3949 min(sds->busiest_load_per_task, sds->max_load - tmp);
3950
3951 /* Amount of load we'd add */
3952 if (sds->max_load * sds->busiest->sgp->power <
3953 sds->busiest_load_per_task * SCHED_POWER_SCALE)
3954 tmp = (sds->max_load * sds->busiest->sgp->power) /
3955 sds->this->sgp->power;
3956 else
3957 tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
3958 sds->this->sgp->power;
3959 pwr_move += sds->this->sgp->power *
3960 min(sds->this_load_per_task, sds->this_load + tmp);
3961 pwr_move /= SCHED_POWER_SCALE;
3962
3963 /* Move if we gain throughput */
3964 if (pwr_move > pwr_now)
3965 env->imbalance = sds->busiest_load_per_task;
3966}
3967
3968/**
3969 * calculate_imbalance - Calculate the amount of imbalance present within the
3970 * groups of a given sched_domain during load balance.
3971 * @env: load balance environment
3972 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3973 */
3974static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
3975{
3976 unsigned long max_pull, load_above_capacity = ~0UL;
3977
3978 sds->busiest_load_per_task /= sds->busiest_nr_running;
3979 if (sds->group_imb) {
3980 sds->busiest_load_per_task =
3981 min(sds->busiest_load_per_task, sds->avg_load);
3982 }
3983
3984 /*
3985 * In the presence of smp nice balancing, certain scenarios can have
3986 * max load less than avg load(as we skip the groups at or below
3987 * its cpu_power, while calculating max_load..)
3988 */
3989 if (sds->max_load < sds->avg_load) {
3990 env->imbalance = 0;
3991 return fix_small_imbalance(env, sds);
3992 }
3993
3994 if (!sds->group_imb) {
3995 /*
3996 * Don't want to pull so many tasks that a group would go idle.
3997 */
3998 load_above_capacity = (sds->busiest_nr_running -
3999 sds->busiest_group_capacity);
4000
4001 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
4002
4003 load_above_capacity /= sds->busiest->sgp->power;
4004 }
4005
4006 /*
4007 * We're trying to get all the cpus to the average_load, so we don't
4008 * want to push ourselves above the average load, nor do we wish to
4009 * reduce the max loaded cpu below the average load. At the same time,
4010 * we also don't want to reduce the group load below the group capacity
4011 * (so that we can implement power-savings policies etc). Thus we look
4012 * for the minimum possible imbalance.
4013 * Be careful of negative numbers as they'll appear as very large values
4014 * with unsigned longs.
4015 */
4016 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4017
4018 /* How much load to actually move to equalise the imbalance */
4019 env->imbalance = min(max_pull * sds->busiest->sgp->power,
4020 (sds->avg_load - sds->this_load) * sds->this->sgp->power)
4021 / SCHED_POWER_SCALE;
4022
4023 /*
4024 * if *imbalance is less than the average load per runnable task
4025 * there is no guarantee that any tasks will be moved so we'll have
4026 * a think about bumping its value to force at least one task to be
4027 * moved
4028 */
4029 if (env->imbalance < sds->busiest_load_per_task)
4030 return fix_small_imbalance(env, sds);
4031
4032}
4033
4034/******* find_busiest_group() helpers end here *********************/
4035
4036/**
4037 * find_busiest_group - Returns the busiest group within the sched_domain
4038 * if there is an imbalance. If there isn't an imbalance, and
4039 * the user has opted for power-savings, it returns a group whose
4040 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4041 * such a group exists.
4042 *
4043 * Also calculates the amount of weighted load which should be moved
4044 * to restore balance.
4045 *
4046 * @env: The load balancing environment.
4047 * @cpus: The set of CPUs under consideration for load-balancing.
4048 * @balance: Pointer to a variable indicating if this_cpu
4049 * is the appropriate cpu to perform load balancing at this_level.
4050 *
4051 * Returns: - the busiest group if imbalance exists.
4052 * - If no imbalance and user has opted for power-savings balance,
4053 * return the least loaded group whose CPUs can be
4054 * put to idle by rebalancing its tasks onto our group.
4055 */
4056static struct sched_group *
4057find_busiest_group(struct lb_env *env, const struct cpumask *cpus, int *balance)
4058{
4059 struct sd_lb_stats sds;
4060
4061 memset(&sds, 0, sizeof(sds));
4062
4063 /*
4064 * Compute the various statistics relavent for load balancing at
4065 * this level.
4066 */
4067 update_sd_lb_stats(env, cpus, balance, &sds);
4068
4069 /*
4070 * this_cpu is not the appropriate cpu to perform load balancing at
4071 * this level.
4072 */
4073 if (!(*balance))
4074 goto ret;
4075
4076 if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
4077 check_asym_packing(env, &sds))
4078 return sds.busiest;
4079
4080 /* There is no busy sibling group to pull tasks from */
4081 if (!sds.busiest || sds.busiest_nr_running == 0)
4082 goto out_balanced;
4083
4084 sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
4085
4086 /*
4087 * If the busiest group is imbalanced the below checks don't
4088 * work because they assumes all things are equal, which typically
4089 * isn't true due to cpus_allowed constraints and the like.
4090 */
4091 if (sds.group_imb)
4092 goto force_balance;
4093
4094 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4095 if (env->idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4096 !sds.busiest_has_capacity)
4097 goto force_balance;
4098
4099 /*
4100 * If the local group is more busy than the selected busiest group
4101 * don't try and pull any tasks.
