Version:  2.0.40 2.2.26 2.4.37 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10

Linux/kernel/sched/fair.c

  1 /*
  2  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
  3  *
  4  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  5  *
  6  *  Interactivity improvements by Mike Galbraith
  7  *  (C) 2007 Mike Galbraith <efault@gmx.de>
  8  *
  9  *  Various enhancements by Dmitry Adamushko.
 10  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
 11  *
 12  *  Group scheduling enhancements by Srivatsa Vaddagiri
 13  *  Copyright IBM Corporation, 2007
 14  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
 15  *
 16  *  Scaled math optimizations by Thomas Gleixner
 17  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
 18  *
 19  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
 20  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
 21  */
 22 
 23 #include <linux/sched.h>
 24 #include <linux/latencytop.h>
 25 #include <linux/cpumask.h>
 26 #include <linux/cpuidle.h>
 27 #include <linux/slab.h>
 28 #include <linux/profile.h>
 29 #include <linux/interrupt.h>
 30 #include <linux/mempolicy.h>
 31 #include <linux/migrate.h>
 32 #include <linux/task_work.h>
 33 
 34 #include <trace/events/sched.h>
 35 
 36 #include "sched.h"
 37 
 38 /*
 39  * Targeted preemption latency for CPU-bound tasks:
 40  *
 41  * NOTE: this latency value is not the same as the concept of
 42  * 'timeslice length' - timeslices in CFS are of variable length
 43  * and have no persistent notion like in traditional, time-slice
 44  * based scheduling concepts.
 45  *
 46  * (to see the precise effective timeslice length of your workload,
 47  *  run vmstat and monitor the context-switches (cs) field)
 48  *
 49  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
 50  */
 51 unsigned int sysctl_sched_latency                       = 6000000ULL;
 52 unsigned int normalized_sysctl_sched_latency            = 6000000ULL;
 53 
 54 /*
 55  * The initial- and re-scaling of tunables is configurable
 56  *
 57  * Options are:
 58  *
 59  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
 60  *   SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
 61  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
 62  *
 63  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
 64  */
 65 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
 66 
 67 /*
 68  * Minimal preemption granularity for CPU-bound tasks:
 69  *
 70  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
 71  */
 72 unsigned int sysctl_sched_min_granularity               = 750000ULL;
 73 unsigned int normalized_sysctl_sched_min_granularity    = 750000ULL;
 74 
 75 /*
 76  * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
 77  */
 78 static unsigned int sched_nr_latency = 8;
 79 
 80 /*
 81  * After fork, child runs first. If set to 0 (default) then
 82  * parent will (try to) run first.
 83  */
 84 unsigned int sysctl_sched_child_runs_first __read_mostly;
 85 
 86 /*
 87  * SCHED_OTHER wake-up granularity.
 88  *
 89  * This option delays the preemption effects of decoupled workloads
 90  * and reduces their over-scheduling. Synchronous workloads will still
 91  * have immediate wakeup/sleep latencies.
 92  *
 93  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
 94  */
 95 unsigned int sysctl_sched_wakeup_granularity            = 1000000UL;
 96 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
 97 
 98 const_debug unsigned int sysctl_sched_migration_cost    = 500000UL;
 99 
100 #ifdef CONFIG_SMP
101 /*
102  * For asym packing, by default the lower numbered cpu has higher priority.
103  */
104 int __weak arch_asym_cpu_priority(int cpu)
105 {
106         return -cpu;
107 }
108 #endif
109 
110 #ifdef CONFIG_CFS_BANDWIDTH
111 /*
112  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
113  * each time a cfs_rq requests quota.
114  *
115  * Note: in the case that the slice exceeds the runtime remaining (either due
116  * to consumption or the quota being specified to be smaller than the slice)
117  * we will always only issue the remaining available time.
118  *
119  * (default: 5 msec, units: microseconds)
120  */
121 unsigned int sysctl_sched_cfs_bandwidth_slice           = 5000UL;
122 #endif
123 
124 /*
125  * The margin used when comparing utilization with CPU capacity:
126  * util * margin < capacity * 1024
127  *
128  * (default: ~20%)
129  */
130 unsigned int capacity_margin                            = 1280;
131 
132 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
133 {
134         lw->weight += inc;
135         lw->inv_weight = 0;
136 }
137 
138 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
139 {
140         lw->weight -= dec;
141         lw->inv_weight = 0;
142 }
143 
144 static inline void update_load_set(struct load_weight *lw, unsigned long w)
145 {
146         lw->weight = w;
147         lw->inv_weight = 0;
148 }
149 
150 /*
151  * Increase the granularity value when there are more CPUs,
152  * because with more CPUs the 'effective latency' as visible
153  * to users decreases. But the relationship is not linear,
154  * so pick a second-best guess by going with the log2 of the
155  * number of CPUs.
156  *
157  * This idea comes from the SD scheduler of Con Kolivas:
158  */
159 static unsigned int get_update_sysctl_factor(void)
160 {
161         unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
162         unsigned int factor;
163 
164         switch (sysctl_sched_tunable_scaling) {
165         case SCHED_TUNABLESCALING_NONE:
166                 factor = 1;
167                 break;
168         case SCHED_TUNABLESCALING_LINEAR:
169                 factor = cpus;
170                 break;
171         case SCHED_TUNABLESCALING_LOG:
172         default:
173                 factor = 1 + ilog2(cpus);
174                 break;
175         }
176 
177         return factor;
178 }
179 
180 static void update_sysctl(void)
181 {
182         unsigned int factor = get_update_sysctl_factor();
183 
184 #define SET_SYSCTL(name) \
185         (sysctl_##name = (factor) * normalized_sysctl_##name)
186         SET_SYSCTL(sched_min_granularity);
187         SET_SYSCTL(sched_latency);
188         SET_SYSCTL(sched_wakeup_granularity);
189 #undef SET_SYSCTL
190 }
191 
192 void sched_init_granularity(void)
193 {
194         update_sysctl();
195 }
196 
197 #define WMULT_CONST     (~0U)
198 #define WMULT_SHIFT     32
199 
200 static void __update_inv_weight(struct load_weight *lw)
201 {
202         unsigned long w;
203 
204         if (likely(lw->inv_weight))
205                 return;
206 
207         w = scale_load_down(lw->weight);
208 
209         if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
210                 lw->inv_weight = 1;
211         else if (unlikely(!w))
212                 lw->inv_weight = WMULT_CONST;
213         else
214                 lw->inv_weight = WMULT_CONST / w;
215 }
216 
217 /*
218  * delta_exec * weight / lw.weight
219  *   OR
220  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
221  *
222  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
223  * we're guaranteed shift stays positive because inv_weight is guaranteed to
224  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
225  *
226  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
227  * weight/lw.weight <= 1, and therefore our shift will also be positive.
228  */
229 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
230 {
231         u64 fact = scale_load_down(weight);
232         int shift = WMULT_SHIFT;
233 
234         __update_inv_weight(lw);
235 
236         if (unlikely(fact >> 32)) {
237                 while (fact >> 32) {
238                         fact >>= 1;
239                         shift--;
240                 }
241         }
242 
243         /* hint to use a 32x32->64 mul */
244         fact = (u64)(u32)fact * lw->inv_weight;
245 
246         while (fact >> 32) {
247                 fact >>= 1;
248                 shift--;
249         }
250 
251         return mul_u64_u32_shr(delta_exec, fact, shift);
252 }
253 
254 
255 const struct sched_class fair_sched_class;
256 
257 /**************************************************************
258  * CFS operations on generic schedulable entities:
259  */
260 
261 #ifdef CONFIG_FAIR_GROUP_SCHED
262 
263 /* cpu runqueue to which this cfs_rq is attached */
264 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
265 {
266         return cfs_rq->rq;
267 }
268 
269 /* An entity is a task if it doesn't "own" a runqueue */
270 #define entity_is_task(se)      (!se->my_q)
271 
272 static inline struct task_struct *task_of(struct sched_entity *se)
273 {
274         SCHED_WARN_ON(!entity_is_task(se));
275         return container_of(se, struct task_struct, se);
276 }
277 
278 /* Walk up scheduling entities hierarchy */
279 #define for_each_sched_entity(se) \
280                 for (; se; se = se->parent)
281 
282 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
283 {
284         return p->se.cfs_rq;
285 }
286 
287 /* runqueue on which this entity is (to be) queued */
288 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
289 {
290         return se->cfs_rq;
291 }
292 
293 /* runqueue "owned" by this group */
294 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
295 {
296         return grp->my_q;
297 }
298 
299 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
300 {
301         if (!cfs_rq->on_list) {
302                 struct rq *rq = rq_of(cfs_rq);
303                 int cpu = cpu_of(rq);
304                 /*
305                  * Ensure we either appear before our parent (if already
306                  * enqueued) or force our parent to appear after us when it is
307                  * enqueued. The fact that we always enqueue bottom-up
308                  * reduces this to two cases and a special case for the root
309                  * cfs_rq. Furthermore, it also means that we will always reset
310                  * tmp_alone_branch either when the branch is connected
311                  * to a tree or when we reach the beg of the tree
312                  */
313                 if (cfs_rq->tg->parent &&
314                     cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
315                         /*
316                          * If parent is already on the list, we add the child
317                          * just before. Thanks to circular linked property of
318                          * the list, this means to put the child at the tail
319                          * of the list that starts by parent.
320                          */
321                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
322                                 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
323                         /*
324                          * The branch is now connected to its tree so we can
325                          * reset tmp_alone_branch to the beginning of the
326                          * list.
327                          */
328                         rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
329                 } else if (!cfs_rq->tg->parent) {
330                         /*
331                          * cfs rq without parent should be put
332                          * at the tail of the list.
333                          */
334                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
335                                 &rq->leaf_cfs_rq_list);
336                         /*
337                          * We have reach the beg of a tree so we can reset
338                          * tmp_alone_branch to the beginning of the list.
339                          */
340                         rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
341                 } else {
342                         /*
343                          * The parent has not already been added so we want to
344                          * make sure that it will be put after us.
345                          * tmp_alone_branch points to the beg of the branch
346                          * where we will add parent.
347                          */
348                         list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
349                                 rq->tmp_alone_branch);
350                         /*
351                          * update tmp_alone_branch to points to the new beg
352                          * of the branch
353                          */
354                         rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
355                 }
356 
357                 cfs_rq->on_list = 1;
358         }
359 }
360 
361 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
362 {
363         if (cfs_rq->on_list) {
364                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
365                 cfs_rq->on_list = 0;
366         }
367 }
368 
369 /* Iterate thr' all leaf cfs_rq's on a runqueue */
370 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
371         list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
372 
373 /* Do the two (enqueued) entities belong to the same group ? */
374 static inline struct cfs_rq *
375 is_same_group(struct sched_entity *se, struct sched_entity *pse)
376 {
377         if (se->cfs_rq == pse->cfs_rq)
378                 return se->cfs_rq;
379 
380         return NULL;
381 }
382 
383 static inline struct sched_entity *parent_entity(struct sched_entity *se)
384 {
385         return se->parent;
386 }
387 
388 static void
389 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
390 {
391         int se_depth, pse_depth;
392 
393         /*
394          * preemption test can be made between sibling entities who are in the
395          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
396          * both tasks until we find their ancestors who are siblings of common
397          * parent.
398          */
399 
400         /* First walk up until both entities are at same depth */
401         se_depth = (*se)->depth;
402         pse_depth = (*pse)->depth;
403 
404         while (se_depth > pse_depth) {
405                 se_depth--;
406                 *se = parent_entity(*se);
407         }
408 
409         while (pse_depth > se_depth) {
410                 pse_depth--;
411                 *pse = parent_entity(*pse);
412         }
413 
414         while (!is_same_group(*se, *pse)) {
415                 *se = parent_entity(*se);
416                 *pse = parent_entity(*pse);
417         }
418 }
419 
420 #else   /* !CONFIG_FAIR_GROUP_SCHED */
421 
422 static inline struct task_struct *task_of(struct sched_entity *se)
423 {
424         return container_of(se, struct task_struct, se);
425 }
426 
427 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
428 {
429         return container_of(cfs_rq, struct rq, cfs);
430 }
431 
432 #define entity_is_task(se)      1
433 
434 #define for_each_sched_entity(se) \
435                 for (; se; se = NULL)
436 
437 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
438 {
439         return &task_rq(p)->cfs;
440 }
441 
442 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
443 {
444         struct task_struct *p = task_of(se);
445         struct rq *rq = task_rq(p);
446 
447         return &rq->cfs;
448 }
449 
450 /* runqueue "owned" by this group */
451 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
452 {
453         return NULL;
454 }
455 
456 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
457 {
458 }
459 
460 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
461 {
462 }
463 
464 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
465                 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
466 
467 static inline struct sched_entity *parent_entity(struct sched_entity *se)
468 {
469         return NULL;
470 }
471 
472 static inline void
473 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
474 {
475 }
476 
477 #endif  /* CONFIG_FAIR_GROUP_SCHED */
478 
479 static __always_inline
480 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
481 
482 /**************************************************************
483  * Scheduling class tree data structure manipulation methods:
484  */
485 
486 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
487 {
488         s64 delta = (s64)(vruntime - max_vruntime);
489         if (delta > 0)
490                 max_vruntime = vruntime;
491 
492         return max_vruntime;
493 }
494 
495 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
496 {
497         s64 delta = (s64)(vruntime - min_vruntime);
498         if (delta < 0)
499                 min_vruntime = vruntime;
500 
501         return min_vruntime;
502 }
503 
504 static inline int entity_before(struct sched_entity *a,
505                                 struct sched_entity *b)
506 {
507         return (s64)(a->vruntime - b->vruntime) < 0;
508 }
509 
510 static void update_min_vruntime(struct cfs_rq *cfs_rq)
511 {
512         struct sched_entity *curr = cfs_rq->curr;
513 
514         u64 vruntime = cfs_rq->min_vruntime;
515 
516         if (curr) {
517                 if (curr->on_rq)
518                         vruntime = curr->vruntime;
519                 else
520                         curr = NULL;
521         }
522 
523         if (cfs_rq->rb_leftmost) {
524                 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
525                                                    struct sched_entity,
526                                                    run_node);
527 
528                 if (!curr)
529                         vruntime = se->vruntime;
530                 else
531                         vruntime = min_vruntime(vruntime, se->vruntime);
532         }
533 
534         /* ensure we never gain time by being placed backwards. */
535         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
536 #ifndef CONFIG_64BIT
537         smp_wmb();
538         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
539 #endif
540 }
541 
542 /*
543  * Enqueue an entity into the rb-tree:
544  */
545 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
546 {
547         struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
548         struct rb_node *parent = NULL;
549         struct sched_entity *entry;
550         int leftmost = 1;
551 
552         /*
553          * Find the right place in the rbtree:
554          */
555         while (*link) {
556                 parent = *link;
557                 entry = rb_entry(parent, struct sched_entity, run_node);
558                 /*
559                  * We dont care about collisions. Nodes with
560                  * the same key stay together.
561                  */
562                 if (entity_before(se, entry)) {
563                         link = &parent->rb_left;
564                 } else {
565                         link = &parent->rb_right;
566                         leftmost = 0;
567                 }
568         }
569 
570         /*
571          * Maintain a cache of leftmost tree entries (it is frequently
572          * used):
573          */
574         if (leftmost)
575                 cfs_rq->rb_leftmost = &se->run_node;
576 
577         rb_link_node(&se->run_node, parent, link);
578         rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
579 }
580 
581 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
582 {
583         if (cfs_rq->rb_leftmost == &se->run_node) {
584                 struct rb_node *next_node;
585 
586                 next_node = rb_next(&se->run_node);
587                 cfs_rq->rb_leftmost = next_node;
588         }
589 
590         rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
591 }
592 
593 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
594 {
595         struct rb_node *left = cfs_rq->rb_leftmost;
596 
597         if (!left)
598                 return NULL;
599 
600         return rb_entry(left, struct sched_entity, run_node);
601 }
602 
603 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
604 {
605         struct rb_node *next = rb_next(&se->run_node);
606 
607         if (!next)
608                 return NULL;
609 
610         return rb_entry(next, struct sched_entity, run_node);
611 }
612 
613 #ifdef CONFIG_SCHED_DEBUG
614 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
615 {
616         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
617 
618         if (!last)
619                 return NULL;
620 
621         return rb_entry(last, struct sched_entity, run_node);
622 }
623 
624 /**************************************************************
625  * Scheduling class statistics methods:
626  */
627 
628 int sched_proc_update_handler(struct ctl_table *table, int write,
629                 void __user *buffer, size_t *lenp,
630                 loff_t *ppos)
631 {
632         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
633         unsigned int factor = get_update_sysctl_factor();
634 
635         if (ret || !write)
636                 return ret;
637 
638         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
639                                         sysctl_sched_min_granularity);
640 
641 #define WRT_SYSCTL(name) \
642         (normalized_sysctl_##name = sysctl_##name / (factor))
643         WRT_SYSCTL(sched_min_granularity);
644         WRT_SYSCTL(sched_latency);
645         WRT_SYSCTL(sched_wakeup_granularity);
646 #undef WRT_SYSCTL
647 
648         return 0;
649 }
650 #endif
651 
652 /*
653  * delta /= w
654  */
655 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
656 {
657         if (unlikely(se->load.weight != NICE_0_LOAD))
658                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
659 
660         return delta;
661 }
662 
663 /*
664  * The idea is to set a period in which each task runs once.
665  *
666  * When there are too many tasks (sched_nr_latency) we have to stretch
667  * this period because otherwise the slices get too small.
668  *
669  * p = (nr <= nl) ? l : l*nr/nl
670  */
671 static u64 __sched_period(unsigned long nr_running)
672 {
673         if (unlikely(nr_running > sched_nr_latency))
674                 return nr_running * sysctl_sched_min_granularity;
675         else
676                 return sysctl_sched_latency;
677 }
678 
679 /*
680  * We calculate the wall-time slice from the period by taking a part
681  * proportional to the weight.
682  *
683  * s = p*P[w/rw]
684  */
685 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
686 {
687         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
688 
689         for_each_sched_entity(se) {
690                 struct load_weight *load;
691                 struct load_weight lw;
692 
693                 cfs_rq = cfs_rq_of(se);
694                 load = &cfs_rq->load;
695 
696                 if (unlikely(!se->on_rq)) {
697                         lw = cfs_rq->load;
698 
699                         update_load_add(&lw, se->load.weight);
700                         load = &lw;
701                 }
702                 slice = __calc_delta(slice, se->load.weight, load);
703         }
704         return slice;
705 }
706 
707 /*
708  * We calculate the vruntime slice of a to-be-inserted task.
709  *
710  * vs = s/w
711  */
712 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
713 {
714         return calc_delta_fair(sched_slice(cfs_rq, se), se);
715 }
716 
717 #ifdef CONFIG_SMP
718 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
719 static unsigned long task_h_load(struct task_struct *p);
720 
721 /*
722  * We choose a half-life close to 1 scheduling period.
723  * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
724  * dependent on this value.
725  */
726 #define LOAD_AVG_PERIOD 32
727 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
728 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
729 
730 /* Give new sched_entity start runnable values to heavy its load in infant time */
731 void init_entity_runnable_average(struct sched_entity *se)
732 {
733         struct sched_avg *sa = &se->avg;
734 
735         sa->last_update_time = 0;
736         /*
737          * sched_avg's period_contrib should be strictly less then 1024, so
738          * we give it 1023 to make sure it is almost a period (1024us), and
739          * will definitely be update (after enqueue).
740          */
741         sa->period_contrib = 1023;
742         /*
743          * Tasks are intialized with full load to be seen as heavy tasks until
744          * they get a chance to stabilize to their real load level.
745          * Group entities are intialized with zero load to reflect the fact that
746          * nothing has been attached to the task group yet.
747          */
748         if (entity_is_task(se))
749                 sa->load_avg = scale_load_down(se->load.weight);
750         sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
751         /*
752          * At this point, util_avg won't be used in select_task_rq_fair anyway
753          */
754         sa->util_avg = 0;
755         sa->util_sum = 0;
756         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
757 }
758 
759 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
760 static void attach_entity_cfs_rq(struct sched_entity *se);
761 
762 /*
763  * With new tasks being created, their initial util_avgs are extrapolated
764  * based on the cfs_rq's current util_avg:
765  *
766  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
767  *
768  * However, in many cases, the above util_avg does not give a desired
769  * value. Moreover, the sum of the util_avgs may be divergent, such
770  * as when the series is a harmonic series.
771  *
772  * To solve this problem, we also cap the util_avg of successive tasks to
773  * only 1/2 of the left utilization budget:
774  *
775  *   util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
776  *
777  * where n denotes the nth task.
778  *
779  * For example, a simplest series from the beginning would be like:
780  *
781  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
782  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
783  *
784  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
785  * if util_avg > util_avg_cap.
786  */
787 void post_init_entity_util_avg(struct sched_entity *se)
788 {
789         struct cfs_rq *cfs_rq = cfs_rq_of(se);
790         struct sched_avg *sa = &se->avg;
791         long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
792 
793         if (cap > 0) {
794                 if (cfs_rq->avg.util_avg != 0) {
795                         sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
796                         sa->util_avg /= (cfs_rq->avg.load_avg + 1);
797 
798                         if (sa->util_avg > cap)
799                                 sa->util_avg = cap;
800                 } else {
801                         sa->util_avg = cap;
802                 }
803                 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
804         }
805 
806         if (entity_is_task(se)) {
807                 struct task_struct *p = task_of(se);
808                 if (p->sched_class != &fair_sched_class) {
809                         /*
810                          * For !fair tasks do:
811                          *
812                         update_cfs_rq_load_avg(now, cfs_rq, false);
813                         attach_entity_load_avg(cfs_rq, se);
814                         switched_from_fair(rq, p);
815                          *
816                          * such that the next switched_to_fair() has the
817                          * expected state.
818                          */
819                         se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
820                         return;
821                 }
822         }
823 
824         attach_entity_cfs_rq(se);
825 }
826 
827 #else /* !CONFIG_SMP */
828 void init_entity_runnable_average(struct sched_entity *se)
829 {
830 }
831 void post_init_entity_util_avg(struct sched_entity *se)
832 {
833 }
834 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
835 {
836 }
837 #endif /* CONFIG_SMP */
838 
839 /*
840  * Update the current task's runtime statistics.
841  */
842 static void update_curr(struct cfs_rq *cfs_rq)
843 {
844         struct sched_entity *curr = cfs_rq->curr;
845         u64 now = rq_clock_task(rq_of(cfs_rq));
846         u64 delta_exec;
847 
848         if (unlikely(!curr))
849                 return;
850 
851         delta_exec = now - curr->exec_start;
852         if (unlikely((s64)delta_exec <= 0))
853                 return;
854 
855         curr->exec_start = now;
856 
857         schedstat_set(curr->statistics.exec_max,
858                       max(delta_exec, curr->statistics.exec_max));
859 
860         curr->sum_exec_runtime += delta_exec;
861         schedstat_add(cfs_rq->exec_clock, delta_exec);
862 
863         curr->vruntime += calc_delta_fair(delta_exec, curr);
864         update_min_vruntime(cfs_rq);
865 
866         if (entity_is_task(curr)) {
867                 struct task_struct *curtask = task_of(curr);
868 
869                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
870                 cpuacct_charge(curtask, delta_exec);
871                 account_group_exec_runtime(curtask, delta_exec);
872         }
873 
874         account_cfs_rq_runtime(cfs_rq, delta_exec);
875 }
876 
877 static void update_curr_fair(struct rq *rq)
878 {
879         update_curr(cfs_rq_of(&rq->curr->se));
880 }
881 
882 static inline void
883 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
884 {
885         u64 wait_start, prev_wait_start;
886 
887         if (!schedstat_enabled())
888                 return;
889 
890         wait_start = rq_clock(rq_of(cfs_rq));
891         prev_wait_start = schedstat_val(se->statistics.wait_start);
892 
893         if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
894             likely(wait_start > prev_wait_start))
895                 wait_start -= prev_wait_start;
896 
897         schedstat_set(se->statistics.wait_start, wait_start);
898 }
899 
900 static inline void
901 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
902 {
903         struct task_struct *p;
904         u64 delta;
905 
906         if (!schedstat_enabled())
907                 return;
908 
909         delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
910 
911         if (entity_is_task(se)) {
912                 p = task_of(se);
913                 if (task_on_rq_migrating(p)) {
914                         /*
915                          * Preserve migrating task's wait time so wait_start
916                          * time stamp can be adjusted to accumulate wait time
917                          * prior to migration.
918                          */
919                         schedstat_set(se->statistics.wait_start, delta);
920                         return;
921                 }
922                 trace_sched_stat_wait(p, delta);
923         }
924 
925         schedstat_set(se->statistics.wait_max,
926                       max(schedstat_val(se->statistics.wait_max), delta));
927         schedstat_inc(se->statistics.wait_count);
928         schedstat_add(se->statistics.wait_sum, delta);
929         schedstat_set(se->statistics.wait_start, 0);
930 }
931 
932 static inline void
933 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
934 {
935         struct task_struct *tsk = NULL;
936         u64 sleep_start, block_start;
937 
938         if (!schedstat_enabled())
939                 return;
940 
941         sleep_start = schedstat_val(se->statistics.sleep_start);
942         block_start = schedstat_val(se->statistics.block_start);
943 
944         if (entity_is_task(se))
945                 tsk = task_of(se);
946 
947         if (sleep_start) {
948                 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
949 
950                 if ((s64)delta < 0)
951                         delta = 0;
952 
953                 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
954                         schedstat_set(se->statistics.sleep_max, delta);
955 
956                 schedstat_set(se->statistics.sleep_start, 0);
957                 schedstat_add(se->statistics.sum_sleep_runtime, delta);
958 
959                 if (tsk) {
960                         account_scheduler_latency(tsk, delta >> 10, 1);
961                         trace_sched_stat_sleep(tsk, delta);
962                 }
963         }
964         if (block_start) {
965                 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
966 
967                 if ((s64)delta < 0)
968                         delta = 0;
969 
970                 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
971                         schedstat_set(se->statistics.block_max, delta);
972 
973                 schedstat_set(se->statistics.block_start, 0);
974                 schedstat_add(se->statistics.sum_sleep_runtime, delta);
975 
976                 if (tsk) {
977                         if (tsk->in_iowait) {
978                                 schedstat_add(se->statistics.iowait_sum, delta);
979                                 schedstat_inc(se->statistics.iowait_count);
980                                 trace_sched_stat_iowait(tsk, delta);
981                         }
982 
983                         trace_sched_stat_blocked(tsk, delta);
984 
985                         /*
986                          * Blocking time is in units of nanosecs, so shift by
987                          * 20 to get a milliseconds-range estimation of the
988                          * amount of time that the task spent sleeping:
989                          */
990                         if (unlikely(prof_on == SLEEP_PROFILING)) {
991                                 profile_hits(SLEEP_PROFILING,
992                                                 (void *)get_wchan(tsk),
993                                                 delta >> 20);
994                         }
995                         account_scheduler_latency(tsk, delta >> 10, 0);
996                 }
997         }
998 }
999 
1000 /*
1001  * Task is being enqueued - update stats:
1002  */
1003 static inline void
1004 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1005 {
1006         if (!schedstat_enabled())
1007                 return;
1008 
1009         /*
1010          * Are we enqueueing a waiting task? (for current tasks
1011          * a dequeue/enqueue event is a NOP)
1012          */
1013         if (se != cfs_rq->curr)
1014                 update_stats_wait_start(cfs_rq, se);
1015 
1016         if (flags & ENQUEUE_WAKEUP)
1017                 update_stats_enqueue_sleeper(cfs_rq, se);
1018 }
1019 
1020 static inline void
1021 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1022 {
1023 
1024         if (!schedstat_enabled())
1025                 return;
1026 
1027         /*
1028          * Mark the end of the wait period if dequeueing a
1029          * waiting task:
1030          */
1031         if (se != cfs_rq->curr)
1032                 update_stats_wait_end(cfs_rq, se);
1033 
1034         if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1035                 struct task_struct *tsk = task_of(se);
1036 
1037                 if (tsk->state & TASK_INTERRUPTIBLE)
1038                         schedstat_set(se->statistics.sleep_start,
1039                                       rq_clock(rq_of(cfs_rq)));
1040                 if (tsk->state & TASK_UNINTERRUPTIBLE)
1041                         schedstat_set(se->statistics.block_start,
1042                                       rq_clock(rq_of(cfs_rq)));
1043         }
1044 }
1045 
1046 /*
1047  * We are picking a new current task - update its stats:
1048  */
1049 static inline void
1050 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1051 {
1052         /*
1053          * We are starting a new run period:
1054          */
1055         se->exec_start = rq_clock_task(rq_of(cfs_rq));
1056 }
1057 
1058 /**************************************************
1059  * Scheduling class queueing methods:
1060  */
1061 
1062 #ifdef CONFIG_NUMA_BALANCING
1063 /*
1064  * Approximate time to scan a full NUMA task in ms. The task scan period is
1065  * calculated based on the tasks virtual memory size and
1066  * numa_balancing_scan_size.
