1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23#include "sched.h"
24
25#include <trace/events/sched.h>
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40unsigned int sysctl_sched_latency = 6000000ULL;
41static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42
43
44
45
46
47
48
49
50
51
52
53
54enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
55
56
57
58
59
60
61unsigned int sysctl_sched_min_granularity = 750000ULL;
62static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
63
64
65
66
67static unsigned int sched_nr_latency = 8;
68
69
70
71
72
73unsigned int sysctl_sched_child_runs_first __read_mostly;
74
75
76
77
78
79
80
81
82
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
92
93int __weak arch_asym_cpu_priority(int cpu)
94{
95 return -cpu;
96}
97
98
99
100
101
102
103
104static unsigned int capacity_margin = 1280;
105#endif
106
107#ifdef CONFIG_CFS_BANDWIDTH
108
109
110
111
112
113
114
115
116
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
141
142
143
144
145
146
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
208
209
210
211
212
213
214
215
216
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
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
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
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
267static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
268{
269 return se->cfs_rq;
270}
271
272
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
303
304
305
306
307
308
309
310 if (cfs_rq->tg->parent &&
311 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
312
313
314
315
316
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
322
323
324
325 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
326 return true;
327 }
328
329 if (!cfs_rq->tg->parent) {
330
331
332
333
334 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
335 &rq->leaf_cfs_rq_list);
336
337
338
339
340 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
341 return true;
342 }
343
344
345
346
347
348
349
350 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
351
352
353
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
366
367
368
369
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
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
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
411
412
413
414
415
416
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
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
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
498
499static __always_inline
500void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
501
502
503
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) {
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
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
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
574
575 while (*link) {
576 parent = *link;
577 entry = rb_entry(parent, struct sched_entity, run_node);
578
579
580
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
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
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
672
673
674
675
676
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
688
689
690
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
716
717
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
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
740
741
742
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
750}
751
752static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
753static void attach_entity_cfs_rq(struct sched_entity *se);
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781void post_init_entity_util_avg(struct task_struct *p)
782{
783 struct sched_entity *se = &p->se;
784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
785 struct sched_avg *sa = &se->avg;
786 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
787 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
788
789 if (cap > 0) {
790 if (cfs_rq->avg.util_avg != 0) {
791 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
792 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
793
794 if (sa->util_avg > cap)
795 sa->util_avg = cap;
796 } else {
797 sa->util_avg = cap;
798 }
799 }
800
801 if (p->sched_class != &fair_sched_class) {
802
803
804
805
806
807
808
809
810
811
812 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
813 return;
814 }
815
816 attach_entity_cfs_rq(se);
817}
818
819#else
820void init_entity_runnable_average(struct sched_entity *se)
821{
822}
823void post_init_entity_util_avg(struct task_struct *p)
824{
825}
826static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
827{
828}
829#endif
830
831
832
833
834static void update_curr(struct cfs_rq *cfs_rq)
835{
836 struct sched_entity *curr = cfs_rq->curr;
837 u64 now = rq_clock_task(rq_of(cfs_rq));
838 u64 delta_exec;
839
840 if (unlikely(!curr))
841 return;
842
843 delta_exec = now - curr->exec_start;
844 if (unlikely((s64)delta_exec <= 0))
845 return;
846
847 curr->exec_start = now;
848
849 schedstat_set(curr->statistics.exec_max,
850 max(delta_exec, curr->statistics.exec_max));
851
852 curr->sum_exec_runtime += delta_exec;
853 schedstat_add(cfs_rq->exec_clock, delta_exec);
854
855 curr->vruntime += calc_delta_fair(delta_exec, curr);
856 update_min_vruntime(cfs_rq);
857
858 if (entity_is_task(curr)) {
859 struct task_struct *curtask = task_of(curr);
860
861 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
862 cgroup_account_cputime(curtask, delta_exec);
863 account_group_exec_runtime(curtask, delta_exec);
864 }
865
866 account_cfs_rq_runtime(cfs_rq, delta_exec);
867}
868
869static void update_curr_fair(struct rq *rq)
870{
871 update_curr(cfs_rq_of(&rq->curr->se));
872}
873
874static inline void
875update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
876{
877 u64 wait_start, prev_wait_start;
878
879 if (!schedstat_enabled())
880 return;
881
882 wait_start = rq_clock(rq_of(cfs_rq));
883 prev_wait_start = schedstat_val(se->statistics.wait_start);
884
885 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
886 likely(wait_start > prev_wait_start))
887 wait_start -= prev_wait_start;
888
889 __schedstat_set(se->statistics.wait_start, wait_start);
890}
891
892static inline void
893update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
894{
895 struct task_struct *p;
896 u64 delta;
897
898 if (!schedstat_enabled())
899 return;
900
901 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
902
903 if (entity_is_task(se)) {
904 p = task_of(se);
905 if (task_on_rq_migrating(p)) {
906
907
908
909
910
911 __schedstat_set(se->statistics.wait_start, delta);
912 return;
913 }
914 trace_sched_stat_wait(p, delta);
915 }
916
917 __schedstat_set(se->statistics.wait_max,
918 max(schedstat_val(se->statistics.wait_max), delta));
919 __schedstat_inc(se->statistics.wait_count);
920 __schedstat_add(se->statistics.wait_sum, delta);
921 __schedstat_set(se->statistics.wait_start, 0);
922}
923
924static inline void
925update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
926{
927 struct task_struct *tsk = NULL;
928 u64 sleep_start, block_start;
929
930 if (!schedstat_enabled())
931 return;
932
933 sleep_start = schedstat_val(se->statistics.sleep_start);
934 block_start = schedstat_val(se->statistics.block_start);
935
936 if (entity_is_task(se))
937 tsk = task_of(se);
938
939 if (sleep_start) {
940 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
941
942 if ((s64)delta < 0)
943 delta = 0;
944
945 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
946 __schedstat_set(se->statistics.sleep_max, delta);
947
948 __schedstat_set(se->statistics.sleep_start, 0);
949 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
950
951 if (tsk) {
952 account_scheduler_latency(tsk, delta >> 10, 1);
953 trace_sched_stat_sleep(tsk, delta);
954 }
955 }
956 if (block_start) {
957 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
958
959 if ((s64)delta < 0)
960 delta = 0;
961
962 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
963 __schedstat_set(se->statistics.block_max, delta);
964
965 __schedstat_set(se->statistics.block_start, 0);
966 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
967
968 if (tsk) {
969 if (tsk->in_iowait) {
970 __schedstat_add(se->statistics.iowait_sum, delta);
971 __schedstat_inc(se->statistics.iowait_count);
972 trace_sched_stat_iowait(tsk, delta);
973 }
974
975 trace_sched_stat_blocked(tsk, delta);
976
977
978
979
980
981
982 if (unlikely(prof_on == SLEEP_PROFILING)) {
983 profile_hits(SLEEP_PROFILING,
984 (void *)get_wchan(tsk),
985 delta >> 20);
986 }
987 account_scheduler_latency(tsk, delta >> 10, 0);
988 }
989 }
990}
991
992
993
994
995static inline void
996update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
997{
998 if (!schedstat_enabled())
999 return;
1000
1001
1002
1003
1004
1005 if (se != cfs_rq->curr)
1006 update_stats_wait_start(cfs_rq, se);
1007
1008 if (flags & ENQUEUE_WAKEUP)
1009 update_stats_enqueue_sleeper(cfs_rq, se);
1010}
1011
1012static inline void
1013update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1014{
1015
1016 if (!schedstat_enabled())
1017 return;
1018
1019
1020
1021
1022
1023 if (se != cfs_rq->curr)
1024 update_stats_wait_end(cfs_rq, se);
1025
1026 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1027 struct task_struct *tsk = task_of(se);
1028
1029 if (tsk->state & TASK_INTERRUPTIBLE)
1030 __schedstat_set(se->statistics.sleep_start,
1031 rq_clock(rq_of(cfs_rq)));
1032 if (tsk->state & TASK_UNINTERRUPTIBLE)
1033 __schedstat_set(se->statistics.block_start,
1034 rq_clock(rq_of(cfs_rq)));
1035 }
1036}
1037
1038
1039
1040
1041static inline void
1042update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1043{
1044
1045
1046
1047 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1048}
1049
1050
1051
1052
1053
1054#ifdef CONFIG_NUMA_BALANCING
1055
1056
1057
1058
1059
1060unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1061unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1062
1063
1064unsigned int sysctl_numa_balancing_scan_size = 256;
1065
1066
1067unsigned int sysctl_numa_balancing_scan_delay = 1000;
1068
1069struct numa_group {
1070 refcount_t refcount;
1071
1072 spinlock_t lock;
1073 int nr_tasks;
1074 pid_t gid;
1075 int active_nodes;
1076
1077 struct rcu_head rcu;
1078 unsigned long total_faults;
1079 unsigned long max_faults_cpu;
1080
1081
1082
1083
1084
1085 unsigned long *faults_cpu;
1086 unsigned long faults[0];
1087};
1088
1089
1090
1091
1092
1093static struct numa_group *deref_task_numa_group(struct task_struct *p)
1094{
1095 return rcu_dereference_check(p->numa_group, p == current ||
1096 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1097}
1098
1099static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1100{
1101 return rcu_dereference_protected(p->numa_group, p == current);
1102}
1103
1104static inline unsigned long group_faults_priv(struct numa_group *ng);
1105static inline unsigned long group_faults_shared(struct numa_group *ng);
1106
1107static unsigned int task_nr_scan_windows(struct task_struct *p)
1108{
1109 unsigned long rss = 0;
1110 unsigned long nr_scan_pages;
1111
1112
1113
1114
1115
1116
1117 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1118 rss = get_mm_rss(p->mm);
1119 if (!rss)
1120 rss = nr_scan_pages;
1121
1122 rss = round_up(rss, nr_scan_pages);
1123 return rss / nr_scan_pages;
1124}
1125
1126
1127#define MAX_SCAN_WINDOW 2560
1128
1129static unsigned int task_scan_min(struct task_struct *p)
1130{
1131 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1132 unsigned int scan, floor;
1133 unsigned int windows = 1;
1134
1135 if (scan_size < MAX_SCAN_WINDOW)
1136 windows = MAX_SCAN_WINDOW / scan_size;
1137 floor = 1000 / windows;
1138
1139 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1140 return max_t(unsigned int, floor, scan);
1141}
1142
1143static unsigned int task_scan_start(struct task_struct *p)
1144{
1145 unsigned long smin = task_scan_min(p);
1146 unsigned long period = smin;
1147 struct numa_group *ng;
1148
1149
1150 rcu_read_lock();
1151 ng = rcu_dereference(p->numa_group);
1152 if (ng) {
1153 unsigned long shared = group_faults_shared(ng);
1154 unsigned long private = group_faults_priv(ng);
1155
1156 period *= refcount_read(&ng->refcount);
1157 period *= shared + 1;
1158 period /= private + shared + 1;
1159 }
1160 rcu_read_unlock();
1161
1162 return max(smin, period);
1163}
1164
1165static unsigned int task_scan_max(struct task_struct *p)
1166{
1167 unsigned long smin = task_scan_min(p);
1168 unsigned long smax;
1169 struct numa_group *ng;
1170
1171
1172 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1173
1174
1175 ng = deref_curr_numa_group(p);
1176 if (ng) {
1177 unsigned long shared = group_faults_shared(ng);
1178 unsigned long private = group_faults_priv(ng);
1179 unsigned long period = smax;
1180
1181 period *= refcount_read(&ng->refcount);
1182 period *= shared + 1;
1183 period /= private + shared + 1;
1184
1185 smax = max(smax, period);
1186 }
1187
1188 return max(smin, smax);
1189}
1190
1191void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1192{
1193 int mm_users = 0;
1194 struct mm_struct *mm = p->mm;
1195
1196 if (mm) {
1197 mm_users = atomic_read(&mm->mm_users);
1198 if (mm_users == 1) {
1199 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1200 mm->numa_scan_seq = 0;
1201 }
1202 }
1203 p->node_stamp = 0;
1204 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1205 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1206 p->numa_work.next = &p->numa_work;
1207 p->numa_faults = NULL;
1208 RCU_INIT_POINTER(p->numa_group, NULL);
1209 p->last_task_numa_placement = 0;
1210 p->last_sum_exec_runtime = 0;
1211
1212
1213 if (!(clone_flags & CLONE_VM)) {
1214 p->numa_preferred_nid = NUMA_NO_NODE;
1215 return;
1216 }
1217
1218
1219
1220
1221
1222 if (mm) {
1223 unsigned int delay;
1224
1225 delay = min_t(unsigned int, task_scan_max(current),
1226 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1227 delay += 2 * TICK_NSEC;
1228 p->node_stamp = delay;
1229 }
1230}
1231
1232static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1233{
1234 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1235 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1236}
1237
1238static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1239{
1240 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1241 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1242}
1243
1244
1245#define NR_NUMA_HINT_FAULT_TYPES 2
1246
1247
1248#define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1249
1250
1251#define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1252
1253pid_t task_numa_group_id(struct task_struct *p)
1254{
1255 struct numa_group *ng;
1256 pid_t gid = 0;
1257
1258 rcu_read_lock();
1259 ng = rcu_dereference(p->numa_group);
1260 if (ng)
1261 gid = ng->gid;
1262 rcu_read_unlock();
1263
1264 return gid;
1265}
1266
1267
1268
1269
1270
1271
1272
1273static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1274{
1275 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1276}
1277
1278static inline unsigned long task_faults(struct task_struct *p, int nid)
1279{
1280 if (!p->numa_faults)
1281 return 0;
1282
1283 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1284 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1285}
1286
1287static inline unsigned long group_faults(struct task_struct *p, int nid)
1288{
1289 struct numa_group *ng = deref_task_numa_group(p);
1290
1291 if (!ng)
1292 return 0;
1293
1294 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1295 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1296}
1297
1298static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1299{
1300 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1301 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1302}
1303
1304static inline unsigned long group_faults_priv(struct numa_group *ng)
1305{
1306 unsigned long faults = 0;
1307 int node;
1308
1309 for_each_online_node(node) {
1310 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1311 }
1312
1313 return faults;
1314}
1315
1316static inline unsigned long group_faults_shared(struct numa_group *ng)
1317{
1318 unsigned long faults = 0;
1319 int node;
1320
1321 for_each_online_node(node) {
1322 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1323 }
1324
1325 return faults;
1326}
1327
1328
1329
1330
1331
1332
1333#define ACTIVE_NODE_FRACTION 3
1334
1335static bool numa_is_active_node(int nid, struct numa_group *ng)
1336{
1337 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1338}
1339
1340
1341static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1342 int maxdist, bool task)
1343{
1344 unsigned long score = 0;
1345 int node;
1346
1347
1348
1349
1350
1351 if (sched_numa_topology_type == NUMA_DIRECT)
1352 return 0;
1353
1354
1355
1356
1357
1358 for_each_online_node(node) {
1359 unsigned long faults;
1360 int dist = node_distance(nid, node);
1361
1362
1363
1364
1365
1366 if (dist == sched_max_numa_distance || node == nid)
1367 continue;
1368
1369
1370
1371
1372
1373
1374
1375
1376 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1377 dist >= maxdist)
1378 continue;
1379
1380
1381 if (task)
1382 faults = task_faults(p, node);
1383 else
1384 faults = group_faults(p, node);
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1395 faults *= (sched_max_numa_distance - dist);
1396 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1397 }
1398
1399 score += faults;
1400 }
1401
1402 return score;
1403}
1404
1405
1406
1407
1408
1409
1410
1411static inline unsigned long task_weight(struct task_struct *p, int nid,
1412 int dist)
1413{
1414 unsigned long faults, total_faults;
1415
1416 if (!p->numa_faults)
1417 return 0;
1418
1419 total_faults = p->total_numa_faults;
1420
1421 if (!total_faults)
1422 return 0;
1423
1424 faults = task_faults(p, nid);
1425 faults += score_nearby_nodes(p, nid, dist, true);
1426
1427 return 1000 * faults / total_faults;
1428}
1429
1430static inline unsigned long group_weight(struct task_struct *p, int nid,
1431 int dist)
1432{
1433 struct numa_group *ng = deref_task_numa_group(p);
1434 unsigned long faults, total_faults;
1435
1436 if (!ng)
1437 return 0;
1438
1439 total_faults = ng->total_faults;
1440
1441 if (!total_faults)
1442 return 0;
1443
1444 faults = group_faults(p, nid);
1445 faults += score_nearby_nodes(p, nid, dist, false);
1446
1447 return 1000 * faults / total_faults;
1448}
1449
1450bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1451 int src_nid, int dst_cpu)
1452{
1453 struct numa_group *ng = deref_curr_numa_group(p);
1454 int dst_nid = cpu_to_node(dst_cpu);
1455 int last_cpupid, this_cpupid;
1456
1457 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1458 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1459
1460
1461
1462
1463
1464
1465
1466 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1467 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1468 return true;
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487 if (!cpupid_pid_unset(last_cpupid) &&
1488 cpupid_to_nid(last_cpupid) != dst_nid)
1489 return false;
1490
1491
1492 if (cpupid_match_pid(p, last_cpupid))
1493 return true;
1494
1495
1496 if (!ng)
1497 return true;
1498
1499
1500
1501
1502
1503 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1504 ACTIVE_NODE_FRACTION)
1505 return true;
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1516 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1517}
1518
1519static unsigned long cpu_runnable_load(struct rq *rq);
1520
1521
1522struct numa_stats {
1523 unsigned long load;
1524
1525
1526 unsigned long compute_capacity;
1527};
1528
1529
1530
1531
1532static void update_numa_stats(struct numa_stats *ns, int nid)
1533{
1534 int cpu;
1535
1536 memset(ns, 0, sizeof(*ns));
1537 for_each_cpu(cpu, cpumask_of_node(nid)) {
1538 struct rq *rq = cpu_rq(cpu);
1539
1540 ns->load += cpu_runnable_load(rq);
1541 ns->compute_capacity += capacity_of(cpu);
1542 }
1543
1544}
1545
1546struct task_numa_env {
1547 struct task_struct *p;
1548
1549 int src_cpu, src_nid;
1550 int dst_cpu, dst_nid;
1551
1552 struct numa_stats src_stats, dst_stats;
1553
1554 int imbalance_pct;
1555 int dist;
1556
1557 struct task_struct *best_task;
1558 long best_imp;
1559 int best_cpu;
1560};
1561
1562static void task_numa_assign(struct task_numa_env *env,
1563 struct task_struct *p, long imp)
1564{
1565 struct rq *rq = cpu_rq(env->dst_cpu);
1566
1567
1568 if (xchg(&rq->numa_migrate_on, 1))
1569 return;
1570
1571
1572
1573
1574
1575 if (env->best_cpu != -1) {
1576 rq = cpu_rq(env->best_cpu);
1577 WRITE_ONCE(rq->numa_migrate_on, 0);
1578 }
1579
1580 if (env->best_task)
1581 put_task_struct(env->best_task);
1582 if (p)
1583 get_task_struct(p);
1584
1585 env->best_task = p;
1586 env->best_imp = imp;
1587 env->best_cpu = env->dst_cpu;
1588}
1589
1590static bool load_too_imbalanced(long src_load, long dst_load,
1591 struct task_numa_env *env)
1592{
1593 long imb, old_imb;
1594 long orig_src_load, orig_dst_load;
1595 long src_capacity, dst_capacity;
1596
1597
1598
1599
1600
1601
1602
1603
1604 src_capacity = env->src_stats.compute_capacity;
1605 dst_capacity = env->dst_stats.compute_capacity;
1606
1607 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1608
1609 orig_src_load = env->src_stats.load;
1610 orig_dst_load = env->dst_stats.load;
1611
1612 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1613
1614
1615 return (imb > old_imb);
1616}
1617
1618
1619
1620
1621
1622
1623#define SMALLIMP 30
1624
1625
1626
1627
1628
1629
1630
1631static void task_numa_compare(struct task_numa_env *env,
1632 long taskimp, long groupimp, bool maymove)
1633{
1634 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1635 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1636 long imp = p_ng ? groupimp : taskimp;
1637 struct task_struct *cur;
1638 long src_load, dst_load;
1639 int dist = env->dist;
1640 long moveimp = imp;
1641 long load;
1642
1643 if (READ_ONCE(dst_rq->numa_migrate_on))
1644 return;
1645
1646 rcu_read_lock();
1647 cur = task_rcu_dereference(&dst_rq->curr);
1648 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1649 cur = NULL;
1650
1651
1652
1653
1654
1655 if (cur == env->p)
1656 goto unlock;
1657
1658 if (!cur) {
1659 if (maymove && moveimp >= env->best_imp)
1660 goto assign;
1661 else
1662 goto unlock;
1663 }
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1674 goto unlock;
1675
1676
1677
1678
1679
1680 cur_ng = rcu_dereference(cur->numa_group);
1681 if (cur_ng == p_ng) {
1682 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1683 task_weight(cur, env->dst_nid, dist);
1684
1685
1686
1687
1688 if (cur_ng)
1689 imp -= imp / 16;
1690 } else {
1691
1692
1693
1694
1695 if (cur_ng && p_ng)
1696 imp += group_weight(cur, env->src_nid, dist) -
1697 group_weight(cur, env->dst_nid, dist);
1698 else
1699 imp += task_weight(cur, env->src_nid, dist) -
1700 task_weight(cur, env->dst_nid, dist);
1701 }
1702
1703 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1704 imp = moveimp;
1705 cur = NULL;
1706 goto assign;
1707 }
1708
1709
1710
1711
1712
1713
1714
1715 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1716 goto unlock;
1717
1718
1719
1720
1721 load = task_h_load(env->p) - task_h_load(cur);
1722 if (!load)
1723 goto assign;
1724
1725 dst_load = env->dst_stats.load + load;
1726 src_load = env->src_stats.load - load;
1727
1728 if (load_too_imbalanced(src_load, dst_load, env))
1729 goto unlock;
1730
1731assign:
1732
1733
1734
1735
1736 if (!cur) {
1737
1738
1739
1740
1741 local_irq_disable();