4102 */
4103 if (sds.this_load >= sds.max_load)
4104 goto out_balanced;
4105
4106 /*
4107 * Don't pull any tasks if this group is already above the domain
4108 * average load.
4109 */
4110 if (sds.this_load >= sds.avg_load)
4111 goto out_balanced;
4112
4113 if (env->idle == CPU_IDLE) {
4114 /*
4115 * This cpu is idle. If the busiest group load doesn't
4116 * have more tasks than the number of available cpu's and
4117 * there is no imbalance between this and busiest group
4118 * wrt to idle cpu's, it is balanced.
4119 */
4120 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4121 sds.busiest_nr_running <= sds.busiest_group_weight)
4122 goto out_balanced;
4123 } else {
4124 /*
4125 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
4126 * imbalance_pct to be conservative.
4127 */
4128 if (100 * sds.max_load <= env->sd->imbalance_pct * sds.this_load)
4129 goto out_balanced;
4130 }
4131
4132force_balance:
4133 /* Looks like there is an imbalance. Compute it */
4134 calculate_imbalance(env, &sds);
4135 return sds.busiest;
4136
4137out_balanced:
4138ret:
4139 env->imbalance = 0;
4140 return NULL;
4141}
4142
4143/*
4144 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4145 */
4146static struct rq *find_busiest_queue(struct lb_env *env,
4147 struct sched_group *group,
4148 const struct cpumask *cpus)
4149{
4150 struct rq *busiest = NULL, *rq;
4151 unsigned long max_load = 0;
4152 int i;
4153
4154 for_each_cpu(i, sched_group_cpus(group)) {
4155 unsigned long power = power_of(i);
4156 unsigned long capacity = DIV_ROUND_CLOSEST(power,
4157 SCHED_POWER_SCALE);
4158 unsigned long wl;
4159
4160 if (!capacity)
4161 capacity = fix_small_capacity(env->sd, group);
4162
4163 if (!cpumask_test_cpu(i, cpus))
4164 continue;
4165
4166 rq = cpu_rq(i);
4167 wl = weighted_cpuload(i);
4168
4169 /*
4170 * When comparing with imbalance, use weighted_cpuload()
4171 * which is not scaled with the cpu power.
4172 */
4173 if (capacity && rq->nr_running == 1 && wl > env->imbalance)
4174 continue;
4175
4176 /*
4177 * For the load comparisons with the other cpu's, consider
4178 * the weighted_cpuload() scaled with the cpu power, so that
4179 * the load can be moved away from the cpu that is potentially
4180 * running at a lower capacity.
4181 */
4182 wl = (wl * SCHED_POWER_SCALE) / power;
4183
4184 if (wl > max_load) {
4185 max_load = wl;
4186 busiest = rq;
4187 }
4188 }
4189
4190 return busiest;
4191}
4192
4193/*
4194 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4195 * so long as it is large enough.
4196 */
4197#define MAX_PINNED_INTERVAL 512
4198
4199/* Working cpumask for load_balance and load_balance_newidle. */
4200DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4201
4202static int need_active_balance(struct lb_env *env)
4203{
4204 struct sched_domain *sd = env->sd;
4205
4206 if (env->idle == CPU_NEWLY_IDLE) {
4207
4208 /*
4209 * ASYM_PACKING needs to force migrate tasks from busy but
4210 * higher numbered CPUs in order to pack all tasks in the
4211 * lowest numbered CPUs.
4212 */
4213 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
4214 return 1;
4215 }
4216
4217 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
4218}
4219
4220static int active_load_balance_cpu_stop(void *data);
4221
4222/*
4223 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4224 * tasks if there is an imbalance.
4225 */
4226static int load_balance(int this_cpu, struct rq *this_rq,
4227 struct sched_domain *sd, enum cpu_idle_type idle,
4228 int *balance)
4229{
4230 int ld_moved, active_balance = 0;
4231 struct sched_group *group;
4232 struct rq *busiest;
4233 unsigned long flags;
4234 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4235
4236 struct lb_env env = {
4237 .sd = sd,
4238 .dst_cpu = this_cpu,
4239 .dst_rq = this_rq,
4240 .idle = idle,
4241 .loop_break = sched_nr_migrate_break,
4242 };
4243
4244 cpumask_copy(cpus, cpu_active_mask);
4245
4246 schedstat_inc(sd, lb_count[idle]);
4247
4248redo:
4249 group = find_busiest_group(&env, cpus, balance);
4250
4251 if (*balance == 0)
4252 goto out_balanced;
4253
4254 if (!group) {
4255 schedstat_inc(sd, lb_nobusyg[idle]);
4256 goto out_balanced;
4257 }
4258
4259 busiest = find_busiest_queue(&env, group, cpus);
4260 if (!busiest) {
4261 schedstat_inc(sd, lb_nobusyq[idle]);
4262 goto out_balanced;
4263 }
4264
4265 BUG_ON(busiest == this_rq);
4266
4267 schedstat_add(sd, lb_imbalance[idle], env.imbalance);
4268
4269 ld_moved = 0;
4270 if (busiest->nr_running > 1) {
4271 /*
4272 * Attempt to move tasks. If find_busiest_group has found
4273 * an imbalance but busiest->nr_running <= 1, the group is
4274 * still unbalanced. ld_moved simply stays zero, so it is
4275 * correctly treated as an imbalance.