1067  */
1068 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1069 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1070 
1071 /* Portion of address space to scan in MB */
1072 unsigned int sysctl_numa_balancing_scan_size = 256;
1073 
1074 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1075 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1076 
1077 static unsigned int task_nr_scan_windows(struct task_struct *p)
1078 {
1079         unsigned long rss = 0;
1080         unsigned long nr_scan_pages;
1081 
1082         /*
1083          * Calculations based on RSS as non-present and empty pages are skipped
1084          * by the PTE scanner and NUMA hinting faults should be trapped based
1085          * on resident pages
1086          */
1087         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1088         rss = get_mm_rss(p->mm);
1089         if (!rss)
1090                 rss = nr_scan_pages;
1091 
1092         rss = round_up(rss, nr_scan_pages);
1093         return rss / nr_scan_pages;
1094 }
1095 
1096 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1097 #define MAX_SCAN_WINDOW 2560
1098 
1099 static unsigned int task_scan_min(struct task_struct *p)
1100 {
1101         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1102         unsigned int scan, floor;
1103         unsigned int windows = 1;
1104 
1105         if (scan_size < MAX_SCAN_WINDOW)
1106                 windows = MAX_SCAN_WINDOW / scan_size;
1107         floor = 1000 / windows;
1108 
1109         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1110         return max_t(unsigned int, floor, scan);
1111 }
1112 
1113 static unsigned int task_scan_max(struct task_struct *p)
1114 {
1115         unsigned int smin = task_scan_min(p);
1116         unsigned int smax;
1117 
1118         /* Watch for min being lower than max due to floor calculations */
1119         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1120         return max(smin, smax);
1121 }
1122 
1123 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1124 {
1125         rq->nr_numa_running += (p->numa_preferred_nid != -1);
1126         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1127 }
1128 
1129 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1130 {
1131         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1132         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1133 }
1134 
1135 struct numa_group {
1136         atomic_t refcount;
1137 
1138         spinlock_t lock; /* nr_tasks, tasks */
1139         int nr_tasks;
1140         pid_t gid;
1141         int active_nodes;
1142 
1143         struct rcu_head rcu;
1144         unsigned long total_faults;
1145         unsigned long max_faults_cpu;
1146         /*
1147          * Faults_cpu is used to decide whether memory should move
1148          * towards the CPU. As a consequence, these stats are weighted
1149          * more by CPU use than by memory faults.
1150          */
1151         unsigned long *faults_cpu;
1152         unsigned long faults[0];
1153 };
1154 
1155 /* Shared or private faults. */
1156 #define NR_NUMA_HINT_FAULT_TYPES 2
1157 
1158 /* Memory and CPU locality */
1159 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1160 
1161 /* Averaged statistics, and temporary buffers. */
1162 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1163 
1164 pid_t task_numa_group_id(struct task_struct *p)
1165 {
1166         return p->numa_group ? p->numa_group->gid : 0;
1167 }
1168 
1169 /*
1170  * The averaged statistics, shared & private, memory & cpu,
1171  * occupy the first half of the array. The second half of the
1172  * array is for current counters, which are averaged into the
1173  * first set by task_numa_placement.
1174  */
1175 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1176 {
1177         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1178 }
1179 
1180 static inline unsigned long task_faults(struct task_struct *p, int nid)
1181 {
1182         if (!p->numa_faults)
1183                 return 0;
1184 
1185         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1186                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1187 }
1188 
1189 static inline unsigned long group_faults(struct task_struct *p, int nid)
1190 {
1191         if (!p->numa_group)
1192                 return 0;
1193 
1194         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1195                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1196 }
1197 
1198 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1199 {
1200         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1201                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1202 }
1203 
1204 /*
1205  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1206  * considered part of a numa group's pseudo-interleaving set. Migrations
1207  * between these nodes are slowed down, to allow things to settle down.
1208  */
1209 #define ACTIVE_NODE_FRACTION 3
1210 
1211 static bool numa_is_active_node(int nid, struct numa_group *ng)
1212 {
1213         return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1214 }
1215 
1216 /* Handle placement on systems where not all nodes are directly connected. */
1217 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1218                                         int maxdist, bool task)
1219 {
1220         unsigned long score = 0;
1221         int node;
1222 
1223         /*
1224          * All nodes are directly connected, and the same distance
1225          * from each other. No need for fancy placement algorithms.
1226          */
1227         if (sched_numa_topology_type == NUMA_DIRECT)
1228                 return 0;
1229 
1230         /*
1231          * This code is called for each node, introducing N^2 complexity,
1232          * which should be ok given the number of nodes rarely exceeds 8.
1233          */
1234         for_each_online_node(node) {
1235                 unsigned long faults;
1236                 int dist = node_distance(nid, node);
1237 
1238                 /*
1239                  * The furthest away nodes in the system are not interesting
1240                  * for placement; nid was already counted.
1241                  */
1242                 if (dist == sched_max_numa_distance || node == nid)
1243                         continue;
1244 
1245                 /*
1246                  * On systems with a backplane NUMA topology, compare groups
1247                  * of nodes, and move tasks towards the group with the most
1248                  * memory accesses. When comparing two nodes at distance
1249                  * "hoplimit", only nodes closer by than "hoplimit" are part
1250                  * of each group. Skip other nodes.
1251                  */
1252                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1253                                         dist > maxdist)
1254                         continue;
1255 
1256                 /* Add up the faults from nearby nodes. */
1257                 if (task)
1258                         faults = task_faults(p, node);
1259                 else
1260                         faults = group_faults(p, node);
1261 
1262                 /*
1263                  * On systems with a glueless mesh NUMA topology, there are
1264                  * no fixed "groups of nodes". Instead, nodes that are not
1265                  * directly connected bounce traffic through intermediate
1266                  * nodes; a numa_group can occupy any set of nodes.
1267                  * The further away a node is, the less the faults count.
1268                  * This seems to result in good task placement.
1269                  */
1270                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1271                         faults *= (sched_max_numa_distance - dist);
1272                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1273                 }
1274 
1275                 score += faults;
1276         }
1277 
1278         return score;
1279 }
1280 
1281 /*
1282  * These return the fraction of accesses done by a particular task, or
1283  * task group, on a particular numa node.  The group weight is given a
1284  * larger multiplier, in order to group tasks together that are almost
1285  * evenly spread out between numa nodes.
1286  */
1287 static inline unsigned long task_weight(struct task_struct *p, int nid,
1288                                         int dist)
1289 {
1290         unsigned long faults, total_faults;
1291 
1292         if (!p->numa_faults)
1293                 return 0;
1294 
1295         total_faults = p->total_numa_faults;
1296 
1297         if (!total_faults)
1298                 return 0;
1299 
1300         faults = task_faults(p, nid);
1301         faults += score_nearby_nodes(p, nid, dist, true);
1302 
1303         return 1000 * faults / total_faults;
1304 }
1305 
1306 static inline unsigned long group_weight(struct task_struct *p, int nid,
1307                                          int dist)
1308 {
1309         unsigned long faults, total_faults;
1310 
1311         if (!p->numa_group)
1312                 return 0;
1313 
1314         total_faults = p->numa_group->total_faults;
1315 
1316         if (!total_faults)
1317                 return 0;
1318 
1319         faults = group_faults(p, nid);
1320         faults += score_nearby_nodes(p, nid, dist, false);
1321 
1322         return 1000 * faults / total_faults;
1323 }
1324 
1325 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1326                                 int src_nid, int dst_cpu)
1327 {
1328         struct numa_group *ng = p->numa_group;
1329         int dst_nid = cpu_to_node(dst_cpu);
1330         int last_cpupid, this_cpupid;
1331 
1332         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1333 
1334         /*
1335          * Multi-stage node selection is used in conjunction with a periodic
1336          * migration fault to build a temporal task<->page relation. By using
1337          * a two-stage filter we remove short/unlikely relations.
1338          *
1339          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1340          * a task's usage of a particular page (n_p) per total usage of this
1341          * page (n_t) (in a given time-span) to a probability.
1342          *
1343          * Our periodic faults will sample this probability and getting the
1344          * same result twice in a row, given these samples are fully
1345          * independent, is then given by P(n)^2, provided our sample period
1346          * is sufficiently short compared to the usage pattern.
1347          *
1348          * This quadric squishes small probabilities, making it less likely we
1349          * act on an unlikely task<->page relation.
1350          */
1351         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1352         if (!cpupid_pid_unset(last_cpupid) &&
1353                                 cpupid_to_nid(last_cpupid) != dst_nid)
1354                 return false;
1355 
1356         /* Always allow migrate on private faults */
1357         if (cpupid_match_pid(p, last_cpupid))
1358                 return true;
1359 
1360         /* A shared fault, but p->numa_group has not been set up yet. */
1361         if (!ng)
1362                 return true;
1363 
1364         /*
1365          * Destination node is much more heavily used than the source
1366          * node? Allow migration.
1367          */
1368         if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1369                                         ACTIVE_NODE_FRACTION)
1370                 return true;
1371 
1372         /*
1373          * Distribute memory according to CPU & memory use on each node,
1374          * with 3/4 hysteresis to avoid unnecessary memory migrations:
1375          *
1376          * faults_cpu(dst)   3   faults_cpu(src)
1377          * --------------- * - > ---------------
1378          * faults_mem(dst)   4   faults_mem(src)
1379          */
1380         return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1381                group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1382 }
1383 
1384 static unsigned long weighted_cpuload(const int cpu);
1385 static unsigned long source_load(int cpu, int type);
1386 static unsigned long target_load(int cpu, int type);
1387 static unsigned long capacity_of(int cpu);
1388 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1389 
1390 /* Cached statistics for all CPUs within a node */
1391 struct numa_stats {
1392         unsigned long nr_running;
1393         unsigned long load;
1394 
1395         /* Total compute capacity of CPUs on a node */
1396         unsigned long compute_capacity;
1397 
1398         /* Approximate capacity in terms of runnable tasks on a node */
1399         unsigned long task_capacity;
1400         int has_free_capacity;
1401 };
1402 
1403 /*
1404  * XXX borrowed from update_sg_lb_stats
1405  */
1406 static void update_numa_stats(struct numa_stats *ns, int nid)
1407 {
1408         int smt, cpu, cpus = 0;
1409         unsigned long capacity;
1410 
1411         memset(ns, 0, sizeof(*ns));
1412         for_each_cpu(cpu, cpumask_of_node(nid)) {
1413                 struct rq *rq = cpu_rq(cpu);
1414 
1415                 ns->nr_running += rq->nr_running;
1416                 ns->load += weighted_cpuload(cpu);
1417                 ns->compute_capacity += capacity_of(cpu);
1418 
1419                 cpus++;
1420         }
1421 
1422         /*
1423          * If we raced with hotplug and there are no CPUs left in our mask
1424          * the @ns structure is NULL'ed and task_numa_compare() will
1425          * not find this node attractive.
1426          *
1427          * We'll either bail at !has_free_capacity, or we'll detect a huge
1428          * imbalance and bail there.
1429          */
1430         if (!cpus)
1431                 return;
1432 
1433         /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1434         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1435         capacity = cpus / smt; /* cores */
1436 
1437         ns->task_capacity = min_t(unsigned, capacity,
1438                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1439         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1440 }
1441 
1442 struct task_numa_env {
1443         struct task_struct *p;
1444 
1445         int src_cpu, src_nid;
1446         int dst_cpu, dst_nid;
1447 
1448         struct numa_stats src_stats, dst_stats;
1449 
1450         int imbalance_pct;
1451         int dist;
1452 
1453         struct task_struct *best_task;
1454         long best_imp;
1455         int best_cpu;
1456 };
1457 
1458 static void task_numa_assign(struct task_numa_env *env,
1459                              struct task_struct *p, long imp)
1460 {
1461         if (env->best_task)
1462                 put_task_struct(env->best_task);
1463         if (p)
1464                 get_task_struct(p);
1465 
1466         env->best_task = p;
1467         env->best_imp = imp;
1468         env->best_cpu = env->dst_cpu;
1469 }
1470 
1471 static bool load_too_imbalanced(long src_load, long dst_load,
1472                                 struct task_numa_env *env)
1473 {
1474         long imb, old_imb;
1475         long orig_src_load, orig_dst_load;
1476         long src_capacity, dst_capacity;
1477 
1478         /*
1479          * The load is corrected for the CPU capacity available on each node.
1480          *
1481          * src_load        dst_load
1482          * ------------ vs ---------
1483          * src_capacity    dst_capacity
1484          */
1485         src_capacity = env->src_stats.compute_capacity;
1486         dst_capacity = env->dst_stats.compute_capacity;
1487 
1488         /* We care about the slope of the imbalance, not the direction. */
1489         if (dst_load < src_load)
1490                 swap(dst_load, src_load);
1491 
1492         /* Is the difference below the threshold? */
1493         imb = dst_load * src_capacity * 100 -
1494               src_load * dst_capacity * env->imbalance_pct;
1495         if (imb <= 0)
1496                 return false;
1497 
1498         /*
1499          * The imbalance is above the allowed threshold.
1500          * Compare it with the old imbalance.
1501          */
1502         orig_src_load = env->src_stats.load;
1503         orig_dst_load = env->dst_stats.load;
1504 
1505         if (orig_dst_load < orig_src_load)
1506                 swap(orig_dst_load, orig_src_load);
1507 
1508         old_imb = orig_dst_load * src_capacity * 100 -
1509                   orig_src_load * dst_capacity * env->imbalance_pct;
1510 
1511         /* Would this change make things worse? */
1512         return (imb > old_imb);
1513 }
1514 
1515 /*
1516  * This checks if the overall compute and NUMA accesses of the system would
1517  * be improved if the source tasks was migrated to the target dst_cpu taking
1518  * into account that it might be best if task running on the dst_cpu should
1519  * be exchanged with the source task
1520  */
1521 static void task_numa_compare(struct task_numa_env *env,
1522                               long taskimp, long groupimp)
1523 {
1524         struct rq *src_rq = cpu_rq(env->src_cpu);
1525         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1526         struct task_struct *cur;
1527         long src_load, dst_load;
1528         long load;
1529         long imp = env->p->numa_group ? groupimp : taskimp;
1530         long moveimp = imp;
1531         int dist = env->dist;
1532 
1533         rcu_read_lock();
1534         cur = task_rcu_dereference(&dst_rq->curr);
1535         if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1536                 cur = NULL;
1537 
1538         /*
1539          * Because we have preemption enabled we can get migrated around and
1540          * end try selecting ourselves (current == env->p) as a swap candidate.
1541          */
1542         if (cur == env->p)
1543                 goto unlock;
1544 
1545         /*
1546          * "imp" is the fault differential for the source task between the
1547          * source and destination node. Calculate the total differential for
1548          * the source task and potential destination task. The more negative
1549          * the value is, the more rmeote accesses that would be expected to
1550          * be incurred if the tasks were swapped.
1551          */
1552         if (cur) {
1553                 /* Skip this swap candidate if cannot move to the source cpu */
1554                 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1555                         goto unlock;
1556 
1557                 /*
1558                  * If dst and source tasks are in the same NUMA group, or not
1559                  * in any group then look only at task weights.
1560                  */
1561                 if (cur->numa_group == env->p->numa_group) {
1562                         imp = taskimp + task_weight(cur, env->src_nid, dist) -
1563                               task_weight(cur, env->dst_nid, dist);
1564                         /*
1565                          * Add some hysteresis to prevent swapping the
1566                          * tasks within a group over tiny differences.
1567                          */
1568                         if (cur->numa_group)
1569                                 imp -= imp/16;
1570                 } else {
1571                         /*
1572                          * Compare the group weights. If a task is all by
1573                          * itself (not part of a group), use the task weight
1574                          * instead.
1575                          */
1576                         if (cur->numa_group)
1577                                 imp += group_weight(cur, env->src_nid, dist) -
1578                                        group_weight(cur, env->dst_nid, dist);
1579                         else
1580                                 imp += task_weight(cur, env->src_nid, dist) -
1581                                        task_weight(cur, env->dst_nid, dist);
1582                 }
1583         }
1584 
1585         if (imp <= env->best_imp && moveimp <= env->best_imp)
1586                 goto unlock;
1587 
1588         if (!cur) {
1589                 /* Is there capacity at our destination? */
1590                 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1591                     !env->dst_stats.has_free_capacity)
1592                         goto unlock;
1593 
1594                 goto balance;
1595         }
1596 
1597         /* Balance doesn't matter much if we're running a task per cpu */
1598         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1599                         dst_rq->nr_running == 1)
1600                 goto assign;
1601 
1602         /*
1603          * In the overloaded case, try and keep the load balanced.
1604          */
1605 balance:
1606         load = task_h_load(env->p);
1607         dst_load = env->dst_stats.load + load;
1608         src_load = env->src_stats.load - load;
1609 
1610         if (moveimp > imp && moveimp > env->best_imp) {
1611                 /*
1612                  * If the improvement from just moving env->p direction is
1613                  * better than swapping tasks around, check if a move is
1614                  * possible. Store a slightly smaller score than moveimp,
1615                  * so an actually idle CPU will win.
1616                  */
1617                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1618                         imp = moveimp - 1;
1619                         cur = NULL;
1620                         goto assign;
1621                 }
1622         }
1623 
1624         if (imp <= env->best_imp)
1625                 goto unlock;
1626 
1627         if (cur) {
1628                 load = task_h_load(cur);
1629                 dst_load -= load;
1630                 src_load += load;
1631         }
1632 
1633         if (load_too_imbalanced(src_load, dst_load, env))
1634                 goto unlock;
1635 
1636         /*
1637          * One idle CPU per node is evaluated for a task numa move.
1638          * Call select_idle_sibling to maybe find a better one.
1639          */
1640         if (!cur) {
1641                 /*
1642                  * select_idle_siblings() uses an per-cpu cpumask that
1643                  * can be used from IRQ context.
1644                  */
1645                 local_irq_disable();
1646                 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1647                                                    env->dst_cpu);
1648                 local_irq_enable();
1649         }
1650 
1651 assign:
1652         task_numa_assign(env, cur, imp);
1653 unlock:
1654         rcu_read_unlock();
1655 }
1656 
1657 static void task_numa_find_cpu(struct task_numa_env *env,
1658                                 long taskimp, long groupimp)
1659 {
1660         int cpu;
1661 
1662         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1663                 /* Skip this CPU if the source task cannot migrate */
1664                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1665                         continue;
1666 
1667                 env->dst_cpu = cpu;
1668                 task_numa_compare(env, taskimp, groupimp);
1669         }
1670 }
1671 
1672 /* Only move tasks to a NUMA node less busy than the current node. */
1673 static bool numa_has_capacity(struct task_numa_env *env)
1674 {
1675         struct numa_stats *src = &env->src_stats;
1676         struct numa_stats *dst = &env->dst_stats;
1677 
1678         if (src->has_free_capacity && !dst->has_free_capacity)
1679                 return false;
1680 
1681         /*
1682          * Only consider a task move if the source has a higher load
1683          * than the destination, corrected for CPU capacity on each node.
1684          *
1685          *      src->load                dst->load
1686          * --------------------- vs ---------------------
1687          * src->compute_capacity    dst->compute_capacity
1688          */
1689         if (src->load * dst->compute_capacity * env->imbalance_pct >
1690 
1691             dst->load * src->compute_capacity * 100)
1692                 return true;
1693 
1694         return false;
1695 }
1696 
1697 static int task_numa_migrate(struct task_struct *p)
1698 {
1699         struct task_numa_env env = {
1700                 .p = p,
1701 
1702                 .src_cpu = task_cpu(p),
1703                 .src_nid = task_node(p),
1704 
1705                 .imbalance_pct = 112,
1706 
1707                 .best_task = NULL,
1708                 .best_imp = 0,
1709                 .best_cpu = -1,
1710         };
1711         struct sched_domain *sd;
1712         unsigned long taskweight, groupweight;
1713         int nid, ret, dist;
1714         long taskimp, groupimp;
1715 
1716         /*
1717          * Pick the lowest SD_NUMA domain, as that would have the smallest
1718          * imbalance and would be the first to start moving tasks about.
1719          *
1720          * And we want to avoid any moving of tasks about, as that would create
1721          * random movement of tasks -- counter the numa conditions we're trying
1722          * to satisfy here.
1723          */
1724         rcu_read_lock();
1725         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1726         if (sd)
1727                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1728         rcu_read_unlock();
1729 
1730         /*
1731          * Cpusets can break the scheduler domain tree into smaller
1732          * balance domains, some of which do not cross NUMA boundaries.
1733          * Tasks that are "trapped" in such domains cannot be migrated
1734          * elsewhere, so there is no point in (re)trying.
1735          */
1736         if (unlikely(!sd)) {
1737                 p->numa_preferred_nid = task_node(p);
1738                 return -EINVAL;
1739         }
1740 
1741         env.dst_nid = p->numa_preferred_nid;
1742         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1743         taskweight = task_weight(p, env.src_nid, dist);
1744         groupweight = group_weight(p, env.src_nid, dist);
1745         update_numa_stats(&env.src_stats, env.src_nid);
1746         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1747         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1748         update_numa_stats(&env.dst_stats, env.dst_nid);
1749 
1750         /* Try to find a spot on the preferred nid. */
1751         if (numa_has_capacity(&env))
1752                 task_numa_find_cpu(&env, taskimp, groupimp);
1753 
1754         /*
1755          * Look at other nodes in these cases:
1756          * - there is no space available on the preferred_nid
1757          * - the task is part of a numa_group that is interleaved across
1758          *   multiple NUMA nodes; in order to better consolidate the group,
1759          *   we need to check other locations.
1760          */
1761         if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1762                 for_each_online_node(nid) {
1763                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1764                                 continue;
1765 
1766                         dist = node_distance(env.src_nid, env.dst_nid);
1767                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1768                                                 dist != env.dist) {
1769                                 taskweight = task_weight(p, env.src_nid, dist);
1770                                 groupweight = group_weight(p, env.src_nid, dist);
1771                         }
1772 
1773                         /* Only consider nodes where both task and groups benefit */
1774                         taskimp = task_weight(p, nid, dist) - taskweight;
1775                         groupimp = group_weight(p, nid, dist) - groupweight;
1776                         if (taskimp < 0 && groupimp < 0)
1777                                 continue;
1778 
1779                         env.dist = dist;
1780                         env.dst_nid = nid;
1781                         update_numa_stats(&env.dst_stats, env.dst_nid);
1782                         if (numa_has_capacity(&env))
1783                                 task_numa_find_cpu(&env, taskimp, groupimp);
1784                 }
1785         }
1786 
1787         /*
1788          * If the task is part of a workload that spans multiple NUMA nodes,
1789          * and is migrating into one of the workload's active nodes, remember
1790          * this node as the task's preferred numa node, so the workload can
1791          * settle down.
1792          * A task that migrated to a second choice node will be better off
1793          * trying for a better one later. Do not set the preferred node here.
1794          */
1795         if (p->numa_group) {
1796                 struct numa_group *ng = p->numa_group;
1797 
1798                 if (env.best_cpu == -1)
1799                         nid = env.src_nid;
1800                 else
1801                         nid = env.dst_nid;
1802 
1803                 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1804                         sched_setnuma(p, env.dst_nid);
1805         }
1806 
1807         /* No better CPU than the current one was found. */
1808         if (env.best_cpu == -1)
1809                 return -EAGAIN;
1810 
1811         /*
1812          * Reset the scan period if the task is being rescheduled on an
1813          * alternative node to recheck if the tasks is now properly placed.
1814          */
1815         p->numa_scan_period = task_scan_min(p);
1816 
1817         if (env.best_task == NULL) {
1818                 ret = migrate_task_to(p, env.best_cpu);
1819                 if (ret != 0)
1820                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1821                 return ret;
1822         }
1823 
1824         ret = migrate_swap(p, env.best_task);
1825         if (ret != 0)
1826                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1827         put_task_struct(env.best_task);
1828         return ret;
1829 }
1830 
1831 /* Attempt to migrate a task to a CPU on the preferred node. */
1832 static void numa_migrate_preferred(struct task_struct *p)
1833 {
1834         unsigned long interval = HZ;
1835 
1836         /* This task has no NUMA fault statistics yet */
1837         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1838                 return;
1839 
1840         /* Periodically retry migrating the task to the preferred node */
1841         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1842         p->numa_migrate_retry = jiffies + interval;
1843 
1844         /* Success if task is already running on preferred CPU */
1845         if (task_node(p) == p->numa_preferred_nid)
1846                 return;
1847 
1848         /* Otherwise, try migrate to a CPU on the preferred node */
1849         task_numa_migrate(p);
1850 }
1851 
1852 /*
1853  * Find out how many nodes on the workload is actively running on. Do this by
1854  * tracking the nodes from which NUMA hinting faults are triggered. This can
1855  * be different from the set of nodes where the workload's memory is currently
1856  * located.
1857  */
1858 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1859 {
1860         unsigned long faults, max_faults = 0;
1861         int nid, active_nodes = 0;
1862 
1863         for_each_online_node(nid) {
1864                 faults = group_faults_cpu(numa_group, nid);
1865                 if (faults > max_faults)
1866                         max_faults = faults;
1867         }
1868 
1869         for_each_online_node(nid) {
1870                 faults = group_faults_cpu(numa_group, nid);
1871                 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1872                         active_nodes++;
1873         }
1874 
1875         numa_group->max_faults_cpu = max_faults;
1876         numa_group->active_nodes = active_nodes;
1877 }
1878 
1879 /*
1880  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1881  * increments. The more local the fault statistics are, the higher the scan
1882  * period will be for the next scan window. If local/(local+remote) ratio is
1883  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1884  * the scan period will decrease. Aim for 70% local accesses.
1885  */
1886 #define NUMA_PERIOD_SLOTS 10
1887 #define NUMA_PERIOD_THRESHOLD 7
1888 
1889 /*
1890  * Increase the scan period (slow down scanning) if the majority of
1891  * our memory is already on our local node, or if the majority of
1892  * the page accesses are shared with other processes.
1893  * Otherwise, decrease the scan period.
1894  */
1895 static void update_task_scan_period(struct task_struct *p,
1896                         unsigned long shared, unsigned long private)
1897 {
1898         unsigned int period_slot;
1899         int ratio;
1900         int diff;
1901 
1902         unsigned long remote = p->numa_faults_locality[0];
1903         unsigned long local = p->numa_faults_locality[1];
1904 
1905         /*
1906          * If there were no record hinting faults then either the task is
1907          * completely idle or all activity is areas that are not of interest
1908          * to automatic numa balancing. Related to that, if there were failed
1909          * migration then it implies we are migrating too quickly or the local
1910          * node is overloaded. In either case, scan slower
1911          */
1912         if (local + shared == 0 || p->numa_faults_locality[2]) {
1913                 p->numa_scan_period = min(p->numa_scan_period_max,
1914                         p->numa_scan_period << 1);
1915 
1916                 p->mm->numa_next_scan = jiffies +
1917                         msecs_to_jiffies(p->numa_scan_period);
1918 
1919                 return;
1920         }
1921 
1922         /*
1923          * Prepare to scale scan period relative to the current period.
1924          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1925          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1926          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1927          */
1928         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1929         ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1930         if (ratio >= NUMA_PERIOD_THRESHOLD) {
1931                 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1932                 if (!slot)
1933                         slot = 1;
1934                 diff = slot * period_slot;
1935         } else {
1936                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1937 
1938                 /*
1939                  * Scale scan rate increases based on sharing. There is an
1940                  * inverse relationship between the degree of sharing and
1941                  * the adjustment made to the scanning period. Broadly
1942                  * speaking the intent is that there is little point
1943                  * scanning faster if shared accesses dominate as it may
1944                  * simply bounce migrations uselessly
1945                  */
1946                 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1947                 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1948         }
1949 
1950         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1951                         task_scan_min(p), task_scan_max(p));
1952         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1953 }
1954 
1955 /*
1956  * Get the fraction of time the task has been running since the last
1957  * NUMA placement cycle. The scheduler keeps similar statistics, but
1958  * decays those on a 32ms period, which is orders of magnitude off
1959  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1960  * stats only if the task is so new there are no NUMA statistics yet.
1961  */
1962 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1963 {
1964         u64 runtime, delta, now;
1965         /* Use the start of this time slice to avoid calculations. */
1966         now = p->se.exec_start;
1967         runtime = p->se.sum_exec_runtime;
1968 
1969         if (p->last_task_numa_placement) {
1970                 delta = runtime - p->last_sum_exec_runtime;
1971                 *period = now - p->last_task_numa_placement;
1972         } else {
1973                 delta = p->se.avg.load_sum / p->se.load.weight;
1974                 *period = LOAD_AVG_MAX;
1975         }
1976 
1977         p->last_sum_exec_runtime = runtime;
1978         p->last_task_numa_placement = now;
1979 
1980         return delta;
1981 }
1982 
1983 /*
1984  * Determine the preferred nid for a task in a numa_group. This needs to
1985  * be done in a way that produces consistent results with group_weight,
1986  * otherwise workloads might not converge.
1987  */
1988 static int preferred_group_nid(struct task_struct *p, int nid)
1989 {
1990         nodemask_t nodes;
1991         int dist;
1992 
1993         /* Direct connections between all NUMA nodes. */
1994         if (sched_numa_topology_type == NUMA_DIRECT)
1995                 return nid;
1996 
1997         /*
1998          * On a system with glueless mesh NUMA topology, group_weight
1999          * scores nodes according to the number of NUMA hinting faults on
2000          * both the node itself, and on nearby nodes.
2001          */
2002         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2003                 unsigned long score, max_score = 0;
2004                 int node, max_node = nid;
2005 
2006                 dist = sched_max_numa_distance;
2007 
2008                 for_each_online_node(node) {
2009                         score = group_weight(p, node, dist);
2010                         if (score > max_score) {
2011                                 max_score = score;
2012                                 max_node = node;
2013                         }
2014                 }
2015                 return max_node;
2016         }
2017 
2018         /*
2019          * Finding the preferred nid in a system with NUMA backplane
2020          * interconnect topology is more involved. The goal is to locate
2021          * tasks from numa_groups near each other in the system, and
2022          * untangle workloads from different sides of the system. This requires
2023          * searching down the hierarchy of node groups, recursively searching
2024          * inside the highest scoring group of nodes. The nodemask tricks
2025          * keep the complexity of the search down.
2026          */
2027         nodes = node_online_map;
2028         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2029                 unsigned long max_faults = 0;
2030                 nodemask_t max_group = NODE_MASK_NONE;
2031                 int a, b;
2032 
2033                 /* Are there nodes at this distance from each other? */
2034                 if (!find_numa_distance(dist))
2035                         continue;
2036 
2037                 for_each_node_mask(a, nodes) {
2038                         unsigned long faults = 0;
2039                         nodemask_t this_group;
2040                         nodes_clear(this_group);
2041 
2042                         /* Sum group's NUMA faults; includes a==b case. */
2043                         for_each_node_mask(b, nodes) {
2044                                 if (node_distance(a, b) < dist) {
2045                                         faults += group_faults(p, b);
2046                                         node_set(b, this_group);
2047                                         node_clear(b, nodes);
2048                                 }
2049                         }
2050 
2051                         /* Remember the top group. */
2052                         if (faults > max_faults) {
2053                                 max_faults = faults;
2054                                 max_group = this_group;
2055                                 /*
2056                                  * subtle: at the smallest distance there is
2057                                  * just one node left in each "group", the
2058                                  * winner is the preferred nid.