1742 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1743 env->dst_cpu);
1744 local_irq_enable();
1745 }
1746
1747 task_numa_assign(env, cur, imp);
1748unlock:
1749 rcu_read_unlock();
1750}
1751
1752static void task_numa_find_cpu(struct task_numa_env *env,
1753 long taskimp, long groupimp)
1754{
1755 long src_load, dst_load, load;
1756 bool maymove = false;
1757 int cpu;
1758
1759 load = task_h_load(env->p);
1760 dst_load = env->dst_stats.load + load;
1761 src_load = env->src_stats.load - load;
1762
1763
1764
1765
1766
1767 maymove = !load_too_imbalanced(src_load, dst_load, env);
1768
1769 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1770
1771 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1772 continue;
1773
1774 env->dst_cpu = cpu;
1775 task_numa_compare(env, taskimp, groupimp, maymove);
1776 }
1777}
1778
1779static int task_numa_migrate(struct task_struct *p)
1780{
1781 struct task_numa_env env = {
1782 .p = p,
1783
1784 .src_cpu = task_cpu(p),
1785 .src_nid = task_node(p),
1786
1787 .imbalance_pct = 112,
1788
1789 .best_task = NULL,
1790 .best_imp = 0,
1791 .best_cpu = -1,
1792 };
1793 unsigned long taskweight, groupweight;
1794 struct sched_domain *sd;
1795 long taskimp, groupimp;
1796 struct numa_group *ng;
1797 struct rq *best_rq;
1798 int nid, ret, dist;
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808 rcu_read_lock();
1809 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1810 if (sd)
1811 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1812 rcu_read_unlock();
1813
1814
1815
1816
1817
1818
1819
1820 if (unlikely(!sd)) {
1821 sched_setnuma(p, task_node(p));
1822 return -EINVAL;
1823 }
1824
1825 env.dst_nid = p->numa_preferred_nid;
1826 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1827 taskweight = task_weight(p, env.src_nid, dist);
1828 groupweight = group_weight(p, env.src_nid, dist);
1829 update_numa_stats(&env.src_stats, env.src_nid);
1830 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1831 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1832 update_numa_stats(&env.dst_stats, env.dst_nid);
1833
1834
1835 task_numa_find_cpu(&env, taskimp, groupimp);
1836
1837
1838
1839
1840
1841
1842
1843
1844 ng = deref_curr_numa_group(p);
1845 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1846 for_each_online_node(nid) {
1847 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1848 continue;
1849
1850 dist = node_distance(env.src_nid, env.dst_nid);
1851 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1852 dist != env.dist) {
1853 taskweight = task_weight(p, env.src_nid, dist);
1854 groupweight = group_weight(p, env.src_nid, dist);
1855 }
1856
1857
1858 taskimp = task_weight(p, nid, dist) - taskweight;
1859 groupimp = group_weight(p, nid, dist) - groupweight;
1860 if (taskimp < 0 && groupimp < 0)
1861 continue;
1862
1863 env.dist = dist;
1864 env.dst_nid = nid;
1865 update_numa_stats(&env.dst_stats, env.dst_nid);
1866 task_numa_find_cpu(&env, taskimp, groupimp);
1867 }
1868 }
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878 if (ng) {
1879 if (env.best_cpu == -1)
1880 nid = env.src_nid;
1881 else
1882 nid = cpu_to_node(env.best_cpu);
1883
1884 if (nid != p->numa_preferred_nid)
1885 sched_setnuma(p, nid);
1886 }
1887
1888
1889 if (env.best_cpu == -1)
1890 return -EAGAIN;
1891
1892 best_rq = cpu_rq(env.best_cpu);
1893 if (env.best_task == NULL) {
1894 ret = migrate_task_to(p, env.best_cpu);
1895 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1896 if (ret != 0)
1897 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1898 return ret;
1899 }
1900
1901 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1902 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1903
1904 if (ret != 0)
1905 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1906 put_task_struct(env.best_task);
1907 return ret;
1908}
1909
1910
1911static void numa_migrate_preferred(struct task_struct *p)
1912{
1913 unsigned long interval = HZ;
1914
1915
1916 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
1917 return;
1918
1919
1920 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1921 p->numa_migrate_retry = jiffies + interval;
1922
1923
1924 if (task_node(p) == p->numa_preferred_nid)
1925 return;
1926
1927
1928 task_numa_migrate(p);
1929}
1930
1931
1932
1933
1934
1935
1936
1937static void numa_group_count_active_nodes(struct numa_group *numa_group)
1938{
1939 unsigned long faults, max_faults = 0;
1940 int nid, active_nodes = 0;
1941
1942 for_each_online_node(nid) {
1943 faults = group_faults_cpu(numa_group, nid);
1944 if (faults > max_faults)
1945 max_faults = faults;
1946 }
1947
1948 for_each_online_node(nid) {
1949 faults = group_faults_cpu(numa_group, nid);
1950 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1951 active_nodes++;
1952 }
1953
1954 numa_group->max_faults_cpu = max_faults;
1955 numa_group->active_nodes = active_nodes;
1956}
1957
1958
1959
1960
1961
1962
1963
1964
1965#define NUMA_PERIOD_SLOTS 10
1966#define NUMA_PERIOD_THRESHOLD 7
1967
1968
1969
1970
1971
1972
1973
1974static void update_task_scan_period(struct task_struct *p,
1975 unsigned long shared, unsigned long private)
1976{
1977 unsigned int period_slot;
1978 int lr_ratio, ps_ratio;
1979 int diff;
1980
1981 unsigned long remote = p->numa_faults_locality[0];
1982 unsigned long local = p->numa_faults_locality[1];
1983
1984
1985
1986
1987
1988
1989
1990
1991 if (local + shared == 0 || p->numa_faults_locality[2]) {
1992 p->numa_scan_period = min(p->numa_scan_period_max,
1993 p->numa_scan_period << 1);
1994
1995 p->mm->numa_next_scan = jiffies +
1996 msecs_to_jiffies(p->numa_scan_period);
1997
1998 return;
1999 }
2000
2001
2002
2003
2004
2005
2006
2007 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2008 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2009 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2010
2011 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2012
2013
2014
2015
2016 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2017 if (!slot)
2018 slot = 1;
2019 diff = slot * period_slot;
2020 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2021
2022
2023
2024
2025
2026 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2027 if (!slot)
2028 slot = 1;
2029 diff = slot * period_slot;
2030 } else {
2031
2032
2033
2034
2035
2036 int ratio = max(lr_ratio, ps_ratio);
2037 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2038 }
2039
2040 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2041 task_scan_min(p), task_scan_max(p));
2042 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2043}
2044
2045
2046
2047
2048
2049
2050
2051
2052static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2053{
2054 u64 runtime, delta, now;
2055
2056 now = p->se.exec_start;
2057 runtime = p->se.sum_exec_runtime;
2058
2059 if (p->last_task_numa_placement) {
2060 delta = runtime - p->last_sum_exec_runtime;
2061 *period = now - p->last_task_numa_placement;
2062
2063
2064 if (unlikely((s64)*period < 0))
2065 *period = 0;
2066 } else {
2067 delta = p->se.avg.load_sum;
2068 *period = LOAD_AVG_MAX;
2069 }
2070
2071 p->last_sum_exec_runtime = runtime;
2072 p->last_task_numa_placement = now;
2073
2074 return delta;
2075}
2076
2077
2078
2079
2080
2081
2082static int preferred_group_nid(struct task_struct *p, int nid)
2083{
2084 nodemask_t nodes;
2085 int dist;
2086
2087
2088 if (sched_numa_topology_type == NUMA_DIRECT)
2089 return nid;
2090
2091
2092
2093
2094
2095
2096 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2097 unsigned long score, max_score = 0;
2098 int node, max_node = nid;
2099
2100 dist = sched_max_numa_distance;
2101
2102 for_each_online_node(node) {
2103 score = group_weight(p, node, dist);
2104 if (score > max_score) {
2105 max_score = score;
2106 max_node = node;
2107 }
2108 }
2109 return max_node;
2110 }
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121 nodes = node_online_map;
2122 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2123 unsigned long max_faults = 0;
2124 nodemask_t max_group = NODE_MASK_NONE;
2125 int a, b;
2126
2127
2128 if (!find_numa_distance(dist))
2129 continue;
2130
2131 for_each_node_mask(a, nodes) {
2132 unsigned long faults = 0;
2133 nodemask_t this_group;
2134 nodes_clear(this_group);
2135
2136
2137 for_each_node_mask(b, nodes) {
2138 if (node_distance(a, b) < dist) {
2139 faults += group_faults(p, b);
2140 node_set(b, this_group);
2141 node_clear(b, nodes);
2142 }
2143 }
2144
2145
2146 if (faults > max_faults) {
2147 max_faults = faults;
2148 max_group = this_group;
2149
2150
2151
2152
2153
2154 nid = a;
2155 }
2156 }
2157
2158 if (!max_faults)
2159 break;
2160 nodes = max_group;
2161 }
2162 return nid;
2163}
2164
2165static void task_numa_placement(struct task_struct *p)
2166{
2167 int seq, nid, max_nid = NUMA_NO_NODE;
2168 unsigned long max_faults = 0;
2169 unsigned long fault_types[2] = { 0, 0 };
2170 unsigned long total_faults;
2171 u64 runtime, period;
2172 spinlock_t *group_lock = NULL;
2173 struct numa_group *ng;
2174
2175
2176
2177
2178
2179
2180 seq = READ_ONCE(p->mm->numa_scan_seq);
2181 if (p->numa_scan_seq == seq)
2182 return;
2183 p->numa_scan_seq = seq;
2184 p->numa_scan_period_max = task_scan_max(p);
2185
2186 total_faults = p->numa_faults_locality[0] +
2187 p->numa_faults_locality[1];
2188 runtime = numa_get_avg_runtime(p, &period);
2189
2190
2191 ng = deref_curr_numa_group(p);
2192 if (ng) {
2193 group_lock = &ng->lock;
2194 spin_lock_irq(group_lock);
2195 }
2196
2197
2198 for_each_online_node(nid) {
2199
2200 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2201 unsigned long faults = 0, group_faults = 0;
2202 int priv;
2203
2204 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2205 long diff, f_diff, f_weight;
2206
2207 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2208 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2209 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2210 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2211
2212
2213 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2214 fault_types[priv] += p->numa_faults[membuf_idx];
2215 p->numa_faults[membuf_idx] = 0;
2216
2217
2218
2219
2220
2221
2222
2223
2224 f_weight = div64_u64(runtime << 16, period + 1);
2225 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2226 (total_faults + 1);
2227 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2228 p->numa_faults[cpubuf_idx] = 0;
2229
2230 p->numa_faults[mem_idx] += diff;
2231 p->numa_faults[cpu_idx] += f_diff;
2232 faults += p->numa_faults[mem_idx];
2233 p->total_numa_faults += diff;
2234 if (ng) {
2235
2236
2237
2238
2239
2240
2241
2242 ng->faults[mem_idx] += diff;
2243 ng->faults_cpu[mem_idx] += f_diff;
2244 ng->total_faults += diff;
2245 group_faults += ng->faults[mem_idx];
2246 }
2247 }
2248
2249 if (!ng) {
2250 if (faults > max_faults) {
2251 max_faults = faults;
2252 max_nid = nid;
2253 }
2254 } else if (group_faults > max_faults) {
2255 max_faults = group_faults;
2256 max_nid = nid;
2257 }
2258 }
2259
2260 if (ng) {
2261 numa_group_count_active_nodes(ng);
2262 spin_unlock_irq(group_lock);
2263 max_nid = preferred_group_nid(p, max_nid);
2264 }
2265
2266 if (max_faults) {
2267
2268 if (max_nid != p->numa_preferred_nid)
2269 sched_setnuma(p, max_nid);
2270 }
2271
2272 update_task_scan_period(p, fault_types[0], fault_types[1]);
2273}
2274
2275static inline int get_numa_group(struct numa_group *grp)
2276{
2277 return refcount_inc_not_zero(&grp->refcount);
2278}
2279
2280static inline void put_numa_group(struct numa_group *grp)
2281{
2282 if (refcount_dec_and_test(&grp->refcount))
2283 kfree_rcu(grp, rcu);
2284}
2285
2286static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2287 int *priv)
2288{
2289 struct numa_group *grp, *my_grp;
2290 struct task_struct *tsk;
2291 bool join = false;
2292 int cpu = cpupid_to_cpu(cpupid);
2293 int i;
2294
2295 if (unlikely(!deref_curr_numa_group(p))) {
2296 unsigned int size = sizeof(struct numa_group) +
2297 4*nr_node_ids*sizeof(unsigned long);
2298
2299 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2300 if (!grp)
2301 return;
2302
2303 refcount_set(&grp->refcount, 1);
2304 grp->active_nodes = 1;
2305 grp->max_faults_cpu = 0;
2306 spin_lock_init(&grp->lock);
2307 grp->gid = p->pid;
2308
2309 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2310 nr_node_ids;
2311
2312 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2313 grp->faults[i] = p->numa_faults[i];
2314
2315 grp->total_faults = p->total_numa_faults;
2316
2317 grp->nr_tasks++;
2318 rcu_assign_pointer(p->numa_group, grp);
2319 }
2320
2321 rcu_read_lock();
2322 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2323
2324 if (!cpupid_match_pid(tsk, cpupid))
2325 goto no_join;
2326
2327 grp = rcu_dereference(tsk->numa_group);
2328 if (!grp)
2329 goto no_join;
2330
2331 my_grp = deref_curr_numa_group(p);
2332 if (grp == my_grp)
2333 goto no_join;
2334
2335
2336
2337
2338
2339 if (my_grp->nr_tasks > grp->nr_tasks)
2340 goto no_join;
2341
2342
2343
2344
2345 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2346 goto no_join;
2347
2348
2349 if (tsk->mm == current->mm)
2350 join = true;
2351
2352
2353 if (flags & TNF_SHARED)
2354 join = true;
2355
2356
2357 *priv = !join;
2358
2359 if (join && !get_numa_group(grp))
2360 goto no_join;
2361
2362 rcu_read_unlock();
2363
2364 if (!join)
2365 return;
2366
2367 BUG_ON(irqs_disabled());
2368 double_lock_irq(&my_grp->lock, &grp->lock);
2369
2370 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2371 my_grp->faults[i] -= p->numa_faults[i];
2372 grp->faults[i] += p->numa_faults[i];
2373 }
2374 my_grp->total_faults -= p->total_numa_faults;
2375 grp->total_faults += p->total_numa_faults;
2376
2377 my_grp->nr_tasks--;
2378 grp->nr_tasks++;
2379
2380 spin_unlock(&my_grp->lock);
2381 spin_unlock_irq(&grp->lock);
2382
2383 rcu_assign_pointer(p->numa_group, grp);
2384
2385 put_numa_group(my_grp);
2386 return;
2387
2388no_join:
2389 rcu_read_unlock();
2390 return;
2391}
2392
2393
2394
2395
2396
2397
2398
2399
2400void task_numa_free(struct task_struct *p, bool final)
2401{
2402
2403 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2404 unsigned long *numa_faults = p->numa_faults;
2405 unsigned long flags;
2406 int i;
2407
2408 if (!numa_faults)
2409 return;
2410
2411 if (grp) {
2412 spin_lock_irqsave(&grp->lock, flags);
2413 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2414 grp->faults[i] -= p->numa_faults[i];
2415 grp->total_faults -= p->total_numa_faults;
2416
2417 grp->nr_tasks--;
2418 spin_unlock_irqrestore(&grp->lock, flags);
2419 RCU_INIT_POINTER(p->numa_group, NULL);
2420 put_numa_group(grp);
2421 }
2422
2423 if (final) {
2424 p->numa_faults = NULL;
2425 kfree(numa_faults);
2426 } else {
2427 p->total_numa_faults = 0;
2428 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2429 numa_faults[i] = 0;
2430 }
2431}
2432
2433
2434
2435
2436void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2437{
2438 struct task_struct *p = current;
2439 bool migrated = flags & TNF_MIGRATED;
2440 int cpu_node = task_node(current);
2441 int local = !!(flags & TNF_FAULT_LOCAL);
2442 struct numa_group *ng;
2443 int priv;
2444
2445 if (!static_branch_likely(&sched_numa_balancing))
2446 return;
2447
2448
2449 if (!p->mm)
2450 return;
2451
2452
2453 if (unlikely(!p->numa_faults)) {
2454 int size = sizeof(*p->numa_faults) *
2455 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2456
2457 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2458 if (!p->numa_faults)
2459 return;
2460
2461 p->total_numa_faults = 0;
2462 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2463 }
2464
2465
2466
2467
2468
2469 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2470 priv = 1;
2471 } else {
2472 priv = cpupid_match_pid(p, last_cpupid);
2473 if (!priv && !(flags & TNF_NO_GROUP))
2474 task_numa_group(p, last_cpupid, flags, &priv);
2475 }
2476
2477
2478
2479
2480
2481
2482
2483 ng = deref_curr_numa_group(p);
2484 if (!priv && !local && ng && ng->active_nodes > 1 &&
2485 numa_is_active_node(cpu_node, ng) &&
2486 numa_is_active_node(mem_node, ng))
2487 local = 1;
2488
2489
2490
2491
2492
2493 if (time_after(jiffies, p->numa_migrate_retry)) {
2494 task_numa_placement(p);
2495 numa_migrate_preferred(p);
2496 }
2497
2498 if (migrated)
2499 p->numa_pages_migrated += pages;
2500 if (flags & TNF_MIGRATE_FAIL)
2501 p->numa_faults_locality[2] += pages;
2502
2503 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2504 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2505 p->numa_faults_locality[local] += pages;
2506}
2507
2508static void reset_ptenuma_scan(struct task_struct *p)
2509{
2510
2511
2512
2513
2514
2515
2516
2517
2518 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2519 p->mm->numa_scan_offset = 0;
2520}
2521
2522
2523
2524
2525
2526void task_numa_work(struct callback_head *work)
2527{
2528 unsigned long migrate, next_scan, now = jiffies;
2529 struct task_struct *p = current;
2530 struct mm_struct *mm = p->mm;
2531 u64 runtime = p->se.sum_exec_runtime;
2532 struct vm_area_struct *vma;
2533 unsigned long start, end;
2534 unsigned long nr_pte_updates = 0;
2535 long pages, virtpages;
2536
2537 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2538
2539 work->next = work;
2540
2541
2542
2543
2544
2545
2546
2547
2548 if (p->flags & PF_EXITING)
2549 return;
2550
2551 if (!mm->numa_next_scan) {
2552 mm->numa_next_scan = now +
2553 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2554 }
2555
2556
2557
2558
2559 migrate = mm->numa_next_scan;
2560 if (time_before(now, migrate))
2561 return;
2562
2563 if (p->numa_scan_period == 0) {
2564 p->numa_scan_period_max = task_scan_max(p);
2565 p->numa_scan_period = task_scan_start(p);
2566 }
2567
2568 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2569 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2570 return;
2571
2572
2573
2574
2575
2576 p->node_stamp += 2 * TICK_NSEC;
2577
2578 start = mm->numa_scan_offset;
2579 pages = sysctl_numa_balancing_scan_size;
2580 pages <<= 20 - PAGE_SHIFT;
2581 virtpages = pages * 8;
2582 if (!pages)
2583 return;
2584
2585
2586 if (!down_read_trylock(&mm->mmap_sem))
2587 return;
2588 vma = find_vma(mm, start);
2589 if (!vma) {
2590 reset_ptenuma_scan(p);
2591 start = 0;
2592 vma = mm->mmap;
2593 }
2594 for (; vma; vma = vma->vm_next) {
2595 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2596 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2597 continue;
2598 }
2599
2600
2601
2602
2603
2604
2605
2606 if (!vma->vm_mm ||
2607 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2608 continue;
2609
2610
2611
2612
2613
2614 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2615 continue;
2616
2617 do {
2618 start = max(start, vma->vm_start);
2619 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2620 end = min(end, vma->vm_end);
2621 nr_pte_updates = change_prot_numa(vma, start, end);
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631 if (nr_pte_updates)
2632 pages -= (end - start) >> PAGE_SHIFT;
2633 virtpages -= (end - start) >> PAGE_SHIFT;
2634
2635 start = end;
2636 if (pages <= 0 || virtpages <= 0)
2637 goto out;
2638
2639 cond_resched();
2640 } while (end != vma->vm_end);
2641 }
2642
2643out:
2644
2645
2646
2647
2648
2649
2650 if (vma)
2651 mm->numa_scan_offset = start;
2652 else
2653 reset_ptenuma_scan(p);
2654 up_read(&mm->mmap_sem);
2655
2656
2657
2658
2659
2660
2661
2662 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2663 u64 diff = p->se.sum_exec_runtime - runtime;
2664 p->node_stamp += 32 * diff;
2665 }
2666}
2667
2668
2669
2670
2671static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2672{
2673 struct callback_head *work = &curr->numa_work;
2674 u64 period, now;
2675
2676
2677
2678
2679 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2680 return;
2681
2682
2683
2684
2685
2686
2687
2688 now = curr->se.sum_exec_runtime;
2689 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2690
2691 if (now > curr->node_stamp + period) {
2692 if (!curr->node_stamp)
2693 curr->numa_scan_period = task_scan_start(curr);
2694 curr->node_stamp += period;
2695
2696 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2697 init_task_work(work, task_numa_work);
2698 task_work_add(curr, work, true);
2699 }
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
2719
2720
2721
2722 if (p->numa_scan_seq) {
2723
2724
2725
2726
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
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
2786
2787
2788
2789
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
2806
2807
2808
2809
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
2823
2824
2825
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
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
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
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
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
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3026}
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
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
3072
3073static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3074
3075
3076
3077
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
3104static inline void update_cfs_group(struct sched_entity *se)
3105{
3106}
3107#endif
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
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128 cpufreq_update_util(rq, flags);
3129 }
3130}
3131
3132#ifdef CONFIG_SMP
3133#ifdef CONFIG_FAIR_GROUP_SCHED
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
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
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
3167
3168
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
3181
3182
3183
3184
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
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
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
3289 if (!delta)
3290 return;
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301 se->avg.util_avg = gcfs_rq->avg.util_avg;
3302 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3303
3304
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
3325
3326
3327 runnable_sum += se->avg.load_sum;
3328 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3329 } else {
3330
3331
3332
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
3340 runnable_sum = min(se->avg.load_sum, load_sum);
3341 }
3342
3343
3344
3345
3346
3347
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
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
3412
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
3420
3421
3422 if (se->avg.load_avg || se->avg.util_avg)
3423 return false;
3424
3425
3426
3427
3428
3429 if (gcfs_rq->propagate)
3430 return false;
3431
3432
3433
3434
3435
3436
3437 return true;
3438}
3439
3440#else
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
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
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
3515
3516
3517
3518
3519
3520
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
3528
3529
3530
3531
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
3538
3539
3540
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
3565
3566
3567
3568
3569
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
3586
3587#define UPDATE_TG 0x1
3588#define SKIP_AGE_LOAD 0x2
3589#define DO_ATTACH 0x4
3590
3591
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
3599
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
3611
3612
3613
3614
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