4276 */
4277 env.flags |= LBF_ALL_PINNED;
4278 env.src_cpu = busiest->cpu;
4279 env.src_rq = busiest;
4280 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
4281
4282more_balance:
4283 local_irq_save(flags);
4284 double_rq_lock(this_rq, busiest);
4285 if (!env.loop)
4286 update_h_load(env.src_cpu);
4287 ld_moved += move_tasks(&env);
4288 double_rq_unlock(this_rq, busiest);
4289 local_irq_restore(flags);
4290
4291 if (env.flags & LBF_NEED_BREAK) {
4292 env.flags &= ~LBF_NEED_BREAK;
4293 goto more_balance;
4294 }
4295
4296 /*
4297 * some other cpu did the load balance for us.
4298 */
4299 if (ld_moved && this_cpu != smp_processor_id())
4300 resched_cpu(this_cpu);
4301
4302 /* All tasks on this runqueue were pinned by CPU affinity */
4303 if (unlikely(env.flags & LBF_ALL_PINNED)) {
4304 cpumask_clear_cpu(cpu_of(busiest), cpus);
4305 if (!cpumask_empty(cpus))
4306 goto redo;
4307 goto out_balanced;
4308 }
4309 }
4310
4311 if (!ld_moved) {
4312 schedstat_inc(sd, lb_failed[idle]);
4313 /*
4314 * Increment the failure counter only on periodic balance.
4315 * We do not want newidle balance, which can be very
4316 * frequent, pollute the failure counter causing
4317 * excessive cache_hot migrations and active balances.
4318 */
4319 if (idle != CPU_NEWLY_IDLE)
4320 sd->nr_balance_failed++;
4321
4322 if (need_active_balance(&env)) {
4323 raw_spin_lock_irqsave(&busiest->lock, flags);
4324
4325 /* don't kick the active_load_balance_cpu_stop,
4326 * if the curr task on busiest cpu can't be
4327 * moved to this_cpu
4328 */
4329 if (!cpumask_test_cpu(this_cpu,
4330 tsk_cpus_allowed(busiest->curr))) {
4331 raw_spin_unlock_irqrestore(&busiest->lock,
4332 flags);
4333 env.flags |= LBF_ALL_PINNED;
4334 goto out_one_pinned;
4335 }
4336
4337 /*
4338 * ->active_balance synchronizes accesses to
4339 * ->active_balance_work. Once set, it's cleared
4340 * only after active load balance is finished.
4341 */
4342 if (!busiest->active_balance) {
4343 busiest->active_balance = 1;
4344 busiest->push_cpu = this_cpu;
4345 active_balance = 1;
4346 }
4347 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4348
4349 if (active_balance) {
4350 stop_one_cpu_nowait(cpu_of(busiest),
4351 active_load_balance_cpu_stop, busiest,
4352 &busiest->active_balance_work);
4353 }
4354
4355 /*
4356 * We've kicked active balancing, reset the failure
4357 * counter.
4358 */
4359 sd->nr_balance_failed = sd->cache_nice_tries+1;
4360 }
4361 } else
4362 sd->nr_balance_failed = 0;
4363
4364 if (likely(!active_balance)) {
4365 /* We were unbalanced, so reset the balancing interval */
4366 sd->balance_interval = sd->min_interval;
4367 } else {
4368 /*
4369 * If we've begun active balancing, start to back off. This
4370 * case may not be covered by the all_pinned logic if there
4371 * is only 1 task on the busy runqueue (because we don't call
4372 * move_tasks).
4373 */
4374 if (sd->balance_interval < sd->max_interval)
4375 sd->balance_interval *= 2;
4376 }
4377
4378 goto out;
4379
4380out_balanced:
4381 schedstat_inc(sd, lb_balanced[idle]);
4382
4383 sd->nr_balance_failed = 0;
4384
4385out_one_pinned:
4386 /* tune up the balancing interval */
4387 if (((env.flags & LBF_ALL_PINNED) &&
4388 sd->balance_interval < MAX_PINNED_INTERVAL) ||
4389 (sd->balance_interval < sd->max_interval))
4390 sd->balance_interval *= 2;
4391
4392 ld_moved = 0;
4393out:
4394 return ld_moved;
4395}
4396
4397/*
4398 * idle_balance is called by schedule() if this_cpu is about to become
4399 * idle. Attempts to pull tasks from other CPUs.
4400 */
4401void idle_balance(int this_cpu, struct rq *this_rq)
4402{
4403 struct sched_domain *sd;
4404 int pulled_task = 0;
4405 unsigned long next_balance = jiffies + HZ;
4406
4407 this_rq->idle_stamp = this_rq->clock;
4408
4409 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4410 return;
4411
4412 /*
4413 * Drop the rq->lock, but keep IRQ/preempt disabled.
4414 */
4415 raw_spin_unlock(&this_rq->lock);
4416
4417 update_shares(this_cpu);
4418 rcu_read_lock();
4419 for_each_domain(this_cpu, sd) {
4420 unsigned long interval;
4421 int balance = 1;
4422
4423 if (!(sd->flags & SD_LOAD_BALANCE))
4424 continue;
4425
4426 if (sd->flags & SD_BALANCE_NEWIDLE) {
4427 /* If we've pulled tasks over stop searching: */
4428 pulled_task = load_balance(this_cpu, this_rq,
4429 sd, CPU_NEWLY_IDLE, &balance);
4430 }
4431
4432 interval = msecs_to_jiffies(sd->balance_interval);
4433 if (time_after(next_balance, sd->last_balance + interval))
4434 next_balance = sd->last_balance + interval;
4435 if (pulled_task) {
4436 this_rq->idle_stamp = 0;
4437 break;
4438 }
4439 }
4440 rcu_read_unlock();
4441
4442 raw_spin_lock(&this_rq->lock);
4443
4444 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4445 /*
4446 * We are going idle. next_balance may be set based on
4447 * a busy processor. So reset next_balance.