2059                                  */
2060                                 nid = a;
2061                         }
2062                 }
2063                 /* Next round, evaluate the nodes within max_group. */
2064                 if (!max_faults)
2065                         break;
2066                 nodes = max_group;
2067         }
2068         return nid;
2069 }
2070 
2071 static void task_numa_placement(struct task_struct *p)
2072 {
2073         int seq, nid, max_nid = -1, max_group_nid = -1;
2074         unsigned long max_faults = 0, max_group_faults = 0;
2075         unsigned long fault_types[2] = { 0, 0 };
2076         unsigned long total_faults;
2077         u64 runtime, period;
2078         spinlock_t *group_lock = NULL;
2079 
2080         /*
2081          * The p->mm->numa_scan_seq field gets updated without
2082          * exclusive access. Use READ_ONCE() here to ensure
2083          * that the field is read in a single access:
2084          */
2085         seq = READ_ONCE(p->mm->numa_scan_seq);
2086         if (p->numa_scan_seq == seq)
2087                 return;
2088         p->numa_scan_seq = seq;
2089         p->numa_scan_period_max = task_scan_max(p);
2090 
2091         total_faults = p->numa_faults_locality[0] +
2092                        p->numa_faults_locality[1];
2093         runtime = numa_get_avg_runtime(p, &period);
2094 
2095         /* If the task is part of a group prevent parallel updates to group stats */
2096         if (p->numa_group) {
2097                 group_lock = &p->numa_group->lock;
2098                 spin_lock_irq(group_lock);
2099         }
2100 
2101         /* Find the node with the highest number of faults */
2102         for_each_online_node(nid) {
2103                 /* Keep track of the offsets in numa_faults array */
2104                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2105                 unsigned long faults = 0, group_faults = 0;
2106                 int priv;
2107 
2108                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2109                         long diff, f_diff, f_weight;
2110 
2111                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2112                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2113                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2114                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2115 
2116                         /* Decay existing window, copy faults since last scan */
2117                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2118                         fault_types[priv] += p->numa_faults[membuf_idx];
2119                         p->numa_faults[membuf_idx] = 0;
2120 
2121                         /*
2122                          * Normalize the faults_from, so all tasks in a group
2123                          * count according to CPU use, instead of by the raw
2124                          * number of faults. Tasks with little runtime have
2125                          * little over-all impact on throughput, and thus their
2126                          * faults are less important.
2127                          */
2128                         f_weight = div64_u64(runtime << 16, period + 1);
2129                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2130                                    (total_faults + 1);
2131                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2132                         p->numa_faults[cpubuf_idx] = 0;
2133 
2134                         p->numa_faults[mem_idx] += diff;
2135                         p->numa_faults[cpu_idx] += f_diff;
2136                         faults += p->numa_faults[mem_idx];
2137                         p->total_numa_faults += diff;
2138                         if (p->numa_group) {
2139                                 /*
2140                                  * safe because we can only change our own group
2141                                  *
2142                                  * mem_idx represents the offset for a given
2143                                  * nid and priv in a specific region because it
2144                                  * is at the beginning of the numa_faults array.
2145                                  */
2146                                 p->numa_group->faults[mem_idx] += diff;
2147                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
2148                                 p->numa_group->total_faults += diff;
2149                                 group_faults += p->numa_group->faults[mem_idx];
2150                         }
2151                 }
2152 
2153                 if (faults > max_faults) {
2154                         max_faults = faults;
2155                         max_nid = nid;
2156                 }
2157 
2158                 if (group_faults > max_group_faults) {
2159                         max_group_faults = group_faults;
2160                         max_group_nid = nid;
2161                 }
2162         }
2163 
2164         update_task_scan_period(p, fault_types[0], fault_types[1]);
2165 
2166         if (p->numa_group) {
2167                 numa_group_count_active_nodes(p->numa_group);
2168                 spin_unlock_irq(group_lock);
2169                 max_nid = preferred_group_nid(p, max_group_nid);
2170         }
2171 
2172         if (max_faults) {
2173                 /* Set the new preferred node */
2174                 if (max_nid != p->numa_preferred_nid)
2175                         sched_setnuma(p, max_nid);
2176 
2177                 if (task_node(p) != p->numa_preferred_nid)
2178                         numa_migrate_preferred(p);
2179         }
2180 }
2181 
2182 static inline int get_numa_group(struct numa_group *grp)
2183 {
2184         return atomic_inc_not_zero(&grp->refcount);
2185 }
2186 
2187 static inline void put_numa_group(struct numa_group *grp)
2188 {
2189         if (atomic_dec_and_test(&grp->refcount))
2190                 kfree_rcu(grp, rcu);
2191 }
2192 
2193 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2194                         int *priv)
2195 {
2196         struct numa_group *grp, *my_grp;
2197         struct task_struct *tsk;
2198         bool join = false;
2199         int cpu = cpupid_to_cpu(cpupid);
2200         int i;
2201 
2202         if (unlikely(!p->numa_group)) {
2203                 unsigned int size = sizeof(struct numa_group) +
2204                                     4*nr_node_ids*sizeof(unsigned long);
2205 
2206                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2207                 if (!grp)
2208                         return;
2209 
2210                 atomic_set(&grp->refcount, 1);
2211                 grp->active_nodes = 1;
2212                 grp->max_faults_cpu = 0;
2213                 spin_lock_init(&grp->lock);
2214                 grp->gid = p->pid;
2215                 /* Second half of the array tracks nids where faults happen */
2216                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2217                                                 nr_node_ids;
2218 
2219                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2220                         grp->faults[i] = p->numa_faults[i];
2221 
2222                 grp->total_faults = p->total_numa_faults;
2223 
2224                 grp->nr_tasks++;
2225                 rcu_assign_pointer(p->numa_group, grp);
2226         }
2227 
2228         rcu_read_lock();
2229         tsk = READ_ONCE(cpu_rq(cpu)->curr);
2230 
2231         if (!cpupid_match_pid(tsk, cpupid))
2232                 goto no_join;
2233 
2234         grp = rcu_dereference(tsk->numa_group);
2235         if (!grp)
2236                 goto no_join;
2237 
2238         my_grp = p->numa_group;
2239         if (grp == my_grp)
2240                 goto no_join;
2241 
2242         /*
2243          * Only join the other group if its bigger; if we're the bigger group,
2244          * the other task will join us.
2245          */
2246         if (my_grp->nr_tasks > grp->nr_tasks)
2247                 goto no_join;
2248 
2249         /*
2250          * Tie-break on the grp address.
2251          */
2252         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2253                 goto no_join;
2254 
2255         /* Always join threads in the same process. */
2256         if (tsk->mm == current->mm)
2257                 join = true;
2258 
2259         /* Simple filter to avoid false positives due to PID collisions */
2260         if (flags & TNF_SHARED)
2261                 join = true;
2262 
2263         /* Update priv based on whether false sharing was detected */
2264         *priv = !join;
2265 
2266         if (join && !get_numa_group(grp))
2267                 goto no_join;
2268 
2269         rcu_read_unlock();
2270 
2271         if (!join)
2272                 return;
2273 
2274         BUG_ON(irqs_disabled());
2275         double_lock_irq(&my_grp->lock, &grp->lock);
2276 
2277         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2278                 my_grp->faults[i] -= p->numa_faults[i];
2279                 grp->faults[i] += p->numa_faults[i];
2280         }
2281         my_grp->total_faults -= p->total_numa_faults;
2282         grp->total_faults += p->total_numa_faults;
2283 
2284         my_grp->nr_tasks--;
2285         grp->nr_tasks++;
2286 
2287         spin_unlock(&my_grp->lock);
2288         spin_unlock_irq(&grp->lock);
2289 
2290         rcu_assign_pointer(p->numa_group, grp);
2291 
2292         put_numa_group(my_grp);
2293         return;
2294 
2295 no_join:
2296         rcu_read_unlock();
2297         return;
2298 }
2299 
2300 void task_numa_free(struct task_struct *p)
2301 {
2302         struct numa_group *grp = p->numa_group;
2303         void *numa_faults = p->numa_faults;
2304         unsigned long flags;
2305         int i;
2306 
2307         if (grp) {
2308                 spin_lock_irqsave(&grp->lock, flags);
2309                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2310                         grp->faults[i] -= p->numa_faults[i];
2311                 grp->total_faults -= p->total_numa_faults;
2312 
2313                 grp->nr_tasks--;
2314                 spin_unlock_irqrestore(&grp->lock, flags);
2315                 RCU_INIT_POINTER(p->numa_group, NULL);
2316                 put_numa_group(grp);
2317         }
2318 
2319         p->numa_faults = NULL;
2320         kfree(numa_faults);
2321 }
2322 
2323 /*
2324  * Got a PROT_NONE fault for a page on @node.
2325  */
2326 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2327 {
2328         struct task_struct *p = current;
2329         bool migrated = flags & TNF_MIGRATED;
2330         int cpu_node = task_node(current);
2331         int local = !!(flags & TNF_FAULT_LOCAL);
2332         struct numa_group *ng;
2333         int priv;
2334 
2335         if (!static_branch_likely(&sched_numa_balancing))
2336                 return;
2337 
2338         /* for example, ksmd faulting in a user's mm */
2339         if (!p->mm)
2340                 return;
2341 
2342         /* Allocate buffer to track faults on a per-node basis */
2343         if (unlikely(!p->numa_faults)) {
2344                 int size = sizeof(*p->numa_faults) *
2345                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2346 
2347                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2348                 if (!p->numa_faults)
2349                         return;
2350 
2351                 p->total_numa_faults = 0;
2352                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2353         }
2354 
2355         /*
2356          * First accesses are treated as private, otherwise consider accesses
2357          * to be private if the accessing pid has not changed
2358          */
2359         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2360                 priv = 1;
2361         } else {
2362                 priv = cpupid_match_pid(p, last_cpupid);
2363                 if (!priv && !(flags & TNF_NO_GROUP))
2364                         task_numa_group(p, last_cpupid, flags, &priv);
2365         }
2366 
2367         /*
2368          * If a workload spans multiple NUMA nodes, a shared fault that
2369          * occurs wholly within the set of nodes that the workload is
2370          * actively using should be counted as local. This allows the
2371          * scan rate to slow down when a workload has settled down.
2372          */
2373         ng = p->numa_group;
2374         if (!priv && !local && ng && ng->active_nodes > 1 &&
2375                                 numa_is_active_node(cpu_node, ng) &&
2376                                 numa_is_active_node(mem_node, ng))
2377                 local = 1;
2378 
2379         task_numa_placement(p);
2380 
2381         /*
2382          * Retry task to preferred node migration periodically, in case it
2383          * case it previously failed, or the scheduler moved us.
2384          */
2385         if (time_after(jiffies, p->numa_migrate_retry))
2386                 numa_migrate_preferred(p);
2387 
2388         if (migrated)
2389                 p->numa_pages_migrated += pages;
2390         if (flags & TNF_MIGRATE_FAIL)
2391                 p->numa_faults_locality[2] += pages;
2392 
2393         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2394         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2395         p->numa_faults_locality[local] += pages;
2396 }
2397 
2398 static void reset_ptenuma_scan(struct task_struct *p)
2399 {
2400         /*
2401          * We only did a read acquisition of the mmap sem, so
2402          * p->mm->numa_scan_seq is written to without exclusive access
2403          * and the update is not guaranteed to be atomic. That's not
2404          * much of an issue though, since this is just used for
2405          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2406          * expensive, to avoid any form of compiler optimizations:
2407          */
2408         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2409         p->mm->numa_scan_offset = 0;
2410 }
2411 
2412 /*
2413  * The expensive part of numa migration is done from task_work context.
2414  * Triggered from task_tick_numa().
2415  */
2416 void task_numa_work(struct callback_head *work)
2417 {
2418         unsigned long migrate, next_scan, now = jiffies;
2419         struct task_struct *p = current;
2420         struct mm_struct *mm = p->mm;
2421         u64 runtime = p->se.sum_exec_runtime;
2422         struct vm_area_struct *vma;
2423         unsigned long start, end;
2424         unsigned long nr_pte_updates = 0;
2425         long pages, virtpages;
2426 
2427         SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2428 
2429         work->next = work; /* protect against double add */
2430         /*
2431          * Who cares about NUMA placement when they're dying.
2432          *
2433          * NOTE: make sure not to dereference p->mm before this check,
2434          * exit_task_work() happens _after_ exit_mm() so we could be called
2435          * without p->mm even though we still had it when we enqueued this
2436          * work.
2437          */
2438         if (p->flags & PF_EXITING)
2439                 return;
2440 
2441         if (!mm->numa_next_scan) {
2442                 mm->numa_next_scan = now +
2443                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2444         }
2445 
2446         /*
2447          * Enforce maximal scan/migration frequency..
2448          */
2449         migrate = mm->numa_next_scan;
2450         if (time_before(now, migrate))
2451                 return;
2452 
2453         if (p->numa_scan_period == 0) {
2454                 p->numa_scan_period_max = task_scan_max(p);
2455                 p->numa_scan_period = task_scan_min(p);
2456         }
2457 
2458         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2459         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2460                 return;
2461 
2462         /*
2463          * Delay this task enough that another task of this mm will likely win
2464          * the next time around.
2465          */
2466         p->node_stamp += 2 * TICK_NSEC;
2467 
2468         start = mm->numa_scan_offset;
2469         pages = sysctl_numa_balancing_scan_size;
2470         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2471         virtpages = pages * 8;     /* Scan up to this much virtual space */
2472         if (!pages)
2473                 return;
2474 
2475 
2476         down_read(&mm->mmap_sem);
2477         vma = find_vma(mm, start);
2478         if (!vma) {
2479                 reset_ptenuma_scan(p);
2480                 start = 0;
2481                 vma = mm->mmap;
2482         }
2483         for (; vma; vma = vma->vm_next) {
2484                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2485                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2486                         continue;
2487                 }
2488 
2489                 /*
2490                  * Shared library pages mapped by multiple processes are not
2491                  * migrated as it is expected they are cache replicated. Avoid
2492                  * hinting faults in read-only file-backed mappings or the vdso
2493                  * as migrating the pages will be of marginal benefit.
2494                  */
2495                 if (!vma->vm_mm ||
2496                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2497                         continue;
2498 
2499                 /*
2500                  * Skip inaccessible VMAs to avoid any confusion between
2501                  * PROT_NONE and NUMA hinting ptes
2502                  */
2503                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2504                         continue;
2505 
2506                 do {
2507                         start = max(start, vma->vm_start);
2508                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2509                         end = min(end, vma->vm_end);
2510                         nr_pte_updates = change_prot_numa(vma, start, end);
2511 
2512                         /*
2513                          * Try to scan sysctl_numa_balancing_size worth of
2514                          * hpages that have at least one present PTE that
2515                          * is not already pte-numa. If the VMA contains
2516                          * areas that are unused or already full of prot_numa
2517                          * PTEs, scan up to virtpages, to skip through those
2518                          * areas faster.
2519                          */
2520                         if (nr_pte_updates)
2521                                 pages -= (end - start) >> PAGE_SHIFT;
2522                         virtpages -= (end - start) >> PAGE_SHIFT;
2523 
2524                         start = end;
2525                         if (pages <= 0 || virtpages <= 0)
2526                                 goto out;
2527 
2528                         cond_resched();
2529                 } while (end != vma->vm_end);
2530         }
2531 
2532 out:
2533         /*
2534          * It is possible to reach the end of the VMA list but the last few
2535          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2536          * would find the !migratable VMA on the next scan but not reset the
2537          * scanner to the start so check it now.
2538          */
2539         if (vma)
2540                 mm->numa_scan_offset = start;
2541         else
2542                 reset_ptenuma_scan(p);
2543         up_read(&mm->mmap_sem);
2544 
2545         /*
2546          * Make sure tasks use at least 32x as much time to run other code
2547          * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2548          * Usually update_task_scan_period slows down scanning enough; on an
2549          * overloaded system we need to limit overhead on a per task basis.
2550          */
2551         if (unlikely(p->se.sum_exec_runtime != runtime)) {
2552                 u64 diff = p->se.sum_exec_runtime - runtime;
2553                 p->node_stamp += 32 * diff;
2554         }
2555 }
2556 
2557 /*
2558  * Drive the periodic memory faults..
2559  */
2560 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2561 {
2562         struct callback_head *work = &curr->numa_work;
2563         u64 period, now;
2564 
2565         /*
2566          * We don't care about NUMA placement if we don't have memory.
2567          */
2568         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2569                 return;
2570 
2571         /*
2572          * Using runtime rather than walltime has the dual advantage that
2573          * we (mostly) drive the selection from busy threads and that the
2574          * task needs to have done some actual work before we bother with
2575          * NUMA placement.
2576          */
2577         now = curr->se.sum_exec_runtime;
2578         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2579 
2580         if (now > curr->node_stamp + period) {
2581                 if (!curr->node_stamp)
2582                         curr->numa_scan_period = task_scan_min(curr);
2583                 curr->node_stamp += period;
2584 
2585                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2586                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2587                         task_work_add(curr, work, true);
2588                 }
2589         }
2590 }
2591 #else
2592 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2593 {
2594 }
2595 
2596 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2597 {
2598 }
2599 
2600 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2601 {
2602 }
2603 #endif /* CONFIG_NUMA_BALANCING */
2604 
2605 static void
2606 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2607 {
2608         update_load_add(&cfs_rq->load, se->load.weight);
2609         if (!parent_entity(se))
2610                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2611 #ifdef CONFIG_SMP
2612         if (entity_is_task(se)) {
2613                 struct rq *rq = rq_of(cfs_rq);
2614 
2615                 account_numa_enqueue(rq, task_of(se));
2616                 list_add(&se->group_node, &rq->cfs_tasks);
2617         }
2618 #endif
2619         cfs_rq->nr_running++;
2620 }
2621 
2622 static void
2623 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2624 {
2625         update_load_sub(&cfs_rq->load, se->load.weight);
2626         if (!parent_entity(se))
2627                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2628 #ifdef CONFIG_SMP
2629         if (entity_is_task(se)) {
2630                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2631                 list_del_init(&se->group_node);
2632         }
2633 #endif
2634         cfs_rq->nr_running--;
2635 }
2636 
2637 #ifdef CONFIG_FAIR_GROUP_SCHED
2638 # ifdef CONFIG_SMP
2639 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2640 {
2641         long tg_weight, load, shares;
2642 
2643         /*
2644          * This really should be: cfs_rq->avg.load_avg, but instead we use
2645          * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2646          * the shares for small weight interactive tasks.
2647          */
2648         load = scale_load_down(cfs_rq->load.weight);
2649 
2650         tg_weight = atomic_long_read(&tg->load_avg);
2651 
2652         /* Ensure tg_weight >= load */
2653         tg_weight -= cfs_rq->tg_load_avg_contrib;
2654         tg_weight += load;
2655 
2656         shares = (tg->shares * load);
2657         if (tg_weight)
2658                 shares /= tg_weight;
2659 
2660         if (shares < MIN_SHARES)
2661                 shares = MIN_SHARES;
2662         if (shares > tg->shares)
2663                 shares = tg->shares;
2664 
2665         return shares;
2666 }
2667 # else /* CONFIG_SMP */
2668 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2669 {
2670         return tg->shares;
2671 }
2672 # endif /* CONFIG_SMP */
2673 
2674 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2675                             unsigned long weight)
2676 {
2677         if (se->on_rq) {
2678                 /* commit outstanding execution time */
2679                 if (cfs_rq->curr == se)
2680                         update_curr(cfs_rq);
2681                 account_entity_dequeue(cfs_rq, se);
2682         }
2683 
2684         update_load_set(&se->load, weight);
2685 
2686         if (se->on_rq)
2687                 account_entity_enqueue(cfs_rq, se);
2688 }
2689 
2690 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2691 
2692 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2693 {
2694         struct task_group *tg;
2695         struct sched_entity *se;
2696         long shares;
2697 
2698         tg = cfs_rq->tg;
2699         se = tg->se[cpu_of(rq_of(cfs_rq))];
2700         if (!se || throttled_hierarchy(cfs_rq))
2701                 return;
2702 #ifndef CONFIG_SMP
2703         if (likely(se->load.weight == tg->shares))
2704                 return;
2705 #endif
2706         shares = calc_cfs_shares(cfs_rq, tg);
2707 
2708         reweight_entity(cfs_rq_of(se), se, shares);
2709 }
2710 #else /* CONFIG_FAIR_GROUP_SCHED */
2711 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2712 {
2713 }
2714 #endif /* CONFIG_FAIR_GROUP_SCHED */
2715 
2716 #ifdef CONFIG_SMP
2717 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2718 static const u32 runnable_avg_yN_inv[] = {
2719         0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2720         0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2721         0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2722         0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2723         0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2724         0x85aac367, 0x82cd8698,
2725 };
2726 
2727 /*
2728  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2729  * over-estimates when re-combining.
2730  */
2731 static const u32 runnable_avg_yN_sum[] = {
2732             0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2733          9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2734         17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2735 };
2736 
2737 /*
2738  * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2739  * lower integers. See Documentation/scheduler/sched-avg.txt how these
2740  * were generated:
2741  */
2742 static const u32 __accumulated_sum_N32[] = {
2743             0, 23371, 35056, 40899, 43820, 45281,
2744         46011, 46376, 46559, 46650, 46696, 46719,
2745 };
2746 
2747 /*
2748  * Approximate:
2749  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2750  */
2751 static __always_inline u64 decay_load(u64 val, u64 n)
2752 {
2753         unsigned int local_n;
2754 
2755         if (!n)
2756                 return val;
2757         else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2758                 return 0;
2759 
2760         /* after bounds checking we can collapse to 32-bit */
2761         local_n = n;
2762 
2763         /*
2764          * As y^PERIOD = 1/2, we can combine
2765          *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2766          * With a look-up table which covers y^n (n<PERIOD)
2767          *
2768          * To achieve constant time decay_load.
2769          */
2770         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2771                 val >>= local_n / LOAD_AVG_PERIOD;
2772                 local_n %= LOAD_AVG_PERIOD;
2773         }
2774 
2775         val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2776         return val;
2777 }
2778 
2779 /*
2780  * For updates fully spanning n periods, the contribution to runnable
2781  * average will be: \Sum 1024*y^n
2782  *
2783  * We can compute this reasonably efficiently by combining:
2784  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2785  */
2786 static u32 __compute_runnable_contrib(u64 n)
2787 {
2788         u32 contrib = 0;
2789 
2790         if (likely(n <= LOAD_AVG_PERIOD))
2791                 return runnable_avg_yN_sum[n];
2792         else if (unlikely(n >= LOAD_AVG_MAX_N))
2793                 return LOAD_AVG_MAX;
2794 
2795         /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2796         contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2797         n %= LOAD_AVG_PERIOD;
2798         contrib = decay_load(contrib, n);
2799         return contrib + runnable_avg_yN_sum[n];
2800 }
2801 
2802 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2803 
2804 /*
2805  * We can represent the historical contribution to runnable average as the
2806  * coefficients of a geometric series.  To do this we sub-divide our runnable
2807  * history into segments of approximately 1ms (1024us); label the segment that
2808  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2809  *
2810  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2811  *      p0            p1           p2
2812  *     (now)       (~1ms ago)  (~2ms ago)
2813  *
2814  * Let u_i denote the fraction of p_i that the entity was runnable.
2815  *
2816  * We then designate the fractions u_i as our co-efficients, yielding the
2817  * following representation of historical load:
2818  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2819  *
2820  * We choose y based on the with of a reasonably scheduling period, fixing:
2821  *   y^32 = 0.5
2822  *
2823  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2824  * approximately half as much as the contribution to load within the last ms
2825  * (u_0).
2826  *
2827  * When a period "rolls over" and we have new u_0`, multiplying the previous
2828  * sum again by y is sufficient to update:
2829  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2830  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2831  */
2832 static __always_inline int
2833 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2834                   unsigned long weight, int running, struct cfs_rq *cfs_rq)
2835 {
2836         u64 delta, scaled_delta, periods;
2837         u32 contrib;
2838         unsigned int delta_w, scaled_delta_w, decayed = 0;
2839         unsigned long scale_freq, scale_cpu;
2840 
2841         delta = now - sa->last_update_time;
2842         /*
2843          * This should only happen when time goes backwards, which it
2844          * unfortunately does during sched clock init when we swap over to TSC.
2845          */
2846         if ((s64)delta < 0) {
2847                 sa->last_update_time = now;
2848                 return 0;
2849         }
2850 
2851         /*
2852          * Use 1024ns as the unit of measurement since it's a reasonable
2853          * approximation of 1us and fast to compute.
2854          */
2855         delta >>= 10;
2856         if (!delta)
2857                 return 0;
2858         sa->last_update_time = now;
2859 
2860         scale_freq = arch_scale_freq_capacity(NULL, cpu);
2861         scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2862 
2863         /* delta_w is the amount already accumulated against our next period */
2864         delta_w = sa->period_contrib;
2865         if (delta + delta_w >= 1024) {
2866                 decayed = 1;
2867 
2868                 /* how much left for next period will start over, we don't know yet */
2869                 sa->period_contrib = 0;
2870 
2871                 /*
2872                  * Now that we know we're crossing a period boundary, figure
2873                  * out how much from delta we need to complete the current
2874                  * period and accrue it.
2875                  */
2876                 delta_w = 1024 - delta_w;
2877                 scaled_delta_w = cap_scale(delta_w, scale_freq);
2878                 if (weight) {
2879                         sa->load_sum += weight * scaled_delta_w;
2880                         if (cfs_rq) {
2881                                 cfs_rq->runnable_load_sum +=
2882                                                 weight * scaled_delta_w;
2883                         }
2884                 }
2885                 if (running)
2886                         sa->util_sum += scaled_delta_w * scale_cpu;
2887 
2888                 delta -= delta_w;
2889 
2890                 /* Figure out how many additional periods this update spans */
2891                 periods = delta / 1024;
2892                 delta %= 1024;
2893 
2894                 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2895                 if (cfs_rq) {
2896                         cfs_rq->runnable_load_sum =
2897                                 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2898                 }
2899                 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2900 
2901                 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2902                 contrib = __compute_runnable_contrib(periods);
2903                 contrib = cap_scale(contrib, scale_freq);
2904                 if (weight) {
2905                         sa->load_sum += weight * contrib;
2906                         if (cfs_rq)
2907                                 cfs_rq->runnable_load_sum += weight * contrib;
2908                 }
2909                 if (running)
2910                         sa->util_sum += contrib * scale_cpu;
2911         }
2912 
2913         /* Remainder of delta accrued against u_0` */
2914         scaled_delta = cap_scale(delta, scale_freq);
2915         if (weight) {
2916                 sa->load_sum += weight * scaled_delta;
2917                 if (cfs_rq)
2918                         cfs_rq->runnable_load_sum += weight * scaled_delta;
2919         }
2920         if (running)
2921                 sa->util_sum += scaled_delta * scale_cpu;
2922 
2923         sa->period_contrib += delta;
2924 
2925         if (decayed) {
2926                 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2927                 if (cfs_rq) {
2928                         cfs_rq->runnable_load_avg =
2929                                 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2930                 }
2931                 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2932         }
2933 
2934         return decayed;
2935 }
2936 
2937 /*
2938  * Signed add and clamp on underflow.
2939  *
2940  * Explicitly do a load-store to ensure the intermediate value never hits
2941  * memory. This allows lockless observations without ever seeing the negative
2942  * values.
2943  */
2944 #define add_positive(_ptr, _val) do {                           \
2945         typeof(_ptr) ptr = (_ptr);                              \
2946         typeof(_val) val = (_val);                              \
2947         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2948                                                                 \
2949         res = var + val;                                        \
2950                                                                 \
2951         if (val < 0 && res > var)                               \
2952                 res = 0;                                        \
2953                                                                 \
2954         WRITE_ONCE(*ptr, res);                                  \
2955 } while (0)
2956 
2957 #ifdef CONFIG_FAIR_GROUP_SCHED
2958 /**
2959  * update_tg_load_avg - update the tg's load avg
2960  * @cfs_rq: the cfs_rq whose avg changed
2961  * @force: update regardless of how small the difference
2962  *
2963  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
2964  * However, because tg->load_avg is a global value there are performance
2965  * considerations.
2966  *
2967  * In order to avoid having to look at the other cfs_rq's, we use a
2968  * differential update where we store the last value we propagated. This in
2969  * turn allows skipping updates if the differential is 'small'.
2970  *
2971  * Updating tg's load_avg is necessary before update_cfs_share() (which is
2972  * done) and effective_load() (which is not done because it is too costly).
2973  */
2974 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2975 {
2976         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2977 
2978         /*
2979          * No need to update load_avg for root_task_group as it is not used.
2980          */
2981         if (cfs_rq->tg == &root_task_group)
2982                 return;
2983 
2984         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2985                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2986                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2987         }
2988 }
2989 
2990 /*
2991  * Called within set_task_rq() right before setting a task's cpu. The
2992  * caller only guarantees p->pi_lock is held; no other assumptions,
2993  * including the state of rq->lock, should be made.
2994  */
2995 void set_task_rq_fair(struct sched_entity *se,
2996                       struct cfs_rq *prev, struct cfs_rq *next)
2997 {
2998         if (!sched_feat(ATTACH_AGE_LOAD))
2999                 return;
3000 
3001         /*
3002          * We are supposed to update the task to "current" time, then its up to
3003          * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3004          * getting what current time is, so simply throw away the out-of-date
3005          * time. This will result in the wakee task is less decayed, but giving
3006          * the wakee more load sounds not bad.