3646
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
3659
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
3668
3669
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;
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 int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3693
3694static inline unsigned long task_util(struct task_struct *p)
3695{
3696 return READ_ONCE(p->se.avg.util_avg);
3697}
3698
3699static inline unsigned long _task_util_est(struct task_struct *p)
3700{
3701 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3702
3703 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3704}
3705
3706static inline unsigned long task_util_est(struct task_struct *p)
3707{
3708 return max(task_util(p), _task_util_est(p));
3709}
3710
3711static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3712 struct task_struct *p)
3713{
3714 unsigned int enqueued;
3715
3716 if (!sched_feat(UTIL_EST))
3717 return;
3718
3719
3720 enqueued = cfs_rq->avg.util_est.enqueued;
3721 enqueued += _task_util_est(p);
3722 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3723}
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733static inline bool within_margin(int value, int margin)
3734{
3735 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3736}
3737
3738static void
3739util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3740{
3741 long last_ewma_diff;
3742 struct util_est ue;
3743 int cpu;
3744
3745 if (!sched_feat(UTIL_EST))
3746 return;
3747
3748
3749 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3750 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3751 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3752
3753
3754
3755
3756
3757 if (!task_sleep)
3758 return;
3759
3760
3761
3762
3763
3764 ue = p->se.avg.util_est;
3765 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3766 return;
3767
3768
3769
3770
3771
3772 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3773 last_ewma_diff = ue.enqueued - ue.ewma;
3774 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3775 return;
3776
3777
3778
3779
3780
3781 cpu = cpu_of(rq_of(cfs_rq));
3782 if (task_util(p) > capacity_orig_of(cpu))
3783 return;
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3803 ue.ewma += last_ewma_diff;
3804 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3805 WRITE_ONCE(p->se.avg.util_est, ue);
3806}
3807
3808static inline int task_fits_capacity(struct task_struct *p, long capacity)
3809{
3810 return capacity * 1024 > task_util_est(p) * capacity_margin;
3811}
3812
3813static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
3814{
3815 if (!static_branch_unlikely(&sched_asym_cpucapacity))
3816 return;
3817
3818 if (!p) {
3819 rq->misfit_task_load = 0;
3820 return;
3821 }
3822
3823 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
3824 rq->misfit_task_load = 0;
3825 return;
3826 }
3827
3828 rq->misfit_task_load = task_h_load(p);
3829}
3830
3831#else
3832
3833#define UPDATE_TG 0x0
3834#define SKIP_AGE_LOAD 0x0
3835#define DO_ATTACH 0x0
3836
3837static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3838{
3839 cfs_rq_util_change(cfs_rq, 0);
3840}
3841
3842static inline void remove_entity_load_avg(struct sched_entity *se) {}
3843
3844static inline void
3845attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3846static inline void
3847detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3848
3849static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3850{
3851 return 0;
3852}
3853
3854static inline void
3855util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3856
3857static inline void
3858util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3859 bool task_sleep) {}
3860static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
3861
3862#endif
3863
3864static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3865{
3866#ifdef CONFIG_SCHED_DEBUG
3867 s64 d = se->vruntime - cfs_rq->min_vruntime;
3868
3869 if (d < 0)
3870 d = -d;
3871
3872 if (d > 3*sysctl_sched_latency)
3873 schedstat_inc(cfs_rq->nr_spread_over);
3874#endif
3875}
3876
3877static void
3878place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3879{
3880 u64 vruntime = cfs_rq->min_vruntime;
3881
3882
3883
3884
3885
3886
3887
3888 if (initial && sched_feat(START_DEBIT))
3889 vruntime += sched_vslice(cfs_rq, se);
3890
3891
3892 if (!initial) {
3893 unsigned long thresh = sysctl_sched_latency;
3894
3895
3896
3897
3898
3899 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3900 thresh >>= 1;
3901
3902 vruntime -= thresh;
3903 }
3904
3905
3906 se->vruntime = max_vruntime(se->vruntime, vruntime);
3907}
3908
3909static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3910
3911static inline void check_schedstat_required(void)
3912{
3913#ifdef CONFIG_SCHEDSTATS
3914 if (schedstat_enabled())
3915 return;
3916
3917
3918 if (trace_sched_stat_wait_enabled() ||
3919 trace_sched_stat_sleep_enabled() ||
3920 trace_sched_stat_iowait_enabled() ||
3921 trace_sched_stat_blocked_enabled() ||
3922 trace_sched_stat_runtime_enabled()) {
3923 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3924 "stat_blocked and stat_runtime require the "
3925 "kernel parameter schedstats=enable or "
3926 "kernel.sched_schedstats=1\n");
3927 }
3928#endif
3929}
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962static void
3963enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3964{
3965 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3966 bool curr = cfs_rq->curr == se;
3967
3968
3969
3970
3971
3972 if (renorm && curr)
3973 se->vruntime += cfs_rq->min_vruntime;
3974
3975 update_curr(cfs_rq);
3976
3977
3978
3979
3980
3981
3982
3983 if (renorm && !curr)
3984 se->vruntime += cfs_rq->min_vruntime;
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3995 update_cfs_group(se);
3996 enqueue_runnable_load_avg(cfs_rq, se);
3997 account_entity_enqueue(cfs_rq, se);
3998
3999 if (flags & ENQUEUE_WAKEUP)
4000 place_entity(cfs_rq, se, 0);
4001
4002 check_schedstat_required();
4003 update_stats_enqueue(cfs_rq, se, flags);
4004 check_spread(cfs_rq, se);
4005 if (!curr)
4006 __enqueue_entity(cfs_rq, se);
4007 se->on_rq = 1;
4008
4009 if (cfs_rq->nr_running == 1) {
4010 list_add_leaf_cfs_rq(cfs_rq);
4011 check_enqueue_throttle(cfs_rq);
4012 }
4013}
4014
4015static void __clear_buddies_last(struct sched_entity *se)
4016{
4017 for_each_sched_entity(se) {
4018 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4019 if (cfs_rq->last != se)
4020 break;
4021
4022 cfs_rq->last = NULL;
4023 }
4024}
4025
4026static void __clear_buddies_next(struct sched_entity *se)
4027{
4028 for_each_sched_entity(se) {
4029 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4030 if (cfs_rq->next != se)
4031 break;
4032
4033 cfs_rq->next = NULL;
4034 }
4035}
4036
4037static void __clear_buddies_skip(struct sched_entity *se)
4038{
4039 for_each_sched_entity(se) {
4040 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4041 if (cfs_rq->skip != se)
4042 break;
4043
4044 cfs_rq->skip = NULL;
4045 }
4046}
4047
4048static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4049{
4050 if (cfs_rq->last == se)
4051 __clear_buddies_last(se);
4052
4053 if (cfs_rq->next == se)
4054 __clear_buddies_next(se);
4055
4056 if (cfs_rq->skip == se)
4057 __clear_buddies_skip(se);
4058}
4059
4060static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4061
4062static void
4063dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4064{
4065
4066
4067
4068 update_curr(cfs_rq);
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078 update_load_avg(cfs_rq, se, UPDATE_TG);
4079 dequeue_runnable_load_avg(cfs_rq, se);
4080
4081 update_stats_dequeue(cfs_rq, se, flags);
4082
4083 clear_buddies(cfs_rq, se);
4084
4085 if (se != cfs_rq->curr)
4086 __dequeue_entity(cfs_rq, se);
4087 se->on_rq = 0;
4088 account_entity_dequeue(cfs_rq, se);
4089
4090
4091
4092
4093
4094
4095
4096 if (!(flags & DEQUEUE_SLEEP))
4097 se->vruntime -= cfs_rq->min_vruntime;
4098
4099
4100 return_cfs_rq_runtime(cfs_rq);
4101
4102 update_cfs_group(se);
4103
4104
4105
4106
4107
4108
4109
4110 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4111 update_min_vruntime(cfs_rq);
4112}
4113
4114
4115
4116
4117static void
4118check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4119{
4120 unsigned long ideal_runtime, delta_exec;
4121 struct sched_entity *se;
4122 s64 delta;
4123
4124 ideal_runtime = sched_slice(cfs_rq, curr);
4125 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4126 if (delta_exec > ideal_runtime) {
4127 resched_curr(rq_of(cfs_rq));
4128
4129
4130
4131
4132 clear_buddies(cfs_rq, curr);
4133 return;
4134 }
4135
4136
4137
4138
4139
4140
4141 if (delta_exec < sysctl_sched_min_granularity)
4142 return;
4143
4144 se = __pick_first_entity(cfs_rq);
4145 delta = curr->vruntime - se->vruntime;
4146
4147 if (delta < 0)
4148 return;
4149
4150 if (delta > ideal_runtime)
4151 resched_curr(rq_of(cfs_rq));
4152}
4153
4154static void
4155set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4156{
4157
4158 if (se->on_rq) {
4159
4160
4161
4162
4163
4164 update_stats_wait_end(cfs_rq, se);
4165 __dequeue_entity(cfs_rq, se);
4166 update_load_avg(cfs_rq, se, UPDATE_TG);
4167 }
4168
4169 update_stats_curr_start(cfs_rq, se);
4170 cfs_rq->curr = se;
4171
4172
4173
4174
4175
4176
4177 if (schedstat_enabled() &&
4178 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4179 schedstat_set(se->statistics.slice_max,
4180 max((u64)schedstat_val(se->statistics.slice_max),
4181 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4182 }
4183
4184 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4185}
4186
4187static int
4188wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4189
4190
4191
4192
4193
4194
4195
4196
4197static struct sched_entity *
4198pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4199{
4200 struct sched_entity *left = __pick_first_entity(cfs_rq);
4201 struct sched_entity *se;
4202
4203
4204
4205
4206
4207 if (!left || (curr && entity_before(curr, left)))
4208 left = curr;
4209
4210 se = left;
4211
4212
4213
4214
4215
4216 if (cfs_rq->skip == se) {
4217 struct sched_entity *second;
4218
4219 if (se == curr) {
4220 second = __pick_first_entity(cfs_rq);
4221 } else {
4222 second = __pick_next_entity(se);
4223 if (!second || (curr && entity_before(curr, second)))
4224 second = curr;
4225 }
4226
4227 if (second && wakeup_preempt_entity(second, left) < 1)
4228 se = second;
4229 }
4230
4231
4232
4233
4234 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4235 se = cfs_rq->last;
4236
4237
4238
4239
4240 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4241 se = cfs_rq->next;
4242
4243 clear_buddies(cfs_rq, se);
4244
4245 return se;
4246}
4247
4248static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4249
4250static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4251{
4252
4253
4254
4255
4256 if (prev->on_rq)
4257 update_curr(cfs_rq);
4258
4259
4260 check_cfs_rq_runtime(cfs_rq);
4261
4262 check_spread(cfs_rq, prev);
4263
4264 if (prev->on_rq) {
4265 update_stats_wait_start(cfs_rq, prev);
4266
4267 __enqueue_entity(cfs_rq, prev);
4268
4269 update_load_avg(cfs_rq, prev, 0);
4270 }
4271 cfs_rq->curr = NULL;
4272}
4273
4274static void
4275entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4276{
4277
4278
4279
4280 update_curr(cfs_rq);
4281
4282
4283
4284
4285 update_load_avg(cfs_rq, curr, UPDATE_TG);
4286 update_cfs_group(curr);
4287
4288#ifdef CONFIG_SCHED_HRTICK
4289
4290
4291
4292
4293 if (queued) {
4294 resched_curr(rq_of(cfs_rq));
4295 return;
4296 }
4297
4298
4299
4300 if (!sched_feat(DOUBLE_TICK) &&
4301 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4302 return;
4303#endif
4304
4305 if (cfs_rq->nr_running > 1)
4306 check_preempt_tick(cfs_rq, curr);
4307}
4308
4309
4310
4311
4312
4313
4314#ifdef CONFIG_CFS_BANDWIDTH
4315
4316#ifdef CONFIG_JUMP_LABEL
4317static struct static_key __cfs_bandwidth_used;
4318
4319static inline bool cfs_bandwidth_used(void)
4320{
4321 return static_key_false(&__cfs_bandwidth_used);
4322}
4323
4324void cfs_bandwidth_usage_inc(void)
4325{
4326 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4327}
4328
4329void cfs_bandwidth_usage_dec(void)
4330{
4331 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4332}
4333#else
4334static bool cfs_bandwidth_used(void)
4335{
4336 return true;
4337}
4338
4339void cfs_bandwidth_usage_inc(void) {}
4340void cfs_bandwidth_usage_dec(void) {}
4341#endif
4342
4343
4344
4345
4346
4347static inline u64 default_cfs_period(void)
4348{
4349 return 100000000ULL;
4350}
4351
4352static inline u64 sched_cfs_bandwidth_slice(void)
4353{
4354 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4355}
4356
4357
4358
4359
4360
4361
4362
4363
4364void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4365{
4366 u64 now;
4367
4368 if (cfs_b->quota == RUNTIME_INF)
4369 return;
4370
4371 now = sched_clock_cpu(smp_processor_id());
4372 cfs_b->runtime = cfs_b->quota;
4373 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4374 cfs_b->expires_seq++;
4375}
4376
4377static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4378{
4379 return &tg->cfs_bandwidth;
4380}
4381
4382
4383static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4384{
4385 if (unlikely(cfs_rq->throttle_count))
4386 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4387
4388 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4389}
4390
4391
4392static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4393{
4394 struct task_group *tg = cfs_rq->tg;
4395 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4396 u64 amount = 0, min_amount, expires;
4397 int expires_seq;
4398
4399
4400 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4401
4402 raw_spin_lock(&cfs_b->lock);
4403 if (cfs_b->quota == RUNTIME_INF)
4404 amount = min_amount;
4405 else {
4406 start_cfs_bandwidth(cfs_b);
4407
4408 if (cfs_b->runtime > 0) {
4409 amount = min(cfs_b->runtime, min_amount);
4410 cfs_b->runtime -= amount;
4411 cfs_b->idle = 0;
4412 }
4413 }
4414 expires_seq = cfs_b->expires_seq;
4415 expires = cfs_b->runtime_expires;
4416 raw_spin_unlock(&cfs_b->lock);
4417
4418 cfs_rq->runtime_remaining += amount;
4419
4420
4421
4422
4423
4424 if (cfs_rq->expires_seq != expires_seq) {
4425 cfs_rq->expires_seq = expires_seq;
4426 cfs_rq->runtime_expires = expires;
4427 }
4428
4429 return cfs_rq->runtime_remaining > 0;
4430}
4431
4432
4433
4434
4435
4436static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4437{
4438 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4439
4440
4441 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4442 return;
4443
4444 if (cfs_rq->runtime_remaining < 0)
4445 return;
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4456
4457 cfs_rq->runtime_expires += TICK_NSEC;
4458 } else {
4459
4460 cfs_rq->runtime_remaining = 0;
4461 }
4462}
4463
4464static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4465{
4466
4467 cfs_rq->runtime_remaining -= delta_exec;
4468 expire_cfs_rq_runtime(cfs_rq);
4469
4470 if (likely(cfs_rq->runtime_remaining > 0))
4471 return;
4472
4473 if (cfs_rq->throttled)
4474 return;
4475
4476
4477
4478
4479 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4480 resched_curr(rq_of(cfs_rq));
4481}
4482
4483static __always_inline
4484void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4485{
4486 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4487 return;
4488
4489 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4490}
4491
4492static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4493{
4494 return cfs_bandwidth_used() && cfs_rq->throttled;
4495}
4496
4497
4498static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4499{
4500 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4501}
4502
4503
4504
4505
4506
4507
4508static inline int throttled_lb_pair(struct task_group *tg,
4509 int src_cpu, int dest_cpu)
4510{
4511 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4512
4513 src_cfs_rq = tg->cfs_rq[src_cpu];
4514 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4515
4516 return throttled_hierarchy(src_cfs_rq) ||
4517 throttled_hierarchy(dest_cfs_rq);
4518}
4519
4520static int tg_unthrottle_up(struct task_group *tg, void *data)
4521{
4522 struct rq *rq = data;
4523 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4524
4525 cfs_rq->throttle_count--;
4526 if (!cfs_rq->throttle_count) {
4527
4528 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4529 cfs_rq->throttled_clock_task;
4530
4531
4532 if (cfs_rq->nr_running >= 1)
4533 list_add_leaf_cfs_rq(cfs_rq);
4534 }
4535
4536 return 0;
4537}
4538
4539static int tg_throttle_down(struct task_group *tg, void *data)
4540{
4541 struct rq *rq = data;
4542 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4543
4544
4545 if (!cfs_rq->throttle_count) {
4546 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4547 list_del_leaf_cfs_rq(cfs_rq);
4548 }
4549 cfs_rq->throttle_count++;
4550
4551 return 0;
4552}
4553
4554static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4555{
4556 struct rq *rq = rq_of(cfs_rq);
4557 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4558 struct sched_entity *se;
4559 long task_delta, dequeue = 1;
4560 bool empty;
4561
4562 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4563
4564
4565 rcu_read_lock();
4566 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4567 rcu_read_unlock();
4568
4569 task_delta = cfs_rq->h_nr_running;
4570 for_each_sched_entity(se) {
4571 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4572
4573 if (!se->on_rq)
4574 break;
4575
4576 if (dequeue)
4577 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4578 qcfs_rq->h_nr_running -= task_delta;
4579
4580 if (qcfs_rq->load.weight)
4581 dequeue = 0;
4582 }
4583
4584 if (!se)
4585 sub_nr_running(rq, task_delta);
4586
4587 cfs_rq->throttled = 1;
4588 cfs_rq->throttled_clock = rq_clock(rq);
4589 raw_spin_lock(&cfs_b->lock);
4590 empty = list_empty(&cfs_b->throttled_cfs_rq);
4591
4592
4593
4594
4595
4596
4597 if (cfs_b->distribute_running)
4598 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4599 else
4600 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4601
4602
4603
4604
4605
4606 if (empty)
4607 start_cfs_bandwidth(cfs_b);
4608
4609 raw_spin_unlock(&cfs_b->lock);
4610}
4611
4612void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4613{
4614 struct rq *rq = rq_of(cfs_rq);
4615 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4616 struct sched_entity *se;
4617 int enqueue = 1;
4618 long task_delta;
4619
4620 se = cfs_rq->tg->se[cpu_of(rq)];
4621
4622 cfs_rq->throttled = 0;
4623
4624 update_rq_clock(rq);
4625
4626 raw_spin_lock(&cfs_b->lock);
4627 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4628 list_del_rcu(&cfs_rq->throttled_list);
4629 raw_spin_unlock(&cfs_b->lock);
4630
4631
4632 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4633
4634 if (!cfs_rq->load.weight)
4635 return;
4636
4637 task_delta = cfs_rq->h_nr_running;
4638 for_each_sched_entity(se) {
4639 if (se->on_rq)
4640 enqueue = 0;
4641
4642 cfs_rq = cfs_rq_of(se);
4643 if (enqueue)
4644 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4645 cfs_rq->h_nr_running += task_delta;
4646
4647 if (cfs_rq_throttled(cfs_rq))
4648 break;
4649 }
4650
4651 assert_list_leaf_cfs_rq(rq);
4652
4653 if (!se)
4654 add_nr_running(rq, task_delta);
4655
4656
4657 if (rq->curr == rq->idle && rq->cfs.nr_running)
4658 resched_curr(rq);
4659}
4660
4661static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4662 u64 remaining, u64 expires)
4663{
4664 struct cfs_rq *cfs_rq;
4665 u64 runtime;
4666 u64 starting_runtime = remaining;
4667
4668 rcu_read_lock();
4669 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4670 throttled_list) {
4671 struct rq *rq = rq_of(cfs_rq);
4672 struct rq_flags rf;
4673
4674 rq_lock_irqsave(rq, &rf);
4675 if (!cfs_rq_throttled(cfs_rq))
4676 goto next;
4677
4678
4679 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4680
4681 runtime = -cfs_rq->runtime_remaining + 1;
4682 if (runtime > remaining)
4683 runtime = remaining;
4684 remaining -= runtime;
4685
4686 cfs_rq->runtime_remaining += runtime;
4687 cfs_rq->runtime_expires = expires;
4688
4689
4690 if (cfs_rq->runtime_remaining > 0)
4691 unthrottle_cfs_rq(cfs_rq);
4692
4693next:
4694 rq_unlock_irqrestore(rq, &rf);
4695
4696 if (!remaining)
4697 break;
4698 }
4699 rcu_read_unlock();
4700
4701 return starting_runtime - remaining;
4702}
4703
4704
4705
4706
4707
4708
4709
4710static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4711{
4712 u64 runtime, runtime_expires;
4713 int throttled;
4714
4715
4716 if (cfs_b->quota == RUNTIME_INF)
4717 goto out_deactivate;
4718
4719 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4720 cfs_b->nr_periods += overrun;
4721
4722
4723
4724
4725
4726 if (cfs_b->idle && !throttled)
4727 goto out_deactivate;
4728
4729 __refill_cfs_bandwidth_runtime(cfs_b);
4730
4731 if (!throttled) {
4732
4733 cfs_b->idle = 1;
4734 return 0;
4735 }
4736
4737
4738 cfs_b->nr_throttled += overrun;
4739
4740 runtime_expires = cfs_b->runtime_expires;
4741
4742
4743
4744
4745
4746
4747
4748
4749 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4750 runtime = cfs_b->runtime;
4751 cfs_b->distribute_running = 1;
4752 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4753
4754 runtime = distribute_cfs_runtime(cfs_b, runtime,
4755 runtime_expires);
4756 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4757
4758 cfs_b->distribute_running = 0;
4759 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4760
4761 lsub_positive(&cfs_b->runtime, runtime);
4762 }
4763
4764
4765
4766
4767
4768
4769
4770 cfs_b->idle = 0;
4771
4772 return 0;
4773
4774out_deactivate:
4775 return 1;
4776}
4777
4778
4779static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4780
4781static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4782
4783static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4784
4785
4786
4787
4788
4789
4790
4791
4792static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4793{
4794 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4795 u64 remaining;
4796
4797
4798 if (hrtimer_callback_running(refresh_timer))
4799 return 1;
4800
4801
4802 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4803 if (remaining < min_expire)
4804 return 1;
4805
4806 return 0;
4807}
4808
4809static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4810{
4811 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4812
4813
4814 if (runtime_refresh_within(cfs_b, min_left))
4815 return;
4816
4817
4818 if (cfs_b->slack_started)
4819 return;
4820 cfs_b->slack_started = true;
4821
4822 hrtimer_start(&cfs_b->slack_timer,
4823 ns_to_ktime(cfs_bandwidth_slack_period),
4824 HRTIMER_MODE_REL);
4825}
4826
4827
4828static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4829{
4830 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4831 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4832
4833 if (slack_runtime <= 0)
4834 return;
4835
4836 raw_spin_lock(&cfs_b->lock);
4837 if (cfs_b->quota != RUNTIME_INF &&
4838 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4839 cfs_b->runtime += slack_runtime;
4840
4841
4842 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4843 !list_empty(&cfs_b->throttled_cfs_rq))
4844 start_cfs_slack_bandwidth(cfs_b);
4845 }
4846 raw_spin_unlock(&cfs_b->lock);
4847
4848
4849 cfs_rq->runtime_remaining -= slack_runtime;
4850}
4851
4852static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4853{
4854 if (!cfs_bandwidth_used())
4855 return;
4856