4448 */
4449 this_rq->next_balance = next_balance;
4450 }
4451}
4452
4453/*
4454 * active_load_balance_cpu_stop is run by cpu stopper. It pushes
4455 * running tasks off the busiest CPU onto idle CPUs. It requires at
4456 * least 1 task to be running on each physical CPU where possible, and
4457 * avoids physical / logical imbalances.
4458 */
4459static int active_load_balance_cpu_stop(void *data)
4460{
4461 struct rq *busiest_rq = data;
4462 int busiest_cpu = cpu_of(busiest_rq);
4463 int target_cpu = busiest_rq->push_cpu;
4464 struct rq *target_rq = cpu_rq(target_cpu);
4465 struct sched_domain *sd;
4466
4467 raw_spin_lock_irq(&busiest_rq->lock);
4468
4469 /* make sure the requested cpu hasn't gone down in the meantime */
4470 if (unlikely(busiest_cpu != smp_processor_id() ||
4471 !busiest_rq->active_balance))
4472 goto out_unlock;
4473
4474 /* Is there any task to move? */
4475 if (busiest_rq->nr_running <= 1)
4476 goto out_unlock;
4477
4478 /*
4479 * This condition is "impossible", if it occurs
4480 * we need to fix it. Originally reported by
4481 * Bjorn Helgaas on a 128-cpu setup.
4482 */
4483 BUG_ON(busiest_rq == target_rq);
4484
4485 /* move a task from busiest_rq to target_rq */
4486 double_lock_balance(busiest_rq, target_rq);
4487
4488 /* Search for an sd spanning us and the target CPU. */
4489 rcu_read_lock();
4490 for_each_domain(target_cpu, sd) {
4491 if ((sd->flags & SD_LOAD_BALANCE) &&
4492 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4493 break;
4494 }
4495
4496 if (likely(sd)) {
4497 struct lb_env env = {
4498 .sd = sd,
4499 .dst_cpu = target_cpu,
4500 .dst_rq = target_rq,
4501 .src_cpu = busiest_rq->cpu,
4502 .src_rq = busiest_rq,
4503 .idle = CPU_IDLE,
4504 };
4505
4506 schedstat_inc(sd, alb_count);
4507
4508 if (move_one_task(&env))
4509 schedstat_inc(sd, alb_pushed);
4510 else
4511 schedstat_inc(sd, alb_failed);
4512 }
4513 rcu_read_unlock();
4514 double_unlock_balance(busiest_rq, target_rq);
4515out_unlock:
4516 busiest_rq->active_balance = 0;
4517 raw_spin_unlock_irq(&busiest_rq->lock);
4518 return 0;
4519}
4520
4521#ifdef CONFIG_NO_HZ
4522/*
4523 * idle load balancing details
4524 * - When one of the busy CPUs notice that there may be an idle rebalancing
4525 * needed, they will kick the idle load balancer, which then does idle
4526 * load balancing for all the idle CPUs.
4527 */
4528static struct {
4529 cpumask_var_t idle_cpus_mask;
4530 atomic_t nr_cpus;
4531 unsigned long next_balance; /* in jiffy units */
4532} nohz ____cacheline_aligned;
4533
4534static inline int find_new_ilb(int call_cpu)
4535{
4536 int ilb = cpumask_first(nohz.idle_cpus_mask);
4537
4538 if (ilb < nr_cpu_ids && idle_cpu(ilb))
4539 return ilb;
4540
4541 return nr_cpu_ids;
4542}
4543
4544/*
4545 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
4546 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
4547 * CPU (if there is one).
4548 */
4549static void nohz_balancer_kick(int cpu)
4550{
4551 int ilb_cpu;
4552
4553 nohz.next_balance++;
4554
4555 ilb_cpu = find_new_ilb(cpu);
4556
4557 if (ilb_cpu >= nr_cpu_ids)
4558 return;
4559
4560 if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
4561 return;
4562 /*
4563 * Use smp_send_reschedule() instead of resched_cpu().
4564 * This way we generate a sched IPI on the target cpu which
4565 * is idle. And the softirq performing nohz idle load balance
4566 * will be run before returning from the IPI.
4567 */
4568 smp_send_reschedule(ilb_cpu);
4569 return;
4570}
4571
4572static inline void clear_nohz_tick_stopped(int cpu)
4573{
4574 if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
4575 cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
4576 atomic_dec(&nohz.nr_cpus);
4577 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4578 }
4579}
4580
4581static inline void set_cpu_sd_state_busy(void)
4582{
4583 struct sched_domain *sd;
4584 int cpu = smp_processor_id();
4585
4586 if (!test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4587 return;
4588 clear_bit(NOHZ_IDLE, nohz_flags(cpu));
4589
4590 rcu_read_lock();
4591 for_each_domain(cpu, sd)
4592 atomic_inc(&sd->groups->sgp->nr_busy_cpus);
4593 rcu_read_unlock();
4594}
4595
4596void set_cpu_sd_state_idle(void)
4597{
4598 struct sched_domain *sd;
4599 int cpu = smp_processor_id();
4600
4601 if (test_bit(NOHZ_IDLE, nohz_flags(cpu)))
4602 return;
4603 set_bit(NOHZ_IDLE, nohz_flags(cpu));
4604
4605 rcu_read_lock();
4606 for_each_domain(cpu, sd)
4607 atomic_dec(&sd->groups->sgp->nr_busy_cpus);
4608 rcu_read_unlock();
4609}
4610
4611/*
4612 * This routine will record that this cpu is going idle with tick stopped.