3007          */
3008         if (se->avg.last_update_time && prev) {
3009                 u64 p_last_update_time;
3010                 u64 n_last_update_time;
3011 
3012 #ifndef CONFIG_64BIT
3013                 u64 p_last_update_time_copy;
3014                 u64 n_last_update_time_copy;
3015 
3016                 do {
3017                         p_last_update_time_copy = prev->load_last_update_time_copy;
3018                         n_last_update_time_copy = next->load_last_update_time_copy;
3019 
3020                         smp_rmb();
3021 
3022                         p_last_update_time = prev->avg.last_update_time;
3023                         n_last_update_time = next->avg.last_update_time;
3024 
3025                 } while (p_last_update_time != p_last_update_time_copy ||
3026                          n_last_update_time != n_last_update_time_copy);
3027 #else
3028                 p_last_update_time = prev->avg.last_update_time;
3029                 n_last_update_time = next->avg.last_update_time;
3030 #endif
3031                 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
3032                                   &se->avg, 0, 0, NULL);
3033                 se->avg.last_update_time = n_last_update_time;
3034         }
3035 }
3036 
3037 /* Take into account change of utilization of a child task group */
3038 static inline void
3039 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se)
3040 {
3041         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3042         long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3043 
3044         /* Nothing to update */
3045         if (!delta)
3046                 return;
3047 
3048         /* Set new sched_entity's utilization */
3049         se->avg.util_avg = gcfs_rq->avg.util_avg;
3050         se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3051 
3052         /* Update parent cfs_rq utilization */
3053         add_positive(&cfs_rq->avg.util_avg, delta);
3054         cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3055 }
3056 
3057 /* Take into account change of load of a child task group */
3058 static inline void
3059 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se)
3060 {
3061         struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3062         long delta, load = gcfs_rq->avg.load_avg;
3063 
3064         /*
3065          * If the load of group cfs_rq is null, the load of the
3066          * sched_entity will also be null so we can skip the formula
3067          */
3068         if (load) {
3069                 long tg_load;
3070 
3071                 /* Get tg's load and ensure tg_load > 0 */
3072                 tg_load = atomic_long_read(&gcfs_rq->tg->load_avg) + 1;
3073 
3074                 /* Ensure tg_load >= load and updated with current load*/
3075                 tg_load -= gcfs_rq->tg_load_avg_contrib;
3076                 tg_load += load;
3077 
3078                 /*
3079                  * We need to compute a correction term in the case that the
3080                  * task group is consuming more CPU than a task of equal
3081                  * weight. A task with a weight equals to tg->shares will have
3082                  * a load less or equal to scale_load_down(tg->shares).
3083                  * Similarly, the sched_entities that represent the task group
3084                  * at parent level, can't have a load higher than
3085                  * scale_load_down(tg->shares). And the Sum of sched_entities'
3086                  * load must be <= scale_load_down(tg->shares).
3087                  */
3088                 if (tg_load > scale_load_down(gcfs_rq->tg->shares)) {
3089                         /* scale gcfs_rq's load into tg's shares*/
3090                         load *= scale_load_down(gcfs_rq->tg->shares);
3091                         load /= tg_load;
3092                 }
3093         }
3094 
3095         delta = load - se->avg.load_avg;
3096 
3097         /* Nothing to update */
3098         if (!delta)
3099                 return;
3100 
3101         /* Set new sched_entity's load */
3102         se->avg.load_avg = load;
3103         se->avg.load_sum = se->avg.load_avg * LOAD_AVG_MAX;
3104 
3105         /* Update parent cfs_rq load */
3106         add_positive(&cfs_rq->avg.load_avg, delta);
3107         cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * LOAD_AVG_MAX;
3108 
3109         /*
3110          * If the sched_entity is already enqueued, we also have to update the
3111          * runnable load avg.
3112          */
3113         if (se->on_rq) {
3114                 /* Update parent cfs_rq runnable_load_avg */
3115                 add_positive(&cfs_rq->runnable_load_avg, delta);
3116                 cfs_rq->runnable_load_sum = cfs_rq->runnable_load_avg * LOAD_AVG_MAX;
3117         }
3118 }
3119 
3120 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq)
3121 {
3122         cfs_rq->propagate_avg = 1;
3123 }
3124 
3125 static inline int test_and_clear_tg_cfs_propagate(struct sched_entity *se)
3126 {
3127         struct cfs_rq *cfs_rq = group_cfs_rq(se);
3128 
3129         if (!cfs_rq->propagate_avg)
3130                 return 0;
3131 
3132         cfs_rq->propagate_avg = 0;
3133         return 1;
3134 }
3135 
3136 /* Update task and its cfs_rq load average */
3137 static inline int propagate_entity_load_avg(struct sched_entity *se)
3138 {
3139         struct cfs_rq *cfs_rq;
3140 
3141         if (entity_is_task(se))
3142                 return 0;
3143 
3144         if (!test_and_clear_tg_cfs_propagate(se))
3145                 return 0;
3146 
3147         cfs_rq = cfs_rq_of(se);
3148 
3149         set_tg_cfs_propagate(cfs_rq);
3150 
3151         update_tg_cfs_util(cfs_rq, se);
3152         update_tg_cfs_load(cfs_rq, se);
3153 
3154         return 1;
3155 }
3156 
3157 #else /* CONFIG_FAIR_GROUP_SCHED */
3158 
3159 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3160 
3161 static inline int propagate_entity_load_avg(struct sched_entity *se)
3162 {
3163         return 0;
3164 }
3165 
3166 static inline void set_tg_cfs_propagate(struct cfs_rq *cfs_rq) {}
3167 
3168 #endif /* CONFIG_FAIR_GROUP_SCHED */
3169 
3170 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
3171 {
3172         if (&this_rq()->cfs == cfs_rq) {
3173                 /*
3174                  * There are a few boundary cases this might miss but it should
3175                  * get called often enough that that should (hopefully) not be
3176                  * a real problem -- added to that it only calls on the local
3177                  * CPU, so if we enqueue remotely we'll miss an update, but
3178                  * the next tick/schedule should update.
3179                  *
3180                  * It will not get called when we go idle, because the idle
3181                  * thread is a different class (!fair), nor will the utilization
3182                  * number include things like RT tasks.
3183                  *
3184                  * As is, the util number is not freq-invariant (we'd have to
3185                  * implement arch_scale_freq_capacity() for that).
3186                  *
3187                  * See cpu_util().
3188                  */
3189                 cpufreq_update_util(rq_of(cfs_rq), 0);
3190         }
3191 }
3192 
3193 /*
3194  * Unsigned subtract and clamp on underflow.
3195  *
3196  * Explicitly do a load-store to ensure the intermediate value never hits
3197  * memory. This allows lockless observations without ever seeing the negative
3198  * values.
3199  */
3200 #define sub_positive(_ptr, _val) do {                           \
3201         typeof(_ptr) ptr = (_ptr);                              \
3202         typeof(*ptr) val = (_val);                              \
3203         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
3204         res = var - val;                                        \
3205         if (res > var)                                          \
3206                 res = 0;                                        \
3207         WRITE_ONCE(*ptr, res);                                  \
3208 } while (0)
3209 
3210 /**
3211  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3212  * @now: current time, as per cfs_rq_clock_task()
3213  * @cfs_rq: cfs_rq to update
3214  * @update_freq: should we call cfs_rq_util_change() or will the call do so
3215  *
3216  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3217  * avg. The immediate corollary is that all (fair) tasks must be attached, see
3218  * post_init_entity_util_avg().
3219  *
3220  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3221  *
3222  * Returns true if the load decayed or we removed load.
3223  *
3224  * Since both these conditions indicate a changed cfs_rq->avg.load we should
3225  * call update_tg_load_avg() when this function returns true.
3226  */
3227 static inline int
3228 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3229 {
3230         struct sched_avg *sa = &cfs_rq->avg;
3231         int decayed, removed_load = 0, removed_util = 0;
3232 
3233         if (atomic_long_read(&cfs_rq->removed_load_avg)) {
3234                 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
3235                 sub_positive(&sa->load_avg, r);
3236                 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
3237                 removed_load = 1;
3238                 set_tg_cfs_propagate(cfs_rq);
3239         }
3240 
3241         if (atomic_long_read(&cfs_rq->removed_util_avg)) {
3242                 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
3243                 sub_positive(&sa->util_avg, r);
3244                 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
3245                 removed_util = 1;
3246                 set_tg_cfs_propagate(cfs_rq);
3247         }
3248 
3249         decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3250                 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
3251 
3252 #ifndef CONFIG_64BIT
3253         smp_wmb();
3254         cfs_rq->load_last_update_time_copy = sa->last_update_time;
3255 #endif
3256 
3257         if (update_freq && (decayed || removed_util))
3258                 cfs_rq_util_change(cfs_rq);
3259 
3260         return decayed || removed_load;
3261 }
3262 
3263 /*
3264  * Optional action to be done while updating the load average
3265  */
3266 #define UPDATE_TG       0x1
3267 #define SKIP_AGE_LOAD   0x2
3268 
3269 /* Update task and its cfs_rq load average */
3270 static inline void update_load_avg(struct sched_entity *se, int flags)
3271 {
3272         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3273         u64 now = cfs_rq_clock_task(cfs_rq);
3274         struct rq *rq = rq_of(cfs_rq);
3275         int cpu = cpu_of(rq);
3276         int decayed;
3277 
3278         /*
3279          * Track task load average for carrying it to new CPU after migrated, and
3280          * track group sched_entity load average for task_h_load calc in migration
3281          */
3282         if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD)) {
3283                 __update_load_avg(now, cpu, &se->avg,
3284                           se->on_rq * scale_load_down(se->load.weight),
3285                           cfs_rq->curr == se, NULL);
3286         }
3287 
3288         decayed  = update_cfs_rq_load_avg(now, cfs_rq, true);
3289         decayed |= propagate_entity_load_avg(se);
3290 
3291         if (decayed && (flags & UPDATE_TG))
3292                 update_tg_load_avg(cfs_rq, 0);
3293 }
3294 
3295 /**
3296  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3297  * @cfs_rq: cfs_rq to attach to
3298  * @se: sched_entity to attach
3299  *
3300  * Must call update_cfs_rq_load_avg() before this, since we rely on
3301  * cfs_rq->avg.last_update_time being current.
3302  */
3303 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3304 {
3305         se->avg.last_update_time = cfs_rq->avg.last_update_time;
3306         cfs_rq->avg.load_avg += se->avg.load_avg;
3307         cfs_rq->avg.load_sum += se->avg.load_sum;
3308         cfs_rq->avg.util_avg += se->avg.util_avg;
3309         cfs_rq->avg.util_sum += se->avg.util_sum;
3310         set_tg_cfs_propagate(cfs_rq);
3311 
3312         cfs_rq_util_change(cfs_rq);
3313 }
3314 
3315 /**
3316  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3317  * @cfs_rq: cfs_rq to detach from
3318  * @se: sched_entity to detach
3319  *
3320  * Must call update_cfs_rq_load_avg() before this, since we rely on
3321  * cfs_rq->avg.last_update_time being current.
3322  */
3323 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3324 {
3325 
3326         sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3327         sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3328         sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3329         sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3330         set_tg_cfs_propagate(cfs_rq);
3331 
3332         cfs_rq_util_change(cfs_rq);
3333 }
3334 
3335 /* Add the load generated by se into cfs_rq's load average */
3336 static inline void
3337 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3338 {
3339         struct sched_avg *sa = &se->avg;
3340 
3341         cfs_rq->runnable_load_avg += sa->load_avg;
3342         cfs_rq->runnable_load_sum += sa->load_sum;
3343 
3344         if (!sa->last_update_time) {
3345                 attach_entity_load_avg(cfs_rq, se);
3346                 update_tg_load_avg(cfs_rq, 0);
3347         }
3348 }
3349 
3350 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3351 static inline void
3352 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3353 {
3354         cfs_rq->runnable_load_avg =
3355                 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3356         cfs_rq->runnable_load_sum =
3357                 max_t(s64,  cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3358 }
3359 
3360 #ifndef CONFIG_64BIT
3361 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3362 {
3363         u64 last_update_time_copy;
3364         u64 last_update_time;
3365 
3366         do {
3367                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3368                 smp_rmb();
3369                 last_update_time = cfs_rq->avg.last_update_time;
3370         } while (last_update_time != last_update_time_copy);
3371 
3372         return last_update_time;
3373 }
3374 #else
3375 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3376 {
3377         return cfs_rq->avg.last_update_time;
3378 }
3379 #endif
3380 
3381 /*
3382  * Synchronize entity load avg of dequeued entity without locking
3383  * the previous rq.
3384  */
3385 void sync_entity_load_avg(struct sched_entity *se)
3386 {
3387         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3388         u64 last_update_time;
3389 
3390         last_update_time = cfs_rq_last_update_time(cfs_rq);
3391         __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3392 }
3393 
3394 /*
3395  * Task first catches up with cfs_rq, and then subtract
3396  * itself from the cfs_rq (task must be off the queue now).
3397  */
3398 void remove_entity_load_avg(struct sched_entity *se)
3399 {
3400         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3401 
3402         /*
3403          * tasks cannot exit without having gone through wake_up_new_task() ->
3404          * post_init_entity_util_avg() which will have added things to the
3405          * cfs_rq, so we can remove unconditionally.
3406          *
3407          * Similarly for groups, they will have passed through
3408          * post_init_entity_util_avg() before unregister_sched_fair_group()
3409          * calls this.
3410          */
3411 
3412         sync_entity_load_avg(se);
3413         atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3414         atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3415 }
3416 
3417 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3418 {
3419         return cfs_rq->runnable_load_avg;
3420 }
3421 
3422 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3423 {
3424         return cfs_rq->avg.load_avg;
3425 }
3426 
3427 static int idle_balance(struct rq *this_rq);
3428 
3429 #else /* CONFIG_SMP */
3430 
3431 static inline int
3432 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3433 {
3434         return 0;
3435 }
3436 
3437 #define UPDATE_TG       0x0
3438 #define SKIP_AGE_LOAD   0x0
3439 
3440 static inline void update_load_avg(struct sched_entity *se, int not_used1)
3441 {
3442         cpufreq_update_util(rq_of(cfs_rq_of(se)), 0);
3443 }
3444 
3445 static inline void
3446 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3447 static inline void
3448 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3449 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3450 
3451 static inline void
3452 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3453 static inline void
3454 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3455 
3456 static inline int idle_balance(struct rq *rq)
3457 {
3458         return 0;
3459 }
3460 
3461 #endif /* CONFIG_SMP */
3462 
3463 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3464 {
3465 #ifdef CONFIG_SCHED_DEBUG
3466         s64 d = se->vruntime - cfs_rq->min_vruntime;
3467 
3468         if (d < 0)
3469                 d = -d;
3470 
3471         if (d > 3*sysctl_sched_latency)
3472                 schedstat_inc(cfs_rq->nr_spread_over);
3473 #endif
3474 }
3475 
3476 static void
3477 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3478 {
3479         u64 vruntime = cfs_rq->min_vruntime;
3480 
3481         /*
3482          * The 'current' period is already promised to the current tasks,
3483          * however the extra weight of the new task will slow them down a
3484          * little, place the new task so that it fits in the slot that
3485          * stays open at the end.
3486          */
3487         if (initial && sched_feat(START_DEBIT))
3488                 vruntime += sched_vslice(cfs_rq, se);
3489 
3490         /* sleeps up to a single latency don't count. */
3491         if (!initial) {
3492                 unsigned long thresh = sysctl_sched_latency;
3493 
3494                 /*
3495                  * Halve their sleep time's effect, to allow
3496                  * for a gentler effect of sleepers:
3497                  */
3498                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3499                         thresh >>= 1;
3500 
3501                 vruntime -= thresh;
3502         }
3503 
3504         /* ensure we never gain time by being placed backwards. */
3505         se->vruntime = max_vruntime(se->vruntime, vruntime);
3506 }
3507 
3508 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3509 
3510 static inline void check_schedstat_required(void)
3511 {
3512 #ifdef CONFIG_SCHEDSTATS
3513         if (schedstat_enabled())
3514                 return;
3515 
3516         /* Force schedstat enabled if a dependent tracepoint is active */
3517         if (trace_sched_stat_wait_enabled()    ||
3518                         trace_sched_stat_sleep_enabled()   ||
3519                         trace_sched_stat_iowait_enabled()  ||
3520                         trace_sched_stat_blocked_enabled() ||
3521                         trace_sched_stat_runtime_enabled())  {
3522                 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3523                              "stat_blocked and stat_runtime require the "
3524                              "kernel parameter schedstats=enabled or "
3525                              "kernel.sched_schedstats=1\n");
3526         }
3527 #endif
3528 }
3529 
3530 
3531 /*
3532  * MIGRATION
3533  *
3534  *      dequeue
3535  *        update_curr()
3536  *          update_min_vruntime()
3537  *        vruntime -= min_vruntime
3538  *
3539  *      enqueue
3540  *        update_curr()
3541  *          update_min_vruntime()
3542  *        vruntime += min_vruntime
3543  *
3544  * this way the vruntime transition between RQs is done when both
3545  * min_vruntime are up-to-date.
3546  *
3547  * WAKEUP (remote)
3548  *
3549  *      ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3550  *        vruntime -= min_vruntime
3551  *
3552  *      enqueue
3553  *        update_curr()
3554  *          update_min_vruntime()
3555  *        vruntime += min_vruntime
3556  *
3557  * this way we don't have the most up-to-date min_vruntime on the originating
3558  * CPU and an up-to-date min_vruntime on the destination CPU.
3559  */
3560 
3561 static void
3562 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3563 {
3564         bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3565         bool curr = cfs_rq->curr == se;
3566 
3567         /*
3568          * If we're the current task, we must renormalise before calling
3569          * update_curr().
3570          */
3571         if (renorm && curr)
3572                 se->vruntime += cfs_rq->min_vruntime;
3573 
3574         update_curr(cfs_rq);
3575 
3576         /*
3577          * Otherwise, renormalise after, such that we're placed at the current
3578          * moment in time, instead of some random moment in the past. Being
3579          * placed in the past could significantly boost this task to the
3580          * fairness detriment of existing tasks.
3581          */
3582         if (renorm && !curr)
3583                 se->vruntime += cfs_rq->min_vruntime;
3584 
3585         update_load_avg(se, UPDATE_TG);
3586         enqueue_entity_load_avg(cfs_rq, se);
3587         account_entity_enqueue(cfs_rq, se);
3588         update_cfs_shares(cfs_rq);
3589 
3590         if (flags & ENQUEUE_WAKEUP)
3591                 place_entity(cfs_rq, se, 0);
3592 
3593         check_schedstat_required();
3594         update_stats_enqueue(cfs_rq, se, flags);
3595         check_spread(cfs_rq, se);
3596         if (!curr)
3597                 __enqueue_entity(cfs_rq, se);
3598         se->on_rq = 1;
3599 
3600         if (cfs_rq->nr_running == 1) {
3601                 list_add_leaf_cfs_rq(cfs_rq);
3602                 check_enqueue_throttle(cfs_rq);
3603         }
3604 }
3605 
3606 static void __clear_buddies_last(struct sched_entity *se)
3607 {
3608         for_each_sched_entity(se) {
3609                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3610                 if (cfs_rq->last != se)
3611                         break;
3612 
3613                 cfs_rq->last = NULL;
3614         }
3615 }
3616 
3617 static void __clear_buddies_next(struct sched_entity *se)
3618 {
3619         for_each_sched_entity(se) {
3620                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3621                 if (cfs_rq->next != se)
3622                         break;
3623 
3624                 cfs_rq->next = NULL;
3625         }
3626 }
3627 
3628 static void __clear_buddies_skip(struct sched_entity *se)
3629 {
3630         for_each_sched_entity(se) {
3631                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3632                 if (cfs_rq->skip != se)
3633                         break;
3634 
3635                 cfs_rq->skip = NULL;
3636         }
3637 }
3638 
3639 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3640 {
3641         if (cfs_rq->last == se)
3642                 __clear_buddies_last(se);
3643 
3644         if (cfs_rq->next == se)
3645                 __clear_buddies_next(se);
3646 
3647         if (cfs_rq->skip == se)
3648                 __clear_buddies_skip(se);
3649 }
3650 
3651 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3652 
3653 static void
3654 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3655 {
3656         /*
3657          * Update run-time statistics of the 'current'.
3658          */
3659         update_curr(cfs_rq);
3660         update_load_avg(se, UPDATE_TG);
3661         dequeue_entity_load_avg(cfs_rq, se);
3662 
3663         update_stats_dequeue(cfs_rq, se, flags);
3664 
3665         clear_buddies(cfs_rq, se);
3666 
3667         if (se != cfs_rq->curr)
3668                 __dequeue_entity(cfs_rq, se);
3669         se->on_rq = 0;
3670         account_entity_dequeue(cfs_rq, se);
3671 
3672         /*
3673          * Normalize after update_curr(); which will also have moved
3674          * min_vruntime if @se is the one holding it back. But before doing
3675          * update_min_vruntime() again, which will discount @se's position and
3676          * can move min_vruntime forward still more.
3677          */
3678         if (!(flags & DEQUEUE_SLEEP))
3679                 se->vruntime -= cfs_rq->min_vruntime;
3680 
3681         /* return excess runtime on last dequeue */
3682         return_cfs_rq_runtime(cfs_rq);
3683 
3684         update_cfs_shares(cfs_rq);
3685 
3686         /*
3687          * Now advance min_vruntime if @se was the entity holding it back,
3688          * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3689          * put back on, and if we advance min_vruntime, we'll be placed back
3690          * further than we started -- ie. we'll be penalized.
3691          */
3692         if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3693                 update_min_vruntime(cfs_rq);
3694 }
3695 
3696 /*
3697  * Preempt the current task with a newly woken task if needed:
3698  */
3699 static void
3700 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3701 {
3702         unsigned long ideal_runtime, delta_exec;
3703         struct sched_entity *se;
3704         s64 delta;
3705 
3706         ideal_runtime = sched_slice(cfs_rq, curr);
3707         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3708         if (delta_exec > ideal_runtime) {
3709                 resched_curr(rq_of(cfs_rq));
3710                 /*
3711                  * The current task ran long enough, ensure it doesn't get
3712                  * re-elected due to buddy favours.
3713                  */
3714                 clear_buddies(cfs_rq, curr);
3715                 return;
3716         }
3717 
3718         /*
3719          * Ensure that a task that missed wakeup preemption by a
3720          * narrow margin doesn't have to wait for a full slice.
3721          * This also mitigates buddy induced latencies under load.
3722          */
3723         if (delta_exec < sysctl_sched_min_granularity)
3724                 return;
3725 
3726         se = __pick_first_entity(cfs_rq);
3727         delta = curr->vruntime - se->vruntime;
3728 
3729         if (delta < 0)
3730                 return;
3731 
3732         if (delta > ideal_runtime)
3733                 resched_curr(rq_of(cfs_rq));
3734 }
3735 
3736 static void
3737 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3738 {
3739         /* 'current' is not kept within the tree. */
3740         if (se->on_rq) {
3741                 /*
3742                  * Any task has to be enqueued before it get to execute on
3743                  * a CPU. So account for the time it spent waiting on the
3744                  * runqueue.
3745                  */
3746                 update_stats_wait_end(cfs_rq, se);
3747                 __dequeue_entity(cfs_rq, se);
3748                 update_load_avg(se, UPDATE_TG);
3749         }
3750 
3751         update_stats_curr_start(cfs_rq, se);
3752         cfs_rq->curr = se;
3753 
3754         /*
3755          * Track our maximum slice length, if the CPU's load is at
3756          * least twice that of our own weight (i.e. dont track it
3757          * when there are only lesser-weight tasks around):
3758          */
3759         if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3760                 schedstat_set(se->statistics.slice_max,
3761                         max((u64)schedstat_val(se->statistics.slice_max),
3762                             se->sum_exec_runtime - se->prev_sum_exec_runtime));
3763         }
3764 
3765         se->prev_sum_exec_runtime = se->sum_exec_runtime;
3766 }
3767 
3768 static int
3769 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3770 
3771 /*
3772  * Pick the next process, keeping these things in mind, in this order:
3773  * 1) keep things fair between processes/task groups
3774  * 2) pick the "next" process, since someone really wants that to run
3775  * 3) pick the "last" process, for cache locality
3776  * 4) do not run the "skip" process, if something else is available
3777  */
3778 static struct sched_entity *
3779 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3780 {
3781         struct sched_entity *left = __pick_first_entity(cfs_rq);
3782         struct sched_entity *se;
3783 
3784         /*
3785          * If curr is set we have to see if its left of the leftmost entity
3786          * still in the tree, provided there was anything in the tree at all.
3787          */
3788         if (!left || (curr && entity_before(curr, left)))
3789                 left = curr;
3790 
3791         se = left; /* ideally we run the leftmost entity */
3792 
3793         /*
3794          * Avoid running the skip buddy, if running something else can
3795          * be done without getting too unfair.
3796          */
3797         if (cfs_rq->skip == se) {
3798                 struct sched_entity *second;
3799 
3800                 if (se == curr) {
3801                         second = __pick_first_entity(cfs_rq);
3802                 } else {
3803                         second = __pick_next_entity(se);
3804                         if (!second || (curr && entity_before(curr, second)))
3805                                 second = curr;
3806                 }
3807 
3808                 if (second && wakeup_preempt_entity(second, left) < 1)
3809                         se = second;
3810         }
3811 
3812         /*
3813          * Prefer last buddy, try to return the CPU to a preempted task.
3814          */
3815         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3816                 se = cfs_rq->last;
3817 
3818         /*
3819          * Someone really wants this to run. If it's not unfair, run it.
3820          */
3821         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3822                 se = cfs_rq->next;
3823 
3824         clear_buddies(cfs_rq, se);
3825 
3826         return se;
3827 }
3828 
3829 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3830 
3831 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3832 {
3833         /*
3834          * If still on the runqueue then deactivate_task()
3835          * was not called and update_curr() has to be done:
3836          */
3837         if (prev->on_rq)
3838                 update_curr(cfs_rq);
3839 
3840         /* throttle cfs_rqs exceeding runtime */
3841         check_cfs_rq_runtime(cfs_rq);
3842 
3843         check_spread(cfs_rq, prev);
3844 
3845         if (prev->on_rq) {
3846                 update_stats_wait_start(cfs_rq, prev);
3847                 /* Put 'current' back into the tree. */
3848                 __enqueue_entity(cfs_rq, prev);
3849                 /* in !on_rq case, update occurred at dequeue */
3850                 update_load_avg(prev, 0);
3851         }
3852         cfs_rq->curr = NULL;
3853 }
3854 
3855 static void
3856 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3857 {
3858         /*
3859          * Update run-time statistics of the 'current'.
3860          */
3861         update_curr(cfs_rq);
3862 
3863         /*
3864          * Ensure that runnable average is periodically updated.
3865          */
3866         update_load_avg(curr, UPDATE_TG);
3867         update_cfs_shares(cfs_rq);
3868 
3869 #ifdef CONFIG_SCHED_HRTICK
3870         /*
3871          * queued ticks are scheduled to match the slice, so don't bother
3872          * validating it and just reschedule.
3873          */
3874         if (queued) {
3875                 resched_curr(rq_of(cfs_rq));
3876                 return;
3877         }
3878         /*
3879          * don't let the period tick interfere with the hrtick preemption
3880          */
3881         if (!sched_feat(DOUBLE_TICK) &&
3882                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3883                 return;
3884 #endif
3885 
3886         if (cfs_rq->nr_running > 1)
3887                 check_preempt_tick(cfs_rq, curr);
3888 }
3889 
3890 
3891 /**************************************************
3892  * CFS bandwidth control machinery
3893  */
3894 
3895 #ifdef CONFIG_CFS_BANDWIDTH
3896 
3897 #ifdef HAVE_JUMP_LABEL
3898 static struct static_key __cfs_bandwidth_used;
3899 
3900 static inline bool cfs_bandwidth_used(void)
3901 {
3902         return static_key_false(&__cfs_bandwidth_used);
3903 }
3904 
3905 void cfs_bandwidth_usage_inc(void)
3906 {
3907         static_key_slow_inc(&__cfs_bandwidth_used);
3908 }
3909 
3910 void cfs_bandwidth_usage_dec(void)
3911 {
3912         static_key_slow_dec(&__cfs_bandwidth_used);
3913 }
3914 #else /* HAVE_JUMP_LABEL */
3915 static bool cfs_bandwidth_used(void)
3916 {
3917         return true;
3918 }
3919 
3920 void cfs_bandwidth_usage_inc(void) {}
3921 void cfs_bandwidth_usage_dec(void) {}
3922 #endif /* HAVE_JUMP_LABEL */
3923 
3924 /*
3925  * default period for cfs group bandwidth.
3926  * default: 0.1s, units: nanoseconds
3927  */
3928 static inline u64 default_cfs_period(void)
3929 {
3930         return 100000000ULL;
3931 }
3932 
3933 static inline u64 sched_cfs_bandwidth_slice(void)
3934 {
3935         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3936 }
3937 
3938 /*
3939  * Replenish runtime according to assigned quota and update expiration time.
3940  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3941  * additional synchronization around rq->lock.