4857 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4858 return;
4859
4860 __return_cfs_rq_runtime(cfs_rq);
4861}
4862
4863
4864
4865
4866
4867static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4868{
4869 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4870 unsigned long flags;
4871 u64 expires;
4872
4873
4874 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4875 cfs_b->slack_started = false;
4876 if (cfs_b->distribute_running) {
4877 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4878 return;
4879 }
4880
4881 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4882 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4883 return;
4884 }
4885
4886 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4887 runtime = cfs_b->runtime;
4888
4889 expires = cfs_b->runtime_expires;
4890 if (runtime)
4891 cfs_b->distribute_running = 1;
4892
4893 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4894
4895 if (!runtime)
4896 return;
4897
4898 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4899
4900 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4901 if (expires == cfs_b->runtime_expires)
4902 lsub_positive(&cfs_b->runtime, runtime);
4903 cfs_b->distribute_running = 0;
4904 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4905}
4906
4907
4908
4909
4910
4911
4912static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4913{
4914 if (!cfs_bandwidth_used())
4915 return;
4916
4917
4918 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4919 return;
4920
4921
4922 if (cfs_rq_throttled(cfs_rq))
4923 return;
4924
4925
4926 account_cfs_rq_runtime(cfs_rq, 0);
4927 if (cfs_rq->runtime_remaining <= 0)
4928 throttle_cfs_rq(cfs_rq);
4929}
4930
4931static void sync_throttle(struct task_group *tg, int cpu)
4932{
4933 struct cfs_rq *pcfs_rq, *cfs_rq;
4934
4935 if (!cfs_bandwidth_used())
4936 return;
4937
4938 if (!tg->parent)
4939 return;
4940
4941 cfs_rq = tg->cfs_rq[cpu];
4942 pcfs_rq = tg->parent->cfs_rq[cpu];
4943
4944 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4945 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4946}
4947
4948
4949static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4950{
4951 if (!cfs_bandwidth_used())
4952 return false;
4953
4954 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4955 return false;
4956
4957
4958
4959
4960
4961 if (cfs_rq_throttled(cfs_rq))
4962 return true;
4963
4964 throttle_cfs_rq(cfs_rq);
4965 return true;
4966}
4967
4968static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4969{
4970 struct cfs_bandwidth *cfs_b =
4971 container_of(timer, struct cfs_bandwidth, slack_timer);
4972
4973 do_sched_cfs_slack_timer(cfs_b);
4974
4975 return HRTIMER_NORESTART;
4976}
4977
4978extern const u64 max_cfs_quota_period;
4979
4980static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4981{
4982 struct cfs_bandwidth *cfs_b =
4983 container_of(timer, struct cfs_bandwidth, period_timer);
4984 unsigned long flags;
4985 int overrun;
4986 int idle = 0;
4987 int count = 0;
4988
4989 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4990 for (;;) {
4991 overrun = hrtimer_forward_now(timer, cfs_b->period);
4992 if (!overrun)
4993 break;
4994
4995 if (++count > 3) {
4996 u64 new, old = ktime_to_ns(cfs_b->period);
4997
4998 new = (old * 147) / 128;
4999 new = min(new, max_cfs_quota_period);
5000
5001 cfs_b->period = ns_to_ktime(new);
5002
5003
5004 cfs_b->quota *= new;
5005 cfs_b->quota = div64_u64(cfs_b->quota, old);
5006
5007 pr_warn_ratelimited(
5008 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
5009 smp_processor_id(),
5010 div_u64(new, NSEC_PER_USEC),
5011 div_u64(cfs_b->quota, NSEC_PER_USEC));
5012
5013
5014 count = 0;
5015 }
5016
5017 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5018 }
5019 if (idle)
5020 cfs_b->period_active = 0;
5021 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5022
5023 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5024}
5025
5026void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5027{
5028 raw_spin_lock_init(&cfs_b->lock);
5029 cfs_b->runtime = 0;
5030 cfs_b->quota = RUNTIME_INF;
5031 cfs_b->period = ns_to_ktime(default_cfs_period());
5032
5033 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5034 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5035 cfs_b->period_timer.function = sched_cfs_period_timer;
5036 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5037 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5038 cfs_b->distribute_running = 0;
5039 cfs_b->slack_started = false;
5040}
5041
5042static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5043{
5044 cfs_rq->runtime_enabled = 0;
5045 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5046}
5047
5048void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5049{
5050 u64 overrun;
5051
5052 lockdep_assert_held(&cfs_b->lock);
5053
5054 if (cfs_b->period_active)
5055 return;
5056
5057 cfs_b->period_active = 1;
5058 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5059 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
5060 cfs_b->expires_seq++;
5061 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5062}
5063
5064static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5065{
5066
5067 if (!cfs_b->throttled_cfs_rq.next)
5068 return;
5069
5070 hrtimer_cancel(&cfs_b->period_timer);
5071 hrtimer_cancel(&cfs_b->slack_timer);
5072}
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082static void __maybe_unused update_runtime_enabled(struct rq *rq)
5083{
5084 struct task_group *tg;
5085
5086 lockdep_assert_held(&rq->lock);
5087
5088 rcu_read_lock();
5089 list_for_each_entry_rcu(tg, &task_groups, list) {
5090 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5091 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5092
5093 raw_spin_lock(&cfs_b->lock);
5094 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5095 raw_spin_unlock(&cfs_b->lock);
5096 }
5097 rcu_read_unlock();
5098}
5099
5100
5101static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5102{
5103 struct task_group *tg;
5104
5105 lockdep_assert_held(&rq->lock);
5106
5107 rcu_read_lock();
5108 list_for_each_entry_rcu(tg, &task_groups, list) {
5109 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5110
5111 if (!cfs_rq->runtime_enabled)
5112 continue;
5113
5114
5115
5116
5117
5118 cfs_rq->runtime_remaining = 1;
5119
5120
5121
5122
5123 cfs_rq->runtime_enabled = 0;
5124
5125 if (cfs_rq_throttled(cfs_rq))
5126 unthrottle_cfs_rq(cfs_rq);
5127 }
5128 rcu_read_unlock();
5129}
5130
5131#else
5132
5133static inline bool cfs_bandwidth_used(void)
5134{
5135 return false;
5136}
5137
5138static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5139{
5140 return rq_clock_task(rq_of(cfs_rq));
5141}
5142
5143static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5144static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5145static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5146static inline void sync_throttle(struct task_group *tg, int cpu) {}
5147static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5148
5149static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5150{
5151 return 0;
5152}
5153
5154static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5155{
5156 return 0;
5157}
5158
5159static inline int throttled_lb_pair(struct task_group *tg,
5160 int src_cpu, int dest_cpu)
5161{
5162 return 0;
5163}
5164
5165void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5166
5167#ifdef CONFIG_FAIR_GROUP_SCHED
5168static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5169#endif
5170
5171static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5172{
5173 return NULL;
5174}
5175static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5176static inline void update_runtime_enabled(struct rq *rq) {}
5177static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5178
5179#endif
5180
5181
5182
5183
5184
5185#ifdef CONFIG_SCHED_HRTICK
5186static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5187{
5188 struct sched_entity *se = &p->se;
5189 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5190
5191 SCHED_WARN_ON(task_rq(p) != rq);
5192
5193 if (rq->cfs.h_nr_running > 1) {
5194 u64 slice = sched_slice(cfs_rq, se);
5195 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5196 s64 delta = slice - ran;
5197
5198 if (delta < 0) {
5199 if (rq->curr == p)
5200 resched_curr(rq);
5201 return;
5202 }
5203 hrtick_start(rq, delta);
5204 }
5205}
5206
5207
5208
5209
5210
5211
5212static void hrtick_update(struct rq *rq)
5213{
5214 struct task_struct *curr = rq->curr;
5215
5216 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5217 return;
5218
5219 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5220 hrtick_start_fair(rq, curr);
5221}
5222#else
5223static inline void
5224hrtick_start_fair(struct rq *rq, struct task_struct *p)
5225{
5226}
5227
5228static inline void hrtick_update(struct rq *rq)
5229{
5230}
5231#endif
5232
5233#ifdef CONFIG_SMP
5234static inline unsigned long cpu_util(int cpu);
5235
5236static inline bool cpu_overutilized(int cpu)
5237{
5238 return (capacity_of(cpu) * 1024) < (cpu_util(cpu) * capacity_margin);
5239}
5240
5241static inline void update_overutilized_status(struct rq *rq)
5242{
5243 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5244 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5245 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5246 }
5247}
5248#else
5249static inline void update_overutilized_status(struct rq *rq) { }
5250#endif
5251
5252
5253
5254
5255
5256
5257static void
5258enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5259{
5260 struct cfs_rq *cfs_rq;
5261 struct sched_entity *se = &p->se;
5262
5263
5264
5265
5266
5267
5268
5269 util_est_enqueue(&rq->cfs, p);
5270
5271
5272
5273
5274
5275
5276 if (p->in_iowait)
5277 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5278
5279 for_each_sched_entity(se) {
5280 if (se->on_rq)
5281 break;
5282 cfs_rq = cfs_rq_of(se);
5283 enqueue_entity(cfs_rq, se, flags);
5284
5285
5286
5287
5288
5289
5290
5291 if (cfs_rq_throttled(cfs_rq))
5292 break;
5293 cfs_rq->h_nr_running++;
5294
5295 flags = ENQUEUE_WAKEUP;
5296 }
5297
5298 for_each_sched_entity(se) {
5299 cfs_rq = cfs_rq_of(se);
5300 cfs_rq->h_nr_running++;
5301
5302 if (cfs_rq_throttled(cfs_rq))
5303 break;
5304
5305 update_load_avg(cfs_rq, se, UPDATE_TG);
5306 update_cfs_group(se);
5307 }
5308
5309 if (!se) {
5310 add_nr_running(rq, 1);
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325 if (flags & ENQUEUE_WAKEUP)
5326 update_overutilized_status(rq);
5327
5328 }
5329
5330 if (cfs_bandwidth_used()) {
5331
5332
5333
5334
5335
5336
5337 for_each_sched_entity(se) {
5338 cfs_rq = cfs_rq_of(se);
5339
5340 if (list_add_leaf_cfs_rq(cfs_rq))
5341 break;
5342 }
5343 }
5344
5345 assert_list_leaf_cfs_rq(rq);
5346
5347 hrtick_update(rq);
5348}
5349
5350static void set_next_buddy(struct sched_entity *se);
5351
5352
5353
5354
5355
5356
5357static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5358{
5359 struct cfs_rq *cfs_rq;
5360 struct sched_entity *se = &p->se;
5361 int task_sleep = flags & DEQUEUE_SLEEP;
5362
5363 for_each_sched_entity(se) {
5364 cfs_rq = cfs_rq_of(se);
5365 dequeue_entity(cfs_rq, se, flags);
5366
5367
5368
5369
5370
5371
5372
5373 if (cfs_rq_throttled(cfs_rq))
5374 break;
5375 cfs_rq->h_nr_running--;
5376
5377
5378 if (cfs_rq->load.weight) {
5379
5380 se = parent_entity(se);
5381
5382
5383
5384
5385 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5386 set_next_buddy(se);
5387 break;
5388 }
5389 flags |= DEQUEUE_SLEEP;
5390 }
5391
5392 for_each_sched_entity(se) {
5393 cfs_rq = cfs_rq_of(se);
5394 cfs_rq->h_nr_running--;
5395
5396 if (cfs_rq_throttled(cfs_rq))
5397 break;
5398
5399 update_load_avg(cfs_rq, se, UPDATE_TG);
5400 update_cfs_group(se);
5401 }
5402
5403 if (!se)
5404 sub_nr_running(rq, 1);
5405
5406 util_est_dequeue(&rq->cfs, p, task_sleep);
5407 hrtick_update(rq);
5408}
5409
5410#ifdef CONFIG_SMP
5411
5412
5413DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5414DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5415
5416#ifdef CONFIG_NO_HZ_COMMON
5417
5418static struct {
5419 cpumask_var_t idle_cpus_mask;
5420 atomic_t nr_cpus;
5421 int has_blocked;
5422 unsigned long next_balance;
5423 unsigned long next_blocked;
5424} nohz ____cacheline_aligned;
5425
5426#endif
5427
5428static unsigned long cpu_runnable_load(struct rq *rq)
5429{
5430 return cfs_rq_runnable_load_avg(&rq->cfs);
5431}
5432
5433static unsigned long capacity_of(int cpu)
5434{
5435 return cpu_rq(cpu)->cpu_capacity;
5436}
5437
5438static unsigned long cpu_avg_load_per_task(int cpu)
5439{
5440 struct rq *rq = cpu_rq(cpu);
5441 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5442 unsigned long load_avg = cpu_runnable_load(rq);
5443
5444 if (nr_running)
5445 return load_avg / nr_running;
5446
5447 return 0;
5448}
5449
5450static void record_wakee(struct task_struct *p)
5451{
5452
5453
5454
5455
5456 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5457 current->wakee_flips >>= 1;
5458 current->wakee_flip_decay_ts = jiffies;
5459 }
5460
5461 if (current->last_wakee != p) {
5462 current->last_wakee = p;
5463 current->wakee_flips++;
5464 }
5465}
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484static int wake_wide(struct task_struct *p)
5485{
5486 unsigned int master = current->wakee_flips;
5487 unsigned int slave = p->wakee_flips;
5488 int factor = this_cpu_read(sd_llc_size);
5489
5490 if (master < slave)
5491 swap(master, slave);
5492 if (slave < factor || master < slave * factor)
5493 return 0;
5494 return 1;
5495}
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509static int
5510wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5511{
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5525 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5526
5527 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5528 return this_cpu;
5529
5530 return nr_cpumask_bits;
5531}
5532
5533static int
5534wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5535 int this_cpu, int prev_cpu, int sync)
5536{
5537 s64 this_eff_load, prev_eff_load;
5538 unsigned long task_load;
5539
5540 this_eff_load = cpu_runnable_load(cpu_rq(this_cpu));
5541
5542 if (sync) {
5543 unsigned long current_load = task_h_load(current);
5544
5545 if (current_load > this_eff_load)
5546 return this_cpu;
5547
5548 this_eff_load -= current_load;
5549 }
5550
5551 task_load = task_h_load(p);
5552
5553 this_eff_load += task_load;
5554 if (sched_feat(WA_BIAS))
5555 this_eff_load *= 100;
5556 this_eff_load *= capacity_of(prev_cpu);
5557
5558 prev_eff_load = cpu_runnable_load(cpu_rq(prev_cpu));
5559 prev_eff_load -= task_load;
5560 if (sched_feat(WA_BIAS))
5561 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5562 prev_eff_load *= capacity_of(this_cpu);
5563
5564
5565
5566
5567
5568
5569
5570 if (sync)
5571 prev_eff_load += 1;
5572
5573 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5574}
5575
5576static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5577 int this_cpu, int prev_cpu, int sync)
5578{
5579 int target = nr_cpumask_bits;
5580
5581 if (sched_feat(WA_IDLE))
5582 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5583
5584 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5585 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5586
5587 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5588 if (target == nr_cpumask_bits)
5589 return prev_cpu;
5590
5591 schedstat_inc(sd->ttwu_move_affine);
5592 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5593 return target;
5594}
5595
5596static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5597
5598static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5599{
5600 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5601}
5602
5603
5604
5605
5606
5607
5608
5609static struct sched_group *
5610find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5611 int this_cpu, int sd_flag)
5612{
5613 struct sched_group *idlest = NULL, *group = sd->groups;
5614 struct sched_group *most_spare_sg = NULL;
5615 unsigned long min_runnable_load = ULONG_MAX;
5616 unsigned long this_runnable_load = ULONG_MAX;
5617 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5618 unsigned long most_spare = 0, this_spare = 0;
5619 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5620 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5621 (sd->imbalance_pct-100) / 100;
5622
5623 do {
5624 unsigned long load, avg_load, runnable_load;
5625 unsigned long spare_cap, max_spare_cap;
5626 int local_group;
5627 int i;
5628
5629
5630 if (!cpumask_intersects(sched_group_span(group),
5631 p->cpus_ptr))
5632 continue;
5633
5634 local_group = cpumask_test_cpu(this_cpu,
5635 sched_group_span(group));
5636
5637
5638
5639
5640
5641 avg_load = 0;
5642 runnable_load = 0;
5643 max_spare_cap = 0;
5644
5645 for_each_cpu(i, sched_group_span(group)) {
5646 load = cpu_runnable_load(cpu_rq(i));
5647 runnable_load += load;
5648
5649 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5650
5651 spare_cap = capacity_spare_without(i, p);
5652
5653 if (spare_cap > max_spare_cap)
5654 max_spare_cap = spare_cap;
5655 }
5656
5657
5658 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5659 group->sgc->capacity;
5660 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5661 group->sgc->capacity;
5662
5663 if (local_group) {
5664 this_runnable_load = runnable_load;
5665 this_avg_load = avg_load;
5666 this_spare = max_spare_cap;
5667 } else {
5668 if (min_runnable_load > (runnable_load + imbalance)) {
5669
5670
5671
5672
5673 min_runnable_load = runnable_load;
5674 min_avg_load = avg_load;
5675 idlest = group;
5676 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5677 (100*min_avg_load > imbalance_scale*avg_load)) {
5678
5679
5680
5681
5682 min_avg_load = avg_load;
5683 idlest = group;
5684 }
5685
5686 if (most_spare < max_spare_cap) {
5687 most_spare = max_spare_cap;
5688 most_spare_sg = group;
5689 }
5690 }
5691 } while (group = group->next, group != sd->groups);
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704 if (sd_flag & SD_BALANCE_FORK)
5705 goto skip_spare;
5706
5707 if (this_spare > task_util(p) / 2 &&
5708 imbalance_scale*this_spare > 100*most_spare)
5709 return NULL;
5710
5711 if (most_spare > task_util(p) / 2)
5712 return most_spare_sg;
5713
5714skip_spare:
5715 if (!idlest)
5716 return NULL;
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726 if ((sd->flags & SD_NUMA) &&
5727 min_runnable_load + imbalance >= this_runnable_load)
5728 return NULL;
5729
5730 if (min_runnable_load > (this_runnable_load + imbalance))
5731 return NULL;
5732
5733 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5734 (100*this_avg_load < imbalance_scale*min_avg_load))
5735 return NULL;
5736
5737 return idlest;
5738}
5739
5740
5741
5742
5743static int
5744find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5745{
5746 unsigned long load, min_load = ULONG_MAX;
5747 unsigned int min_exit_latency = UINT_MAX;
5748 u64 latest_idle_timestamp = 0;
5749 int least_loaded_cpu = this_cpu;
5750 int shallowest_idle_cpu = -1;
5751 int i;
5752
5753
5754 if (group->group_weight == 1)
5755 return cpumask_first(sched_group_span(group));
5756
5757
5758 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5759 if (available_idle_cpu(i)) {
5760 struct rq *rq = cpu_rq(i);
5761 struct cpuidle_state *idle = idle_get_state(rq);
5762 if (idle && idle->exit_latency < min_exit_latency) {
5763
5764
5765
5766
5767
5768 min_exit_latency = idle->exit_latency;
5769 latest_idle_timestamp = rq->idle_stamp;
5770 shallowest_idle_cpu = i;
5771 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5772 rq->idle_stamp > latest_idle_timestamp) {
5773
5774
5775
5776
5777
5778 latest_idle_timestamp = rq->idle_stamp;
5779 shallowest_idle_cpu = i;
5780 }
5781 } else if (shallowest_idle_cpu == -1) {
5782 load = cpu_runnable_load(cpu_rq(i));
5783 if (load < min_load) {
5784 min_load = load;
5785 least_loaded_cpu = i;
5786 }
5787 }
5788 }
5789
5790 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5791}
5792
5793static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5794 int cpu, int prev_cpu, int sd_flag)
5795{
5796 int new_cpu = cpu;
5797
5798 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5799 return prev_cpu;
5800
5801
5802
5803
5804
5805 if (!(sd_flag & SD_BALANCE_FORK))
5806 sync_entity_load_avg(&p->se);
5807
5808 while (sd) {
5809 struct sched_group *group;
5810 struct sched_domain *tmp;
5811 int weight;
5812
5813 if (!(sd->flags & sd_flag)) {
5814 sd = sd->child;
5815 continue;
5816 }
5817
5818 group = find_idlest_group(sd, p, cpu, sd_flag);
5819 if (!group) {
5820 sd = sd->child;
5821 continue;
5822 }
5823
5824 new_cpu = find_idlest_group_cpu(group, p, cpu);
5825 if (new_cpu == cpu) {
5826
5827 sd = sd->child;
5828 continue;
5829 }
5830
5831
5832 cpu = new_cpu;
5833 weight = sd->span_weight;
5834 sd = NULL;
5835 for_each_domain(cpu, tmp) {
5836 if (weight <= tmp->span_weight)
5837 break;
5838 if (tmp->flags & sd_flag)
5839 sd = tmp;
5840 }
5841 }
5842
5843 return new_cpu;
5844}
5845
5846#ifdef CONFIG_SCHED_SMT
5847DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5848EXPORT_SYMBOL_GPL(sched_smt_present);
5849
5850static inline void set_idle_cores(int cpu, int val)
5851{
5852 struct sched_domain_shared *sds;
5853
5854 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5855 if (sds)
5856 WRITE_ONCE(sds->has_idle_cores, val);
5857}
5858
5859static inline bool test_idle_cores(int cpu, bool def)
5860{
5861 struct sched_domain_shared *sds;
5862
5863 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5864 if (sds)
5865 return READ_ONCE(sds->has_idle_cores);
5866
5867 return def;
5868}
5869
5870
5871
5872
5873
5874
5875
5876
5877void __update_idle_core(struct rq *rq)
5878{
5879 int core = cpu_of(rq);
5880 int cpu;
5881
5882 rcu_read_lock();
5883 if (test_idle_cores(core, true))
5884 goto unlock;
5885
5886 for_each_cpu(cpu, cpu_smt_mask(core)) {
5887 if (cpu == core)
5888 continue;
5889
5890 if (!available_idle_cpu(cpu))
5891 goto unlock;
5892 }
5893
5894 set_idle_cores(core, 1);
5895unlock:
5896 rcu_read_unlock();
5897}
5898
5899
5900
5901
5902
5903
5904static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5905{
5906 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5907 int core, cpu;
5908
5909 if (!static_branch_likely(&sched_smt_present))
5910 return -1;
5911
5912 if (!test_idle_cores(target, false))
5913 return -1;
5914
5915 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
5916
5917 for_each_cpu_wrap(core, cpus, target) {
5918 bool idle = true;
5919
5920 for_each_cpu(cpu, cpu_smt_mask(core)) {
5921 __cpumask_clear_cpu(cpu, cpus);
5922 if (!available_idle_cpu(cpu))
5923 idle = false;
5924 }
5925
5926 if (idle)
5927 return core;
5928 }
5929
5930
5931
5932
5933 set_idle_cores(target, 0);
5934
5935 return -1;
5936}
5937
5938
5939
5940
5941static int select_idle_smt(struct task_struct *p, int target)
5942{
5943 int cpu;
5944
5945 if (!static_branch_likely(&sched_smt_present))
5946 return -1;
5947
5948 for_each_cpu(cpu, cpu_smt_mask(target)) {
5949 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
5950 continue;
5951 if (available_idle_cpu(cpu))
5952 return cpu;
5953 }
5954
5955 return -1;
5956}
5957
5958#else
5959
5960static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5961{
5962 return -1;
5963}
5964
5965static inline int select_idle_smt(struct task_struct *p, int target)
5966{
5967 return -1;
5968}
5969
5970#endif
5971
5972
5973
5974
5975
5976
5977static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5978{
5979 struct sched_domain *this_sd;