4613 * This info will be used in performing idle load balancing in the future.
4614 */
4615void select_nohz_load_balancer(int stop_tick)
4616{
4617 int cpu = smp_processor_id();
4618
4619 /*
4620 * If this cpu is going down, then nothing needs to be done.
4621 */
4622 if (!cpu_active(cpu))
4623 return;
4624
4625 if (stop_tick) {
4626 if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
4627 return;
4628
4629 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
4630 atomic_inc(&nohz.nr_cpus);
4631 set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
4632 }
4633 return;
4634}
4635
4636static int __cpuinit sched_ilb_notifier(struct notifier_block *nfb,
4637 unsigned long action, void *hcpu)
4638{
4639 switch (action & ~CPU_TASKS_FROZEN) {
4640 case CPU_DYING:
4641 clear_nohz_tick_stopped(smp_processor_id());
4642 return NOTIFY_OK;
4643 default:
4644 return NOTIFY_DONE;
4645 }
4646}
4647#endif
4648
4649static DEFINE_SPINLOCK(balancing);
4650
4651/*
4652 * Scale the max load_balance interval with the number of CPUs in the system.
4653 * This trades load-balance latency on larger machines for less cross talk.
4654 */
4655void update_max_interval(void)
4656{
4657 max_load_balance_interval = HZ*num_online_cpus()/10;
4658}
4659
4660/*
4661 * It checks each scheduling domain to see if it is due to be balanced,
4662 * and initiates a balancing operation if so.
4663 *
4664 * Balancing parameters are set up in arch_init_sched_domains.
4665 */
4666static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4667{
4668 int balance = 1;
4669 struct rq *rq = cpu_rq(cpu);
4670 unsigned long interval;
4671 struct sched_domain *sd;
4672 /* Earliest time when we have to do rebalance again */
4673 unsigned long next_balance = jiffies + 60*HZ;
4674 int update_next_balance = 0;
4675 int need_serialize;
4676
4677 update_shares(cpu);
4678
4679 rcu_read_lock();
4680 for_each_domain(cpu, sd) {
4681 if (!(sd->flags & SD_LOAD_BALANCE))
4682 continue;
4683
4684 interval = sd->balance_interval;
4685 if (idle != CPU_IDLE)
4686 interval *= sd->busy_factor;
4687
4688 /* scale ms to jiffies */
4689 interval = msecs_to_jiffies(interval);
4690 interval = clamp(interval, 1UL, max_load_balance_interval);
4691
4692 need_serialize = sd->flags & SD_SERIALIZE;
4693
4694 if (need_serialize) {
4695 if (!spin_trylock(&balancing))
4696 goto out;
4697 }
4698
4699 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4700 if (load_balance(cpu, rq, sd, idle, &balance)) {
4701 /*
4702 * We've pulled tasks over so either we're no
4703 * longer idle.
4704 */
4705 idle = CPU_NOT_IDLE;
4706 }
4707 sd->last_balance = jiffies;
4708 }
4709 if (need_serialize)
4710 spin_unlock(&balancing);
4711out:
4712 if (time_after(next_balance, sd->last_balance + interval)) {
4713 next_balance = sd->last_balance + interval;
4714 update_next_balance = 1;
4715 }
4716
4717 /*
4718 * Stop the load balance at this level. There is another
4719 * CPU in our sched group which is doing load balancing more
4720 * actively.
4721 */
4722 if (!balance)
4723 break;
4724 }
4725 rcu_read_unlock();
4726
4727 /*
4728 * next_balance will be updated only when there is a need.
4729 * When the cpu is attached to null domain for ex, it will not be
4730 * updated.
4731 */
4732 if (likely(update_next_balance))
4733 rq->next_balance = next_balance;
4734}
4735
4736#ifdef CONFIG_NO_HZ
4737/*
4738 * In CONFIG_NO_HZ case, the idle balance kickee will do the
4739 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4740 */
4741static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
4742{
4743 struct rq *this_rq = cpu_rq(this_cpu);
4744 struct rq *rq;
4745 int balance_cpu;
4746
4747 if (idle != CPU_IDLE ||
4748 !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
4749 goto end;
4750
4751 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
4752 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
4753 continue;
4754
4755 /*
4756 * If this cpu gets work to do, stop the load balancing
4757 * work being done for other cpus. Next load
4758 * balancing owner will pick it up.
4759 */
4760 if (need_resched())
4761 break;
4762
4763 raw_spin_lock_irq(&this_rq->lock);
4764 update_rq_clock(this_rq);
4765 update_idle_cpu_load(this_rq);
4766 raw_spin_unlock_irq(&this_rq->lock);
4767
4768 rebalance_domains(balance_cpu, CPU_IDLE);
4769
4770 rq = cpu_rq(balance_cpu);
4771 if (time_after(this_rq->next_balance, rq->next_balance))
4772 this_rq->next_balance = rq->next_balance;
4773 }
4774 nohz.next_balance = this_rq->next_balance;
4775end:
4776 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
4777}
4778
4779/*
4780 * Current heuristic for kicking the idle load balancer in the presence
4781 * of an idle cpu is the system.
4782 * - This rq has more than one task.
4783 * - At any scheduler domain level, this cpu's scheduler group has multiple
4784 * busy cpu's exceeding the group's power.