3942  *
3943  * requires cfs_b->lock
3944  */
3945 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3946 {
3947         u64 now;
3948 
3949         if (cfs_b->quota == RUNTIME_INF)
3950                 return;
3951 
3952         now = sched_clock_cpu(smp_processor_id());
3953         cfs_b->runtime = cfs_b->quota;
3954         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3955 }
3956 
3957 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3958 {
3959         return &tg->cfs_bandwidth;
3960 }
3961 
3962 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3963 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3964 {
3965         if (unlikely(cfs_rq->throttle_count))
3966                 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3967 
3968         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3969 }
3970 
3971 /* returns 0 on failure to allocate runtime */
3972 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3973 {
3974         struct task_group *tg = cfs_rq->tg;
3975         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3976         u64 amount = 0, min_amount, expires;
3977 
3978         /* note: this is a positive sum as runtime_remaining <= 0 */
3979         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3980 
3981         raw_spin_lock(&cfs_b->lock);
3982         if (cfs_b->quota == RUNTIME_INF)
3983                 amount = min_amount;
3984         else {
3985                 start_cfs_bandwidth(cfs_b);
3986 
3987                 if (cfs_b->runtime > 0) {
3988                         amount = min(cfs_b->runtime, min_amount);
3989                         cfs_b->runtime -= amount;
3990                         cfs_b->idle = 0;
3991                 }
3992         }
3993         expires = cfs_b->runtime_expires;
3994         raw_spin_unlock(&cfs_b->lock);
3995 
3996         cfs_rq->runtime_remaining += amount;
3997         /*
3998          * we may have advanced our local expiration to account for allowed
3999          * spread between our sched_clock and the one on which runtime was
4000          * issued.
4001          */
4002         if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4003                 cfs_rq->runtime_expires = expires;
4004 
4005         return cfs_rq->runtime_remaining > 0;
4006 }
4007 
4008 /*
4009  * Note: This depends on the synchronization provided by sched_clock and the
4010  * fact that rq->clock snapshots this value.
4011  */
4012 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4013 {
4014         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4015 
4016         /* if the deadline is ahead of our clock, nothing to do */
4017         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4018                 return;
4019 
4020         if (cfs_rq->runtime_remaining < 0)
4021                 return;
4022 
4023         /*
4024          * If the local deadline has passed we have to consider the
4025          * possibility that our sched_clock is 'fast' and the global deadline
4026          * has not truly expired.
4027          *
4028          * Fortunately we can check determine whether this the case by checking
4029          * whether the global deadline has advanced. It is valid to compare
4030          * cfs_b->runtime_expires without any locks since we only care about
4031          * exact equality, so a partial write will still work.
4032          */
4033 
4034         if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4035                 /* extend local deadline, drift is bounded above by 2 ticks */
4036                 cfs_rq->runtime_expires += TICK_NSEC;
4037         } else {
4038                 /* global deadline is ahead, expiration has passed */
4039                 cfs_rq->runtime_remaining = 0;
4040         }
4041 }
4042 
4043 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4044 {
4045         /* dock delta_exec before expiring quota (as it could span periods) */
4046         cfs_rq->runtime_remaining -= delta_exec;
4047         expire_cfs_rq_runtime(cfs_rq);
4048 
4049         if (likely(cfs_rq->runtime_remaining > 0))
4050                 return;
4051 
4052         /*
4053          * if we're unable to extend our runtime we resched so that the active
4054          * hierarchy can be throttled
4055          */
4056         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4057                 resched_curr(rq_of(cfs_rq));
4058 }
4059 
4060 static __always_inline
4061 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4062 {
4063         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4064                 return;
4065 
4066         __account_cfs_rq_runtime(cfs_rq, delta_exec);
4067 }
4068 
4069 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4070 {
4071         return cfs_bandwidth_used() && cfs_rq->throttled;
4072 }
4073 
4074 /* check whether cfs_rq, or any parent, is throttled */
4075 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4076 {
4077         return cfs_bandwidth_used() && cfs_rq->throttle_count;
4078 }
4079 
4080 /*
4081  * Ensure that neither of the group entities corresponding to src_cpu or
4082  * dest_cpu are members of a throttled hierarchy when performing group
4083  * load-balance operations.
4084  */
4085 static inline int throttled_lb_pair(struct task_group *tg,
4086                                     int src_cpu, int dest_cpu)
4087 {
4088         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4089 
4090         src_cfs_rq = tg->cfs_rq[src_cpu];
4091         dest_cfs_rq = tg->cfs_rq[dest_cpu];
4092 
4093         return throttled_hierarchy(src_cfs_rq) ||
4094                throttled_hierarchy(dest_cfs_rq);
4095 }
4096 
4097 /* updated child weight may affect parent so we have to do this bottom up */
4098 static int tg_unthrottle_up(struct task_group *tg, void *data)
4099 {
4100         struct rq *rq = data;
4101         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4102 
4103         cfs_rq->throttle_count--;
4104         if (!cfs_rq->throttle_count) {
4105                 /* adjust cfs_rq_clock_task() */
4106                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4107                                              cfs_rq->throttled_clock_task;
4108         }
4109 
4110         return 0;
4111 }
4112 
4113 static int tg_throttle_down(struct task_group *tg, void *data)
4114 {
4115         struct rq *rq = data;
4116         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4117 
4118         /* group is entering throttled state, stop time */
4119         if (!cfs_rq->throttle_count)
4120                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4121         cfs_rq->throttle_count++;
4122 
4123         return 0;
4124 }
4125 
4126 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4127 {
4128         struct rq *rq = rq_of(cfs_rq);
4129         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4130         struct sched_entity *se;
4131         long task_delta, dequeue = 1;
4132         bool empty;
4133 
4134         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4135 
4136         /* freeze hierarchy runnable averages while throttled */
4137         rcu_read_lock();
4138         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4139         rcu_read_unlock();
4140 
4141         task_delta = cfs_rq->h_nr_running;
4142         for_each_sched_entity(se) {
4143                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4144                 /* throttled entity or throttle-on-deactivate */
4145                 if (!se->on_rq)
4146                         break;
4147 
4148                 if (dequeue)
4149                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4150                 qcfs_rq->h_nr_running -= task_delta;
4151 
4152                 if (qcfs_rq->load.weight)
4153                         dequeue = 0;
4154         }
4155 
4156         if (!se)
4157                 sub_nr_running(rq, task_delta);
4158 
4159         cfs_rq->throttled = 1;
4160         cfs_rq->throttled_clock = rq_clock(rq);
4161         raw_spin_lock(&cfs_b->lock);
4162         empty = list_empty(&cfs_b->throttled_cfs_rq);
4163 
4164         /*
4165          * Add to the _head_ of the list, so that an already-started
4166          * distribute_cfs_runtime will not see us
4167          */
4168         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4169 
4170         /*
4171          * If we're the first throttled task, make sure the bandwidth
4172          * timer is running.
4173          */
4174         if (empty)
4175                 start_cfs_bandwidth(cfs_b);
4176 
4177         raw_spin_unlock(&cfs_b->lock);
4178 }
4179 
4180 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4181 {
4182         struct rq *rq = rq_of(cfs_rq);
4183         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4184         struct sched_entity *se;
4185         int enqueue = 1;
4186         long task_delta;
4187 
4188         se = cfs_rq->tg->se[cpu_of(rq)];
4189 
4190         cfs_rq->throttled = 0;
4191 
4192         update_rq_clock(rq);
4193 
4194         raw_spin_lock(&cfs_b->lock);
4195         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4196         list_del_rcu(&cfs_rq->throttled_list);
4197         raw_spin_unlock(&cfs_b->lock);
4198 
4199         /* update hierarchical throttle state */
4200         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4201 
4202         if (!cfs_rq->load.weight)
4203                 return;
4204 
4205         task_delta = cfs_rq->h_nr_running;
4206         for_each_sched_entity(se) {
4207                 if (se->on_rq)
4208                         enqueue = 0;
4209 
4210                 cfs_rq = cfs_rq_of(se);
4211                 if (enqueue)
4212                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4213                 cfs_rq->h_nr_running += task_delta;
4214 
4215                 if (cfs_rq_throttled(cfs_rq))
4216                         break;
4217         }
4218 
4219         if (!se)
4220                 add_nr_running(rq, task_delta);
4221 
4222         /* determine whether we need to wake up potentially idle cpu */
4223         if (rq->curr == rq->idle && rq->cfs.nr_running)
4224                 resched_curr(rq);
4225 }
4226 
4227 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4228                 u64 remaining, u64 expires)
4229 {
4230         struct cfs_rq *cfs_rq;
4231         u64 runtime;
4232         u64 starting_runtime = remaining;
4233 
4234         rcu_read_lock();
4235         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4236                                 throttled_list) {
4237                 struct rq *rq = rq_of(cfs_rq);
4238 
4239                 raw_spin_lock(&rq->lock);
4240                 if (!cfs_rq_throttled(cfs_rq))
4241                         goto next;
4242 
4243                 runtime = -cfs_rq->runtime_remaining + 1;
4244                 if (runtime > remaining)
4245                         runtime = remaining;
4246                 remaining -= runtime;
4247 
4248                 cfs_rq->runtime_remaining += runtime;
4249                 cfs_rq->runtime_expires = expires;
4250 
4251                 /* we check whether we're throttled above */
4252                 if (cfs_rq->runtime_remaining > 0)
4253                         unthrottle_cfs_rq(cfs_rq);
4254 
4255 next:
4256                 raw_spin_unlock(&rq->lock);
4257 
4258                 if (!remaining)
4259                         break;
4260         }
4261         rcu_read_unlock();
4262 
4263         return starting_runtime - remaining;
4264 }
4265 
4266 /*
4267  * Responsible for refilling a task_group's bandwidth and unthrottling its
4268  * cfs_rqs as appropriate. If there has been no activity within the last
4269  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4270  * used to track this state.
4271  */
4272 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4273 {
4274         u64 runtime, runtime_expires;
4275         int throttled;
4276 
4277         /* no need to continue the timer with no bandwidth constraint */
4278         if (cfs_b->quota == RUNTIME_INF)
4279                 goto out_deactivate;
4280 
4281         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4282         cfs_b->nr_periods += overrun;
4283 
4284         /*
4285          * idle depends on !throttled (for the case of a large deficit), and if
4286          * we're going inactive then everything else can be deferred
4287          */
4288         if (cfs_b->idle && !throttled)
4289                 goto out_deactivate;
4290 
4291         __refill_cfs_bandwidth_runtime(cfs_b);
4292 
4293         if (!throttled) {
4294                 /* mark as potentially idle for the upcoming period */
4295                 cfs_b->idle = 1;
4296                 return 0;
4297         }
4298 
4299         /* account preceding periods in which throttling occurred */
4300         cfs_b->nr_throttled += overrun;
4301 
4302         runtime_expires = cfs_b->runtime_expires;
4303 
4304         /*
4305          * This check is repeated as we are holding onto the new bandwidth while
4306          * we unthrottle. This can potentially race with an unthrottled group
4307          * trying to acquire new bandwidth from the global pool. This can result
4308          * in us over-using our runtime if it is all used during this loop, but
4309          * only by limited amounts in that extreme case.
4310          */
4311         while (throttled && cfs_b->runtime > 0) {
4312                 runtime = cfs_b->runtime;
4313                 raw_spin_unlock(&cfs_b->lock);
4314                 /* we can't nest cfs_b->lock while distributing bandwidth */
4315                 runtime = distribute_cfs_runtime(cfs_b, runtime,
4316                                                  runtime_expires);
4317                 raw_spin_lock(&cfs_b->lock);
4318 
4319                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4320 
4321                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4322         }
4323 
4324         /*
4325          * While we are ensured activity in the period following an
4326          * unthrottle, this also covers the case in which the new bandwidth is
4327          * insufficient to cover the existing bandwidth deficit.  (Forcing the
4328          * timer to remain active while there are any throttled entities.)
4329          */
4330         cfs_b->idle = 0;
4331 
4332         return 0;
4333 
4334 out_deactivate:
4335         return 1;
4336 }
4337 
4338 /* a cfs_rq won't donate quota below this amount */
4339 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4340 /* minimum remaining period time to redistribute slack quota */
4341 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4342 /* how long we wait to gather additional slack before distributing */
4343 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4344 
4345 /*
4346  * Are we near the end of the current quota period?
4347  *
4348  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4349  * hrtimer base being cleared by hrtimer_start. In the case of
4350  * migrate_hrtimers, base is never cleared, so we are fine.
4351  */
4352 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4353 {
4354         struct hrtimer *refresh_timer = &cfs_b->period_timer;
4355         u64 remaining;
4356 
4357         /* if the call-back is running a quota refresh is already occurring */
4358         if (hrtimer_callback_running(refresh_timer))
4359                 return 1;
4360 
4361         /* is a quota refresh about to occur? */
4362         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4363         if (remaining < min_expire)
4364                 return 1;
4365 
4366         return 0;
4367 }
4368 
4369 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4370 {
4371         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4372 
4373         /* if there's a quota refresh soon don't bother with slack */
4374         if (runtime_refresh_within(cfs_b, min_left))
4375                 return;
4376 
4377         hrtimer_start(&cfs_b->slack_timer,
4378                         ns_to_ktime(cfs_bandwidth_slack_period),
4379                         HRTIMER_MODE_REL);
4380 }
4381 
4382 /* we know any runtime found here is valid as update_curr() precedes return */
4383 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4384 {
4385         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4386         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4387 
4388         if (slack_runtime <= 0)
4389                 return;
4390 
4391         raw_spin_lock(&cfs_b->lock);
4392         if (cfs_b->quota != RUNTIME_INF &&
4393             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4394                 cfs_b->runtime += slack_runtime;
4395 
4396                 /* we are under rq->lock, defer unthrottling using a timer */
4397                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4398                     !list_empty(&cfs_b->throttled_cfs_rq))
4399                         start_cfs_slack_bandwidth(cfs_b);
4400         }
4401         raw_spin_unlock(&cfs_b->lock);
4402 
4403         /* even if it's not valid for return we don't want to try again */
4404         cfs_rq->runtime_remaining -= slack_runtime;
4405 }
4406 
4407 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4408 {
4409         if (!cfs_bandwidth_used())
4410                 return;
4411 
4412         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4413                 return;
4414 
4415         __return_cfs_rq_runtime(cfs_rq);
4416 }
4417 
4418 /*
4419  * This is done with a timer (instead of inline with bandwidth return) since
4420  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4421  */
4422 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4423 {
4424         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4425         u64 expires;
4426 
4427         /* confirm we're still not at a refresh boundary */
4428         raw_spin_lock(&cfs_b->lock);
4429         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4430                 raw_spin_unlock(&cfs_b->lock);
4431                 return;
4432         }
4433 
4434         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4435                 runtime = cfs_b->runtime;
4436 
4437         expires = cfs_b->runtime_expires;
4438         raw_spin_unlock(&cfs_b->lock);
4439 
4440         if (!runtime)
4441                 return;
4442 
4443         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4444 
4445         raw_spin_lock(&cfs_b->lock);
4446         if (expires == cfs_b->runtime_expires)
4447                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4448         raw_spin_unlock(&cfs_b->lock);
4449 }
4450 
4451 /*
4452  * When a group wakes up we want to make sure that its quota is not already
4453  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4454  * runtime as update_curr() throttling can not not trigger until it's on-rq.
4455  */
4456 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4457 {
4458         if (!cfs_bandwidth_used())
4459                 return;
4460 
4461         /* an active group must be handled by the update_curr()->put() path */
4462         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4463                 return;
4464 
4465         /* ensure the group is not already throttled */
4466         if (cfs_rq_throttled(cfs_rq))
4467                 return;
4468 
4469         /* update runtime allocation */
4470         account_cfs_rq_runtime(cfs_rq, 0);
4471         if (cfs_rq->runtime_remaining <= 0)
4472                 throttle_cfs_rq(cfs_rq);
4473 }
4474 
4475 static void sync_throttle(struct task_group *tg, int cpu)
4476 {
4477         struct cfs_rq *pcfs_rq, *cfs_rq;
4478 
4479         if (!cfs_bandwidth_used())
4480                 return;
4481 
4482         if (!tg->parent)
4483                 return;
4484 
4485         cfs_rq = tg->cfs_rq[cpu];
4486         pcfs_rq = tg->parent->cfs_rq[cpu];
4487 
4488         cfs_rq->throttle_count = pcfs_rq->throttle_count;
4489         cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4490 }
4491 
4492 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4493 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4494 {
4495         if (!cfs_bandwidth_used())
4496                 return false;
4497 
4498         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4499                 return false;
4500 
4501         /*
4502          * it's possible for a throttled entity to be forced into a running
4503          * state (e.g. set_curr_task), in this case we're finished.
4504          */
4505         if (cfs_rq_throttled(cfs_rq))
4506                 return true;
4507 
4508         throttle_cfs_rq(cfs_rq);
4509         return true;
4510 }
4511 
4512 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4513 {
4514         struct cfs_bandwidth *cfs_b =
4515                 container_of(timer, struct cfs_bandwidth, slack_timer);
4516 
4517         do_sched_cfs_slack_timer(cfs_b);
4518 
4519         return HRTIMER_NORESTART;
4520 }
4521 
4522 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4523 {
4524         struct cfs_bandwidth *cfs_b =
4525                 container_of(timer, struct cfs_bandwidth, period_timer);
4526         int overrun;
4527         int idle = 0;
4528 
4529         raw_spin_lock(&cfs_b->lock);
4530         for (;;) {
4531                 overrun = hrtimer_forward_now(timer, cfs_b->period);
4532                 if (!overrun)
4533                         break;
4534 
4535                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4536         }
4537         if (idle)
4538                 cfs_b->period_active = 0;
4539         raw_spin_unlock(&cfs_b->lock);
4540 
4541         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4542 }
4543 
4544 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4545 {
4546         raw_spin_lock_init(&cfs_b->lock);
4547         cfs_b->runtime = 0;
4548         cfs_b->quota = RUNTIME_INF;
4549         cfs_b->period = ns_to_ktime(default_cfs_period());
4550 
4551         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4552         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4553         cfs_b->period_timer.function = sched_cfs_period_timer;
4554         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4555         cfs_b->slack_timer.function = sched_cfs_slack_timer;
4556 }
4557 
4558 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4559 {
4560         cfs_rq->runtime_enabled = 0;
4561         INIT_LIST_HEAD(&cfs_rq->throttled_list);
4562 }
4563 
4564 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4565 {
4566         lockdep_assert_held(&cfs_b->lock);
4567 
4568         if (!cfs_b->period_active) {
4569                 cfs_b->period_active = 1;
4570                 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4571                 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4572         }
4573 }
4574 
4575 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4576 {
4577         /* init_cfs_bandwidth() was not called */
4578         if (!cfs_b->throttled_cfs_rq.next)
4579                 return;
4580 
4581         hrtimer_cancel(&cfs_b->period_timer);
4582         hrtimer_cancel(&cfs_b->slack_timer);
4583 }
4584 
4585 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4586 {
4587         struct cfs_rq *cfs_rq;
4588 
4589         for_each_leaf_cfs_rq(rq, cfs_rq) {
4590                 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4591 
4592                 raw_spin_lock(&cfs_b->lock);
4593                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4594                 raw_spin_unlock(&cfs_b->lock);
4595         }
4596 }
4597 
4598 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4599 {
4600         struct cfs_rq *cfs_rq;
4601 
4602         for_each_leaf_cfs_rq(rq, cfs_rq) {
4603                 if (!cfs_rq->runtime_enabled)
4604                         continue;
4605 
4606                 /*
4607                  * clock_task is not advancing so we just need to make sure
4608                  * there's some valid quota amount
4609                  */
4610                 cfs_rq->runtime_remaining = 1;
4611                 /*
4612                  * Offline rq is schedulable till cpu is completely disabled
4613                  * in take_cpu_down(), so we prevent new cfs throttling here.
4614                  */
4615                 cfs_rq->runtime_enabled = 0;
4616 
4617                 if (cfs_rq_throttled(cfs_rq))
4618                         unthrottle_cfs_rq(cfs_rq);
4619         }
4620 }
4621 
4622 #else /* CONFIG_CFS_BANDWIDTH */
4623 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4624 {
4625         return rq_clock_task(rq_of(cfs_rq));
4626 }
4627 
4628 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4629 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4630 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4631 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4632 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4633 
4634 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4635 {
4636         return 0;
4637 }
4638 
4639 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4640 {
4641         return 0;
4642 }
4643 
4644 static inline int throttled_lb_pair(struct task_group *tg,
4645                                     int src_cpu, int dest_cpu)
4646 {
4647         return 0;
4648 }
4649 
4650 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4651 
4652 #ifdef CONFIG_FAIR_GROUP_SCHED
4653 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4654 #endif
4655 
4656 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4657 {
4658         return NULL;
4659 }
4660 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4661 static inline void update_runtime_enabled(struct rq *rq) {}
4662 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4663 
4664 #endif /* CONFIG_CFS_BANDWIDTH */
4665 
4666 /**************************************************
4667  * CFS operations on tasks:
4668  */
4669 
4670 #ifdef CONFIG_SCHED_HRTICK
4671 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4672 {
4673         struct sched_entity *se = &p->se;
4674         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4675 
4676         SCHED_WARN_ON(task_rq(p) != rq);
4677 
4678         if (rq->cfs.h_nr_running > 1) {
4679                 u64 slice = sched_slice(cfs_rq, se);
4680                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4681                 s64 delta = slice - ran;
4682 
4683                 if (delta < 0) {
4684                         if (rq->curr == p)
4685                                 resched_curr(rq);
4686                         return;
4687                 }
4688                 hrtick_start(rq, delta);
4689         }
4690 }
4691 
4692 /*
4693  * called from enqueue/dequeue and updates the hrtick when the
4694  * current task is from our class and nr_running is low enough
4695  * to matter.
4696  */
4697 static void hrtick_update(struct rq *rq)
4698 {
4699         struct task_struct *curr = rq->curr;
4700 
4701         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4702                 return;
4703 
4704         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4705                 hrtick_start_fair(rq, curr);
4706 }
4707 #else /* !CONFIG_SCHED_HRTICK */
4708 static inline void
4709 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4710 {
4711 }
4712 
4713 static inline void hrtick_update(struct rq *rq)
4714 {
4715 }
4716 #endif
4717 
4718 /*
4719  * The enqueue_task method is called before nr_running is
4720  * increased. Here we update the fair scheduling stats and
4721  * then put the task into the rbtree:
4722  */
4723 static void
4724 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4725 {
4726         struct cfs_rq *cfs_rq;
4727         struct sched_entity *se = &p->se;
4728 
4729         /*
4730          * If in_iowait is set, the code below may not trigger any cpufreq
4731          * utilization updates, so do it here explicitly with the IOWAIT flag
4732          * passed.
4733          */
4734         if (p->in_iowait)
4735                 cpufreq_update_this_cpu(rq, SCHED_CPUFREQ_IOWAIT);
4736 
4737         for_each_sched_entity(se) {
4738                 if (se->on_rq)
4739                         break;
4740                 cfs_rq = cfs_rq_of(se);
4741                 enqueue_entity(cfs_rq, se, flags);
4742 
4743                 /*
4744                  * end evaluation on encountering a throttled cfs_rq
4745                  *
4746                  * note: in the case of encountering a throttled cfs_rq we will
4747                  * post the final h_nr_running increment below.
4748                  */
4749                 if (cfs_rq_throttled(cfs_rq))
4750                         break;
4751                 cfs_rq->h_nr_running++;
4752 
4753                 flags = ENQUEUE_WAKEUP;
4754         }
4755 
4756         for_each_sched_entity(se) {
4757                 cfs_rq = cfs_rq_of(se);
4758                 cfs_rq->h_nr_running++;
4759 
4760                 if (cfs_rq_throttled(cfs_rq))
4761                         break;
4762 
4763                 update_load_avg(se, UPDATE_TG);
4764                 update_cfs_shares(cfs_rq);
4765         }
4766 
4767         if (!se)
4768                 add_nr_running(rq, 1);
4769 
4770         hrtick_update(rq);
4771 }
4772 
4773 static void set_next_buddy(struct sched_entity *se);
4774 
4775 /*
4776  * The dequeue_task method is called before nr_running is
4777  * decreased. We remove the task from the rbtree and
4778  * update the fair scheduling stats:
4779  */
4780 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4781 {
4782         struct cfs_rq *cfs_rq;
4783         struct sched_entity *se = &p->se;
4784         int task_sleep = flags & DEQUEUE_SLEEP;
4785 
4786         for_each_sched_entity(se) {
4787                 cfs_rq = cfs_rq_of(se);
4788                 dequeue_entity(cfs_rq, se, flags);
4789 
4790                 /*
4791                  * end evaluation on encountering a throttled cfs_rq
4792                  *
4793                  * note: in the case of encountering a throttled cfs_rq we will
4794                  * post the final h_nr_running decrement below.
4795                 */
4796                 if (cfs_rq_throttled(cfs_rq))
4797                         break;
4798                 cfs_rq->h_nr_running--;
4799 
4800                 /* Don't dequeue parent if it has other entities besides us */
4801                 if (cfs_rq->load.weight) {
4802                         /* Avoid re-evaluating load for this entity: */
4803                         se = parent_entity(se);
4804                         /*
4805                          * Bias pick_next to pick a task from this cfs_rq, as
4806                          * p is sleeping when it is within its sched_slice.
4807                          */
4808                         if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4809                                 set_next_buddy(se);
4810                         break;
4811                 }
4812                 flags |= DEQUEUE_SLEEP;
4813         }
4814 
4815         for_each_sched_entity(se) {
4816                 cfs_rq = cfs_rq_of(se);
4817                 cfs_rq->h_nr_running--;
4818 
4819                 if (cfs_rq_throttled(cfs_rq))
4820                         break;
4821 
4822                 update_load_avg(se, UPDATE_TG);
4823                 update_cfs_shares(cfs_rq);
4824         }
4825 
4826         if (!se)
4827                 sub_nr_running(rq, 1);
4828 
4829         hrtick_update(rq);
4830 }
4831 
4832 #ifdef CONFIG_SMP
4833 
4834 /* Working cpumask for: load_balance, load_balance_newidle. */
4835 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
4836 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
4837 
4838 #ifdef CONFIG_NO_HZ_COMMON
4839 /*
4840  * per rq 'load' arrray crap; XXX kill this.
4841  */
4842 
4843 /*
4844  * The exact cpuload calculated at every tick would be:
4845  *
4846  *   load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4847  *
4848  * If a cpu misses updates for n ticks (as it was idle) and update gets
4849  * called on the n+1-th tick when cpu may be busy, then we have:
4850  *
4851  *   load_n   = (1 - 1/2^i)^n * load_0
4852  *   load_n+1 = (1 - 1/2^i)   * load_n + (1/2^i) * cur_load
4853  *
4854  * decay_load_missed() below does efficient calculation of
4855  *
4856  *   load' = (1 - 1/2^i)^n * load
4857  *
4858  * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4859  * This allows us to precompute the above in said factors, thereby allowing the
4860  * reduction of an arbitrary n in O(log_2 n) steps. (See also
4861  * fixed_power_int())
4862  *
4863  * The calculation is approximated on a 128 point scale.
4864  */
4865 #define DEGRADE_SHIFT           7
4866 
4867 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4868 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4869         {   0,   0,  0,  0,  0,  0, 0, 0 },
4870         {  64,  32,  8,  0,  0,  0, 0, 0 },
4871         {  96,  72, 40, 12,  1,  0, 0, 0 },
4872         { 112,  98, 75, 43, 15,  1, 0, 0 },
4873         { 120, 112, 98, 76, 45, 16, 2, 0 }
4874 };
4875 
4876 /*
4877  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4878  * would be when CPU is idle and so we just decay the old load without
4879  * adding any new load.
4880  */
4881 static unsigned long
4882 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4883 {
4884         int j = 0;
4885 
4886         if (!missed_updates)
4887                 return load;
4888 
4889         if (missed_updates >= degrade_zero_ticks[idx])
4890                 return 0;
4891 
4892         if (idx == 1)
4893                 return load >> missed_updates;
4894 
4895         while (missed_updates) {
4896                 if (missed_updates % 2)
4897                         load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4898 
4899                 missed_updates >>= 1;
4900                 j++;
4901         }
4902         return load;
4903 }
4904 #endif /* CONFIG_NO_HZ_COMMON */
4905 
4906 /**
4907  * __cpu_load_update - update the rq->cpu_load[] statistics
4908  * @this_rq: The rq to update statistics for
4909  * @this_load: The current load
4910  * @pending_updates: The number of missed updates
4911  *
4912  * Update rq->cpu_load[] statistics. This function is usually called every
4913  * scheduler tick (TICK_NSEC).
4914  *
4915  * This function computes a decaying average:
4916  *
4917  *   load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4918  *
4919  * Because of NOHZ it might not get called on every tick which gives need for
4920  * the @pending_updates argument.
4921  *
4922  *   load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4923  *             = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4924  *             = A * (A * load[i]_n-2 + B) + B
4925  *             = A * (A * (A * load[i]_n-3 + B) + B) + B
4926  *             = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4927  *             = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4928  *             = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4929  *             = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4930  *
4931  * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4932  * any change in load would have resulted in the tick being turned back on.
4933  *
4934  * For regular NOHZ, this reduces to:
4935  *
4936  *   load[i]_n = (1 - 1/2^i)^n * load[i]_0
4937  *
4938  * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4939  * term.
4940  */
4941 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4942                             unsigned long pending_updates)
4943 {
4944         unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4945         int i, scale;
4946 
4947         this_rq->nr_load_updates++;
4948 
4949         /* Update our load: */
4950         this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4951         for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4952                 unsigned long old_load, new_load;
4953 
4954                 /* scale is effectively 1 << i now, and >> i divides by scale */
4955 
4956                 old_load = this_rq->cpu_load[i];
4957 #ifdef CONFIG_NO_HZ_COMMON
4958                 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4959                 if (tickless_load) {
4960                         old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4961                         /*
4962                          * old_load can never be a negative value because a
4963                          * decayed tickless_load cannot be greater than the
4964                          * original tickless_load.
4965                          */
4966                         old_load += tickless_load;
4967                 }
4968 #endif
4969                 new_load = this_load;
4970                 /*
4971                  * Round up the averaging division if load is increasing. This
4972                  * prevents us from getting stuck on 9 if the load is 10, for
4973                  * example.
4974                  */
4975                 if (new_load > old_load)
4976                         new_load += scale - 1;
4977 
4978                 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4979         }
4980 
4981         sched_avg_update(this_rq);
4982 }
4983 
4984 /* Used instead of source_load when we know the type == 0 */
4985 static unsigned long weighted_cpuload(const int cpu)
4986 {
4987         return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4988 }
4989 
4990 #ifdef CONFIG_NO_HZ_COMMON
4991 /*
4992  * There is no sane way to deal with nohz on smp when using jiffies because the
4993  * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4994  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4995  *
4996  * Therefore we need to avoid the delta approach from the regular tick when
4997  * possible since that would seriously skew the load calculation. This is why we
4998  * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4999  * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5000  * loop exit, nohz_idle_balance, nohz full exit...)