5980 u64 avg_cost, avg_idle;
5981 u64 time, cost;
5982 s64 delta;
5983 int cpu, nr = INT_MAX;
5984 int this = smp_processor_id();
5985
5986 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5987 if (!this_sd)
5988 return -1;
5989
5990
5991
5992
5993
5994 avg_idle = this_rq()->avg_idle / 512;
5995 avg_cost = this_sd->avg_scan_cost + 1;
5996
5997 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
5998 return -1;
5999
6000 if (sched_feat(SIS_PROP)) {
6001 u64 span_avg = sd->span_weight * avg_idle;
6002 if (span_avg > 4*avg_cost)
6003 nr = div_u64(span_avg, avg_cost);
6004 else
6005 nr = 4;
6006 }
6007
6008 time = cpu_clock(this);
6009
6010 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6011 if (!--nr)
6012 return -1;
6013 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6014 continue;
6015 if (available_idle_cpu(cpu))
6016 break;
6017 }
6018
6019 time = cpu_clock(this) - time;
6020 cost = this_sd->avg_scan_cost;
6021 delta = (s64)(time - cost) / 8;
6022 this_sd->avg_scan_cost += delta;
6023
6024 return cpu;
6025}
6026
6027
6028
6029
6030static int select_idle_sibling(struct task_struct *p, int prev, int target)
6031{
6032 struct sched_domain *sd;
6033 int i, recent_used_cpu;
6034
6035 if (available_idle_cpu(target))
6036 return target;
6037
6038
6039
6040
6041 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6042 return prev;
6043
6044
6045 recent_used_cpu = p->recent_used_cpu;
6046 if (recent_used_cpu != prev &&
6047 recent_used_cpu != target &&
6048 cpus_share_cache(recent_used_cpu, target) &&
6049 available_idle_cpu(recent_used_cpu) &&
6050 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6051
6052
6053
6054
6055 p->recent_used_cpu = prev;
6056 return recent_used_cpu;
6057 }
6058
6059 sd = rcu_dereference(per_cpu(sd_llc, target));
6060 if (!sd)
6061 return target;
6062
6063 i = select_idle_core(p, sd, target);
6064 if ((unsigned)i < nr_cpumask_bits)
6065 return i;
6066
6067 i = select_idle_cpu(p, sd, target);
6068 if ((unsigned)i < nr_cpumask_bits)
6069 return i;
6070
6071 i = select_idle_smt(p, target);
6072 if ((unsigned)i < nr_cpumask_bits)
6073 return i;
6074
6075 return target;
6076}
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116static inline unsigned long cpu_util(int cpu)
6117{
6118 struct cfs_rq *cfs_rq;
6119 unsigned int util;
6120
6121 cfs_rq = &cpu_rq(cpu)->cfs;
6122 util = READ_ONCE(cfs_rq->avg.util_avg);
6123
6124 if (sched_feat(UTIL_EST))
6125 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6126
6127 return min_t(unsigned long, util, capacity_orig_of(cpu));
6128}
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6144{
6145 struct cfs_rq *cfs_rq;
6146 unsigned int util;
6147
6148
6149 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6150 return cpu_util(cpu);
6151
6152 cfs_rq = &cpu_rq(cpu)->cfs;
6153 util = READ_ONCE(cfs_rq->avg.util_avg);
6154
6155
6156 lsub_positive(&util, task_util(p));
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184 if (sched_feat(UTIL_EST)) {
6185 unsigned int estimated =
6186 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205 if (unlikely(task_on_rq_queued(p) || current == p))
6206 lsub_positive(&estimated, _task_util_est(p));
6207
6208 util = max(util, estimated);
6209 }
6210
6211
6212
6213
6214
6215
6216 return min_t(unsigned long, util, capacity_orig_of(cpu));
6217}
6218
6219
6220
6221
6222
6223
6224
6225
6226static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6227{
6228 long min_cap, max_cap;
6229
6230 if (!static_branch_unlikely(&sched_asym_cpucapacity))
6231 return 0;
6232
6233 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6234 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6235
6236
6237 if (max_cap - min_cap < max_cap >> 3)
6238 return 0;
6239
6240
6241 sync_entity_load_avg(&p->se);
6242
6243 return !task_fits_capacity(p, min_cap);
6244}
6245
6246
6247
6248
6249
6250static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6251{
6252 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6253 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6254
6255
6256
6257
6258
6259
6260
6261 if (task_cpu(p) == cpu && dst_cpu != cpu)
6262 sub_positive(&util, task_util(p));
6263 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6264 util += task_util(p);
6265
6266 if (sched_feat(UTIL_EST)) {
6267 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6268
6269
6270
6271
6272
6273
6274
6275 if (dst_cpu == cpu)
6276 util_est += _task_util_est(p);
6277
6278 util = max(util, util_est);
6279 }
6280
6281 return min(util, capacity_orig_of(cpu));
6282}
6283
6284
6285
6286
6287
6288
6289
6290
6291static long
6292compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6293{
6294 unsigned int max_util, util_cfs, cpu_util, cpu_cap;
6295 unsigned long sum_util, energy = 0;
6296 struct task_struct *tsk;
6297 int cpu;
6298
6299 for (; pd; pd = pd->next) {
6300 struct cpumask *pd_mask = perf_domain_span(pd);
6301
6302
6303
6304
6305
6306 cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6307 max_util = sum_util = 0;
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6320 util_cfs = cpu_util_next(cpu, p, dst_cpu);
6321
6322
6323
6324
6325
6326
6327
6328 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6329 ENERGY_UTIL, NULL);
6330
6331
6332
6333
6334
6335
6336
6337
6338 tsk = cpu == dst_cpu ? p : NULL;
6339 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6340 FREQUENCY_UTIL, tsk);
6341 max_util = max(max_util, cpu_util);
6342 }
6343
6344 energy += em_pd_energy(pd->em_pd, max_util, sum_util);
6345 }
6346
6347 return energy;
6348}
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6391{
6392 unsigned long prev_energy = ULONG_MAX, best_energy = ULONG_MAX;
6393 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6394 int cpu, best_energy_cpu = prev_cpu;
6395 struct perf_domain *head, *pd;
6396 unsigned long cpu_cap, util;
6397 struct sched_domain *sd;
6398
6399 rcu_read_lock();
6400 pd = rcu_dereference(rd->pd);
6401 if (!pd || READ_ONCE(rd->overutilized))
6402 goto fail;
6403 head = pd;
6404
6405
6406
6407
6408
6409 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6410 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6411 sd = sd->parent;
6412 if (!sd)
6413 goto fail;
6414
6415 sync_entity_load_avg(&p->se);
6416 if (!task_util_est(p))
6417 goto unlock;
6418
6419 for (; pd; pd = pd->next) {
6420 unsigned long cur_energy, spare_cap, max_spare_cap = 0;
6421 int max_spare_cap_cpu = -1;
6422
6423 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6424 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6425 continue;
6426
6427
6428 util = cpu_util_next(cpu, p, cpu);
6429 cpu_cap = capacity_of(cpu);
6430 if (cpu_cap * 1024 < util * capacity_margin)
6431 continue;
6432
6433
6434 if (cpu == prev_cpu) {
6435 prev_energy = compute_energy(p, prev_cpu, head);
6436 best_energy = min(best_energy, prev_energy);
6437 continue;
6438 }
6439
6440
6441
6442
6443
6444 spare_cap = cpu_cap - util;
6445 if (spare_cap > max_spare_cap) {
6446 max_spare_cap = spare_cap;
6447 max_spare_cap_cpu = cpu;
6448 }
6449 }
6450
6451
6452 if (max_spare_cap_cpu >= 0) {
6453 cur_energy = compute_energy(p, max_spare_cap_cpu, head);
6454 if (cur_energy < best_energy) {
6455 best_energy = cur_energy;
6456 best_energy_cpu = max_spare_cap_cpu;
6457 }
6458 }
6459 }
6460unlock:
6461 rcu_read_unlock();
6462
6463
6464
6465
6466
6467 if (prev_energy == ULONG_MAX)
6468 return best_energy_cpu;
6469
6470 if ((prev_energy - best_energy) > (prev_energy >> 4))
6471 return best_energy_cpu;
6472
6473 return prev_cpu;
6474
6475fail:
6476 rcu_read_unlock();
6477
6478 return -1;
6479}
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493static int
6494select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6495{
6496 struct sched_domain *tmp, *sd = NULL;
6497 int cpu = smp_processor_id();
6498 int new_cpu = prev_cpu;
6499 int want_affine = 0;
6500 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6501
6502 if (sd_flag & SD_BALANCE_WAKE) {
6503 record_wakee(p);
6504
6505 if (sched_energy_enabled()) {
6506 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6507 if (new_cpu >= 0)
6508 return new_cpu;
6509 new_cpu = prev_cpu;
6510 }
6511
6512 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu) &&
6513 cpumask_test_cpu(cpu, p->cpus_ptr);
6514 }
6515
6516 rcu_read_lock();
6517 for_each_domain(cpu, tmp) {
6518 if (!(tmp->flags & SD_LOAD_BALANCE))
6519 break;
6520
6521
6522
6523
6524
6525 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6526 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6527 if (cpu != prev_cpu)
6528 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6529
6530 sd = NULL;
6531 break;
6532 }
6533
6534 if (tmp->flags & sd_flag)
6535 sd = tmp;
6536 else if (!want_affine)
6537 break;
6538 }
6539
6540 if (unlikely(sd)) {
6541
6542 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6543 } else if (sd_flag & SD_BALANCE_WAKE) {
6544
6545
6546 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6547
6548 if (want_affine)
6549 current->recent_used_cpu = cpu;
6550 }
6551 rcu_read_unlock();
6552
6553 return new_cpu;
6554}
6555
6556static void detach_entity_cfs_rq(struct sched_entity *se);
6557
6558
6559
6560
6561
6562
6563static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6564{
6565
6566
6567
6568
6569
6570
6571 if (p->state == TASK_WAKING) {
6572 struct sched_entity *se = &p->se;
6573 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6574 u64 min_vruntime;
6575
6576#ifndef CONFIG_64BIT
6577 u64 min_vruntime_copy;
6578
6579 do {
6580 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6581 smp_rmb();
6582 min_vruntime = cfs_rq->min_vruntime;
6583 } while (min_vruntime != min_vruntime_copy);
6584#else
6585 min_vruntime = cfs_rq->min_vruntime;
6586#endif
6587
6588 se->vruntime -= min_vruntime;
6589 }
6590
6591 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6592
6593
6594
6595
6596 lockdep_assert_held(&task_rq(p)->lock);
6597 detach_entity_cfs_rq(&p->se);
6598
6599 } else {
6600
6601
6602
6603
6604
6605
6606
6607
6608 remove_entity_load_avg(&p->se);
6609 }
6610
6611
6612 p->se.avg.last_update_time = 0;
6613
6614
6615 p->se.exec_start = 0;
6616
6617 update_scan_period(p, new_cpu);
6618}
6619
6620static void task_dead_fair(struct task_struct *p)
6621{
6622 remove_entity_load_avg(&p->se);
6623}
6624#endif
6625
6626static unsigned long wakeup_gran(struct sched_entity *se)
6627{
6628 unsigned long gran = sysctl_sched_wakeup_granularity;
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643 return calc_delta_fair(gran, se);
6644}
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660static int
6661wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6662{
6663 s64 gran, vdiff = curr->vruntime - se->vruntime;
6664
6665 if (vdiff <= 0)
6666 return -1;
6667
6668 gran = wakeup_gran(se);
6669 if (vdiff > gran)
6670 return 1;
6671
6672 return 0;
6673}
6674
6675static void set_last_buddy(struct sched_entity *se)
6676{
6677 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6678 return;
6679
6680 for_each_sched_entity(se) {
6681 if (SCHED_WARN_ON(!se->on_rq))
6682 return;
6683 cfs_rq_of(se)->last = se;
6684 }
6685}
6686
6687static void set_next_buddy(struct sched_entity *se)
6688{
6689 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6690 return;
6691
6692 for_each_sched_entity(se) {
6693 if (SCHED_WARN_ON(!se->on_rq))
6694 return;
6695 cfs_rq_of(se)->next = se;
6696 }
6697}
6698
6699static void set_skip_buddy(struct sched_entity *se)
6700{
6701 for_each_sched_entity(se)
6702 cfs_rq_of(se)->skip = se;
6703}
6704
6705
6706
6707
6708static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6709{
6710 struct task_struct *curr = rq->curr;
6711 struct sched_entity *se = &curr->se, *pse = &p->se;
6712 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6713 int scale = cfs_rq->nr_running >= sched_nr_latency;
6714 int next_buddy_marked = 0;
6715
6716 if (unlikely(se == pse))
6717 return;
6718
6719
6720
6721
6722
6723
6724
6725 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6726 return;
6727
6728 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6729 set_next_buddy(pse);
6730 next_buddy_marked = 1;
6731 }
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743 if (test_tsk_need_resched(curr))
6744 return;
6745
6746
6747 if (unlikely(task_has_idle_policy(curr)) &&
6748 likely(!task_has_idle_policy(p)))
6749 goto preempt;
6750
6751
6752
6753
6754
6755 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6756 return;
6757
6758 find_matching_se(&se, &pse);
6759 update_curr(cfs_rq_of(se));
6760 BUG_ON(!pse);
6761 if (wakeup_preempt_entity(se, pse) == 1) {
6762
6763
6764
6765
6766 if (!next_buddy_marked)
6767 set_next_buddy(pse);
6768 goto preempt;
6769 }
6770
6771 return;
6772
6773preempt:
6774 resched_curr(rq);
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784 if (unlikely(!se->on_rq || curr == rq->idle))
6785 return;
6786
6787 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6788 set_last_buddy(se);
6789}
6790
6791static struct task_struct *
6792pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6793{
6794 struct cfs_rq *cfs_rq = &rq->cfs;
6795 struct sched_entity *se;
6796 struct task_struct *p;
6797 int new_tasks;
6798
6799again:
6800 if (!cfs_rq->nr_running)
6801 goto idle;
6802
6803#ifdef CONFIG_FAIR_GROUP_SCHED
6804 if (prev->sched_class != &fair_sched_class)
6805 goto simple;
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815 do {
6816 struct sched_entity *curr = cfs_rq->curr;
6817
6818
6819
6820
6821
6822
6823
6824 if (curr) {
6825 if (curr->on_rq)
6826 update_curr(cfs_rq);
6827 else
6828 curr = NULL;
6829
6830
6831
6832
6833
6834
6835
6836 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6837 cfs_rq = &rq->cfs;
6838
6839 if (!cfs_rq->nr_running)
6840 goto idle;
6841
6842 goto simple;
6843 }
6844 }
6845
6846 se = pick_next_entity(cfs_rq, curr);
6847 cfs_rq = group_cfs_rq(se);
6848 } while (cfs_rq);
6849
6850 p = task_of(se);
6851
6852
6853
6854
6855
6856
6857 if (prev != p) {
6858 struct sched_entity *pse = &prev->se;
6859
6860 while (!(cfs_rq = is_same_group(se, pse))) {
6861 int se_depth = se->depth;
6862 int pse_depth = pse->depth;
6863
6864 if (se_depth <= pse_depth) {
6865 put_prev_entity(cfs_rq_of(pse), pse);
6866 pse = parent_entity(pse);
6867 }
6868 if (se_depth >= pse_depth) {
6869 set_next_entity(cfs_rq_of(se), se);
6870 se = parent_entity(se);
6871 }
6872 }
6873
6874 put_prev_entity(cfs_rq, pse);
6875 set_next_entity(cfs_rq, se);
6876 }
6877
6878 goto done;
6879simple:
6880#endif
6881
6882 put_prev_task(rq, prev);
6883
6884 do {
6885 se = pick_next_entity(cfs_rq, NULL);
6886 set_next_entity(cfs_rq, se);
6887 cfs_rq = group_cfs_rq(se);
6888 } while (cfs_rq);
6889
6890 p = task_of(se);
6891
6892done: __maybe_unused;
6893#ifdef CONFIG_SMP
6894
6895
6896
6897
6898
6899 list_move(&p->se.group_node, &rq->cfs_tasks);
6900#endif
6901
6902 if (hrtick_enabled(rq))
6903 hrtick_start_fair(rq, p);
6904
6905 update_misfit_status(p, rq);
6906
6907 return p;
6908
6909idle:
6910 update_misfit_status(NULL, rq);
6911 new_tasks = idle_balance(rq, rf);
6912
6913
6914
6915
6916
6917
6918 if (new_tasks < 0)
6919 return RETRY_TASK;
6920
6921 if (new_tasks > 0)
6922 goto again;
6923
6924
6925
6926
6927
6928 update_idle_rq_clock_pelt(rq);
6929
6930 return NULL;
6931}
6932
6933
6934
6935
6936static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6937{
6938 struct sched_entity *se = &prev->se;
6939 struct cfs_rq *cfs_rq;
6940
6941 for_each_sched_entity(se) {
6942 cfs_rq = cfs_rq_of(se);
6943 put_prev_entity(cfs_rq, se);
6944 }
6945}
6946
6947
6948
6949
6950
6951
6952static void yield_task_fair(struct rq *rq)
6953{
6954 struct task_struct *curr = rq->curr;
6955 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6956 struct sched_entity *se = &curr->se;
6957
6958
6959
6960
6961 if (unlikely(rq->nr_running == 1))
6962 return;
6963
6964 clear_buddies(cfs_rq, se);
6965
6966 if (curr->policy != SCHED_BATCH) {
6967 update_rq_clock(rq);
6968
6969
6970
6971 update_curr(cfs_rq);
6972
6973
6974
6975
6976
6977 rq_clock_skip_update(rq);
6978 }
6979
6980 set_skip_buddy(se);
6981}
6982
6983static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6984{
6985 struct sched_entity *se = &p->se;
6986
6987
6988 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6989 return false;
6990
6991
6992 set_next_buddy(se);
6993
6994 yield_task_fair(rq);
6995
6996 return true;
6997}
6998
6999#ifdef CONFIG_SMP
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7119
7120enum fbq_type { regular, remote, all };
7121
7122enum group_type {
7123 group_other = 0,
7124 group_misfit_task,
7125 group_imbalanced,
7126 group_overloaded,
7127};
7128
7129#define LBF_ALL_PINNED 0x01
7130#define LBF_NEED_BREAK 0x02
7131#define LBF_DST_PINNED 0x04
7132#define LBF_SOME_PINNED 0x08
7133#define LBF_NOHZ_STATS 0x10
7134#define LBF_NOHZ_AGAIN 0x20
7135
7136struct lb_env {
7137 struct sched_domain *sd;
7138
7139 struct rq *src_rq;
7140 int src_cpu;
7141
7142 int dst_cpu;
7143 struct rq *dst_rq;
7144
7145 struct cpumask *dst_grpmask;
7146 int new_dst_cpu;
7147 enum cpu_idle_type idle;
7148 long imbalance;
7149
7150 struct cpumask *cpus;
7151
7152 unsigned int flags;
7153
7154 unsigned int loop;
7155 unsigned int loop_break;
7156 unsigned int loop_max;
7157
7158 enum fbq_type fbq_type;
7159 enum group_type src_grp_type;
7160 struct list_head tasks;
7161};
7162
7163
7164
7165
7166static int task_hot(struct task_struct *p, struct lb_env *env)
7167{
7168 s64 delta;
7169
7170 lockdep_assert_held(&env->src_rq->lock);
7171
7172 if (p->sched_class != &fair_sched_class)
7173 return 0;
7174
7175 if (unlikely(task_has_idle_policy(p)))
7176 return 0;
7177
7178
7179
7180
7181 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7182 (&p->se == cfs_rq_of(&p->se)->next ||
7183 &p->se == cfs_rq_of(&p->se)->last))
7184 return 1;
7185
7186 if (sysctl_sched_migration_cost == -1)
7187 return 1;
7188 if (sysctl_sched_migration_cost == 0)
7189 return 0;
7190
7191 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7192
7193 return delta < (s64)sysctl_sched_migration_cost;
7194}
7195
7196#ifdef CONFIG_NUMA_BALANCING
7197
7198
7199
7200
7201
7202static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7203{
7204 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7205 unsigned long src_weight, dst_weight;
7206 int src_nid, dst_nid, dist;
7207
7208 if (!static_branch_likely(&sched_numa_balancing))
7209 return -1;
7210
7211 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7212 return -1;
7213
7214 src_nid = cpu_to_node(env->src_cpu);
7215 dst_nid = cpu_to_node(env->dst_cpu);
7216
7217 if (src_nid == dst_nid)
7218 return -1;
7219
7220
7221 if (src_nid == p->numa_preferred_nid) {
7222 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7223 return 1;
7224 else
7225 return -1;
7226 }
7227
7228
7229 if (dst_nid == p->numa_preferred_nid)
7230 return 0;
7231
7232
7233 if (env->idle == CPU_IDLE)
7234 return -1;
7235
7236 dist = node_distance(src_nid, dst_nid);
7237 if (numa_group) {
7238 src_weight = group_weight(p, src_nid, dist);
7239 dst_weight = group_weight(p, dst_nid, dist);
7240 } else {
7241 src_weight = task_weight(p, src_nid, dist);
7242 dst_weight = task_weight(p, dst_nid, dist);
7243 }
7244
7245 return dst_weight < src_weight;
7246}
7247
7248#else
7249static inline int migrate_degrades_locality(struct task_struct *p,
7250 struct lb_env *env)
7251{
7252 return -1;
7253}
7254#endif
7255
7256
7257
7258
7259static
7260int can_migrate_task(struct task_struct *p, struct lb_env *env)
7261{
7262 int tsk_cache_hot;
7263
7264 lockdep_assert_held(&env->src_rq->lock);
7265
7266
7267
7268
7269
7270
7271
7272
7273 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7274 return 0;
7275
7276 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7277 int cpu;
7278
7279 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7280
7281 env->flags |= LBF_SOME_PINNED;
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7292 return 0;
7293
7294
7295 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7296 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7297 env->flags |= LBF_DST_PINNED;
7298 env->new_dst_cpu = cpu;
7299 break;
7300 }
7301 }
7302
7303 return 0;
7304 }
7305
7306
7307 env->flags &= ~LBF_ALL_PINNED;
7308
7309 if (task_running(env->src_rq, p)) {
7310 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7311 return 0;
7312 }
7313
7314
7315
7316
7317
7318
7319
7320 tsk_cache_hot = migrate_degrades_locality(p, env);
7321 if (tsk_cache_hot == -1)
7322 tsk_cache_hot = task_hot(p, env);
7323
7324 if (tsk_cache_hot <= 0 ||
7325 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7326 if (tsk_cache_hot == 1) {
7327 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7328 schedstat_inc(p->se.statistics.nr_forced_migrations);
7329 }
7330 return 1;
7331 }
7332
7333 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7334 return 0;
7335}
7336
7337
7338
7339
7340static void detach_task(struct task_struct *p, struct lb_env *env)
7341{
7342 lockdep_assert_held(&env->src_rq->lock);
7343
7344 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7345 set_task_cpu(p, env->dst_cpu);
7346}
7347
7348
7349
7350
7351
7352
7353
7354static struct task_struct *detach_one_task(struct lb_env *env)
7355{
7356 struct task_struct *p;
7357
7358 lockdep_assert_held(&env->src_rq->lock);
7359
7360 list_for_each_entry_reverse(p,
7361 &env->src_rq->cfs_tasks, se.group_node) {
7362 if (!can_migrate_task(p, env))
7363 continue;
7364
7365 detach_task(p, env);
7366
7367
7368
7369
7370
7371
7372
7373 schedstat_inc(env->sd->lb_gained[env->idle]);
7374 return p;
7375 }
7376 return NULL;
7377}
7378
7379static const unsigned int sched_nr_migrate_break = 32;
7380
7381
7382
7383
7384
7385
7386
7387static int detach_tasks(struct lb_env *env)
7388{
7389 struct list_head *tasks = &env->src_rq->cfs_tasks;
7390 struct task_struct *p;
7391 unsigned long load;
7392 int detached = 0;
7393
7394 lockdep_assert_held(&env->src_rq->lock);
7395
7396 if (env->imbalance <= 0)
7397 return 0;
7398
7399 while (!list_empty(tasks)) {
7400
7401
7402
7403
7404 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7405 break;
7406
7407 p = list_last_entry(tasks, struct task_struct, se.group_node);
7408
7409 env->loop++;
7410
7411 if (env->loop > env->loop_max)
7412 break;
7413
7414
7415 if (env->loop > env->loop_break) {
7416 env->loop_break += sched_nr_migrate_break;
7417 env->flags |= LBF_NEED_BREAK;
7418 break;
7419 }
7420
7421 if (!can_migrate_task(p, env))
7422 goto next;
7423
7424 load = task_h_load(p);
7425
7426 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7427 goto next;
7428
7429 if ((load / 2) > env->imbalance)
7430 goto next;
7431
7432 detach_task(p, env);
7433 list_add(&p->se.group_node, &env->tasks);
7434
7435 detached++;
7436 env->imbalance -= load;
7437
7438#ifdef CONFIG_PREEMPT
7439
7440
7441
7442
7443
7444 if (env->idle == CPU_NEWLY_IDLE)
7445 break;
7446#endif
7447
7448
7449
7450
7451
7452 if (env->imbalance <= 0)
7453 break;
7454
7455 continue;
7456next:
7457 list_move(&p->se.group_node, tasks);
7458 }
7459
7460
7461
7462
7463
7464
7465 schedstat_add(env->sd->lb_gained[env->idle], detached);
7466
7467 return detached;
7468}
7469
7470
7471
7472
7473static void attach_task(struct rq *rq, struct task_struct *p)
7474{
7475 lockdep_assert_held(&rq->lock);
7476
7477 BUG_ON(task_rq(p) != rq);
7478 activate_task(rq, p, ENQUEUE_NOCLOCK);
7479 check_preempt_curr(rq, p, 0);
7480}
7481
7482
7483
7484
7485
7486static void attach_one_task(struct rq *rq, struct task_struct *p)