4785 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
4786 * domain span are idle.
4787 */
4788static inline int nohz_kick_needed(struct rq *rq, int cpu)
4789{
4790 unsigned long now = jiffies;
4791 struct sched_domain *sd;
4792
4793 if (unlikely(idle_cpu(cpu)))
4794 return 0;
4795
4796 /*
4797 * We may be recently in ticked or tickless idle mode. At the first
4798 * busy tick after returning from idle, we will update the busy stats.
4799 */
4800 set_cpu_sd_state_busy();
4801 clear_nohz_tick_stopped(cpu);
4802
4803 /*
4804 * None are in tickless mode and hence no need for NOHZ idle load
4805 * balancing.
4806 */
4807 if (likely(!atomic_read(&nohz.nr_cpus)))
4808 return 0;
4809
4810 if (time_before(now, nohz.next_balance))
4811 return 0;
4812
4813 if (rq->nr_running >= 2)
4814 goto need_kick;
4815
4816 rcu_read_lock();
4817 for_each_domain(cpu, sd) {
4818 struct sched_group *sg = sd->groups;
4819 struct sched_group_power *sgp = sg->sgp;
4820 int nr_busy = atomic_read(&sgp->nr_busy_cpus);
4821
4822 if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
4823 goto need_kick_unlock;
4824
4825 if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
4826 && (cpumask_first_and(nohz.idle_cpus_mask,
4827 sched_domain_span(sd)) < cpu))
4828 goto need_kick_unlock;
4829
4830 if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
4831 break;
4832 }
4833 rcu_read_unlock();
4834 return 0;
4835
4836need_kick_unlock:
4837 rcu_read_unlock();
4838need_kick:
4839 return 1;
4840}
4841#else
4842static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
4843#endif
4844
4845/*
4846 * run_rebalance_domains is triggered when needed from the scheduler tick.
4847 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
4848 */
4849static void run_rebalance_domains(struct softirq_action *h)
4850{
4851 int this_cpu = smp_processor_id();
4852 struct rq *this_rq = cpu_rq(this_cpu);
4853 enum cpu_idle_type idle = this_rq->idle_balance ?
4854 CPU_IDLE : CPU_NOT_IDLE;
4855
4856 rebalance_domains(this_cpu, idle);
4857
4858 /*
4859 * If this cpu has a pending nohz_balance_kick, then do the
4860 * balancing on behalf of the other idle cpus whose ticks are
4861 * stopped.
4862 */
4863 nohz_idle_balance(this_cpu, idle);
4864}
4865
4866static inline int on_null_domain(int cpu)
4867{
4868 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4869}
4870
4871/*
4872 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4873 */
4874void trigger_load_balance(struct rq *rq, int cpu)
4875{
4876 /* Don't need to rebalance while attached to NULL domain */
4877 if (time_after_eq(jiffies, rq->next_balance) &&
4878 likely(!on_null_domain(cpu)))
4879 raise_softirq(SCHED_SOFTIRQ);
4880#ifdef CONFIG_NO_HZ
4881 if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
4882 nohz_balancer_kick(cpu);
4883#endif
4884}
4885
4886static void rq_online_fair(struct rq *rq)
4887{
4888 update_sysctl();
4889}
4890
4891static void rq_offline_fair(struct rq *rq)
4892{
4893 update_sysctl();
4894}
4895
4896#endif /* CONFIG_SMP */
4897
4898/*
4899 * scheduler tick hitting a task of our scheduling class:
4900 */
4901static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
4902{
4903 struct cfs_rq *cfs_rq;
4904 struct sched_entity *se = &curr->se;
4905
4906 for_each_sched_entity(se) {
4907 cfs_rq = cfs_rq_of(se);
4908 entity_tick(cfs_rq, se, queued);
4909 }
4910}
4911
4912/*
4913 * called on fork with the child task as argument from the parent's context
4914 * - child not yet on the tasklist
4915 * - preemption disabled
4916 */
4917static void task_fork_fair(struct task_struct *p)
4918{
4919 struct cfs_rq *cfs_rq;
4920 struct sched_entity *se = &p->se, *curr;
4921 int this_cpu = smp_processor_id();
4922 struct rq *rq = this_rq();
4923 unsigned long flags;
4924
4925 raw_spin_lock_irqsave(&rq->lock, flags);
4926
4927 update_rq_clock(rq);
4928
4929 cfs_rq = task_cfs_rq(current);
4930 curr = cfs_rq->curr;
4931
4932 if (unlikely(task_cpu(p) != this_cpu)) {
4933 rcu_read_lock();
4934 __set_task_cpu(p, this_cpu);
4935 rcu_read_unlock();
4936 }
4937
4938 update_curr(cfs_rq);
4939
4940 if (curr)
4941 se->vruntime = curr->vruntime;
4942 place_entity(cfs_rq, se, 1);
4943
4944 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
4945 /*
4946 * Upon rescheduling, sched_class::put_prev_task() will place
4947 * 'current' within the tree based on its new key value.
4948 */
4949 swap(curr->vruntime, se->vruntime);
4950 resched_task(rq->curr);
4951 }
4952
4953 se->vruntime -= cfs_rq->min_vruntime;
4954
4955 raw_spin_unlock_irqrestore(&rq->lock, flags);
4956}
4957
4958/*
4959 * Priority of the task has changed. Check to see if we preempt
4960 * the current task.