5001  *
5002  * This means we might still be one tick off for nohz periods.
5003  */
5004 
5005 static void cpu_load_update_nohz(struct rq *this_rq,
5006                                  unsigned long curr_jiffies,
5007                                  unsigned long load)
5008 {
5009         unsigned long pending_updates;
5010 
5011         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5012         if (pending_updates) {
5013                 this_rq->last_load_update_tick = curr_jiffies;
5014                 /*
5015                  * In the regular NOHZ case, we were idle, this means load 0.
5016                  * In the NOHZ_FULL case, we were non-idle, we should consider
5017                  * its weighted load.
5018                  */
5019                 cpu_load_update(this_rq, load, pending_updates);
5020         }
5021 }
5022 
5023 /*
5024  * Called from nohz_idle_balance() to update the load ratings before doing the
5025  * idle balance.
5026  */
5027 static void cpu_load_update_idle(struct rq *this_rq)
5028 {
5029         /*
5030          * bail if there's load or we're actually up-to-date.
5031          */
5032         if (weighted_cpuload(cpu_of(this_rq)))
5033                 return;
5034 
5035         cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5036 }
5037 
5038 /*
5039  * Record CPU load on nohz entry so we know the tickless load to account
5040  * on nohz exit. cpu_load[0] happens then to be updated more frequently
5041  * than other cpu_load[idx] but it should be fine as cpu_load readers
5042  * shouldn't rely into synchronized cpu_load[*] updates.
5043  */
5044 void cpu_load_update_nohz_start(void)
5045 {
5046         struct rq *this_rq = this_rq();
5047 
5048         /*
5049          * This is all lockless but should be fine. If weighted_cpuload changes
5050          * concurrently we'll exit nohz. And cpu_load write can race with
5051          * cpu_load_update_idle() but both updater would be writing the same.
5052          */
5053         this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
5054 }
5055 
5056 /*
5057  * Account the tickless load in the end of a nohz frame.
5058  */
5059 void cpu_load_update_nohz_stop(void)
5060 {
5061         unsigned long curr_jiffies = READ_ONCE(jiffies);
5062         struct rq *this_rq = this_rq();
5063         unsigned long load;
5064 
5065         if (curr_jiffies == this_rq->last_load_update_tick)
5066                 return;
5067 
5068         load = weighted_cpuload(cpu_of(this_rq));
5069         raw_spin_lock(&this_rq->lock);
5070         update_rq_clock(this_rq);
5071         cpu_load_update_nohz(this_rq, curr_jiffies, load);
5072         raw_spin_unlock(&this_rq->lock);
5073 }
5074 #else /* !CONFIG_NO_HZ_COMMON */
5075 static inline void cpu_load_update_nohz(struct rq *this_rq,
5076                                         unsigned long curr_jiffies,
5077                                         unsigned long load) { }
5078 #endif /* CONFIG_NO_HZ_COMMON */
5079 
5080 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5081 {
5082 #ifdef CONFIG_NO_HZ_COMMON
5083         /* See the mess around cpu_load_update_nohz(). */
5084         this_rq->last_load_update_tick = READ_ONCE(jiffies);
5085 #endif
5086         cpu_load_update(this_rq, load, 1);
5087 }
5088 
5089 /*
5090  * Called from scheduler_tick()
5091  */
5092 void cpu_load_update_active(struct rq *this_rq)
5093 {
5094         unsigned long load = weighted_cpuload(cpu_of(this_rq));
5095 
5096         if (tick_nohz_tick_stopped())
5097                 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5098         else
5099                 cpu_load_update_periodic(this_rq, load);
5100 }
5101 
5102 /*
5103  * Return a low guess at the load of a migration-source cpu weighted
5104  * according to the scheduling class and "nice" value.
5105  *
5106  * We want to under-estimate the load of migration sources, to
5107  * balance conservatively.
5108  */
5109 static unsigned long source_load(int cpu, int type)
5110 {
5111         struct rq *rq = cpu_rq(cpu);
5112         unsigned long total = weighted_cpuload(cpu);
5113 
5114         if (type == 0 || !sched_feat(LB_BIAS))
5115                 return total;
5116 
5117         return min(rq->cpu_load[type-1], total);
5118 }
5119 
5120 /*
5121  * Return a high guess at the load of a migration-target cpu weighted
5122  * according to the scheduling class and "nice" value.
5123  */
5124 static unsigned long target_load(int cpu, int type)
5125 {
5126         struct rq *rq = cpu_rq(cpu);
5127         unsigned long total = weighted_cpuload(cpu);
5128 
5129         if (type == 0 || !sched_feat(LB_BIAS))
5130                 return total;
5131 
5132         return max(rq->cpu_load[type-1], total);
5133 }
5134 
5135 static unsigned long capacity_of(int cpu)
5136 {
5137         return cpu_rq(cpu)->cpu_capacity;
5138 }
5139 
5140 static unsigned long capacity_orig_of(int cpu)
5141 {
5142         return cpu_rq(cpu)->cpu_capacity_orig;
5143 }
5144 
5145 static unsigned long cpu_avg_load_per_task(int cpu)
5146 {
5147         struct rq *rq = cpu_rq(cpu);
5148         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5149         unsigned long load_avg = weighted_cpuload(cpu);
5150 
5151         if (nr_running)
5152                 return load_avg / nr_running;
5153 
5154         return 0;
5155 }
5156 
5157 #ifdef CONFIG_FAIR_GROUP_SCHED
5158 /*
5159  * effective_load() calculates the load change as seen from the root_task_group
5160  *
5161  * Adding load to a group doesn't make a group heavier, but can cause movement
5162  * of group shares between cpus. Assuming the shares were perfectly aligned one
5163  * can calculate the shift in shares.
5164  *
5165  * Calculate the effective load difference if @wl is added (subtracted) to @tg
5166  * on this @cpu and results in a total addition (subtraction) of @wg to the
5167  * total group weight.
5168  *
5169  * Given a runqueue weight distribution (rw_i) we can compute a shares
5170  * distribution (s_i) using:
5171  *
5172  *   s_i = rw_i / \Sum rw_j                                             (1)
5173  *
5174  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
5175  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
5176  * shares distribution (s_i):
5177  *
5178  *   rw_i = {   2,   4,   1,   0 }
5179  *   s_i  = { 2/7, 4/7, 1/7,   0 }
5180  *
5181  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
5182  * task used to run on and the CPU the waker is running on), we need to
5183  * compute the effect of waking a task on either CPU and, in case of a sync
5184  * wakeup, compute the effect of the current task going to sleep.
5185  *
5186  * So for a change of @wl to the local @cpu with an overall group weight change
5187  * of @wl we can compute the new shares distribution (s'_i) using:
5188  *
5189  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)                            (2)
5190  *
5191  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
5192  * differences in waking a task to CPU 0. The additional task changes the
5193  * weight and shares distributions like:
5194  *
5195  *   rw'_i = {   3,   4,   1,   0 }
5196  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
5197  *
5198  * We can then compute the difference in effective weight by using:
5199  *
5200  *   dw_i = S * (s'_i - s_i)                                            (3)
5201  *
5202  * Where 'S' is the group weight as seen by its parent.
5203  *
5204  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
5205  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
5206  * 4/7) times the weight of the group.
5207  */
5208 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5209 {
5210         struct sched_entity *se = tg->se[cpu];
5211 
5212         if (!tg->parent)        /* the trivial, non-cgroup case */
5213                 return wl;
5214 
5215         for_each_sched_entity(se) {
5216                 struct cfs_rq *cfs_rq = se->my_q;
5217                 long W, w = cfs_rq_load_avg(cfs_rq);
5218 
5219                 tg = cfs_rq->tg;
5220 
5221                 /*
5222                  * W = @wg + \Sum rw_j
5223                  */
5224                 W = wg + atomic_long_read(&tg->load_avg);
5225 
5226                 /* Ensure \Sum rw_j >= rw_i */
5227                 W -= cfs_rq->tg_load_avg_contrib;
5228                 W += w;
5229 
5230                 /*
5231                  * w = rw_i + @wl
5232                  */
5233                 w += wl;
5234 
5235                 /*
5236                  * wl = S * s'_i; see (2)
5237                  */
5238                 if (W > 0 && w < W)
5239                         wl = (w * (long)scale_load_down(tg->shares)) / W;
5240                 else
5241                         wl = scale_load_down(tg->shares);
5242 
5243                 /*
5244                  * Per the above, wl is the new se->load.weight value; since
5245                  * those are clipped to [MIN_SHARES, ...) do so now. See
5246                  * calc_cfs_shares().
5247                  */
5248                 if (wl < MIN_SHARES)
5249                         wl = MIN_SHARES;
5250 
5251                 /*
5252                  * wl = dw_i = S * (s'_i - s_i); see (3)
5253                  */
5254                 wl -= se->avg.load_avg;
5255 
5256                 /*
5257                  * Recursively apply this logic to all parent groups to compute
5258                  * the final effective load change on the root group. Since
5259                  * only the @tg group gets extra weight, all parent groups can
5260                  * only redistribute existing shares. @wl is the shift in shares
5261                  * resulting from this level per the above.
5262                  */
5263                 wg = 0;
5264         }
5265 
5266         return wl;
5267 }
5268 #else
5269 
5270 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5271 {
5272         return wl;
5273 }
5274 
5275 #endif
5276 
5277 static void record_wakee(struct task_struct *p)
5278 {
5279         /*
5280          * Only decay a single time; tasks that have less then 1 wakeup per
5281          * jiffy will not have built up many flips.
5282          */
5283         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5284                 current->wakee_flips >>= 1;
5285                 current->wakee_flip_decay_ts = jiffies;
5286         }
5287 
5288         if (current->last_wakee != p) {
5289                 current->last_wakee = p;
5290                 current->wakee_flips++;
5291         }
5292 }
5293 
5294 /*
5295  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5296  *
5297  * A waker of many should wake a different task than the one last awakened
5298  * at a frequency roughly N times higher than one of its wakees.
5299  *
5300  * In order to determine whether we should let the load spread vs consolidating
5301  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5302  * partner, and a factor of lls_size higher frequency in the other.
5303  *
5304  * With both conditions met, we can be relatively sure that the relationship is
5305  * non-monogamous, with partner count exceeding socket size.
5306  *
5307  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5308  * whatever is irrelevant, spread criteria is apparent partner count exceeds
5309  * socket size.
5310  */
5311 static int wake_wide(struct task_struct *p)
5312 {
5313         unsigned int master = current->wakee_flips;
5314         unsigned int slave = p->wakee_flips;
5315         int factor = this_cpu_read(sd_llc_size);
5316 
5317         if (master < slave)
5318                 swap(master, slave);
5319         if (slave < factor || master < slave * factor)
5320                 return 0;
5321         return 1;
5322 }
5323 
5324 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5325                        int prev_cpu, int sync)
5326 {
5327         s64 this_load, load;
5328         s64 this_eff_load, prev_eff_load;
5329         int idx, this_cpu;
5330         struct task_group *tg;
5331         unsigned long weight;
5332         int balanced;
5333 
5334         idx       = sd->wake_idx;
5335         this_cpu  = smp_processor_id();
5336         load      = source_load(prev_cpu, idx);
5337         this_load = target_load(this_cpu, idx);
5338 
5339         /*
5340          * If sync wakeup then subtract the (maximum possible)
5341          * effect of the currently running task from the load
5342          * of the current CPU:
5343          */
5344         if (sync) {
5345                 tg = task_group(current);
5346                 weight = current->se.avg.load_avg;
5347 
5348                 this_load += effective_load(tg, this_cpu, -weight, -weight);
5349                 load += effective_load(tg, prev_cpu, 0, -weight);
5350         }
5351 
5352         tg = task_group(p);
5353         weight = p->se.avg.load_avg;
5354 
5355         /*
5356          * In low-load situations, where prev_cpu is idle and this_cpu is idle
5357          * due to the sync cause above having dropped this_load to 0, we'll
5358          * always have an imbalance, but there's really nothing you can do
5359          * about that, so that's good too.
5360          *
5361          * Otherwise check if either cpus are near enough in load to allow this
5362          * task to be woken on this_cpu.
5363          */
5364         this_eff_load = 100;
5365         this_eff_load *= capacity_of(prev_cpu);
5366 
5367         prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5368         prev_eff_load *= capacity_of(this_cpu);
5369 
5370         if (this_load > 0) {
5371                 this_eff_load *= this_load +
5372                         effective_load(tg, this_cpu, weight, weight);
5373 
5374                 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5375         }
5376 
5377         balanced = this_eff_load <= prev_eff_load;
5378 
5379         schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5380 
5381         if (!balanced)
5382                 return 0;
5383 
5384         schedstat_inc(sd->ttwu_move_affine);
5385         schedstat_inc(p->se.statistics.nr_wakeups_affine);
5386 
5387         return 1;
5388 }
5389 
5390 static inline int task_util(struct task_struct *p);
5391 static int cpu_util_wake(int cpu, struct task_struct *p);
5392 
5393 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5394 {
5395         return capacity_orig_of(cpu) - cpu_util_wake(cpu, p);
5396 }
5397 
5398 /*
5399  * find_idlest_group finds and returns the least busy CPU group within the
5400  * domain.
5401  */
5402 static struct sched_group *
5403 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5404                   int this_cpu, int sd_flag)
5405 {
5406         struct sched_group *idlest = NULL, *group = sd->groups;
5407         struct sched_group *most_spare_sg = NULL;
5408         unsigned long min_runnable_load = ULONG_MAX, this_runnable_load = 0;
5409         unsigned long min_avg_load = ULONG_MAX, this_avg_load = 0;
5410         unsigned long most_spare = 0, this_spare = 0;
5411         int load_idx = sd->forkexec_idx;
5412         int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5413         unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5414                                 (sd->imbalance_pct-100) / 100;
5415 
5416         if (sd_flag & SD_BALANCE_WAKE)
5417                 load_idx = sd->wake_idx;
5418 
5419         do {
5420                 unsigned long load, avg_load, runnable_load;
5421                 unsigned long spare_cap, max_spare_cap;
5422                 int local_group;
5423                 int i;
5424 
5425                 /* Skip over this group if it has no CPUs allowed */
5426                 if (!cpumask_intersects(sched_group_cpus(group),
5427                                         tsk_cpus_allowed(p)))
5428                         continue;
5429 
5430                 local_group = cpumask_test_cpu(this_cpu,
5431                                                sched_group_cpus(group));
5432 
5433                 /*
5434                  * Tally up the load of all CPUs in the group and find
5435                  * the group containing the CPU with most spare capacity.
5436                  */
5437                 avg_load = 0;
5438                 runnable_load = 0;
5439                 max_spare_cap = 0;
5440 
5441                 for_each_cpu(i, sched_group_cpus(group)) {
5442                         /* Bias balancing toward cpus of our domain */
5443                         if (local_group)
5444                                 load = source_load(i, load_idx);
5445                         else
5446                                 load = target_load(i, load_idx);
5447 
5448                         runnable_load += load;
5449 
5450                         avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5451 
5452                         spare_cap = capacity_spare_wake(i, p);
5453 
5454                         if (spare_cap > max_spare_cap)
5455                                 max_spare_cap = spare_cap;
5456                 }
5457 
5458                 /* Adjust by relative CPU capacity of the group */
5459                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5460                                         group->sgc->capacity;
5461                 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5462                                         group->sgc->capacity;
5463 
5464                 if (local_group) {
5465                         this_runnable_load = runnable_load;
5466                         this_avg_load = avg_load;
5467                         this_spare = max_spare_cap;
5468                 } else {
5469                         if (min_runnable_load > (runnable_load + imbalance)) {
5470                                 /*
5471                                  * The runnable load is significantly smaller
5472                                  * so we can pick this new cpu
5473                                  */
5474                                 min_runnable_load = runnable_load;
5475                                 min_avg_load = avg_load;
5476                                 idlest = group;
5477                         } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5478                                    (100*min_avg_load > imbalance_scale*avg_load)) {
5479                                 /*
5480                                  * The runnable loads are close so take the
5481                                  * blocked load into account through avg_load.
5482                                  */
5483                                 min_avg_load = avg_load;
5484                                 idlest = group;
5485                         }
5486 
5487                         if (most_spare < max_spare_cap) {
5488                                 most_spare = max_spare_cap;
5489                                 most_spare_sg = group;
5490                         }
5491                 }
5492         } while (group = group->next, group != sd->groups);
5493 
5494         /*
5495          * The cross-over point between using spare capacity or least load
5496          * is too conservative for high utilization tasks on partially
5497          * utilized systems if we require spare_capacity > task_util(p),
5498          * so we allow for some task stuffing by using
5499          * spare_capacity > task_util(p)/2.
5500          *
5501          * Spare capacity can't be used for fork because the utilization has
5502          * not been set yet, we must first select a rq to compute the initial
5503          * utilization.
5504          */
5505         if (sd_flag & SD_BALANCE_FORK)
5506                 goto skip_spare;
5507 
5508         if (this_spare > task_util(p) / 2 &&
5509             imbalance_scale*this_spare > 100*most_spare)
5510                 return NULL;
5511 
5512         if (most_spare > task_util(p) / 2)
5513                 return most_spare_sg;
5514 
5515 skip_spare:
5516         if (!idlest)
5517                 return NULL;
5518 
5519         if (min_runnable_load > (this_runnable_load + imbalance))
5520                 return NULL;
5521 
5522         if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5523              (100*this_avg_load < imbalance_scale*min_avg_load))
5524                 return NULL;
5525 
5526         return idlest;
5527 }
5528 
5529 /*
5530  * find_idlest_cpu - find the idlest cpu among the cpus in group.
5531  */
5532 static int
5533 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5534 {
5535         unsigned long load, min_load = ULONG_MAX;
5536         unsigned int min_exit_latency = UINT_MAX;
5537         u64 latest_idle_timestamp = 0;
5538         int least_loaded_cpu = this_cpu;
5539         int shallowest_idle_cpu = -1;
5540         int i;
5541 
5542         /* Check if we have any choice: */
5543         if (group->group_weight == 1)
5544                 return cpumask_first(sched_group_cpus(group));
5545 
5546         /* Traverse only the allowed CPUs */
5547         for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5548                 if (idle_cpu(i)) {
5549                         struct rq *rq = cpu_rq(i);
5550                         struct cpuidle_state *idle = idle_get_state(rq);
5551                         if (idle && idle->exit_latency < min_exit_latency) {
5552                                 /*
5553                                  * We give priority to a CPU whose idle state
5554                                  * has the smallest exit latency irrespective
5555                                  * of any idle timestamp.
5556                                  */
5557                                 min_exit_latency = idle->exit_latency;
5558                                 latest_idle_timestamp = rq->idle_stamp;
5559                                 shallowest_idle_cpu = i;
5560                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5561                                    rq->idle_stamp > latest_idle_timestamp) {
5562                                 /*
5563                                  * If equal or no active idle state, then
5564                                  * the most recently idled CPU might have
5565                                  * a warmer cache.
5566                                  */
5567                                 latest_idle_timestamp = rq->idle_stamp;
5568                                 shallowest_idle_cpu = i;
5569                         }
5570                 } else if (shallowest_idle_cpu == -1) {
5571                         load = weighted_cpuload(i);
5572                         if (load < min_load || (load == min_load && i == this_cpu)) {
5573                                 min_load = load;
5574                                 least_loaded_cpu = i;
5575                         }
5576                 }
5577         }
5578 
5579         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5580 }
5581 
5582 /*
5583  * Implement a for_each_cpu() variant that starts the scan at a given cpu
5584  * (@start), and wraps around.
5585  *
5586  * This is used to scan for idle CPUs; such that not all CPUs looking for an
5587  * idle CPU find the same CPU. The down-side is that tasks tend to cycle
5588  * through the LLC domain.
5589  *
5590  * Especially tbench is found sensitive to this.
5591  */
5592 
5593 static int cpumask_next_wrap(int n, const struct cpumask *mask, int start, int *wrapped)
5594 {
5595         int next;
5596 
5597 again:
5598         next = find_next_bit(cpumask_bits(mask), nr_cpumask_bits, n+1);
5599 
5600         if (*wrapped) {
5601                 if (next >= start)
5602                         return nr_cpumask_bits;
5603         } else {
5604                 if (next >= nr_cpumask_bits) {
5605                         *wrapped = 1;
5606                         n = -1;
5607                         goto again;
5608                 }
5609         }
5610 
5611         return next;
5612 }
5613 
5614 #define for_each_cpu_wrap(cpu, mask, start, wrap)                               \
5615         for ((wrap) = 0, (cpu) = (start)-1;                                     \
5616                 (cpu) = cpumask_next_wrap((cpu), (mask), (start), &(wrap)),     \
5617                 (cpu) < nr_cpumask_bits; )
5618 
5619 #ifdef CONFIG_SCHED_SMT
5620 
5621 static inline void set_idle_cores(int cpu, int val)
5622 {
5623         struct sched_domain_shared *sds;
5624 
5625         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5626         if (sds)
5627                 WRITE_ONCE(sds->has_idle_cores, val);
5628 }
5629 
5630 static inline bool test_idle_cores(int cpu, bool def)
5631 {
5632         struct sched_domain_shared *sds;
5633 
5634         sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5635         if (sds)
5636                 return READ_ONCE(sds->has_idle_cores);
5637 
5638         return def;
5639 }
5640 
5641 /*
5642  * Scans the local SMT mask to see if the entire core is idle, and records this
5643  * information in sd_llc_shared->has_idle_cores.
5644  *
5645  * Since SMT siblings share all cache levels, inspecting this limited remote
5646  * state should be fairly cheap.
5647  */
5648 void __update_idle_core(struct rq *rq)
5649 {
5650         int core = cpu_of(rq);
5651         int cpu;
5652 
5653         rcu_read_lock();
5654         if (test_idle_cores(core, true))
5655                 goto unlock;
5656 
5657         for_each_cpu(cpu, cpu_smt_mask(core)) {
5658                 if (cpu == core)
5659                         continue;
5660 
5661                 if (!idle_cpu(cpu))
5662                         goto unlock;
5663         }
5664 
5665         set_idle_cores(core, 1);
5666 unlock:
5667         rcu_read_unlock();
5668 }
5669 
5670 /*
5671  * Scan the entire LLC domain for idle cores; this dynamically switches off if
5672  * there are no idle cores left in the system; tracked through
5673  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5674  */
5675 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5676 {
5677         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5678         int core, cpu, wrap;
5679 
5680         if (!static_branch_likely(&sched_smt_present))
5681                 return -1;
5682 
5683         if (!test_idle_cores(target, false))
5684                 return -1;
5685 
5686         cpumask_and(cpus, sched_domain_span(sd), tsk_cpus_allowed(p));
5687 
5688         for_each_cpu_wrap(core, cpus, target, wrap) {
5689                 bool idle = true;
5690 
5691                 for_each_cpu(cpu, cpu_smt_mask(core)) {
5692                         cpumask_clear_cpu(cpu, cpus);
5693                         if (!idle_cpu(cpu))
5694                                 idle = false;
5695                 }
5696 
5697                 if (idle)
5698                         return core;
5699         }
5700 
5701         /*
5702          * Failed to find an idle core; stop looking for one.
5703          */
5704         set_idle_cores(target, 0);
5705 
5706         return -1;
5707 }
5708 
5709 /*
5710  * Scan the local SMT mask for idle CPUs.
5711  */
5712 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5713 {
5714         int cpu;
5715 
5716         if (!static_branch_likely(&sched_smt_present))
5717                 return -1;
5718 
5719         for_each_cpu(cpu, cpu_smt_mask(target)) {
5720                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5721                         continue;
5722                 if (idle_cpu(cpu))
5723                         return cpu;
5724         }
5725 
5726         return -1;
5727 }
5728 
5729 #else /* CONFIG_SCHED_SMT */
5730 
5731 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5732 {
5733         return -1;
5734 }
5735 
5736 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5737 {
5738         return -1;
5739 }
5740 
5741 #endif /* CONFIG_SCHED_SMT */
5742 
5743 /*
5744  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5745  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5746  * average idle time for this rq (as found in rq->avg_idle).
5747  */
5748 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5749 {
5750         struct sched_domain *this_sd;
5751         u64 avg_cost, avg_idle = this_rq()->avg_idle;
5752         u64 time, cost;
5753         s64 delta;
5754         int cpu, wrap;
5755 
5756         this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
5757         if (!this_sd)
5758                 return -1;
5759 
5760         avg_cost = this_sd->avg_scan_cost;
5761 
5762         /*
5763          * Due to large variance we need a large fuzz factor; hackbench in
5764          * particularly is sensitive here.
5765          */
5766         if ((avg_idle / 512) < avg_cost)
5767                 return -1;
5768 
5769         time = local_clock();
5770 
5771         for_each_cpu_wrap(cpu, sched_domain_span(sd), target, wrap) {
5772                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)))
5773                         continue;
5774                 if (idle_cpu(cpu))
5775                         break;
5776         }
5777 
5778         time = local_clock() - time;
5779         cost = this_sd->avg_scan_cost;
5780         delta = (s64)(time - cost) / 8;
5781         this_sd->avg_scan_cost += delta;
5782 
5783         return cpu;
5784 }
5785 
5786 /*
5787  * Try and locate an idle core/thread in the LLC cache domain.
5788  */
5789 static int select_idle_sibling(struct task_struct *p, int prev, int target)
5790 {
5791         struct sched_domain *sd;
5792         int i;
5793 
5794         if (idle_cpu(target))
5795                 return target;
5796 
5797         /*
5798          * If the previous cpu is cache affine and idle, don't be stupid.
5799          */
5800         if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
5801                 return prev;
5802 
5803         sd = rcu_dereference(per_cpu(sd_llc, target));
5804         if (!sd)
5805                 return target;
5806 
5807         i = select_idle_core(p, sd, target);
5808         if ((unsigned)i < nr_cpumask_bits)
5809                 return i;
5810 
5811         i = select_idle_cpu(p, sd, target);
5812         if ((unsigned)i < nr_cpumask_bits)
5813                 return i;
5814 
5815         i = select_idle_smt(p, sd, target);
5816         if ((unsigned)i < nr_cpumask_bits)
5817                 return i;
5818 
5819         return target;
5820 }
5821 
5822 /*
5823  * cpu_util returns the amount of capacity of a CPU that is used by CFS
5824  * tasks. The unit of the return value must be the one of capacity so we can
5825  * compare the utilization with the capacity of the CPU that is available for
5826  * CFS task (ie cpu_capacity).
5827  *
5828  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5829  * recent utilization of currently non-runnable tasks on a CPU. It represents
5830  * the amount of utilization of a CPU in the range [0..capacity_orig] where
5831  * capacity_orig is the cpu_capacity available at the highest frequency
5832  * (arch_scale_freq_capacity()).
5833  * The utilization of a CPU converges towards a sum equal to or less than the
5834  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5835  * the running time on this CPU scaled by capacity_curr.
5836  *
5837  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5838  * higher than capacity_orig because of unfortunate rounding in
5839  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5840  * the average stabilizes with the new running time. We need to check that the
5841  * utilization stays within the range of [0..capacity_orig] and cap it if
5842  * necessary. Without utilization capping, a group could be seen as overloaded
5843  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5844  * available capacity. We allow utilization to overshoot capacity_curr (but not
5845  * capacity_orig) as it useful for predicting the capacity required after task
5846  * migrations (scheduler-driven DVFS).
5847  */
5848 static int cpu_util(int cpu)
5849 {
5850         unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5851         unsigned long capacity = capacity_orig_of(cpu);
5852 
5853         return (util >= capacity) ? capacity : util;
5854 }
5855 
5856 static inline int task_util(struct task_struct *p)
5857 {
5858         return p->se.avg.util_avg;
5859 }
5860 
5861 /*
5862  * cpu_util_wake: Compute cpu utilization with any contributions from
5863  * the waking task p removed.
5864  */
5865 static int cpu_util_wake(int cpu, struct task_struct *p)
5866 {
5867         unsigned long util, capacity;
5868 
5869         /* Task has no contribution or is new */
5870         if (cpu != task_cpu(p) || !p->se.avg.last_update_time)
5871                 return cpu_util(cpu);
5872 
5873         capacity = capacity_orig_of(cpu);
5874         util = max_t(long, cpu_rq(cpu)->cfs.avg.util_avg - task_util(p), 0);
5875 
5876         return (util >= capacity) ? capacity : util;
5877 }
5878 
5879 /*
5880  * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
5881  * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
5882  *
5883  * In that case WAKE_AFFINE doesn't make sense and we'll let
5884  * BALANCE_WAKE sort things out.
5885  */
5886 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
5887 {
5888         long min_cap, max_cap;
5889 
5890         min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
5891         max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
5892 
5893         /* Minimum capacity is close to max, no need to abort wake_affine */
5894         if (max_cap - min_cap < max_cap >> 3)
5895                 return 0;
5896 
5897         /* Bring task utilization in sync with prev_cpu */
5898         sync_entity_load_avg(&p->se);
5899 
5900         return min_cap * 1024 < task_util(p) * capacity_margin;
5901 }
5902 
5903 /*
5904  * select_task_rq_fair: Select target runqueue for the waking task in domains
5905  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5906  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5907  *
5908  * Balances load by selecting the idlest cpu in the idlest group, or under
5909  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5910  *
5911  * Returns the target cpu number.
5912  *
5913  * preempt must be disabled.