7487{
7488 struct rq_flags rf;
7489
7490 rq_lock(rq, &rf);
7491 update_rq_clock(rq);
7492 attach_task(rq, p);
7493 rq_unlock(rq, &rf);
7494}
7495
7496
7497
7498
7499
7500static void attach_tasks(struct lb_env *env)
7501{
7502 struct list_head *tasks = &env->tasks;
7503 struct task_struct *p;
7504 struct rq_flags rf;
7505
7506 rq_lock(env->dst_rq, &rf);
7507 update_rq_clock(env->dst_rq);
7508
7509 while (!list_empty(tasks)) {
7510 p = list_first_entry(tasks, struct task_struct, se.group_node);
7511 list_del_init(&p->se.group_node);
7512
7513 attach_task(env->dst_rq, p);
7514 }
7515
7516 rq_unlock(env->dst_rq, &rf);
7517}
7518
7519#ifdef CONFIG_NO_HZ_COMMON
7520static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7521{
7522 if (cfs_rq->avg.load_avg)
7523 return true;
7524
7525 if (cfs_rq->avg.util_avg)
7526 return true;
7527
7528 return false;
7529}
7530
7531static inline bool others_have_blocked(struct rq *rq)
7532{
7533 if (READ_ONCE(rq->avg_rt.util_avg))
7534 return true;
7535
7536 if (READ_ONCE(rq->avg_dl.util_avg))
7537 return true;
7538
7539#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7540 if (READ_ONCE(rq->avg_irq.util_avg))
7541 return true;
7542#endif
7543
7544 return false;
7545}
7546
7547static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7548{
7549 rq->last_blocked_load_update_tick = jiffies;
7550
7551 if (!has_blocked)
7552 rq->has_blocked_load = 0;
7553}
7554#else
7555static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7556static inline bool others_have_blocked(struct rq *rq) { return false; }
7557static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7558#endif
7559
7560#ifdef CONFIG_FAIR_GROUP_SCHED
7561
7562static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7563{
7564 if (cfs_rq->load.weight)
7565 return false;
7566
7567 if (cfs_rq->avg.load_sum)
7568 return false;
7569
7570 if (cfs_rq->avg.util_sum)
7571 return false;
7572
7573 if (cfs_rq->avg.runnable_load_sum)
7574 return false;
7575
7576 return true;
7577}
7578
7579static void update_blocked_averages(int cpu)
7580{
7581 struct rq *rq = cpu_rq(cpu);
7582 struct cfs_rq *cfs_rq, *pos;
7583 const struct sched_class *curr_class;
7584 struct rq_flags rf;
7585 bool done = true;
7586
7587 rq_lock_irqsave(rq, &rf);
7588 update_rq_clock(rq);
7589
7590
7591
7592
7593
7594 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7595 struct sched_entity *se;
7596
7597 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq))
7598 update_tg_load_avg(cfs_rq, 0);
7599
7600
7601 se = cfs_rq->tg->se[cpu];
7602 if (se && !skip_blocked_update(se))
7603 update_load_avg(cfs_rq_of(se), se, 0);
7604
7605
7606
7607
7608
7609 if (cfs_rq_is_decayed(cfs_rq))
7610 list_del_leaf_cfs_rq(cfs_rq);
7611
7612
7613 if (cfs_rq_has_blocked(cfs_rq))
7614 done = false;
7615 }
7616
7617 curr_class = rq->curr->sched_class;
7618 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7619 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7620 update_irq_load_avg(rq, 0);
7621
7622 if (others_have_blocked(rq))
7623 done = false;
7624
7625 update_blocked_load_status(rq, !done);
7626 rq_unlock_irqrestore(rq, &rf);
7627}
7628
7629
7630
7631
7632
7633
7634static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7635{
7636 struct rq *rq = rq_of(cfs_rq);
7637 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7638 unsigned long now = jiffies;
7639 unsigned long load;
7640
7641 if (cfs_rq->last_h_load_update == now)
7642 return;
7643
7644 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7645 for_each_sched_entity(se) {
7646 cfs_rq = cfs_rq_of(se);
7647 WRITE_ONCE(cfs_rq->h_load_next, se);
7648 if (cfs_rq->last_h_load_update == now)
7649 break;
7650 }
7651
7652 if (!se) {
7653 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7654 cfs_rq->last_h_load_update = now;
7655 }
7656
7657 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7658 load = cfs_rq->h_load;
7659 load = div64_ul(load * se->avg.load_avg,
7660 cfs_rq_load_avg(cfs_rq) + 1);
7661 cfs_rq = group_cfs_rq(se);
7662 cfs_rq->h_load = load;
7663 cfs_rq->last_h_load_update = now;
7664 }
7665}
7666
7667static unsigned long task_h_load(struct task_struct *p)
7668{
7669 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7670
7671 update_cfs_rq_h_load(cfs_rq);
7672 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7673 cfs_rq_load_avg(cfs_rq) + 1);
7674}
7675#else
7676static inline void update_blocked_averages(int cpu)
7677{
7678 struct rq *rq = cpu_rq(cpu);
7679 struct cfs_rq *cfs_rq = &rq->cfs;
7680 const struct sched_class *curr_class;
7681 struct rq_flags rf;
7682
7683 rq_lock_irqsave(rq, &rf);
7684 update_rq_clock(rq);
7685 update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7686
7687 curr_class = rq->curr->sched_class;
7688 update_rt_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &rt_sched_class);
7689 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, curr_class == &dl_sched_class);
7690 update_irq_load_avg(rq, 0);
7691 update_blocked_load_status(rq, cfs_rq_has_blocked(cfs_rq) || others_have_blocked(rq));
7692 rq_unlock_irqrestore(rq, &rf);
7693}
7694
7695static unsigned long task_h_load(struct task_struct *p)
7696{
7697 return p->se.avg.load_avg;
7698}
7699#endif
7700
7701
7702
7703
7704
7705
7706struct sg_lb_stats {
7707 unsigned long avg_load;
7708 unsigned long group_load;
7709 unsigned long load_per_task;
7710 unsigned long group_capacity;
7711 unsigned long group_util;
7712 unsigned int sum_nr_running;
7713 unsigned int idle_cpus;
7714 unsigned int group_weight;
7715 enum group_type group_type;
7716 int group_no_capacity;
7717 unsigned long group_misfit_task_load;
7718#ifdef CONFIG_NUMA_BALANCING
7719 unsigned int nr_numa_running;
7720 unsigned int nr_preferred_running;
7721#endif
7722};
7723
7724
7725
7726
7727
7728struct sd_lb_stats {
7729 struct sched_group *busiest;
7730 struct sched_group *local;
7731 unsigned long total_running;
7732 unsigned long total_load;
7733 unsigned long total_capacity;
7734 unsigned long avg_load;
7735
7736 struct sg_lb_stats busiest_stat;
7737 struct sg_lb_stats local_stat;
7738};
7739
7740static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7741{
7742
7743
7744
7745
7746
7747
7748 *sds = (struct sd_lb_stats){
7749 .busiest = NULL,
7750 .local = NULL,
7751 .total_running = 0UL,
7752 .total_load = 0UL,
7753 .total_capacity = 0UL,
7754 .busiest_stat = {
7755 .avg_load = 0UL,
7756 .sum_nr_running = 0,
7757 .group_type = group_other,
7758 },
7759 };
7760}
7761
7762static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7763{
7764 struct rq *rq = cpu_rq(cpu);
7765 unsigned long max = arch_scale_cpu_capacity(cpu);
7766 unsigned long used, free;
7767 unsigned long irq;
7768
7769 irq = cpu_util_irq(rq);
7770
7771 if (unlikely(irq >= max))
7772 return 1;
7773
7774 used = READ_ONCE(rq->avg_rt.util_avg);
7775 used += READ_ONCE(rq->avg_dl.util_avg);
7776
7777 if (unlikely(used >= max))
7778 return 1;
7779
7780 free = max - used;
7781
7782 return scale_irq_capacity(free, irq, max);
7783}
7784
7785static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7786{
7787 unsigned long capacity = scale_rt_capacity(sd, cpu);
7788 struct sched_group *sdg = sd->groups;
7789
7790 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
7791
7792 if (!capacity)
7793 capacity = 1;
7794
7795 cpu_rq(cpu)->cpu_capacity = capacity;
7796 sdg->sgc->capacity = capacity;
7797 sdg->sgc->min_capacity = capacity;
7798 sdg->sgc->max_capacity = capacity;
7799}
7800
7801void update_group_capacity(struct sched_domain *sd, int cpu)
7802{
7803 struct sched_domain *child = sd->child;
7804 struct sched_group *group, *sdg = sd->groups;
7805 unsigned long capacity, min_capacity, max_capacity;
7806 unsigned long interval;
7807
7808 interval = msecs_to_jiffies(sd->balance_interval);
7809 interval = clamp(interval, 1UL, max_load_balance_interval);
7810 sdg->sgc->next_update = jiffies + interval;
7811
7812 if (!child) {
7813 update_cpu_capacity(sd, cpu);
7814 return;
7815 }
7816
7817 capacity = 0;
7818 min_capacity = ULONG_MAX;
7819 max_capacity = 0;
7820
7821 if (child->flags & SD_OVERLAP) {
7822
7823
7824
7825
7826
7827 for_each_cpu(cpu, sched_group_span(sdg)) {
7828 struct sched_group_capacity *sgc;
7829 struct rq *rq = cpu_rq(cpu);
7830
7831
7832
7833
7834
7835
7836
7837
7838
7839
7840
7841
7842 if (unlikely(!rq->sd)) {
7843 capacity += capacity_of(cpu);
7844 } else {
7845 sgc = rq->sd->groups->sgc;
7846 capacity += sgc->capacity;
7847 }
7848
7849 min_capacity = min(capacity, min_capacity);
7850 max_capacity = max(capacity, max_capacity);
7851 }
7852 } else {
7853
7854
7855
7856
7857
7858 group = child->groups;
7859 do {
7860 struct sched_group_capacity *sgc = group->sgc;
7861
7862 capacity += sgc->capacity;
7863 min_capacity = min(sgc->min_capacity, min_capacity);
7864 max_capacity = max(sgc->max_capacity, max_capacity);
7865 group = group->next;
7866 } while (group != child->groups);
7867 }
7868
7869 sdg->sgc->capacity = capacity;
7870 sdg->sgc->min_capacity = min_capacity;
7871 sdg->sgc->max_capacity = max_capacity;
7872}
7873
7874
7875
7876
7877
7878
7879static inline int
7880check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7881{
7882 return ((rq->cpu_capacity * sd->imbalance_pct) <
7883 (rq->cpu_capacity_orig * 100));
7884}
7885
7886
7887
7888
7889
7890
7891static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
7892{
7893 return rq->misfit_task_load &&
7894 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
7895 check_cpu_capacity(rq, sd));
7896}
7897
7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
7920
7921
7922
7923
7924
7925
7926
7927static inline int sg_imbalanced(struct sched_group *group)
7928{
7929 return group->sgc->imbalance;
7930}
7931
7932
7933
7934
7935
7936
7937
7938
7939
7940
7941
7942
7943
7944static inline bool
7945group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7946{
7947 if (sgs->sum_nr_running < sgs->group_weight)
7948 return true;
7949
7950 if ((sgs->group_capacity * 100) >
7951 (sgs->group_util * env->sd->imbalance_pct))
7952 return true;
7953
7954 return false;
7955}
7956
7957
7958
7959
7960
7961
7962
7963
7964
7965static inline bool
7966group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7967{
7968 if (sgs->sum_nr_running <= sgs->group_weight)
7969 return false;
7970
7971 if ((sgs->group_capacity * 100) <
7972 (sgs->group_util * env->sd->imbalance_pct))
7973 return true;
7974
7975 return false;
7976}
7977
7978
7979
7980
7981
7982static inline bool
7983group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7984{
7985 return sg->sgc->min_capacity * capacity_margin <
7986 ref->sgc->min_capacity * 1024;
7987}
7988
7989
7990
7991
7992
7993static inline bool
7994group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7995{
7996 return sg->sgc->max_capacity * capacity_margin <
7997 ref->sgc->max_capacity * 1024;
7998}
7999
8000static inline enum
8001group_type group_classify(struct sched_group *group,
8002 struct sg_lb_stats *sgs)
8003{
8004 if (sgs->group_no_capacity)
8005 return group_overloaded;
8006
8007 if (sg_imbalanced(group))
8008 return group_imbalanced;
8009
8010 if (sgs->group_misfit_task_load)
8011 return group_misfit_task;
8012
8013 return group_other;
8014}
8015
8016static bool update_nohz_stats(struct rq *rq, bool force)
8017{
8018#ifdef CONFIG_NO_HZ_COMMON
8019 unsigned int cpu = rq->cpu;
8020
8021 if (!rq->has_blocked_load)
8022 return false;
8023
8024 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8025 return false;
8026
8027 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8028 return true;
8029
8030 update_blocked_averages(cpu);
8031
8032 return rq->has_blocked_load;
8033#else
8034 return false;
8035#endif
8036}
8037
8038
8039
8040
8041
8042
8043
8044
8045static inline void update_sg_lb_stats(struct lb_env *env,
8046 struct sched_group *group,
8047 struct sg_lb_stats *sgs,
8048 int *sg_status)
8049{
8050 int i, nr_running;
8051
8052 memset(sgs, 0, sizeof(*sgs));
8053
8054 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8055 struct rq *rq = cpu_rq(i);
8056
8057 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8058 env->flags |= LBF_NOHZ_AGAIN;
8059
8060 sgs->group_load += cpu_runnable_load(rq);
8061 sgs->group_util += cpu_util(i);
8062 sgs->sum_nr_running += rq->cfs.h_nr_running;
8063
8064 nr_running = rq->nr_running;
8065 if (nr_running > 1)
8066 *sg_status |= SG_OVERLOAD;
8067
8068 if (cpu_overutilized(i))
8069 *sg_status |= SG_OVERUTILIZED;
8070
8071#ifdef CONFIG_NUMA_BALANCING
8072 sgs->nr_numa_running += rq->nr_numa_running;
8073 sgs->nr_preferred_running += rq->nr_preferred_running;
8074#endif
8075
8076
8077
8078 if (!nr_running && idle_cpu(i))
8079 sgs->idle_cpus++;
8080
8081 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8082 sgs->group_misfit_task_load < rq->misfit_task_load) {
8083 sgs->group_misfit_task_load = rq->misfit_task_load;
8084 *sg_status |= SG_OVERLOAD;
8085 }
8086 }
8087
8088
8089 sgs->group_capacity = group->sgc->capacity;
8090 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8091
8092 if (sgs->sum_nr_running)
8093 sgs->load_per_task = sgs->group_load / sgs->sum_nr_running;
8094
8095 sgs->group_weight = group->group_weight;
8096
8097 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8098 sgs->group_type = group_classify(group, sgs);
8099}
8100
8101
8102
8103
8104
8105
8106
8107
8108
8109
8110
8111
8112
8113
8114static bool update_sd_pick_busiest(struct lb_env *env,
8115 struct sd_lb_stats *sds,
8116 struct sched_group *sg,
8117 struct sg_lb_stats *sgs)
8118{
8119 struct sg_lb_stats *busiest = &sds->busiest_stat;
8120
8121
8122
8123
8124
8125
8126
8127 if (sgs->group_type == group_misfit_task &&
8128 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8129 !group_has_capacity(env, &sds->local_stat)))
8130 return false;
8131
8132 if (sgs->group_type > busiest->group_type)
8133 return true;
8134
8135 if (sgs->group_type < busiest->group_type)
8136 return false;
8137
8138 if (sgs->avg_load <= busiest->avg_load)
8139 return false;
8140
8141 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8142 goto asym_packing;
8143
8144
8145
8146
8147
8148
8149
8150 if (sgs->sum_nr_running <= sgs->group_weight &&
8151 group_smaller_min_cpu_capacity(sds->local, sg))
8152 return false;
8153
8154
8155
8156
8157 if (sgs->group_type == group_misfit_task &&
8158 sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8159 return false;
8160
8161asym_packing:
8162
8163 if (!(env->sd->flags & SD_ASYM_PACKING))
8164 return true;
8165
8166
8167 if (env->idle == CPU_NOT_IDLE)
8168 return true;
8169
8170
8171
8172
8173
8174 if (sgs->sum_nr_running &&
8175 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8176 if (!sds->busiest)
8177 return true;
8178
8179
8180 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8181 sg->asym_prefer_cpu))
8182 return true;
8183 }
8184
8185 return false;
8186}
8187
8188#ifdef CONFIG_NUMA_BALANCING
8189static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8190{
8191 if (sgs->sum_nr_running > sgs->nr_numa_running)
8192 return regular;
8193 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8194 return remote;
8195 return all;
8196}
8197
8198static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8199{
8200 if (rq->nr_running > rq->nr_numa_running)
8201 return regular;
8202 if (rq->nr_running > rq->nr_preferred_running)
8203 return remote;
8204 return all;
8205}
8206#else
8207static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8208{
8209 return all;
8210}
8211
8212static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8213{
8214 return regular;
8215}
8216#endif
8217
8218
8219
8220
8221
8222
8223static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8224{
8225 struct sched_domain *child = env->sd->child;
8226 struct sched_group *sg = env->sd->groups;
8227 struct sg_lb_stats *local = &sds->local_stat;
8228 struct sg_lb_stats tmp_sgs;
8229 bool prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8230 int sg_status = 0;
8231
8232#ifdef CONFIG_NO_HZ_COMMON
8233 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8234 env->flags |= LBF_NOHZ_STATS;
8235#endif
8236
8237 do {
8238 struct sg_lb_stats *sgs = &tmp_sgs;
8239 int local_group;
8240
8241 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8242 if (local_group) {
8243 sds->local = sg;
8244 sgs = local;
8245
8246 if (env->idle != CPU_NEWLY_IDLE ||
8247 time_after_eq(jiffies, sg->sgc->next_update))
8248 update_group_capacity(env->sd, env->dst_cpu);
8249 }
8250
8251 update_sg_lb_stats(env, sg, sgs, &sg_status);
8252
8253 if (local_group)
8254 goto next_group;
8255
8256
8257
8258
8259
8260
8261
8262
8263
8264
8265
8266 if (prefer_sibling && sds->local &&
8267 group_has_capacity(env, local) &&
8268 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8269 sgs->group_no_capacity = 1;
8270 sgs->group_type = group_classify(sg, sgs);
8271 }
8272
8273 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8274 sds->busiest = sg;
8275 sds->busiest_stat = *sgs;
8276 }
8277
8278next_group:
8279
8280 sds->total_running += sgs->sum_nr_running;
8281 sds->total_load += sgs->group_load;
8282 sds->total_capacity += sgs->group_capacity;
8283
8284 sg = sg->next;
8285 } while (sg != env->sd->groups);
8286
8287#ifdef CONFIG_NO_HZ_COMMON
8288 if ((env->flags & LBF_NOHZ_AGAIN) &&
8289 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8290
8291 WRITE_ONCE(nohz.next_blocked,
8292 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8293 }
8294#endif
8295
8296 if (env->sd->flags & SD_NUMA)
8297 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8298
8299 if (!env->sd->parent) {
8300 struct root_domain *rd = env->dst_rq->rd;
8301
8302
8303 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8304
8305
8306 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8307 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8308 } else if (sg_status & SG_OVERUTILIZED) {
8309 struct root_domain *rd = env->dst_rq->rd;
8310
8311 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8312 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8313 }
8314}
8315
8316
8317
8318
8319
8320
8321
8322
8323
8324
8325
8326
8327
8328
8329
8330
8331
8332
8333
8334
8335
8336
8337
8338
8339static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8340{
8341 int busiest_cpu;
8342
8343 if (!(env->sd->flags & SD_ASYM_PACKING))
8344 return 0;
8345
8346 if (env->idle == CPU_NOT_IDLE)
8347 return 0;
8348
8349 if (!sds->busiest)
8350 return 0;
8351
8352 busiest_cpu = sds->busiest->asym_prefer_cpu;
8353 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8354 return 0;
8355
8356 env->imbalance = sds->busiest_stat.group_load;
8357
8358 return 1;
8359}
8360
8361
8362
8363
8364
8365
8366
8367
8368static inline
8369void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8370{
8371 unsigned long tmp, capa_now = 0, capa_move = 0;
8372 unsigned int imbn = 2;
8373 unsigned long scaled_busy_load_per_task;
8374 struct sg_lb_stats *local, *busiest;
8375
8376 local = &sds->local_stat;
8377 busiest = &sds->busiest_stat;
8378
8379 if (!local->sum_nr_running)
8380 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8381 else if (busiest->load_per_task > local->load_per_task)
8382 imbn = 1;
8383
8384 scaled_busy_load_per_task =
8385 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8386 busiest->group_capacity;
8387
8388 if (busiest->avg_load + scaled_busy_load_per_task >=
8389 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8390 env->imbalance = busiest->load_per_task;
8391 return;
8392 }
8393
8394
8395
8396
8397
8398
8399
8400 capa_now += busiest->group_capacity *
8401 min(busiest->load_per_task, busiest->avg_load);
8402 capa_now += local->group_capacity *
8403 min(local->load_per_task, local->avg_load);
8404 capa_now /= SCHED_CAPACITY_SCALE;
8405
8406
8407 if (busiest->avg_load > scaled_busy_load_per_task) {
8408 capa_move += busiest->group_capacity *
8409 min(busiest->load_per_task,
8410 busiest->avg_load - scaled_busy_load_per_task);
8411 }
8412
8413
8414 if (busiest->avg_load * busiest->group_capacity <
8415 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8416 tmp = (busiest->avg_load * busiest->group_capacity) /
8417 local->group_capacity;
8418 } else {
8419 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8420 local->group_capacity;
8421 }
8422 capa_move += local->group_capacity *
8423 min(local->load_per_task, local->avg_load + tmp);
8424 capa_move /= SCHED_CAPACITY_SCALE;
8425
8426
8427 if (capa_move > capa_now)
8428 env->imbalance = busiest->load_per_task;
8429}
8430
8431
8432
8433
8434
8435
8436
8437static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8438{
8439 unsigned long max_pull, load_above_capacity = ~0UL;
8440 struct sg_lb_stats *local, *busiest;
8441
8442 local = &sds->local_stat;
8443 busiest = &sds->busiest_stat;
8444
8445 if (busiest->group_type == group_imbalanced) {
8446
8447
8448
8449
8450 busiest->load_per_task =
8451 min(busiest->load_per_task, sds->avg_load);
8452 }
8453
8454
8455
8456
8457
8458
8459
8460 if (busiest->group_type != group_misfit_task &&
8461 (busiest->avg_load <= sds->avg_load ||
8462 local->avg_load >= sds->avg_load)) {
8463 env->imbalance = 0;
8464 return fix_small_imbalance(env, sds);
8465 }
8466
8467
8468
8469
8470 if (busiest->group_type == group_overloaded &&
8471 local->group_type == group_overloaded) {
8472 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8473 if (load_above_capacity > busiest->group_capacity) {
8474 load_above_capacity -= busiest->group_capacity;
8475 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8476 load_above_capacity /= busiest->group_capacity;
8477 } else
8478 load_above_capacity = ~0UL;
8479 }
8480
8481
8482
8483
8484
8485
8486
8487
8488 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8489
8490
8491 env->imbalance = min(
8492 max_pull * busiest->group_capacity,
8493 (sds->avg_load - local->avg_load) * local->group_capacity
8494 ) / SCHED_CAPACITY_SCALE;
8495
8496
8497 if (busiest->group_type == group_misfit_task) {
8498 env->imbalance = max_t(long, env->imbalance,
8499 busiest->group_misfit_task_load);
8500 }
8501
8502
8503
8504
8505
8506
8507
8508 if (env->imbalance < busiest->load_per_task)
8509 return fix_small_imbalance(env, sds);
8510}
8511
8512
8513
8514
8515
8516
8517
8518
8519
8520
8521
8522
8523
8524
8525static struct sched_group *find_busiest_group(struct lb_env *env)
8526{
8527 struct sg_lb_stats *local, *busiest;
8528 struct sd_lb_stats sds;
8529
8530 init_sd_lb_stats(&sds);
8531
8532
8533
8534
8535
8536 update_sd_lb_stats(env, &sds);
8537
8538 if (sched_energy_enabled()) {
8539 struct root_domain *rd = env->dst_rq->rd;
8540
8541 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
8542 goto out_balanced;
8543 }
8544
8545 local = &sds.local_stat;
8546 busiest = &sds.busiest_stat;
8547
8548
8549 if (check_asym_packing(env, &sds))
8550 return sds.busiest;
8551
8552
8553 if (!sds.busiest || busiest->sum_nr_running == 0)
8554 goto out_balanced;
8555
8556
8557 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8558 / sds.total_capacity;
8559
8560
8561
8562
8563
8564
8565 if (busiest->group_type == group_imbalanced)
8566 goto force_balance;
8567
8568
8569
8570
8571
8572 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8573 busiest->group_no_capacity)
8574 goto force_balance;
8575
8576
8577 if (busiest->group_type == group_misfit_task)
8578 goto force_balance;
8579
8580
8581
8582
8583
8584 if (local->avg_load >= busiest->avg_load)
8585 goto out_balanced;
8586
8587
8588
8589
8590
8591 if (local->avg_load >= sds.avg_load)
8592 goto out_balanced;
8593
8594 if (env->idle == CPU_IDLE) {