4961 */
4962static void
4963prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
4964{
4965 if (!p->se.on_rq)
4966 return;
4967
4968 /*
4969 * Reschedule if we are currently running on this runqueue and
4970 * our priority decreased, or if we are not currently running on
4971 * this runqueue and our priority is higher than the current's
4972 */
4973 if (rq->curr == p) {
4974 if (p->prio > oldprio)
4975 resched_task(rq->curr);
4976 } else
4977 check_preempt_curr(rq, p, 0);
4978}
4979
4980static void switched_from_fair(struct rq *rq, struct task_struct *p)
4981{
4982 struct sched_entity *se = &p->se;
4983 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4984
4985 /*
4986 * Ensure the task's vruntime is normalized, so that when its
4987 * switched back to the fair class the enqueue_entity(.flags=0) will
4988 * do the right thing.
4989 *
4990 * If it was on_rq, then the dequeue_entity(.flags=0) will already
4991 * have normalized the vruntime, if it was !on_rq, then only when
4992 * the task is sleeping will it still have non-normalized vruntime.
4993 */
4994 if (!se->on_rq && p->state != TASK_RUNNING) {
4995 /*
4996 * Fix up our vruntime so that the current sleep doesn't
4997 * cause 'unlimited' sleep bonus.
4998 */
4999 place_entity(cfs_rq, se, 0);
5000 se->vruntime -= cfs_rq->min_vruntime;
5001 }
5002}
5003
5004/*
5005 * We switched to the sched_fair class.
5006 */
5007static void switched_to_fair(struct rq *rq, struct task_struct *p)
5008{
5009 if (!p->se.on_rq)
5010 return;
5011
5012 /*
5013 * We were most likely switched from sched_rt, so
5014 * kick off the schedule if running, otherwise just see
5015 * if we can still preempt the current task.
5016 */
5017 if (rq->curr == p)
5018 resched_task(rq->curr);
5019 else
5020 check_preempt_curr(rq, p, 0);
5021}
5022
5023/* Account for a task changing its policy or group.
5024 *
5025 * This routine is mostly called to set cfs_rq->curr field when a task
5026 * migrates between groups/classes.
5027 */
5028static void set_curr_task_fair(struct rq *rq)
5029{
5030 struct sched_entity *se = &rq->curr->se;
5031
5032 for_each_sched_entity(se) {
5033 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5034
5035 set_next_entity(cfs_rq, se);
5036 /* ensure bandwidth has been allocated on our new cfs_rq */
5037 account_cfs_rq_runtime(cfs_rq, 0);
5038 }
5039}
5040
5041void init_cfs_rq(struct cfs_rq *cfs_rq)
5042{
5043 cfs_rq->tasks_timeline = RB_ROOT;
5044 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
5045#ifndef CONFIG_64BIT
5046 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
5047#endif
5048}
5049
5050#ifdef CONFIG_FAIR_GROUP_SCHED
5051static void task_move_group_fair(struct task_struct *p, int on_rq)
5052{
5053 /*
5054 * If the task was not on the rq at the time of this cgroup movement
5055 * it must have been asleep, sleeping tasks keep their ->vruntime
5056 * absolute on their old rq until wakeup (needed for the fair sleeper
5057 * bonus in place_entity()).
5058 *
5059 * If it was on the rq, we've just 'preempted' it, which does convert
5060 * ->vruntime to a relative base.
5061 *
5062 * Make sure both cases convert their relative position when migrating
5063 * to another cgroup's rq. This does somewhat interfere with the
5064 * fair sleeper stuff for the first placement, but who cares.
5065 */
5066 /*
5067 * When !on_rq, vruntime of the task has usually NOT been normalized.
5068 * But there are some cases where it has already been normalized:
5069 *
5070 * - Moving a forked child which is waiting for being woken up by
5071 * wake_up_new_task().
5072 * - Moving a task which has been woken up by try_to_wake_up() and
5073 * waiting for actually being woken up by sched_ttwu_pending().
5074 *
5075 * To prevent boost or penalty in the new cfs_rq caused by delta
5076 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
5077 */
5078 if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
5079 on_rq = 1;
5080
5081 if (!on_rq)
5082 p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
5083 set_task_rq(p, task_cpu(p));
5084 if (!on_rq)
5085 p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
5086}
5087
5088void free_fair_sched_group(struct task_group *tg)
5089{
5090 int i;
5091
5092 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
5093
5094 for_each_possible_cpu(i) {
5095 if (tg->cfs_rq)
5096 kfree(tg->cfs_rq[i]);
5097 if (tg->se)
5098 kfree(tg->se[i]);
5099 }
5100
5101 kfree(tg->cfs_rq);
5102 kfree(tg->se);
5103}
5104
5105int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5106{
5107 struct cfs_rq *cfs_rq;
5108 struct sched_entity *se;
5109 int i;
5110
5111 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
5112 if (!tg->cfs_rq)
5113 goto err;
5114 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
5115 if (!tg->se)
5116 goto err;
5117
5118 tg->shares = NICE_0_LOAD;
5119
5120 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
5121
5122 for_each_possible_cpu(i) {
5123 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
5124 GFP_KERNEL, cpu_to_node(i));
5125 if (!cfs_rq)
5126 goto err;
5127
5128 se = kzalloc_node(sizeof(struct sched_entity),
5129 GFP_KERNEL, cpu_to_node(i));
5130 if (!se)
5131 goto err_free_rq;
5132
5133 init_cfs_rq(cfs_rq);
5134 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
5135 }
5136
5137 return 1;
5138
5139err_free_rq:
5140 kfree(cfs_rq);
5141err:
5142 return 0;
5143}
5144
5145void unregister_fair_sched_group(struct task_group *tg, int cpu)
5146{
5147 struct rq *rq = cpu_rq(cpu);
5148 unsigned long flags;
5149
5150 /*
5151 * Only empty task groups can be destroyed; so we can speculatively
5152 * check on_list without danger of it being re-added.