5914  */
5915 static int
5916 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5917 {
5918         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5919         int cpu = smp_processor_id();
5920         int new_cpu = prev_cpu;
5921         int want_affine = 0;
5922         int sync = wake_flags & WF_SYNC;
5923 
5924         if (sd_flag & SD_BALANCE_WAKE) {
5925                 record_wakee(p);
5926                 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
5927                               && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5928         }
5929 
5930         rcu_read_lock();
5931         for_each_domain(cpu, tmp) {
5932                 if (!(tmp->flags & SD_LOAD_BALANCE))
5933                         break;
5934 
5935                 /*
5936                  * If both cpu and prev_cpu are part of this domain,
5937                  * cpu is a valid SD_WAKE_AFFINE target.
5938                  */
5939                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5940                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5941                         affine_sd = tmp;
5942                         break;
5943                 }
5944 
5945                 if (tmp->flags & sd_flag)
5946                         sd = tmp;
5947                 else if (!want_affine)
5948                         break;
5949         }
5950 
5951         if (affine_sd) {
5952                 sd = NULL; /* Prefer wake_affine over balance flags */
5953                 if (cpu != prev_cpu && wake_affine(affine_sd, p, prev_cpu, sync))
5954                         new_cpu = cpu;
5955         }
5956 
5957         if (!sd) {
5958                 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5959                         new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
5960 
5961         } else while (sd) {
5962                 struct sched_group *group;
5963                 int weight;
5964 
5965                 if (!(sd->flags & sd_flag)) {
5966                         sd = sd->child;
5967                         continue;
5968                 }
5969 
5970                 group = find_idlest_group(sd, p, cpu, sd_flag);
5971                 if (!group) {
5972                         sd = sd->child;
5973                         continue;
5974                 }
5975 
5976                 new_cpu = find_idlest_cpu(group, p, cpu);
5977                 if (new_cpu == -1 || new_cpu == cpu) {
5978                         /* Now try balancing at a lower domain level of cpu */
5979                         sd = sd->child;
5980                         continue;
5981                 }
5982 
5983                 /* Now try balancing at a lower domain level of new_cpu */
5984                 cpu = new_cpu;
5985                 weight = sd->span_weight;
5986                 sd = NULL;
5987                 for_each_domain(cpu, tmp) {
5988                         if (weight <= tmp->span_weight)
5989                                 break;
5990                         if (tmp->flags & sd_flag)
5991                                 sd = tmp;
5992                 }
5993                 /* while loop will break here if sd == NULL */
5994         }
5995         rcu_read_unlock();
5996 
5997         return new_cpu;
5998 }
5999 
6000 /*
6001  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
6002  * cfs_rq_of(p) references at time of call are still valid and identify the
6003  * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6004  */
6005 static void migrate_task_rq_fair(struct task_struct *p)
6006 {
6007         /*
6008          * As blocked tasks retain absolute vruntime the migration needs to
6009          * deal with this by subtracting the old and adding the new
6010          * min_vruntime -- the latter is done by enqueue_entity() when placing
6011          * the task on the new runqueue.
6012          */
6013         if (p->state == TASK_WAKING) {
6014                 struct sched_entity *se = &p->se;
6015                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6016                 u64 min_vruntime;
6017 
6018 #ifndef CONFIG_64BIT
6019                 u64 min_vruntime_copy;
6020 
6021                 do {
6022                         min_vruntime_copy = cfs_rq->min_vruntime_copy;
6023                         smp_rmb();
6024                         min_vruntime = cfs_rq->min_vruntime;
6025                 } while (min_vruntime != min_vruntime_copy);
6026 #else
6027                 min_vruntime = cfs_rq->min_vruntime;
6028 #endif
6029 
6030                 se->vruntime -= min_vruntime;
6031         }
6032 
6033         /*
6034          * We are supposed to update the task to "current" time, then its up to date
6035          * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
6036          * what current time is, so simply throw away the out-of-date time. This
6037          * will result in the wakee task is less decayed, but giving the wakee more
6038          * load sounds not bad.
6039          */
6040         remove_entity_load_avg(&p->se);
6041 
6042         /* Tell new CPU we are migrated */
6043         p->se.avg.last_update_time = 0;
6044 
6045         /* We have migrated, no longer consider this task hot */
6046         p->se.exec_start = 0;
6047 }
6048 
6049 static void task_dead_fair(struct task_struct *p)
6050 {
6051         remove_entity_load_avg(&p->se);
6052 }
6053 #endif /* CONFIG_SMP */
6054 
6055 static unsigned long
6056 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
6057 {
6058         unsigned long gran = sysctl_sched_wakeup_granularity;
6059 
6060         /*
6061          * Since its curr running now, convert the gran from real-time
6062          * to virtual-time in his units.
6063          *
6064          * By using 'se' instead of 'curr' we penalize light tasks, so
6065          * they get preempted easier. That is, if 'se' < 'curr' then
6066          * the resulting gran will be larger, therefore penalizing the
6067          * lighter, if otoh 'se' > 'curr' then the resulting gran will
6068          * be smaller, again penalizing the lighter task.
6069          *
6070          * This is especially important for buddies when the leftmost
6071          * task is higher priority than the buddy.
6072          */
6073         return calc_delta_fair(gran, se);
6074 }
6075 
6076 /*
6077  * Should 'se' preempt 'curr'.
6078  *
6079  *             |s1
6080  *        |s2
6081  *   |s3
6082  *         g
6083  *      |<--->|c
6084  *
6085  *  w(c, s1) = -1
6086  *  w(c, s2) =  0
6087  *  w(c, s3) =  1
6088  *
6089  */
6090 static int
6091 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6092 {
6093         s64 gran, vdiff = curr->vruntime - se->vruntime;
6094 
6095         if (vdiff <= 0)
6096                 return -1;
6097 
6098         gran = wakeup_gran(curr, se);
6099         if (vdiff > gran)
6100                 return 1;
6101 
6102         return 0;
6103 }
6104 
6105 static void set_last_buddy(struct sched_entity *se)
6106 {
6107         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6108                 return;
6109 
6110         for_each_sched_entity(se)
6111                 cfs_rq_of(se)->last = se;
6112 }
6113 
6114 static void set_next_buddy(struct sched_entity *se)
6115 {
6116         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6117                 return;
6118 
6119         for_each_sched_entity(se)
6120                 cfs_rq_of(se)->next = se;
6121 }
6122 
6123 static void set_skip_buddy(struct sched_entity *se)
6124 {
6125         for_each_sched_entity(se)
6126                 cfs_rq_of(se)->skip = se;
6127 }
6128 
6129 /*
6130  * Preempt the current task with a newly woken task if needed:
6131  */
6132 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6133 {
6134         struct task_struct *curr = rq->curr;
6135         struct sched_entity *se = &curr->se, *pse = &p->se;
6136         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6137         int scale = cfs_rq->nr_running >= sched_nr_latency;
6138         int next_buddy_marked = 0;
6139 
6140         if (unlikely(se == pse))
6141                 return;
6142 
6143         /*
6144          * This is possible from callers such as attach_tasks(), in which we
6145          * unconditionally check_prempt_curr() after an enqueue (which may have
6146          * lead to a throttle).  This both saves work and prevents false
6147          * next-buddy nomination below.
6148          */
6149         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6150                 return;
6151 
6152         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6153                 set_next_buddy(pse);
6154                 next_buddy_marked = 1;
6155         }
6156 
6157         /*
6158          * We can come here with TIF_NEED_RESCHED already set from new task
6159          * wake up path.
6160          *
6161          * Note: this also catches the edge-case of curr being in a throttled
6162          * group (e.g. via set_curr_task), since update_curr() (in the
6163          * enqueue of curr) will have resulted in resched being set.  This
6164          * prevents us from potentially nominating it as a false LAST_BUDDY
6165          * below.
6166          */
6167         if (test_tsk_need_resched(curr))
6168                 return;
6169 
6170         /* Idle tasks are by definition preempted by non-idle tasks. */
6171         if (unlikely(curr->policy == SCHED_IDLE) &&
6172             likely(p->policy != SCHED_IDLE))
6173                 goto preempt;
6174 
6175         /*
6176          * Batch and idle tasks do not preempt non-idle tasks (their preemption
6177          * is driven by the tick):
6178          */
6179         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6180                 return;
6181 
6182         find_matching_se(&se, &pse);
6183         update_curr(cfs_rq_of(se));
6184         BUG_ON(!pse);
6185         if (wakeup_preempt_entity(se, pse) == 1) {
6186                 /*
6187                  * Bias pick_next to pick the sched entity that is
6188                  * triggering this preemption.
6189                  */
6190                 if (!next_buddy_marked)
6191                         set_next_buddy(pse);
6192                 goto preempt;
6193         }
6194 
6195         return;
6196 
6197 preempt:
6198         resched_curr(rq);
6199         /*
6200          * Only set the backward buddy when the current task is still
6201          * on the rq. This can happen when a wakeup gets interleaved
6202          * with schedule on the ->pre_schedule() or idle_balance()
6203          * point, either of which can * drop the rq lock.
6204          *
6205          * Also, during early boot the idle thread is in the fair class,
6206          * for obvious reasons its a bad idea to schedule back to it.
6207          */
6208         if (unlikely(!se->on_rq || curr == rq->idle))
6209                 return;
6210 
6211         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6212                 set_last_buddy(se);
6213 }
6214 
6215 static struct task_struct *
6216 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
6217 {
6218         struct cfs_rq *cfs_rq = &rq->cfs;
6219         struct sched_entity *se;
6220         struct task_struct *p;
6221         int new_tasks;
6222 
6223 again:
6224 #ifdef CONFIG_FAIR_GROUP_SCHED
6225         if (!cfs_rq->nr_running)
6226                 goto idle;
6227 
6228         if (prev->sched_class != &fair_sched_class)
6229                 goto simple;
6230 
6231         /*
6232          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6233          * likely that a next task is from the same cgroup as the current.
6234          *
6235          * Therefore attempt to avoid putting and setting the entire cgroup
6236          * hierarchy, only change the part that actually changes.
6237          */
6238 
6239         do {
6240                 struct sched_entity *curr = cfs_rq->curr;
6241 
6242                 /*
6243                  * Since we got here without doing put_prev_entity() we also
6244                  * have to consider cfs_rq->curr. If it is still a runnable
6245                  * entity, update_curr() will update its vruntime, otherwise
6246                  * forget we've ever seen it.
6247                  */
6248                 if (curr) {
6249                         if (curr->on_rq)
6250                                 update_curr(cfs_rq);
6251                         else
6252                                 curr = NULL;
6253 
6254                         /*
6255                          * This call to check_cfs_rq_runtime() will do the
6256                          * throttle and dequeue its entity in the parent(s).
6257                          * Therefore the 'simple' nr_running test will indeed
6258                          * be correct.
6259                          */
6260                         if (unlikely(check_cfs_rq_runtime(cfs_rq)))
6261                                 goto simple;
6262                 }
6263 
6264                 se = pick_next_entity(cfs_rq, curr);
6265                 cfs_rq = group_cfs_rq(se);
6266         } while (cfs_rq);
6267 
6268         p = task_of(se);
6269 
6270         /*
6271          * Since we haven't yet done put_prev_entity and if the selected task
6272          * is a different task than we started out with, try and touch the
6273          * least amount of cfs_rqs.
6274          */
6275         if (prev != p) {
6276                 struct sched_entity *pse = &prev->se;
6277 
6278                 while (!(cfs_rq = is_same_group(se, pse))) {
6279                         int se_depth = se->depth;
6280                         int pse_depth = pse->depth;
6281 
6282                         if (se_depth <= pse_depth) {
6283                                 put_prev_entity(cfs_rq_of(pse), pse);
6284                                 pse = parent_entity(pse);
6285                         }
6286                         if (se_depth >= pse_depth) {
6287                                 set_next_entity(cfs_rq_of(se), se);
6288                                 se = parent_entity(se);
6289                         }
6290                 }
6291 
6292                 put_prev_entity(cfs_rq, pse);
6293                 set_next_entity(cfs_rq, se);
6294         }
6295 
6296         if (hrtick_enabled(rq))
6297                 hrtick_start_fair(rq, p);
6298 
6299         return p;
6300 simple:
6301         cfs_rq = &rq->cfs;
6302 #endif
6303 
6304         if (!cfs_rq->nr_running)
6305                 goto idle;
6306 
6307         put_prev_task(rq, prev);
6308 
6309         do {
6310                 se = pick_next_entity(cfs_rq, NULL);
6311                 set_next_entity(cfs_rq, se);
6312                 cfs_rq = group_cfs_rq(se);
6313         } while (cfs_rq);
6314 
6315         p = task_of(se);
6316 
6317         if (hrtick_enabled(rq))
6318                 hrtick_start_fair(rq, p);
6319 
6320         return p;
6321 
6322 idle:
6323         /*
6324          * This is OK, because current is on_cpu, which avoids it being picked
6325          * for load-balance and preemption/IRQs are still disabled avoiding
6326          * further scheduler activity on it and we're being very careful to
6327          * re-start the picking loop.
6328          */
6329         lockdep_unpin_lock(&rq->lock, cookie);
6330         new_tasks = idle_balance(rq);
6331         lockdep_repin_lock(&rq->lock, cookie);
6332         /*
6333          * Because idle_balance() releases (and re-acquires) rq->lock, it is
6334          * possible for any higher priority task to appear. In that case we
6335          * must re-start the pick_next_entity() loop.
6336          */
6337         if (new_tasks < 0)
6338                 return RETRY_TASK;
6339 
6340         if (new_tasks > 0)
6341                 goto again;
6342 
6343         return NULL;
6344 }
6345 
6346 /*
6347  * Account for a descheduled task:
6348  */
6349 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6350 {
6351         struct sched_entity *se = &prev->se;
6352         struct cfs_rq *cfs_rq;
6353 
6354         for_each_sched_entity(se) {
6355                 cfs_rq = cfs_rq_of(se);
6356                 put_prev_entity(cfs_rq, se);
6357         }
6358 }
6359 
6360 /*
6361  * sched_yield() is very simple
6362  *
6363  * The magic of dealing with the ->skip buddy is in pick_next_entity.
6364  */
6365 static void yield_task_fair(struct rq *rq)
6366 {
6367         struct task_struct *curr = rq->curr;
6368         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6369         struct sched_entity *se = &curr->se;
6370 
6371         /*
6372          * Are we the only task in the tree?
6373          */
6374         if (unlikely(rq->nr_running == 1))
6375                 return;
6376 
6377         clear_buddies(cfs_rq, se);
6378 
6379         if (curr->policy != SCHED_BATCH) {
6380                 update_rq_clock(rq);
6381                 /*
6382                  * Update run-time statistics of the 'current'.
6383                  */
6384                 update_curr(cfs_rq);
6385                 /*
6386                  * Tell update_rq_clock() that we've just updated,
6387                  * so we don't do microscopic update in schedule()
6388                  * and double the fastpath cost.
6389                  */
6390                 rq_clock_skip_update(rq, true);
6391         }
6392 
6393         set_skip_buddy(se);
6394 }
6395 
6396 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6397 {
6398         struct sched_entity *se = &p->se;
6399 
6400         /* throttled hierarchies are not runnable */
6401         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6402                 return false;
6403 
6404         /* Tell the scheduler that we'd really like pse to run next. */
6405         set_next_buddy(se);
6406 
6407         yield_task_fair(rq);
6408 
6409         return true;
6410 }
6411 
6412 #ifdef CONFIG_SMP
6413 /**************************************************
6414  * Fair scheduling class load-balancing methods.
6415  *
6416  * BASICS
6417  *
6418  * The purpose of load-balancing is to achieve the same basic fairness the
6419  * per-cpu scheduler provides, namely provide a proportional amount of compute
6420  * time to each task. This is expressed in the following equation:
6421  *
6422  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
6423  *
6424  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
6425  * W_i,0 is defined as:
6426  *
6427  *   W_i,0 = \Sum_j w_i,j                                             (2)
6428  *
6429  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
6430  * is derived from the nice value as per sched_prio_to_weight[].
6431  *
6432  * The weight average is an exponential decay average of the instantaneous
6433  * weight:
6434  *
6435  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
6436  *
6437  * C_i is the compute capacity of cpu i, typically it is the
6438  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6439  * can also include other factors [XXX].
6440  *
6441  * To achieve this balance we define a measure of imbalance which follows
6442  * directly from (1):
6443  *
6444  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
6445  *
6446  * We them move tasks around to minimize the imbalance. In the continuous
6447  * function space it is obvious this converges, in the discrete case we get
6448  * a few fun cases generally called infeasible weight scenarios.
6449  *
6450  * [XXX expand on:
6451  *     - infeasible weights;
6452  *     - local vs global optima in the discrete case. ]
6453  *
6454  *
6455  * SCHED DOMAINS
6456  *
6457  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6458  * for all i,j solution, we create a tree of cpus that follows the hardware
6459  * topology where each level pairs two lower groups (or better). This results
6460  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
6461  * tree to only the first of the previous level and we decrease the frequency
6462  * of load-balance at each level inv. proportional to the number of cpus in
6463  * the groups.
6464  *
6465  * This yields:
6466  *
6467  *     log_2 n     1     n
6468  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
6469  *     i = 0      2^i   2^i
6470  *                               `- size of each group
6471  *         |         |     `- number of cpus doing load-balance
6472  *         |         `- freq
6473  *         `- sum over all levels
6474  *
6475  * Coupled with a limit on how many tasks we can migrate every balance pass,
6476  * this makes (5) the runtime complexity of the balancer.
6477  *
6478  * An important property here is that each CPU is still (indirectly) connected
6479  * to every other cpu in at most O(log n) steps:
6480  *
6481  * The adjacency matrix of the resulting graph is given by:
6482  *
6483  *             log_2 n
6484  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
6485  *             k = 0
6486  *
6487  * And you'll find that:
6488  *
6489  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
6490  *
6491  * Showing there's indeed a path between every cpu in at most O(log n) steps.
6492  * The task movement gives a factor of O(m), giving a convergence complexity
6493  * of:
6494  *
6495  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
6496  *
6497  *
6498  * WORK CONSERVING
6499  *
6500  * In order to avoid CPUs going idle while there's still work to do, new idle
6501  * balancing is more aggressive and has the newly idle cpu iterate up the domain
6502  * tree itself instead of relying on other CPUs to bring it work.
6503  *
6504  * This adds some complexity to both (5) and (8) but it reduces the total idle
6505  * time.
6506  *
6507  * [XXX more?]
6508  *
6509  *
6510  * CGROUPS
6511  *
6512  * Cgroups make a horror show out of (2), instead of a simple sum we get:
6513  *
6514  *                                s_k,i
6515  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
6516  *                                 S_k
6517  *
6518  * Where
6519  *
6520  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
6521  *
6522  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
6523  *
6524  * The big problem is S_k, its a global sum needed to compute a local (W_i)
6525  * property.
6526  *
6527  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6528  *      rewrite all of this once again.]
6529  */
6530 
6531 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6532 
6533 enum fbq_type { regular, remote, all };
6534 
6535 #define LBF_ALL_PINNED  0x01
6536 #define LBF_NEED_BREAK  0x02
6537 #define LBF_DST_PINNED  0x04
6538 #define LBF_SOME_PINNED 0x08
6539 
6540 struct lb_env {
6541         struct sched_domain     *sd;
6542 
6543         struct rq               *src_rq;
6544         int                     src_cpu;
6545 
6546         int                     dst_cpu;
6547         struct rq               *dst_rq;
6548 
6549         struct cpumask          *dst_grpmask;
6550         int                     new_dst_cpu;
6551         enum cpu_idle_type      idle;
6552         long                    imbalance;
6553         /* The set of CPUs under consideration for load-balancing */
6554         struct cpumask          *cpus;
6555 
6556         unsigned int            flags;
6557 
6558         unsigned int            loop;
6559         unsigned int            loop_break;
6560         unsigned int            loop_max;
6561 
6562         enum fbq_type           fbq_type;
6563         struct list_head        tasks;
6564 };
6565 
6566 /*
6567  * Is this task likely cache-hot:
6568  */
6569 static int task_hot(struct task_struct *p, struct lb_env *env)
6570 {
6571         s64 delta;
6572 
6573         lockdep_assert_held(&env->src_rq->lock);
6574 
6575         if (p->sched_class != &fair_sched_class)
6576                 return 0;
6577 
6578         if (unlikely(p->policy == SCHED_IDLE))
6579                 return 0;
6580 
6581         /*
6582          * Buddy candidates are cache hot:
6583          */
6584         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6585                         (&p->se == cfs_rq_of(&p->se)->next ||
6586                          &p->se == cfs_rq_of(&p->se)->last))
6587                 return 1;
6588 
6589         if (sysctl_sched_migration_cost == -1)
6590                 return 1;
6591         if (sysctl_sched_migration_cost == 0)
6592                 return 0;
6593 
6594         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6595 
6596         return delta < (s64)sysctl_sched_migration_cost;
6597 }
6598 
6599 #ifdef CONFIG_NUMA_BALANCING
6600 /*
6601  * Returns 1, if task migration degrades locality
6602  * Returns 0, if task migration improves locality i.e migration preferred.
6603  * Returns -1, if task migration is not affected by locality.
6604  */
6605 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6606 {
6607         struct numa_group *numa_group = rcu_dereference(p->numa_group);
6608         unsigned long src_faults, dst_faults;
6609         int src_nid, dst_nid;
6610 
6611         if (!static_branch_likely(&sched_numa_balancing))
6612                 return -1;
6613 
6614         if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6615                 return -1;
6616 
6617         src_nid = cpu_to_node(env->src_cpu);
6618         dst_nid = cpu_to_node(env->dst_cpu);
6619 
6620         if (src_nid == dst_nid)
6621                 return -1;
6622 
6623         /* Migrating away from the preferred node is always bad. */
6624         if (src_nid == p->numa_preferred_nid) {
6625                 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6626                         return 1;
6627                 else
6628                         return -1;
6629         }
6630 
6631         /* Encourage migration to the preferred node. */
6632         if (dst_nid == p->numa_preferred_nid)
6633                 return 0;
6634 
6635         if (numa_group) {
6636                 src_faults = group_faults(p, src_nid);
6637                 dst_faults = group_faults(p, dst_nid);
6638         } else {
6639                 src_faults = task_faults(p, src_nid);
6640                 dst_faults = task_faults(p, dst_nid);
6641         }
6642 
6643         return dst_faults < src_faults;
6644 }
6645 
6646 #else
6647 static inline int migrate_degrades_locality(struct task_struct *p,
6648                                              struct lb_env *env)
6649 {
6650         return -1;
6651 }
6652 #endif
6653 
6654 /*
6655  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6656  */
6657 static
6658 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6659 {
6660         int tsk_cache_hot;
6661 
6662         lockdep_assert_held(&env->src_rq->lock);
6663 
6664         /*
6665          * We do not migrate tasks that are:
6666          * 1) throttled_lb_pair, or
6667          * 2) cannot be migrated to this CPU due to cpus_allowed, or
6668          * 3) running (obviously), or
6669          * 4) are cache-hot on their current CPU.
6670          */
6671         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6672                 return 0;
6673 
6674         if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6675                 int cpu;
6676 
6677                 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6678 
6679                 env->flags |= LBF_SOME_PINNED;
6680 
6681                 /*
6682                  * Remember if this task can be migrated to any other cpu in
6683                  * our sched_group. We may want to revisit it if we couldn't
6684                  * meet load balance goals by pulling other tasks on src_cpu.
6685                  *
6686                  * Also avoid computing new_dst_cpu if we have already computed
6687                  * one in current iteration.
6688                  */
6689                 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6690                         return 0;
6691 
6692                 /* Prevent to re-select dst_cpu via env's cpus */
6693                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6694                         if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6695                                 env->flags |= LBF_DST_PINNED;
6696                                 env->new_dst_cpu = cpu;
6697                                 break;
6698                         }
6699                 }
6700 
6701                 return 0;
6702         }
6703 
6704         /* Record that we found atleast one task that could run on dst_cpu */
6705         env->flags &= ~LBF_ALL_PINNED;
6706 
6707         if (task_running(env->src_rq, p)) {
6708                 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
6709                 return 0;
6710         }
6711 
6712         /*
6713          * Aggressive migration if:
6714          * 1) destination numa is preferred
6715          * 2) task is cache cold, or
6716          * 3) too many balance attempts have failed.
6717          */
6718         tsk_cache_hot = migrate_degrades_locality(p, env);
6719         if (tsk_cache_hot == -1)
6720                 tsk_cache_hot = task_hot(p, env);
6721 
6722         if (tsk_cache_hot <= 0 ||
6723             env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6724                 if (tsk_cache_hot == 1) {
6725                         schedstat_inc(env->sd->lb_hot_gained[env->idle]);
6726                         schedstat_inc(p->se.statistics.nr_forced_migrations);
6727                 }
6728                 return 1;
6729         }
6730 
6731         schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
6732         return 0;
6733 }
6734 
6735 /*
6736  * detach_task() -- detach the task for the migration specified in env
6737  */
6738 static void detach_task(struct task_struct *p, struct lb_env *env)
6739 {
6740         lockdep_assert_held(&env->src_rq->lock);
6741 
6742         p->on_rq = TASK_ON_RQ_MIGRATING;
6743         deactivate_task(env->src_rq, p, 0);
6744         set_task_cpu(p, env->dst_cpu);
6745 }
6746 
6747 /*
6748  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6749  * part of active balancing operations within "domain".
6750  *
6751  * Returns a task if successful and NULL otherwise.
6752  */
6753 static struct task_struct *detach_one_task(struct lb_env *env)
6754 {
6755         struct task_struct *p, *n;
6756 
6757         lockdep_assert_held(&env->src_rq->lock);
6758 
6759         list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6760                 if (!can_migrate_task(p, env))
6761                         continue;
6762 
6763                 detach_task(p, env);
6764 
6765                 /*
6766                  * Right now, this is only the second place where
6767                  * lb_gained[env->idle] is updated (other is detach_tasks)
6768                  * so we can safely collect stats here rather than
6769                  * inside detach_tasks().
6770                  */
6771                 schedstat_inc(env->sd->lb_gained[env->idle]);
6772                 return p;
6773         }
6774         return NULL;
6775 }
6776 
6777 static const unsigned int sched_nr_migrate_break = 32;
6778 
6779 /*
6780  * detach_tasks() -- tries to detach up to imbalance weighted load from
6781  * busiest_rq, as part of a balancing operation within domain "sd".
6782  *
6783  * Returns number of detached tasks if successful and 0 otherwise.
6784  */
6785 static int detach_tasks(struct lb_env *env)
6786 {
6787         struct list_head *tasks = &env->src_rq->cfs_tasks;
6788         struct task_struct *p;
6789         unsigned long load;
6790         int detached = 0;
6791 
6792         lockdep_assert_held(&env->src_rq->lock);
6793 
6794         if (env->imbalance <= 0)
6795                 return 0;
6796 
6797         while (!list_empty(tasks)) {
6798                 /*
6799                  * We don't want to steal all, otherwise we may be treated likewise,
6800                  * which could at worst lead to a livelock crash.
6801                  */
6802                 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6803                         break;
6804 
6805                 p = list_first_entry(tasks, struct task_struct, se.group_node);
6806 
6807                 env->loop++;
6808                 /* We've more or less seen every task there is, call it quits */
6809                 if (env->loop > env->loop_max)
6810                         break;
6811 
6812                 /* take a breather every nr_migrate tasks */
6813                 if (env->loop > env->loop_break) {
6814                         env->loop_break += sched_nr_migrate_break;
6815                         env->flags |= LBF_NEED_BREAK;
6816                         break;
6817                 }
6818 
6819                 if (!can_migrate_task(p, env))
6820                         goto next;
6821 
6822                 load = task_h_load(p);
6823 
6824                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6825                         goto next;
6826 
6827                 if ((load / 2) > env->imbalance)
6828                         goto next;
6829 
6830                 detach_task(p, env);
6831                 list_add(&p->se.group_node, &env->tasks);
6832 
6833                 detached++;
6834                 env->imbalance -= load;
6835 
6836 #ifdef CONFIG_PREEMPT
6837                 /*
6838                  * NEWIDLE balancing is a source of latency, so preemptible
6839                  * kernels will stop after the first task is detached to minimize
6840                  * the critical section.
6841                  */
6842                 if (env->idle == CPU_NEWLY_IDLE)
6843                         break;
6844 #endif
6845 
6846                 /*
6847                  * We only want to steal up to the prescribed amount of
6848                  * weighted load.
6849                  */
6850                 if (env->imbalance <= 0)
6851                         break;
6852 
6853                 continue;
6854 next:
6855                 list_move_tail(&p->se.group_node, tasks);
6856         }
6857 
6858         /*
6859          * Right now, this is one of only two places we collect this stat
6860          * so we can safely collect detach_one_task() stats here rather
6861          * than inside detach_one_task().
6862          */
6863         schedstat_add(env->sd->lb_gained[env->idle], detached);
6864 
6865         return detached;
6866 }
6867 
6868 /*
6869  * attach_task() -- attach the task detached by detach_task() to its new rq.
6870  */
6871 static void attach_task(struct rq *rq, struct task_struct *p)
6872 {
6873         lockdep_assert_held(&rq->lock);
6874 
6875         BUG_ON(task_rq(p) != rq);
6876         activate_task(rq, p, 0);
6877         p->on_rq = TASK_ON_RQ_QUEUED;
6878         check_preempt_curr(rq, p, 0);
6879 }
6880 
6881 /*
6882  * attach_one_task() -- attaches the task returned from detach_one_task() to
6883  * its new rq.
6884  */
6885 static void attach_one_task(struct rq *rq, struct task_struct *p)
6886 {
6887         raw_spin_lock(&rq->lock);
6888         attach_task(rq, p);
6889         raw_spin_unlock(&rq->lock);
6890 }
6891 
6892 /*
6893  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6894  * new rq.