8595
8596
8597
8598
8599
8600
8601
8602 if ((busiest->group_type != group_overloaded) &&
8603 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8604 goto out_balanced;
8605 } else {
8606
8607
8608
8609
8610 if (100 * busiest->avg_load <=
8611 env->sd->imbalance_pct * local->avg_load)
8612 goto out_balanced;
8613 }
8614
8615force_balance:
8616
8617 env->src_grp_type = busiest->group_type;
8618 calculate_imbalance(env, &sds);
8619 return env->imbalance ? sds.busiest : NULL;
8620
8621out_balanced:
8622 env->imbalance = 0;
8623 return NULL;
8624}
8625
8626
8627
8628
8629static struct rq *find_busiest_queue(struct lb_env *env,
8630 struct sched_group *group)
8631{
8632 struct rq *busiest = NULL, *rq;
8633 unsigned long busiest_load = 0, busiest_capacity = 1;
8634 int i;
8635
8636 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8637 unsigned long capacity, load;
8638 enum fbq_type rt;
8639
8640 rq = cpu_rq(i);
8641 rt = fbq_classify_rq(rq);
8642
8643
8644
8645
8646
8647
8648
8649
8650
8651
8652
8653
8654
8655
8656
8657
8658
8659
8660
8661
8662 if (rt > env->fbq_type)
8663 continue;
8664
8665
8666
8667
8668
8669 if (env->src_grp_type == group_misfit_task) {
8670 if (rq->misfit_task_load > busiest_load) {
8671 busiest_load = rq->misfit_task_load;
8672 busiest = rq;
8673 }
8674
8675 continue;
8676 }
8677
8678 capacity = capacity_of(i);
8679
8680
8681
8682
8683
8684
8685
8686 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8687 capacity_of(env->dst_cpu) < capacity &&
8688 rq->nr_running == 1)
8689 continue;
8690
8691 load = cpu_runnable_load(rq);
8692
8693
8694
8695
8696
8697
8698 if (rq->nr_running == 1 && load > env->imbalance &&
8699 !check_cpu_capacity(rq, env->sd))
8700 continue;
8701
8702
8703
8704
8705
8706
8707
8708
8709
8710
8711
8712
8713 if (load * busiest_capacity > busiest_load * capacity) {
8714 busiest_load = load;
8715 busiest_capacity = capacity;
8716 busiest = rq;
8717 }
8718 }
8719
8720 return busiest;
8721}
8722
8723
8724
8725
8726
8727#define MAX_PINNED_INTERVAL 512
8728
8729static inline bool
8730asym_active_balance(struct lb_env *env)
8731{
8732
8733
8734
8735
8736
8737 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
8738 sched_asym_prefer(env->dst_cpu, env->src_cpu);
8739}
8740
8741static inline bool
8742voluntary_active_balance(struct lb_env *env)
8743{
8744 struct sched_domain *sd = env->sd;
8745
8746 if (asym_active_balance(env))
8747 return 1;
8748
8749
8750
8751
8752
8753
8754
8755 if ((env->idle != CPU_NOT_IDLE) &&
8756 (env->src_rq->cfs.h_nr_running == 1)) {
8757 if ((check_cpu_capacity(env->src_rq, sd)) &&
8758 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8759 return 1;
8760 }
8761
8762 if (env->src_grp_type == group_misfit_task)
8763 return 1;
8764
8765 return 0;
8766}
8767
8768static int need_active_balance(struct lb_env *env)
8769{
8770 struct sched_domain *sd = env->sd;
8771
8772 if (voluntary_active_balance(env))
8773 return 1;
8774
8775 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8776}
8777
8778static int active_load_balance_cpu_stop(void *data);
8779
8780static int should_we_balance(struct lb_env *env)
8781{
8782 struct sched_group *sg = env->sd->groups;
8783 int cpu, balance_cpu = -1;
8784
8785
8786
8787
8788
8789 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8790 return 0;
8791
8792
8793
8794
8795
8796 if (env->idle == CPU_NEWLY_IDLE)
8797 return 1;
8798
8799
8800 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8801 if (!idle_cpu(cpu))
8802 continue;
8803
8804 balance_cpu = cpu;
8805 break;
8806 }
8807
8808 if (balance_cpu == -1)
8809 balance_cpu = group_balance_cpu(sg);
8810
8811
8812
8813
8814
8815 return balance_cpu == env->dst_cpu;
8816}
8817
8818
8819
8820
8821
8822static int load_balance(int this_cpu, struct rq *this_rq,
8823 struct sched_domain *sd, enum cpu_idle_type idle,
8824 int *continue_balancing)
8825{
8826 int ld_moved, cur_ld_moved, active_balance = 0;
8827 struct sched_domain *sd_parent = sd->parent;
8828 struct sched_group *group;
8829 struct rq *busiest;
8830 struct rq_flags rf;
8831 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8832
8833 struct lb_env env = {
8834 .sd = sd,
8835 .dst_cpu = this_cpu,
8836 .dst_rq = this_rq,
8837 .dst_grpmask = sched_group_span(sd->groups),
8838 .idle = idle,
8839 .loop_break = sched_nr_migrate_break,
8840 .cpus = cpus,
8841 .fbq_type = all,
8842 .tasks = LIST_HEAD_INIT(env.tasks),
8843 };
8844
8845 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8846
8847 schedstat_inc(sd->lb_count[idle]);
8848
8849redo:
8850 if (!should_we_balance(&env)) {
8851 *continue_balancing = 0;
8852 goto out_balanced;
8853 }
8854
8855 group = find_busiest_group(&env);
8856 if (!group) {
8857 schedstat_inc(sd->lb_nobusyg[idle]);
8858 goto out_balanced;
8859 }
8860
8861 busiest = find_busiest_queue(&env, group);
8862 if (!busiest) {
8863 schedstat_inc(sd->lb_nobusyq[idle]);
8864 goto out_balanced;
8865 }
8866
8867 BUG_ON(busiest == env.dst_rq);
8868
8869 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8870
8871 env.src_cpu = busiest->cpu;
8872 env.src_rq = busiest;
8873
8874 ld_moved = 0;
8875 if (busiest->nr_running > 1) {
8876
8877
8878
8879
8880
8881
8882 env.flags |= LBF_ALL_PINNED;
8883 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8884
8885more_balance:
8886 rq_lock_irqsave(busiest, &rf);
8887 update_rq_clock(busiest);
8888
8889
8890
8891
8892
8893 cur_ld_moved = detach_tasks(&env);
8894
8895
8896
8897
8898
8899
8900
8901
8902
8903 rq_unlock(busiest, &rf);
8904
8905 if (cur_ld_moved) {
8906 attach_tasks(&env);
8907 ld_moved += cur_ld_moved;
8908 }
8909
8910 local_irq_restore(rf.flags);
8911
8912 if (env.flags & LBF_NEED_BREAK) {
8913 env.flags &= ~LBF_NEED_BREAK;
8914 goto more_balance;
8915 }
8916
8917
8918
8919
8920
8921
8922
8923
8924
8925
8926
8927
8928
8929
8930
8931
8932
8933
8934
8935
8936 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8937
8938
8939 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
8940
8941 env.dst_rq = cpu_rq(env.new_dst_cpu);
8942 env.dst_cpu = env.new_dst_cpu;
8943 env.flags &= ~LBF_DST_PINNED;
8944 env.loop = 0;
8945 env.loop_break = sched_nr_migrate_break;
8946
8947
8948
8949
8950
8951 goto more_balance;
8952 }
8953
8954
8955
8956
8957 if (sd_parent) {
8958 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8959
8960 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8961 *group_imbalance = 1;
8962 }
8963
8964
8965 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8966 __cpumask_clear_cpu(cpu_of(busiest), cpus);
8967
8968
8969
8970
8971
8972
8973
8974
8975 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8976 env.loop = 0;
8977 env.loop_break = sched_nr_migrate_break;
8978 goto redo;
8979 }
8980 goto out_all_pinned;
8981 }
8982 }
8983
8984 if (!ld_moved) {
8985 schedstat_inc(sd->lb_failed[idle]);
8986
8987
8988
8989
8990
8991
8992 if (idle != CPU_NEWLY_IDLE)
8993 sd->nr_balance_failed++;
8994
8995 if (need_active_balance(&env)) {
8996 unsigned long flags;
8997
8998 raw_spin_lock_irqsave(&busiest->lock, flags);
8999
9000
9001
9002
9003
9004
9005 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9006 raw_spin_unlock_irqrestore(&busiest->lock,
9007 flags);
9008 env.flags |= LBF_ALL_PINNED;
9009 goto out_one_pinned;
9010 }
9011
9012
9013
9014
9015
9016
9017 if (!busiest->active_balance) {
9018 busiest->active_balance = 1;
9019 busiest->push_cpu = this_cpu;
9020 active_balance = 1;
9021 }
9022 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9023
9024 if (active_balance) {
9025 stop_one_cpu_nowait(cpu_of(busiest),
9026 active_load_balance_cpu_stop, busiest,
9027 &busiest->active_balance_work);
9028 }
9029
9030
9031 sd->nr_balance_failed = sd->cache_nice_tries+1;
9032 }
9033 } else
9034 sd->nr_balance_failed = 0;
9035
9036 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9037
9038 sd->balance_interval = sd->min_interval;
9039 } else {
9040
9041
9042
9043
9044
9045
9046 if (sd->balance_interval < sd->max_interval)
9047 sd->balance_interval *= 2;
9048 }
9049
9050 goto out;
9051
9052out_balanced:
9053
9054
9055
9056
9057 if (sd_parent) {
9058 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9059
9060 if (*group_imbalance)
9061 *group_imbalance = 0;
9062 }
9063
9064out_all_pinned:
9065
9066
9067
9068
9069
9070 schedstat_inc(sd->lb_balanced[idle]);
9071
9072 sd->nr_balance_failed = 0;
9073
9074out_one_pinned:
9075 ld_moved = 0;
9076
9077
9078
9079
9080
9081
9082
9083 if (env.idle == CPU_NEWLY_IDLE)
9084 goto out;
9085
9086
9087 if ((env.flags & LBF_ALL_PINNED &&
9088 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9089 sd->balance_interval < sd->max_interval)
9090 sd->balance_interval *= 2;
9091out:
9092 return ld_moved;
9093}
9094
9095static inline unsigned long
9096get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9097{
9098 unsigned long interval = sd->balance_interval;
9099
9100 if (cpu_busy)
9101 interval *= sd->busy_factor;
9102
9103
9104 interval = msecs_to_jiffies(interval);
9105 interval = clamp(interval, 1UL, max_load_balance_interval);
9106
9107 return interval;
9108}
9109
9110static inline void
9111update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9112{
9113 unsigned long interval, next;
9114
9115
9116 interval = get_sd_balance_interval(sd, 0);
9117 next = sd->last_balance + interval;
9118
9119 if (time_after(*next_balance, next))
9120 *next_balance = next;
9121}
9122
9123
9124
9125
9126
9127
9128
9129static int active_load_balance_cpu_stop(void *data)
9130{
9131 struct rq *busiest_rq = data;
9132 int busiest_cpu = cpu_of(busiest_rq);
9133 int target_cpu = busiest_rq->push_cpu;
9134 struct rq *target_rq = cpu_rq(target_cpu);
9135 struct sched_domain *sd;
9136 struct task_struct *p = NULL;
9137 struct rq_flags rf;
9138
9139 rq_lock_irq(busiest_rq, &rf);
9140
9141
9142
9143
9144
9145 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9146 goto out_unlock;
9147
9148
9149 if (unlikely(busiest_cpu != smp_processor_id() ||
9150 !busiest_rq->active_balance))
9151 goto out_unlock;
9152
9153
9154 if (busiest_rq->nr_running <= 1)
9155 goto out_unlock;
9156
9157
9158
9159
9160
9161
9162 BUG_ON(busiest_rq == target_rq);
9163
9164
9165 rcu_read_lock();
9166 for_each_domain(target_cpu, sd) {
9167 if ((sd->flags & SD_LOAD_BALANCE) &&
9168 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9169 break;
9170 }
9171
9172 if (likely(sd)) {
9173 struct lb_env env = {
9174 .sd = sd,
9175 .dst_cpu = target_cpu,
9176 .dst_rq = target_rq,
9177 .src_cpu = busiest_rq->cpu,
9178 .src_rq = busiest_rq,
9179 .idle = CPU_IDLE,
9180
9181
9182
9183
9184
9185
9186 .flags = LBF_DST_PINNED,
9187 };
9188
9189 schedstat_inc(sd->alb_count);
9190 update_rq_clock(busiest_rq);
9191
9192 p = detach_one_task(&env);
9193 if (p) {
9194 schedstat_inc(sd->alb_pushed);
9195
9196 sd->nr_balance_failed = 0;
9197 } else {
9198 schedstat_inc(sd->alb_failed);
9199 }
9200 }
9201 rcu_read_unlock();
9202out_unlock:
9203 busiest_rq->active_balance = 0;
9204 rq_unlock(busiest_rq, &rf);
9205
9206 if (p)
9207 attach_one_task(target_rq, p);
9208
9209 local_irq_enable();
9210
9211 return 0;
9212}
9213
9214static DEFINE_SPINLOCK(balancing);
9215
9216
9217
9218
9219
9220void update_max_interval(void)
9221{
9222 max_load_balance_interval = HZ*num_online_cpus()/10;
9223}
9224
9225
9226
9227
9228
9229
9230
9231static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9232{
9233 int continue_balancing = 1;
9234 int cpu = rq->cpu;
9235 unsigned long interval;
9236 struct sched_domain *sd;
9237
9238 unsigned long next_balance = jiffies + 60*HZ;
9239 int update_next_balance = 0;
9240 int need_serialize, need_decay = 0;
9241 u64 max_cost = 0;
9242
9243 rcu_read_lock();
9244 for_each_domain(cpu, sd) {
9245
9246
9247
9248
9249 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9250 sd->max_newidle_lb_cost =
9251 (sd->max_newidle_lb_cost * 253) / 256;
9252 sd->next_decay_max_lb_cost = jiffies + HZ;
9253 need_decay = 1;
9254 }
9255 max_cost += sd->max_newidle_lb_cost;
9256
9257 if (!(sd->flags & SD_LOAD_BALANCE))
9258 continue;
9259
9260
9261
9262
9263
9264
9265 if (!continue_balancing) {
9266 if (need_decay)
9267 continue;
9268 break;
9269 }
9270
9271 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9272
9273 need_serialize = sd->flags & SD_SERIALIZE;
9274 if (need_serialize) {
9275 if (!spin_trylock(&balancing))
9276 goto out;
9277 }
9278
9279 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9280 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9281
9282
9283
9284
9285
9286 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9287 }
9288 sd->last_balance = jiffies;
9289 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9290 }
9291 if (need_serialize)
9292 spin_unlock(&balancing);
9293out:
9294 if (time_after(next_balance, sd->last_balance + interval)) {
9295 next_balance = sd->last_balance + interval;
9296 update_next_balance = 1;
9297 }
9298 }
9299 if (need_decay) {
9300
9301
9302
9303
9304 rq->max_idle_balance_cost =
9305 max((u64)sysctl_sched_migration_cost, max_cost);
9306 }
9307 rcu_read_unlock();
9308
9309
9310
9311
9312
9313
9314 if (likely(update_next_balance)) {
9315 rq->next_balance = next_balance;
9316
9317#ifdef CONFIG_NO_HZ_COMMON
9318
9319
9320
9321
9322
9323
9324
9325
9326 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9327 nohz.next_balance = rq->next_balance;
9328#endif
9329 }
9330}
9331
9332static inline int on_null_domain(struct rq *rq)
9333{
9334 return unlikely(!rcu_dereference_sched(rq->sd));
9335}
9336
9337#ifdef CONFIG_NO_HZ_COMMON
9338
9339
9340
9341
9342
9343
9344
9345
9346
9347static inline int find_new_ilb(void)
9348{
9349 int ilb;
9350
9351 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9352 housekeeping_cpumask(HK_FLAG_MISC)) {
9353 if (idle_cpu(ilb))
9354 return ilb;
9355 }
9356
9357 return nr_cpu_ids;
9358}
9359
9360
9361
9362
9363
9364static void kick_ilb(unsigned int flags)
9365{
9366 int ilb_cpu;
9367
9368 nohz.next_balance++;
9369
9370 ilb_cpu = find_new_ilb();
9371
9372 if (ilb_cpu >= nr_cpu_ids)
9373 return;
9374
9375 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9376 if (flags & NOHZ_KICK_MASK)
9377 return;
9378
9379
9380
9381
9382
9383
9384
9385 smp_send_reschedule(ilb_cpu);
9386}
9387
9388
9389
9390
9391
9392static void nohz_balancer_kick(struct rq *rq)
9393{
9394 unsigned long now = jiffies;
9395 struct sched_domain_shared *sds;
9396 struct sched_domain *sd;
9397 int nr_busy, i, cpu = rq->cpu;
9398 unsigned int flags = 0;
9399
9400 if (unlikely(rq->idle_balance))
9401 return;
9402
9403
9404
9405
9406
9407 nohz_balance_exit_idle(rq);
9408
9409
9410
9411
9412
9413 if (likely(!atomic_read(&nohz.nr_cpus)))
9414 return;
9415
9416 if (READ_ONCE(nohz.has_blocked) &&
9417 time_after(now, READ_ONCE(nohz.next_blocked)))
9418 flags = NOHZ_STATS_KICK;
9419
9420 if (time_before(now, nohz.next_balance))
9421 goto out;
9422
9423 if (rq->nr_running >= 2) {
9424 flags = NOHZ_KICK_MASK;
9425 goto out;
9426 }
9427
9428 rcu_read_lock();
9429
9430 sd = rcu_dereference(rq->sd);
9431 if (sd) {
9432
9433
9434
9435
9436
9437 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
9438 flags = NOHZ_KICK_MASK;
9439 goto unlock;
9440 }
9441 }
9442
9443 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
9444 if (sd) {
9445
9446
9447
9448
9449
9450 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
9451 if (sched_asym_prefer(i, cpu)) {
9452 flags = NOHZ_KICK_MASK;
9453 goto unlock;
9454 }
9455 }
9456 }
9457
9458 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
9459 if (sd) {
9460
9461
9462
9463
9464 if (check_misfit_status(rq, sd)) {
9465 flags = NOHZ_KICK_MASK;
9466 goto unlock;
9467 }
9468
9469
9470
9471
9472
9473
9474
9475
9476 goto unlock;
9477 }
9478
9479 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9480 if (sds) {
9481
9482
9483
9484
9485
9486
9487
9488
9489
9490 nr_busy = atomic_read(&sds->nr_busy_cpus);
9491 if (nr_busy > 1) {
9492 flags = NOHZ_KICK_MASK;
9493 goto unlock;
9494 }
9495 }
9496unlock:
9497 rcu_read_unlock();
9498out:
9499 if (flags)
9500 kick_ilb(flags);
9501}
9502
9503static void set_cpu_sd_state_busy(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 = 0;
9513
9514 atomic_inc(&sd->shared->nr_busy_cpus);
9515unlock:
9516 rcu_read_unlock();
9517}
9518
9519void nohz_balance_exit_idle(struct rq *rq)
9520{
9521 SCHED_WARN_ON(rq != this_rq());
9522
9523 if (likely(!rq->nohz_tick_stopped))
9524 return;
9525
9526 rq->nohz_tick_stopped = 0;
9527 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9528 atomic_dec(&nohz.nr_cpus);
9529
9530 set_cpu_sd_state_busy(rq->cpu);
9531}
9532
9533static void set_cpu_sd_state_idle(int cpu)
9534{
9535 struct sched_domain *sd;
9536
9537 rcu_read_lock();
9538 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9539
9540 if (!sd || sd->nohz_idle)
9541 goto unlock;
9542 sd->nohz_idle = 1;
9543
9544 atomic_dec(&sd->shared->nr_busy_cpus);
9545unlock:
9546 rcu_read_unlock();
9547}
9548
9549
9550
9551
9552
9553void nohz_balance_enter_idle(int cpu)
9554{
9555 struct rq *rq = cpu_rq(cpu);
9556
9557 SCHED_WARN_ON(cpu != smp_processor_id());
9558
9559
9560 if (!cpu_active(cpu))
9561 return;
9562
9563
9564 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9565 return;
9566
9567
9568
9569
9570
9571
9572 rq->has_blocked_load = 1;
9573
9574
9575
9576
9577
9578
9579
9580 if (rq->nohz_tick_stopped)
9581 goto out;
9582
9583
9584 if (on_null_domain(rq))
9585 return;
9586
9587 rq->nohz_tick_stopped = 1;
9588
9589 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9590 atomic_inc(&nohz.nr_cpus);
9591
9592
9593
9594
9595
9596
9597 smp_mb__after_atomic();
9598
9599 set_cpu_sd_state_idle(cpu);
9600
9601out:
9602
9603
9604
9605
9606 WRITE_ONCE(nohz.has_blocked, 1);
9607}
9608
9609
9610
9611
9612
9613
9614
9615
9616static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9617 enum cpu_idle_type idle)
9618{
9619
9620 unsigned long now = jiffies;
9621 unsigned long next_balance = now + 60*HZ;
9622 bool has_blocked_load = false;
9623 int update_next_balance = 0;
9624 int this_cpu = this_rq->cpu;
9625 int balance_cpu;
9626 int ret = false;
9627 struct rq *rq;
9628
9629 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9630
9631
9632
9633
9634
9635
9636
9637
9638
9639 WRITE_ONCE(nohz.has_blocked, 0);
9640
9641
9642
9643
9644
9645 smp_mb();
9646
9647 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9648 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9649 continue;
9650
9651
9652
9653
9654
9655
9656 if (need_resched()) {
9657 has_blocked_load = true;
9658 goto abort;
9659 }
9660
9661 rq = cpu_rq(balance_cpu);
9662
9663 has_blocked_load |= update_nohz_stats(rq, true);
9664
9665
9666
9667
9668
9669 if (time_after_eq(jiffies, rq->next_balance)) {
9670 struct rq_flags rf;
9671
9672 rq_lock_irqsave(rq, &rf);
9673 update_rq_clock(rq);
9674 rq_unlock_irqrestore(rq, &rf);
9675
9676 if (flags & NOHZ_BALANCE_KICK)
9677 rebalance_domains(rq, CPU_IDLE);
9678 }
9679
9680 if (time_after(next_balance, rq->next_balance)) {
9681 next_balance = rq->next_balance;
9682 update_next_balance = 1;
9683 }
9684 }
9685
9686
9687 if (idle != CPU_NEWLY_IDLE) {
9688 update_blocked_averages(this_cpu);
9689 has_blocked_load |= this_rq->has_blocked_load;
9690 }
9691
9692 if (flags & NOHZ_BALANCE_KICK)
9693 rebalance_domains(this_rq, CPU_IDLE);
9694
9695 WRITE_ONCE(nohz.next_blocked,
9696 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9697
9698
9699 ret = true;
9700
9701abort:
9702
9703 if (has_blocked_load)
9704 WRITE_ONCE(nohz.has_blocked, 1);
9705
9706
9707
9708
9709
9710
9711 if (likely(update_next_balance))
9712 nohz.next_balance = next_balance;
9713
9714 return ret;
9715}
9716
9717
9718
9719
9720
9721static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9722{
9723 int this_cpu = this_rq->cpu;
9724 unsigned int flags;
9725
9726 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9727 return false;
9728
9729 if (idle != CPU_IDLE) {
9730 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9731 return false;
9732 }
9733
9734
9735 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9736 if (!(flags & NOHZ_KICK_MASK))
9737 return false;
9738
9739 _nohz_idle_balance(this_rq, flags, idle);
9740
9741 return true;
9742}
9743
9744static void nohz_newidle_balance(struct rq *this_rq)
9745{
9746 int this_cpu = this_rq->cpu;
9747
9748
9749
9750
9751
9752 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9753 return;
9754
9755
9756 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9757 return;
9758
9759
9760 if (!READ_ONCE(nohz.has_blocked) ||
9761 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9762 return;
9763
9764 raw_spin_unlock(&this_rq->lock);
9765
9766
9767
9768
9769
9770
9771 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9772 kick_ilb(NOHZ_STATS_KICK);
9773 raw_spin_lock(&this_rq->lock);
9774}
9775
9776#else
9777static inline void nohz_balancer_kick(struct rq *rq) { }
9778
9779static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9780{
9781 return false;
9782}
9783
9784static inline void nohz_newidle_balance(struct rq *this_rq) { }
9785#endif
9786
9787
9788
9789
9790
9791static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9792{
9793 unsigned long next_balance = jiffies + HZ;
9794 int this_cpu = this_rq->cpu;
9795 struct sched_domain *sd;
9796 int pulled_task = 0;
9797 u64 curr_cost = 0;
9798
9799
9800
9801
9802
9803 this_rq->idle_stamp = rq_clock(this_rq);
9804
9805
9806
9807
9808 if (!cpu_active(this_cpu))
9809 return 0;
9810
9811
9812
9813
9814
9815
9816
9817 rq_unpin_lock(this_rq, rf);
9818
9819 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9820 !READ_ONCE(this_rq->rd->overload)) {
9821
9822 rcu_read_lock();
9823 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9824 if (sd)
9825 update_next_balance(sd, &next_balance);
9826 rcu_read_unlock();
9827
9828 nohz_newidle_balance(this_rq);
9829
9830 goto out;
9831 }
9832
9833 raw_spin_unlock(&this_rq->lock);
9834
9835 update_blocked_averages(this_cpu);
9836 rcu_read_lock();
9837 for_each_domain(this_cpu, sd) {
9838 int continue_balancing = 1;
9839 u64 t0, domain_cost;
9840
9841 if (!(sd->flags & SD_LOAD_BALANCE))
9842 continue;
9843
9844 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9845 update_next_balance(sd, &next_balance);
9846 break;
9847 }
9848
9849 if (sd->flags & SD_BALANCE_NEWIDLE) {
9850 t0 = sched_clock_cpu(this_cpu);
9851
9852 pulled_task = load_balance(this_cpu, this_rq,
9853 sd, CPU_NEWLY_IDLE,
9854 &continue_balancing);
9855
9856 domain_cost = sched_clock_cpu(this_cpu) - t0;
9857 if (domain_cost > sd->max_newidle_lb_cost)
9858 sd->max_newidle_lb_cost = domain_cost;
9859
9860 curr_cost += domain_cost;
9861 }
9862
9863 update_next_balance(sd, &next_balance);
9864
9865
9866
9867
9868
9869 if (pulled_task || this_rq->nr_running > 0)
9870 break;
9871 }
9872 rcu_read_unlock();
9873
9874 raw_spin_lock(&this_rq->lock);
9875
9876 if (curr_cost > this_rq->max_idle_balance_cost)
9877 this_rq->max_idle_balance_cost = curr_cost;
9878
9879out:
9880
9881
9882
9883
9884
9885 if (this_rq->cfs.h_nr_running && !pulled_task)
9886 pulled_task = 1;
9887
9888
9889 if (time_after(this_rq->next_balance, next_balance))
9890 this_rq->next_balance = next_balance;
9891
9892
9893 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9894 pulled_task = -1;
9895
9896 if (pulled_task)
9897 this_rq->idle_stamp = 0;
9898
9899 rq_repin_lock(this_rq, rf);
9900
9901 return pulled_task;
9902}
9903
9904
9905
9906
9907
9908static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9909{
9910 struct rq *this_rq = this_rq();
9911 enum cpu_idle_type idle = this_rq->idle_balance ?