5153 */
5154 if (!tg->cfs_rq[cpu]->on_list)
5155 return;
5156
5157 raw_spin_lock_irqsave(&rq->lock, flags);
5158 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
5159 raw_spin_unlock_irqrestore(&rq->lock, flags);
5160}
5161
5162void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
5163 struct sched_entity *se, int cpu,
5164 struct sched_entity *parent)
5165{
5166 struct rq *rq = cpu_rq(cpu);
5167
5168 cfs_rq->tg = tg;
5169 cfs_rq->rq = rq;
5170#ifdef CONFIG_SMP
5171 /* allow initial update_cfs_load() to truncate */
5172 cfs_rq->load_stamp = 1;
5173#endif
5174 init_cfs_rq_runtime(cfs_rq);
5175
5176 tg->cfs_rq[cpu] = cfs_rq;
5177 tg->se[cpu] = se;
5178
5179 /* se could be NULL for root_task_group */
5180 if (!se)
5181 return;
5182
5183 if (!parent)
5184 se->cfs_rq = &rq->cfs;
5185 else
5186 se->cfs_rq = parent->my_q;
5187
5188 se->my_q = cfs_rq;
5189 update_load_set(&se->load, 0);
5190 se->parent = parent;
5191}
5192
5193static DEFINE_MUTEX(shares_mutex);
5194
5195int sched_group_set_shares(struct task_group *tg, unsigned long shares)
5196{
5197 int i;
5198 unsigned long flags;
5199
5200 /*
5201 * We can't change the weight of the root cgroup.
5202 */
5203 if (!tg->se[0])
5204 return -EINVAL;
5205
5206 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
5207
5208 mutex_lock(&shares_mutex);
5209 if (tg->shares == shares)
5210 goto done;
5211
5212 tg->shares = shares;
5213 for_each_possible_cpu(i) {
5214 struct rq *rq = cpu_rq(i);
5215 struct sched_entity *se;
5216
5217 se = tg->se[i];
5218 /* Propagate contribution to hierarchy */
5219 raw_spin_lock_irqsave(&rq->lock, flags);
5220 for_each_sched_entity(se)
5221 update_cfs_shares(group_cfs_rq(se));
5222 raw_spin_unlock_irqrestore(&rq->lock, flags);
5223 }
5224
5225done:
5226 mutex_unlock(&shares_mutex);
5227 return 0;
5228}
5229#else /* CONFIG_FAIR_GROUP_SCHED */
5230
5231void free_fair_sched_group(struct task_group *tg) { }
5232
5233int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
5234{
5235 return 1;
5236}
5237
5238void unregister_fair_sched_group(struct task_group *tg, int cpu) { }
5239
5240#endif /* CONFIG_FAIR_GROUP_SCHED */
5241
5242
5243static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
5244{
5245 struct sched_entity *se = &task->se;
5246 unsigned int rr_interval = 0;
5247
5248 /*
5249 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
5250 * idle runqueue:
5251 */
5252 if (rq->cfs.load.weight)
5253 rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5254
5255 return rr_interval;
5256}
5257
5258/*
5259 * All the scheduling class methods:
5260 */
5261const struct sched_class fair_sched_class = {
5262 .next = &idle_sched_class,
5263 .enqueue_task = enqueue_task_fair,
5264 .dequeue_task = dequeue_task_fair,
5265 .yield_task = yield_task_fair,
5266 .yield_to_task = yield_to_task_fair,
5267
5268 .check_preempt_curr = check_preempt_wakeup,
5269
5270 .pick_next_task = pick_next_task_fair,
5271 .put_prev_task = put_prev_task_fair,
5272
5273#ifdef CONFIG_SMP
5274 .select_task_rq = select_task_rq_fair,
5275
5276 .rq_online = rq_online_fair,
5277 .rq_offline = rq_offline_fair,
5278
5279 .task_waking = task_waking_fair,
5280#endif
5281
5282 .set_curr_task = set_curr_task_fair,
5283 .task_tick = task_tick_fair,
5284 .task_fork = task_fork_fair,
5285
5286 .prio_changed = prio_changed_fair,
5287 .switched_from = switched_from_fair,
5288 .switched_to = switched_to_fair,
5289
5290 .get_rr_interval = get_rr_interval_fair,
5291
5292#ifdef CONFIG_FAIR_GROUP_SCHED
5293 .task_move_group = task_move_group_fair,
5294#endif
5295};
5296
5297#ifdef CONFIG_SCHED_DEBUG
5298void print_cfs_stats(struct seq_file *m, int cpu)
5299{
5300 struct cfs_rq *cfs_rq;
5301
5302 rcu_read_lock();
5303 for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
5304 print_cfs_rq(m, cpu, cfs_rq);
5305 rcu_read_unlock();
5306}
5307#endif
5308
5309__init void init_sched_fair_class(void)
5310{
5311#ifdef CONFIG_SMP
5312 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
5313
5314#ifdef CONFIG_NO_HZ
5315 nohz.next_balance = jiffies;
5316 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
5317 cpu_notifier(sched_ilb_notifier, 0);
5318#endif
5319#endif /* SMP */
5320
5321}