9912 CPU_IDLE : CPU_NOT_IDLE;
9913
9914
9915
9916
9917
9918
9919
9920
9921
9922 if (nohz_idle_balance(this_rq, idle))
9923 return;
9924
9925
9926 update_blocked_averages(this_rq->cpu);
9927 rebalance_domains(this_rq, idle);
9928}
9929
9930
9931
9932
9933void trigger_load_balance(struct rq *rq)
9934{
9935
9936 if (unlikely(on_null_domain(rq)))
9937 return;
9938
9939 if (time_after_eq(jiffies, rq->next_balance))
9940 raise_softirq(SCHED_SOFTIRQ);
9941
9942 nohz_balancer_kick(rq);
9943}
9944
9945static void rq_online_fair(struct rq *rq)
9946{
9947 update_sysctl();
9948
9949 update_runtime_enabled(rq);
9950}
9951
9952static void rq_offline_fair(struct rq *rq)
9953{
9954 update_sysctl();
9955
9956
9957 unthrottle_offline_cfs_rqs(rq);
9958}
9959
9960#endif
9961
9962
9963
9964
9965
9966
9967
9968
9969
9970static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9971{
9972 struct cfs_rq *cfs_rq;
9973 struct sched_entity *se = &curr->se;
9974
9975 for_each_sched_entity(se) {
9976 cfs_rq = cfs_rq_of(se);
9977 entity_tick(cfs_rq, se, queued);
9978 }
9979
9980 if (static_branch_unlikely(&sched_numa_balancing))
9981 task_tick_numa(rq, curr);
9982
9983 update_misfit_status(curr, rq);
9984 update_overutilized_status(task_rq(curr));
9985}
9986
9987
9988
9989
9990
9991
9992static void task_fork_fair(struct task_struct *p)
9993{
9994 struct cfs_rq *cfs_rq;
9995 struct sched_entity *se = &p->se, *curr;
9996 struct rq *rq = this_rq();
9997 struct rq_flags rf;
9998
9999 rq_lock(rq, &rf);
10000 update_rq_clock(rq);
10001
10002 cfs_rq = task_cfs_rq(current);
10003 curr = cfs_rq->curr;
10004 if (curr) {
10005 update_curr(cfs_rq);
10006 se->vruntime = curr->vruntime;
10007 }
10008 place_entity(cfs_rq, se, 1);
10009
10010 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10011
10012
10013
10014
10015 swap(curr->vruntime, se->vruntime);
10016 resched_curr(rq);
10017 }
10018
10019 se->vruntime -= cfs_rq->min_vruntime;
10020 rq_unlock(rq, &rf);
10021}
10022
10023
10024
10025
10026
10027static void
10028prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10029{
10030 if (!task_on_rq_queued(p))
10031 return;
10032
10033
10034
10035
10036
10037
10038 if (rq->curr == p) {
10039 if (p->prio > oldprio)
10040 resched_curr(rq);
10041 } else
10042 check_preempt_curr(rq, p, 0);
10043}
10044
10045static inline bool vruntime_normalized(struct task_struct *p)
10046{
10047 struct sched_entity *se = &p->se;
10048
10049
10050
10051
10052
10053
10054 if (p->on_rq)
10055 return true;
10056
10057
10058
10059
10060
10061
10062
10063
10064
10065
10066 if (!se->sum_exec_runtime ||
10067 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10068 return true;
10069
10070 return false;
10071}
10072
10073#ifdef CONFIG_FAIR_GROUP_SCHED
10074
10075
10076
10077
10078static void propagate_entity_cfs_rq(struct sched_entity *se)
10079{
10080 struct cfs_rq *cfs_rq;
10081
10082
10083 se = se->parent;
10084
10085 for_each_sched_entity(se) {
10086 cfs_rq = cfs_rq_of(se);
10087
10088 if (cfs_rq_throttled(cfs_rq))
10089 break;
10090
10091 update_load_avg(cfs_rq, se, UPDATE_TG);
10092 }
10093}
10094#else
10095static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10096#endif
10097
10098static void detach_entity_cfs_rq(struct sched_entity *se)
10099{
10100 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10101
10102
10103 update_load_avg(cfs_rq, se, 0);
10104 detach_entity_load_avg(cfs_rq, se);
10105 update_tg_load_avg(cfs_rq, false);
10106 propagate_entity_cfs_rq(se);
10107}
10108
10109static void attach_entity_cfs_rq(struct sched_entity *se)
10110{
10111 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10112
10113#ifdef CONFIG_FAIR_GROUP_SCHED
10114
10115
10116
10117
10118 se->depth = se->parent ? se->parent->depth + 1 : 0;
10119#endif
10120
10121
10122 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10123 attach_entity_load_avg(cfs_rq, se, 0);
10124 update_tg_load_avg(cfs_rq, false);
10125 propagate_entity_cfs_rq(se);
10126}
10127
10128static void detach_task_cfs_rq(struct task_struct *p)
10129{
10130 struct sched_entity *se = &p->se;
10131 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10132
10133 if (!vruntime_normalized(p)) {
10134
10135
10136
10137
10138 place_entity(cfs_rq, se, 0);
10139 se->vruntime -= cfs_rq->min_vruntime;
10140 }
10141
10142 detach_entity_cfs_rq(se);
10143}
10144
10145static void attach_task_cfs_rq(struct task_struct *p)
10146{
10147 struct sched_entity *se = &p->se;
10148 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10149
10150 attach_entity_cfs_rq(se);
10151
10152 if (!vruntime_normalized(p))
10153 se->vruntime += cfs_rq->min_vruntime;
10154}
10155
10156static void switched_from_fair(struct rq *rq, struct task_struct *p)
10157{
10158 detach_task_cfs_rq(p);
10159}
10160
10161static void switched_to_fair(struct rq *rq, struct task_struct *p)
10162{
10163 attach_task_cfs_rq(p);
10164
10165 if (task_on_rq_queued(p)) {
10166
10167
10168
10169
10170
10171 if (rq->curr == p)
10172 resched_curr(rq);
10173 else
10174 check_preempt_curr(rq, p, 0);
10175 }
10176}
10177
10178
10179
10180
10181
10182
10183static void set_curr_task_fair(struct rq *rq)
10184{
10185 struct sched_entity *se = &rq->curr->se;
10186
10187 for_each_sched_entity(se) {
10188 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10189
10190 set_next_entity(cfs_rq, se);
10191
10192 account_cfs_rq_runtime(cfs_rq, 0);
10193 }
10194}
10195
10196void init_cfs_rq(struct cfs_rq *cfs_rq)
10197{
10198 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10199 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10200#ifndef CONFIG_64BIT
10201 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10202#endif
10203#ifdef CONFIG_SMP
10204 raw_spin_lock_init(&cfs_rq->removed.lock);
10205#endif
10206}
10207
10208#ifdef CONFIG_FAIR_GROUP_SCHED
10209static void task_set_group_fair(struct task_struct *p)
10210{
10211 struct sched_entity *se = &p->se;
10212
10213 set_task_rq(p, task_cpu(p));
10214 se->depth = se->parent ? se->parent->depth + 1 : 0;
10215}
10216
10217static void task_move_group_fair(struct task_struct *p)
10218{
10219 detach_task_cfs_rq(p);
10220 set_task_rq(p, task_cpu(p));
10221
10222#ifdef CONFIG_SMP
10223
10224 p->se.avg.last_update_time = 0;
10225#endif
10226 attach_task_cfs_rq(p);
10227}
10228
10229static void task_change_group_fair(struct task_struct *p, int type)
10230{
10231 switch (type) {
10232 case TASK_SET_GROUP:
10233 task_set_group_fair(p);
10234 break;
10235
10236 case TASK_MOVE_GROUP:
10237 task_move_group_fair(p);
10238 break;
10239 }
10240}
10241
10242void free_fair_sched_group(struct task_group *tg)
10243{
10244 int i;
10245
10246 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10247
10248 for_each_possible_cpu(i) {
10249 if (tg->cfs_rq)
10250 kfree(tg->cfs_rq[i]);
10251 if (tg->se)
10252 kfree(tg->se[i]);
10253 }
10254
10255 kfree(tg->cfs_rq);
10256 kfree(tg->se);
10257}
10258
10259int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10260{
10261 struct sched_entity *se;
10262 struct cfs_rq *cfs_rq;
10263 int i;
10264
10265 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10266 if (!tg->cfs_rq)
10267 goto err;
10268 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10269 if (!tg->se)
10270 goto err;
10271
10272 tg->shares = NICE_0_LOAD;
10273
10274 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10275
10276 for_each_possible_cpu(i) {
10277 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10278 GFP_KERNEL, cpu_to_node(i));
10279 if (!cfs_rq)
10280 goto err;
10281
10282 se = kzalloc_node(sizeof(struct sched_entity),
10283 GFP_KERNEL, cpu_to_node(i));
10284 if (!se)
10285 goto err_free_rq;
10286
10287 init_cfs_rq(cfs_rq);
10288 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10289 init_entity_runnable_average(se);
10290 }
10291
10292 return 1;
10293
10294err_free_rq:
10295 kfree(cfs_rq);
10296err:
10297 return 0;
10298}
10299
10300void online_fair_sched_group(struct task_group *tg)
10301{
10302 struct sched_entity *se;
10303 struct rq *rq;
10304 int i;
10305
10306 for_each_possible_cpu(i) {
10307 rq = cpu_rq(i);
10308 se = tg->se[i];
10309
10310 raw_spin_lock_irq(&rq->lock);
10311 update_rq_clock(rq);
10312 attach_entity_cfs_rq(se);
10313 sync_throttle(tg, i);
10314 raw_spin_unlock_irq(&rq->lock);
10315 }
10316}
10317
10318void unregister_fair_sched_group(struct task_group *tg)
10319{
10320 unsigned long flags;
10321 struct rq *rq;
10322 int cpu;
10323
10324 for_each_possible_cpu(cpu) {
10325 if (tg->se[cpu])
10326 remove_entity_load_avg(tg->se[cpu]);
10327
10328
10329
10330
10331
10332 if (!tg->cfs_rq[cpu]->on_list)
10333 continue;
10334
10335 rq = cpu_rq(cpu);
10336
10337 raw_spin_lock_irqsave(&rq->lock, flags);
10338 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10339 raw_spin_unlock_irqrestore(&rq->lock, flags);
10340 }
10341}
10342
10343void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10344 struct sched_entity *se, int cpu,
10345 struct sched_entity *parent)
10346{
10347 struct rq *rq = cpu_rq(cpu);
10348
10349 cfs_rq->tg = tg;
10350 cfs_rq->rq = rq;
10351 init_cfs_rq_runtime(cfs_rq);
10352
10353 tg->cfs_rq[cpu] = cfs_rq;
10354 tg->se[cpu] = se;
10355
10356
10357 if (!se)
10358 return;
10359
10360 if (!parent) {
10361 se->cfs_rq = &rq->cfs;
10362 se->depth = 0;
10363 } else {
10364 se->cfs_rq = parent->my_q;
10365 se->depth = parent->depth + 1;
10366 }
10367
10368 se->my_q = cfs_rq;
10369
10370 update_load_set(&se->load, NICE_0_LOAD);
10371 se->parent = parent;
10372}
10373
10374static DEFINE_MUTEX(shares_mutex);
10375
10376int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10377{
10378 int i;
10379
10380
10381
10382
10383 if (!tg->se[0])
10384 return -EINVAL;
10385
10386 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10387
10388 mutex_lock(&shares_mutex);
10389 if (tg->shares == shares)
10390 goto done;
10391
10392 tg->shares = shares;
10393 for_each_possible_cpu(i) {
10394 struct rq *rq = cpu_rq(i);
10395 struct sched_entity *se = tg->se[i];
10396 struct rq_flags rf;
10397
10398
10399 rq_lock_irqsave(rq, &rf);
10400 update_rq_clock(rq);
10401 for_each_sched_entity(se) {
10402 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10403 update_cfs_group(se);
10404 }
10405 rq_unlock_irqrestore(rq, &rf);
10406 }
10407
10408done:
10409 mutex_unlock(&shares_mutex);
10410 return 0;
10411}
10412#else
10413
10414void free_fair_sched_group(struct task_group *tg) { }
10415
10416int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10417{
10418 return 1;
10419}
10420
10421void online_fair_sched_group(struct task_group *tg) { }
10422
10423void unregister_fair_sched_group(struct task_group *tg) { }
10424
10425#endif
10426
10427
10428static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10429{
10430 struct sched_entity *se = &task->se;
10431 unsigned int rr_interval = 0;
10432
10433
10434
10435
10436
10437 if (rq->cfs.load.weight)
10438 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10439
10440 return rr_interval;
10441}
10442
10443
10444
10445
10446const struct sched_class fair_sched_class = {
10447 .next = &idle_sched_class,
10448 .enqueue_task = enqueue_task_fair,
10449 .dequeue_task = dequeue_task_fair,
10450 .yield_task = yield_task_fair,
10451 .yield_to_task = yield_to_task_fair,
10452
10453 .check_preempt_curr = check_preempt_wakeup,
10454
10455 .pick_next_task = pick_next_task_fair,
10456 .put_prev_task = put_prev_task_fair,
10457
10458#ifdef CONFIG_SMP
10459 .select_task_rq = select_task_rq_fair,
10460 .migrate_task_rq = migrate_task_rq_fair,
10461
10462 .rq_online = rq_online_fair,
10463 .rq_offline = rq_offline_fair,
10464
10465 .task_dead = task_dead_fair,
10466 .set_cpus_allowed = set_cpus_allowed_common,
10467#endif
10468
10469 .set_curr_task = set_curr_task_fair,
10470 .task_tick = task_tick_fair,
10471 .task_fork = task_fork_fair,
10472
10473 .prio_changed = prio_changed_fair,
10474 .switched_from = switched_from_fair,
10475 .switched_to = switched_to_fair,
10476
10477 .get_rr_interval = get_rr_interval_fair,
10478
10479 .update_curr = update_curr_fair,
10480
10481#ifdef CONFIG_FAIR_GROUP_SCHED
10482 .task_change_group = task_change_group_fair,
10483#endif
10484
10485#ifdef CONFIG_UCLAMP_TASK
10486 .uclamp_enabled = 1,
10487#endif
10488};
10489
10490#ifdef CONFIG_SCHED_DEBUG
10491void print_cfs_stats(struct seq_file *m, int cpu)
10492{
10493 struct cfs_rq *cfs_rq, *pos;
10494
10495 rcu_read_lock();
10496 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10497 print_cfs_rq(m, cpu, cfs_rq);
10498 rcu_read_unlock();
10499}
10500
10501#ifdef CONFIG_NUMA_BALANCING
10502void show_numa_stats(struct task_struct *p, struct seq_file *m)
10503{
10504 int node;
10505 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10506 struct numa_group *ng;
10507
10508 rcu_read_lock();
10509 ng = rcu_dereference(p->numa_group);
10510 for_each_online_node(node) {
10511 if (p->numa_faults) {
10512 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10513 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10514 }
10515 if (ng) {
10516 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10517 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10518 }
10519 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10520 }
10521 rcu_read_unlock();
10522}
10523#endif
10524#endif
10525
10526__init void init_sched_fair_class(void)
10527{
10528#ifdef CONFIG_SMP
10529 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10530
10531#ifdef CONFIG_NO_HZ_COMMON
10532 nohz.next_balance = jiffies;
10533 nohz.next_blocked = jiffies;
10534 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
10535#endif
10536#endif
10537
10538}
10539
10540
10541
10542
10543
10544const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
10545{
10546#ifdef CONFIG_SMP
10547 return cfs_rq ? &cfs_rq->avg : NULL;
10548#else
10549 return NULL;
10550#endif
10551}
10552EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
10553
10554char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
10555{
10556 if (!cfs_rq) {
10557 if (str)
10558 strlcpy(str, "(null)", len);
10559 else
10560 return NULL;
10561 }
10562
10563 cfs_rq_tg_path(cfs_rq, str, len);
10564 return str;
10565}
10566EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
10567
10568int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
10569{
10570 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
10571}
10572EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
10573
10574const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
10575{
10576#ifdef CONFIG_SMP
10577 return rq ? &rq->avg_rt : NULL;
10578#else
10579 return NULL;
10580#endif
10581}
10582EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
10583
10584const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
10585{
10586#ifdef CONFIG_SMP
10587 return rq ? &rq->avg_dl : NULL;
10588#else
10589 return NULL;
10590#endif
10591}
10592EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
10593
10594const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
10595{
10596#if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
10597 return rq ? &rq->avg_irq : NULL;
10598#else
10599 return NULL;
10600#endif
10601}
10602EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
10603
10604int sched_trace_rq_cpu(struct rq *rq)
10605{
10606 return rq ? cpu_of(rq) : -1;
10607}
10608EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
10609
10610const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
10611{
10612#ifdef CONFIG_SMP
10613 return rd ? rd->span : NULL;
10614#else
10615 return NULL;
10616#endif
10617}
10618EXPORT_SYMBOL_GPL(sched_trace_rd_span);
10619