6895  */
6896 static void attach_tasks(struct lb_env *env)
6897 {
6898         struct list_head *tasks = &env->tasks;
6899         struct task_struct *p;
6900 
6901         raw_spin_lock(&env->dst_rq->lock);
6902 
6903         while (!list_empty(tasks)) {
6904                 p = list_first_entry(tasks, struct task_struct, se.group_node);
6905                 list_del_init(&p->se.group_node);
6906 
6907                 attach_task(env->dst_rq, p);
6908         }
6909 
6910         raw_spin_unlock(&env->dst_rq->lock);
6911 }
6912 
6913 #ifdef CONFIG_FAIR_GROUP_SCHED
6914 static void update_blocked_averages(int cpu)
6915 {
6916         struct rq *rq = cpu_rq(cpu);
6917         struct cfs_rq *cfs_rq;
6918         unsigned long flags;
6919 
6920         raw_spin_lock_irqsave(&rq->lock, flags);
6921         update_rq_clock(rq);
6922 
6923         /*
6924          * Iterates the task_group tree in a bottom up fashion, see
6925          * list_add_leaf_cfs_rq() for details.
6926          */
6927         for_each_leaf_cfs_rq(rq, cfs_rq) {
6928                 /* throttled entities do not contribute to load */
6929                 if (throttled_hierarchy(cfs_rq))
6930                         continue;
6931 
6932                 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6933                         update_tg_load_avg(cfs_rq, 0);
6934 
6935                 /* Propagate pending load changes to the parent */
6936                 if (cfs_rq->tg->se[cpu])
6937                         update_load_avg(cfs_rq->tg->se[cpu], 0);
6938         }
6939         raw_spin_unlock_irqrestore(&rq->lock, flags);
6940 }
6941 
6942 /*
6943  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6944  * This needs to be done in a top-down fashion because the load of a child
6945  * group is a fraction of its parents load.
6946  */
6947 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6948 {
6949         struct rq *rq = rq_of(cfs_rq);
6950         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6951         unsigned long now = jiffies;
6952         unsigned long load;
6953 
6954         if (cfs_rq->last_h_load_update == now)
6955                 return;
6956 
6957         cfs_rq->h_load_next = NULL;
6958         for_each_sched_entity(se) {
6959                 cfs_rq = cfs_rq_of(se);
6960                 cfs_rq->h_load_next = se;
6961                 if (cfs_rq->last_h_load_update == now)
6962                         break;
6963         }
6964 
6965         if (!se) {
6966                 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6967                 cfs_rq->last_h_load_update = now;
6968         }
6969 
6970         while ((se = cfs_rq->h_load_next) != NULL) {
6971                 load = cfs_rq->h_load;
6972                 load = div64_ul(load * se->avg.load_avg,
6973                         cfs_rq_load_avg(cfs_rq) + 1);
6974                 cfs_rq = group_cfs_rq(se);
6975                 cfs_rq->h_load = load;
6976                 cfs_rq->last_h_load_update = now;
6977         }
6978 }
6979 
6980 static unsigned long task_h_load(struct task_struct *p)
6981 {
6982         struct cfs_rq *cfs_rq = task_cfs_rq(p);
6983 
6984         update_cfs_rq_h_load(cfs_rq);
6985         return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6986                         cfs_rq_load_avg(cfs_rq) + 1);
6987 }
6988 #else
6989 static inline void update_blocked_averages(int cpu)
6990 {
6991         struct rq *rq = cpu_rq(cpu);
6992         struct cfs_rq *cfs_rq = &rq->cfs;
6993         unsigned long flags;
6994 
6995         raw_spin_lock_irqsave(&rq->lock, flags);
6996         update_rq_clock(rq);
6997         update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6998         raw_spin_unlock_irqrestore(&rq->lock, flags);
6999 }
7000 
7001 static unsigned long task_h_load(struct task_struct *p)
7002 {
7003         return p->se.avg.load_avg;
7004 }
7005 #endif
7006 
7007 /********** Helpers for find_busiest_group ************************/
7008 
7009 enum group_type {
7010         group_other = 0,
7011         group_imbalanced,
7012         group_overloaded,
7013 };
7014 
7015 /*
7016  * sg_lb_stats - stats of a sched_group required for load_balancing
7017  */
7018 struct sg_lb_stats {
7019         unsigned long avg_load; /*Avg load across the CPUs of the group */
7020         unsigned long group_load; /* Total load over the CPUs of the group */
7021         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7022         unsigned long load_per_task;
7023         unsigned long group_capacity;
7024         unsigned long group_util; /* Total utilization of the group */
7025         unsigned int sum_nr_running; /* Nr tasks running in the group */
7026         unsigned int idle_cpus;
7027         unsigned int group_weight;
7028         enum group_type group_type;
7029         int group_no_capacity;
7030 #ifdef CONFIG_NUMA_BALANCING
7031         unsigned int nr_numa_running;
7032         unsigned int nr_preferred_running;
7033 #endif
7034 };
7035 
7036 /*
7037  * sd_lb_stats - Structure to store the statistics of a sched_domain
7038  *               during load balancing.
7039  */
7040 struct sd_lb_stats {
7041         struct sched_group *busiest;    /* Busiest group in this sd */
7042         struct sched_group *local;      /* Local group in this sd */
7043         unsigned long total_load;       /* Total load of all groups in sd */
7044         unsigned long total_capacity;   /* Total capacity of all groups in sd */
7045         unsigned long avg_load; /* Average load across all groups in sd */
7046 
7047         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7048         struct sg_lb_stats local_stat;  /* Statistics of the local group */
7049 };
7050 
7051 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7052 {
7053         /*
7054          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7055          * local_stat because update_sg_lb_stats() does a full clear/assignment.
7056          * We must however clear busiest_stat::avg_load because
7057          * update_sd_pick_busiest() reads this before assignment.
7058          */
7059         *sds = (struct sd_lb_stats){
7060                 .busiest = NULL,
7061                 .local = NULL,
7062                 .total_load = 0UL,
7063                 .total_capacity = 0UL,
7064                 .busiest_stat = {
7065                         .avg_load = 0UL,
7066                         .sum_nr_running = 0,
7067                         .group_type = group_other,
7068                 },
7069         };
7070 }
7071 
7072 /**
7073  * get_sd_load_idx - Obtain the load index for a given sched domain.
7074  * @sd: The sched_domain whose load_idx is to be obtained.
7075  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7076  *
7077  * Return: The load index.
7078  */
7079 static inline int get_sd_load_idx(struct sched_domain *sd,
7080                                         enum cpu_idle_type idle)
7081 {
7082         int load_idx;
7083 
7084         switch (idle) {
7085         case CPU_NOT_IDLE:
7086                 load_idx = sd->busy_idx;
7087                 break;
7088 
7089         case CPU_NEWLY_IDLE:
7090                 load_idx = sd->newidle_idx;
7091                 break;
7092         default:
7093                 load_idx = sd->idle_idx;
7094                 break;
7095         }
7096 
7097         return load_idx;
7098 }
7099 
7100 static unsigned long scale_rt_capacity(int cpu)
7101 {
7102         struct rq *rq = cpu_rq(cpu);
7103         u64 total, used, age_stamp, avg;
7104         s64 delta;
7105 
7106         /*
7107          * Since we're reading these variables without serialization make sure
7108          * we read them once before doing sanity checks on them.
7109          */
7110         age_stamp = READ_ONCE(rq->age_stamp);
7111         avg = READ_ONCE(rq->rt_avg);
7112         delta = __rq_clock_broken(rq) - age_stamp;
7113 
7114         if (unlikely(delta < 0))
7115                 delta = 0;
7116 
7117         total = sched_avg_period() + delta;
7118 
7119         used = div_u64(avg, total);
7120 
7121         if (likely(used < SCHED_CAPACITY_SCALE))
7122                 return SCHED_CAPACITY_SCALE - used;
7123 
7124         return 1;
7125 }
7126 
7127 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7128 {
7129         unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7130         struct sched_group *sdg = sd->groups;
7131 
7132         cpu_rq(cpu)->cpu_capacity_orig = capacity;
7133 
7134         capacity *= scale_rt_capacity(cpu);
7135         capacity >>= SCHED_CAPACITY_SHIFT;
7136 
7137         if (!capacity)
7138                 capacity = 1;
7139 
7140         cpu_rq(cpu)->cpu_capacity = capacity;
7141         sdg->sgc->capacity = capacity;
7142         sdg->sgc->min_capacity = capacity;
7143 }
7144 
7145 void update_group_capacity(struct sched_domain *sd, int cpu)
7146 {
7147         struct sched_domain *child = sd->child;
7148         struct sched_group *group, *sdg = sd->groups;
7149         unsigned long capacity, min_capacity;
7150         unsigned long interval;
7151 
7152         interval = msecs_to_jiffies(sd->balance_interval);
7153         interval = clamp(interval, 1UL, max_load_balance_interval);
7154         sdg->sgc->next_update = jiffies + interval;
7155 
7156         if (!child) {
7157                 update_cpu_capacity(sd, cpu);
7158                 return;
7159         }
7160 
7161         capacity = 0;
7162         min_capacity = ULONG_MAX;
7163 
7164         if (child->flags & SD_OVERLAP) {
7165                 /*
7166                  * SD_OVERLAP domains cannot assume that child groups
7167                  * span the current group.
7168                  */
7169 
7170                 for_each_cpu(cpu, sched_group_cpus(sdg)) {
7171                         struct sched_group_capacity *sgc;
7172                         struct rq *rq = cpu_rq(cpu);
7173 
7174                         /*
7175                          * build_sched_domains() -> init_sched_groups_capacity()
7176                          * gets here before we've attached the domains to the
7177                          * runqueues.
7178                          *
7179                          * Use capacity_of(), which is set irrespective of domains
7180                          * in update_cpu_capacity().
7181                          *
7182                          * This avoids capacity from being 0 and
7183                          * causing divide-by-zero issues on boot.
7184                          */
7185                         if (unlikely(!rq->sd)) {
7186                                 capacity += capacity_of(cpu);
7187                         } else {
7188                                 sgc = rq->sd->groups->sgc;
7189                                 capacity += sgc->capacity;
7190                         }
7191 
7192                         min_capacity = min(capacity, min_capacity);
7193                 }
7194         } else  {
7195                 /*
7196                  * !SD_OVERLAP domains can assume that child groups
7197                  * span the current group.
7198                  */
7199 
7200                 group = child->groups;
7201                 do {
7202                         struct sched_group_capacity *sgc = group->sgc;
7203 
7204                         capacity += sgc->capacity;
7205                         min_capacity = min(sgc->min_capacity, min_capacity);
7206                         group = group->next;
7207                 } while (group != child->groups);
7208         }
7209 
7210         sdg->sgc->capacity = capacity;
7211         sdg->sgc->min_capacity = min_capacity;
7212 }
7213 
7214 /*
7215  * Check whether the capacity of the rq has been noticeably reduced by side
7216  * activity. The imbalance_pct is used for the threshold.
7217  * Return true is the capacity is reduced
7218  */
7219 static inline int
7220 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7221 {
7222         return ((rq->cpu_capacity * sd->imbalance_pct) <
7223                                 (rq->cpu_capacity_orig * 100));
7224 }
7225 
7226 /*
7227  * Group imbalance indicates (and tries to solve) the problem where balancing
7228  * groups is inadequate due to tsk_cpus_allowed() constraints.
7229  *
7230  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
7231  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
7232  * Something like:
7233  *
7234  *      { 0 1 2 3 } { 4 5 6 7 }
7235  *              *     * * *
7236  *
7237  * If we were to balance group-wise we'd place two tasks in the first group and
7238  * two tasks in the second group. Clearly this is undesired as it will overload
7239  * cpu 3 and leave one of the cpus in the second group unused.
7240  *
7241  * The current solution to this issue is detecting the skew in the first group
7242  * by noticing the lower domain failed to reach balance and had difficulty
7243  * moving tasks due to affinity constraints.
7244  *
7245  * When this is so detected; this group becomes a candidate for busiest; see
7246  * update_sd_pick_busiest(). And calculate_imbalance() and
7247  * find_busiest_group() avoid some of the usual balance conditions to allow it
7248  * to create an effective group imbalance.
7249  *
7250  * This is a somewhat tricky proposition since the next run might not find the
7251  * group imbalance and decide the groups need to be balanced again. A most
7252  * subtle and fragile situation.
7253  */
7254 
7255 static inline int sg_imbalanced(struct sched_group *group)
7256 {
7257         return group->sgc->imbalance;
7258 }
7259 
7260 /*
7261  * group_has_capacity returns true if the group has spare capacity that could
7262  * be used by some tasks.
7263  * We consider that a group has spare capacity if the  * number of task is
7264  * smaller than the number of CPUs or if the utilization is lower than the
7265  * available capacity for CFS tasks.
7266  * For the latter, we use a threshold to stabilize the state, to take into
7267  * account the variance of the tasks' load and to return true if the available
7268  * capacity in meaningful for the load balancer.
7269  * As an example, an available capacity of 1% can appear but it doesn't make
7270  * any benefit for the load balance.
7271  */
7272 static inline bool
7273 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7274 {
7275         if (sgs->sum_nr_running < sgs->group_weight)
7276                 return true;
7277 
7278         if ((sgs->group_capacity * 100) >
7279                         (sgs->group_util * env->sd->imbalance_pct))
7280                 return true;
7281 
7282         return false;
7283 }
7284 
7285 /*
7286  *  group_is_overloaded returns true if the group has more tasks than it can
7287  *  handle.
7288  *  group_is_overloaded is not equals to !group_has_capacity because a group
7289  *  with the exact right number of tasks, has no more spare capacity but is not
7290  *  overloaded so both group_has_capacity and group_is_overloaded return
7291  *  false.
7292  */
7293 static inline bool
7294 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7295 {
7296         if (sgs->sum_nr_running <= sgs->group_weight)
7297                 return false;
7298 
7299         if ((sgs->group_capacity * 100) <
7300                         (sgs->group_util * env->sd->imbalance_pct))
7301                 return true;
7302 
7303         return false;
7304 }
7305 
7306 /*
7307  * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7308  * per-CPU capacity than sched_group ref.
7309  */
7310 static inline bool
7311 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7312 {
7313         return sg->sgc->min_capacity * capacity_margin <
7314                                                 ref->sgc->min_capacity * 1024;
7315 }
7316 
7317 static inline enum
7318 group_type group_classify(struct sched_group *group,
7319                           struct sg_lb_stats *sgs)
7320 {
7321         if (sgs->group_no_capacity)
7322                 return group_overloaded;
7323 
7324         if (sg_imbalanced(group))
7325                 return group_imbalanced;
7326 
7327         return group_other;
7328 }
7329 
7330 /**
7331  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7332  * @env: The load balancing environment.
7333  * @group: sched_group whose statistics are to be updated.
7334  * @load_idx: Load index of sched_domain of this_cpu for load calc.
7335  * @local_group: Does group contain this_cpu.
7336  * @sgs: variable to hold the statistics for this group.
7337  * @overload: Indicate more than one runnable task for any CPU.
7338  */
7339 static inline void update_sg_lb_stats(struct lb_env *env,
7340                         struct sched_group *group, int load_idx,
7341                         int local_group, struct sg_lb_stats *sgs,
7342                         bool *overload)
7343 {
7344         unsigned long load;
7345         int i, nr_running;
7346 
7347         memset(sgs, 0, sizeof(*sgs));
7348 
7349         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7350                 struct rq *rq = cpu_rq(i);
7351 
7352                 /* Bias balancing toward cpus of our domain */
7353                 if (local_group)
7354                         load = target_load(i, load_idx);
7355                 else
7356                         load = source_load(i, load_idx);
7357 
7358                 sgs->group_load += load;
7359                 sgs->group_util += cpu_util(i);
7360                 sgs->sum_nr_running += rq->cfs.h_nr_running;
7361 
7362                 nr_running = rq->nr_running;
7363                 if (nr_running > 1)
7364                         *overload = true;
7365 
7366 #ifdef CONFIG_NUMA_BALANCING
7367                 sgs->nr_numa_running += rq->nr_numa_running;
7368                 sgs->nr_preferred_running += rq->nr_preferred_running;
7369 #endif
7370                 sgs->sum_weighted_load += weighted_cpuload(i);
7371                 /*
7372                  * No need to call idle_cpu() if nr_running is not 0
7373                  */
7374                 if (!nr_running && idle_cpu(i))
7375                         sgs->idle_cpus++;
7376         }
7377 
7378         /* Adjust by relative CPU capacity of the group */
7379         sgs->group_capacity = group->sgc->capacity;
7380         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7381 
7382         if (sgs->sum_nr_running)
7383                 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7384 
7385         sgs->group_weight = group->group_weight;
7386 
7387         sgs->group_no_capacity = group_is_overloaded(env, sgs);
7388         sgs->group_type = group_classify(group, sgs);
7389 }
7390 
7391 /**
7392  * update_sd_pick_busiest - return 1 on busiest group
7393  * @env: The load balancing environment.
7394  * @sds: sched_domain statistics
7395  * @sg: sched_group candidate to be checked for being the busiest
7396  * @sgs: sched_group statistics
7397  *
7398  * Determine if @sg is a busier group than the previously selected
7399  * busiest group.
7400  *
7401  * Return: %true if @sg is a busier group than the previously selected
7402  * busiest group. %false otherwise.
7403  */
7404 static bool update_sd_pick_busiest(struct lb_env *env,
7405                                    struct sd_lb_stats *sds,
7406                                    struct sched_group *sg,
7407                                    struct sg_lb_stats *sgs)
7408 {
7409         struct sg_lb_stats *busiest = &sds->busiest_stat;
7410 
7411         if (sgs->group_type > busiest->group_type)
7412                 return true;
7413 
7414         if (sgs->group_type < busiest->group_type)
7415                 return false;
7416 
7417         if (sgs->avg_load <= busiest->avg_load)
7418                 return false;
7419 
7420         if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7421                 goto asym_packing;
7422 
7423         /*
7424          * Candidate sg has no more than one task per CPU and
7425          * has higher per-CPU capacity. Migrating tasks to less
7426          * capable CPUs may harm throughput. Maximize throughput,
7427          * power/energy consequences are not considered.
7428          */
7429         if (sgs->sum_nr_running <= sgs->group_weight &&
7430             group_smaller_cpu_capacity(sds->local, sg))
7431                 return false;
7432 
7433 asym_packing:
7434         /* This is the busiest node in its class. */
7435         if (!(env->sd->flags & SD_ASYM_PACKING))
7436                 return true;
7437 
7438         /* No ASYM_PACKING if target cpu is already busy */
7439         if (env->idle == CPU_NOT_IDLE)
7440                 return true;
7441         /*
7442          * ASYM_PACKING needs to move all the work to the highest
7443          * prority CPUs in the group, therefore mark all groups
7444          * of lower priority than ourself as busy.
7445          */
7446         if (sgs->sum_nr_running &&
7447             sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7448                 if (!sds->busiest)
7449                         return true;
7450 
7451                 /* Prefer to move from lowest priority cpu's work */
7452                 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7453                                       sg->asym_prefer_cpu))
7454                         return true;
7455         }
7456 
7457         return false;
7458 }
7459 
7460 #ifdef CONFIG_NUMA_BALANCING
7461 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7462 {
7463         if (sgs->sum_nr_running > sgs->nr_numa_running)
7464                 return regular;
7465         if (sgs->sum_nr_running > sgs->nr_preferred_running)
7466                 return remote;
7467         return all;
7468 }
7469 
7470 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7471 {
7472         if (rq->nr_running > rq->nr_numa_running)
7473                 return regular;
7474         if (rq->nr_running > rq->nr_preferred_running)
7475                 return remote;
7476         return all;
7477 }
7478 #else
7479 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7480 {
7481         return all;
7482 }
7483 
7484 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7485 {
7486         return regular;
7487 }
7488 #endif /* CONFIG_NUMA_BALANCING */
7489 
7490 /**
7491  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7492  * @env: The load balancing environment.
7493  * @sds: variable to hold the statistics for this sched_domain.
7494  */
7495 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7496 {
7497         struct sched_domain *child = env->sd->child;
7498         struct sched_group *sg = env->sd->groups;
7499         struct sg_lb_stats tmp_sgs;
7500         int load_idx, prefer_sibling = 0;
7501         bool overload = false;
7502 
7503         if (child && child->flags & SD_PREFER_SIBLING)
7504                 prefer_sibling = 1;
7505 
7506         load_idx = get_sd_load_idx(env->sd, env->idle);
7507 
7508         do {
7509                 struct sg_lb_stats *sgs = &tmp_sgs;
7510                 int local_group;
7511 
7512                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
7513                 if (local_group) {
7514                         sds->local = sg;
7515                         sgs = &sds->local_stat;
7516 
7517                         if (env->idle != CPU_NEWLY_IDLE ||
7518                             time_after_eq(jiffies, sg->sgc->next_update))
7519                                 update_group_capacity(env->sd, env->dst_cpu);
7520                 }
7521 
7522                 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7523                                                 &overload);
7524 
7525                 if (local_group)
7526                         goto next_group;
7527 
7528                 /*
7529                  * In case the child domain prefers tasks go to siblings
7530                  * first, lower the sg capacity so that we'll try
7531                  * and move all the excess tasks away. We lower the capacity
7532                  * of a group only if the local group has the capacity to fit
7533                  * these excess tasks. The extra check prevents the case where
7534                  * you always pull from the heaviest group when it is already
7535                  * under-utilized (possible with a large weight task outweighs
7536                  * the tasks on the system).
7537                  */
7538                 if (prefer_sibling && sds->local &&
7539                     group_has_capacity(env, &sds->local_stat) &&
7540                     (sgs->sum_nr_running > 1)) {
7541                         sgs->group_no_capacity = 1;
7542                         sgs->group_type = group_classify(sg, sgs);
7543                 }
7544 
7545                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7546                         sds->busiest = sg;
7547                         sds->busiest_stat = *sgs;
7548                 }
7549 
7550 next_group:
7551                 /* Now, start updating sd_lb_stats */
7552                 sds->total_load += sgs->group_load;
7553                 sds->total_capacity += sgs->group_capacity;
7554 
7555                 sg = sg->next;
7556         } while (sg != env->sd->groups);
7557 
7558         if (env->sd->flags & SD_NUMA)
7559                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
7560 
7561         if (!env->sd->parent) {
7562                 /* update overload indicator if we are at root domain */
7563                 if (env->dst_rq->rd->overload != overload)
7564                         env->dst_rq->rd->overload = overload;
7565         }
7566 
7567 }
7568 
7569 /**
7570  * check_asym_packing - Check to see if the group is packed into the
7571  *                      sched doman.
7572  *
7573  * This is primarily intended to used at the sibling level.  Some
7574  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
7575  * case of POWER7, it can move to lower SMT modes only when higher
7576  * threads are idle.  When in lower SMT modes, the threads will
7577  * perform better since they share less core resources.  Hence when we
7578  * have idle threads, we want them to be the higher ones.
7579  *
7580  * This packing function is run on idle threads.  It checks to see if
7581  * the busiest CPU in this domain (core in the P7 case) has a higher
7582  * CPU number than the packing function is being run on.  Here we are
7583  * assuming lower CPU number will be equivalent to lower a SMT thread
7584  * number.
7585  *
7586  * Return: 1 when packing is required and a task should be moved to
7587  * this CPU.  The amount of the imbalance is returned in *imbalance.
7588  *
7589  * @env: The load balancing environment.
7590  * @sds: Statistics of the sched_domain which is to be packed
7591  */
7592 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7593 {
7594         int busiest_cpu;
7595 
7596         if (!(env->sd->flags & SD_ASYM_PACKING))
7597                 return 0;
7598 
7599         if (env->idle == CPU_NOT_IDLE)
7600                 return 0;
7601 
7602         if (!sds->busiest)
7603                 return 0;
7604 
7605         busiest_cpu = sds->busiest->asym_prefer_cpu;
7606         if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
7607                 return 0;
7608 
7609         env->imbalance = DIV_ROUND_CLOSEST(
7610                 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7611                 SCHED_CAPACITY_SCALE);
7612 
7613         return 1;
7614 }
7615 
7616 /**
7617  * fix_small_imbalance - Calculate the minor imbalance that exists
7618  *                      amongst the groups of a sched_domain, during
7619  *                      load balancing.
7620  * @env: The load balancing environment.
7621  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7622  */
7623 static inline
7624 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7625 {
7626         unsigned long tmp, capa_now = 0, capa_move = 0;
7627         unsigned int imbn = 2;
7628         unsigned long scaled_busy_load_per_task;
7629         struct sg_lb_stats *local, *busiest;
7630 
7631         local = &sds->local_stat;
7632         busiest = &sds->busiest_stat;
7633 
7634         if (!local->sum_nr_running)
7635                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7636         else if (busiest->load_per_task > local->load_per_task)
7637                 imbn = 1;
7638 
7639         scaled_busy_load_per_task =
7640                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7641                 busiest->group_capacity;
7642 
7643         if (busiest->avg_load + scaled_busy_load_per_task >=
7644             local->avg_load + (scaled_busy_load_per_task * imbn)) {
7645                 env->imbalance = busiest->load_per_task;
7646                 return;
7647         }
7648 
7649         /*
7650          * OK, we don't have enough imbalance to justify moving tasks,
7651          * however we may be able to increase total CPU capacity used by
7652          * moving them.
7653          */
7654 
7655         capa_now += busiest->group_capacity *
7656                         min(busiest->load_per_task, busiest->avg_load);
7657         capa_now += local->group_capacity *
7658                         min(local->load_per_task, local->avg_load);
7659         capa_now /= SCHED_CAPACITY_SCALE;
7660 
7661         /* Amount of load we'd subtract */
7662         if (busiest->avg_load > scaled_busy_load_per_task) {
7663                 capa_move += busiest->group_capacity *
7664                             min(busiest->load_per_task,
7665                                 busiest->avg_load - scaled_busy_load_per_task);
7666         }
7667 
7668         /* Amount of load we'd add */
7669         if (busiest->avg_load * busiest->group_capacity <
7670             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7671                 tmp = (busiest->avg_load * busiest->group_capacity) /
7672                       local->group_capacity;
7673         } else {
7674                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7675                       local->group_capacity;
7676         }
7677         capa_move += local->group_capacity *
7678                     min(local->load_per_task, local->avg_load + tmp);
7679         capa_move /= SCHED_CAPACITY_SCALE;
7680 
7681         /* Move if we gain throughput */
7682         if (capa_move > capa_now)
7683                 env->imbalance = busiest->load_per_task;
7684 }
7685 
7686 /**
7687  * calculate_imbalance - Calculate the amount of imbalance present within the
7688  *                       groups of a given sched_domain during load balance.
7689  * @env: load balance environment
7690  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7691  */
7692 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7693 {
7694         unsigned long max_pull, load_above_capacity = ~0UL;
7695         struct sg_lb_stats *local, *busiest;
7696 
7697         local = &sds->local_stat;
7698         busiest = &sds->busiest_stat;
7699 
7700         if (busiest->group_type == group_imbalanced) {
7701                 /*
7702                  * In the group_imb case we cannot rely on group-wide averages
7703                  * to ensure cpu-load equilibrium, look at wider averages. XXX
7704                  */
7705                 busiest->load_per_task =
7706                         min(busiest->load_per_task, sds->avg_load);
7707         }
7708 
7709         /*
7710          * Avg load of busiest sg can be less and avg load of local sg can
7711          * be greater than avg load across all sgs of sd because avg load
7712          * factors in sg capacity and sgs with smaller group_type are
7713          * skipped when updating the busiest sg:
7714          */
7715         if (busiest->avg_load <= sds->avg_load ||
7716             local->avg_load >= sds->avg_load) {
7717                 env->imbalance = 0;
7718                 return fix_small_imbalance(env, sds);
7719         }
7720 
7721         /*
7722          * If there aren't any idle cpus, avoid creating some.
7723          */
7724         if (busiest->group_type == group_overloaded &&
7725             local->group_type   == group_overloaded) {
7726                 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7727                 if (load_above_capacity > busiest->group_capacity) {
7728                         load_above_capacity -= busiest->group_capacity;
7729                         load_above_capacity *= scale_load_down(NICE_0_LOAD);
7730                         load_above_capacity /= busiest->group_capacity;
7731                 } else
7732                         load_above_capacity = ~0UL;
7733         }
7734 
7735         /*
7736          * We're trying to get all the cpus to the average_load, so we don't
7737          * want to push ourselves above the average load, nor do we wish to
7738          * reduce the max loaded cpu below the average load. At the same time,
7739          * we also don't want to reduce the group load below the group
7740          * capacity. Thus we look for the minimum possible imbalance.
7741          */
7742         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7743 
7744         /* How much load to actually move to equalise the imbalance */
7745         env->imbalance = min(
7746                 max_pull * busiest->group_capacity,
7747                 (sds->avg_load - local->avg_load) * local->group_capacity
7748         ) / SCHED_CAPACITY_SCALE;
7749 
7750         /*
7751          * if *imbalance is less than the average load per runnable task
7752          * there is no guarantee that any tasks will be moved so we'll have
7753          * a think about bumping its value to force at least one task to be
7754          * moved
7755          */
7756         if (env->imbalance < busiest->load_per_task)
7757                 return fix_small_imbalance(env, sds);
7758 }
7759 
7760 /******* find_busiest_group() helpers end here *********************/
7761 
7762 /**
7763  * find_busiest_group - Returns the busiest group within the sched_domain
7764  * if there is an imbalance.
7765  *
7766  * Also calculates the amount of weighted load which should be moved
7767  * to restore balance.
7768  *
7769  * @env: The load balancing environment.
7770  *
7771  * Return:      - The busiest group if imbalance exists.
7772  */
7773 static struct sched_group *find_busiest_group(struct lb_env *env)
7774 {
7775         struct sg_lb_stats *local, *busiest;
7776         struct sd_lb_stats sds;
7777 
7778         init_sd_lb_stats(&sds);
7779 
7780         /*
7781          * Compute the various statistics relavent for load balancing at
7782          * this level.
7783          */
7784         update_sd_lb_stats(env, &sds);
7785         local = &sds.local_stat;
7786         busiest = &sds.busiest_stat;
7787 
7788         /* ASYM feature bypasses nice load balance check */
7789         if (check_asym_packing(env, &sds))
7790                 return sds.busiest;
7791 
7792         /* There is no busy sibling group to pull tasks from */
7793         if (!sds.busiest || busiest->sum_nr_running == 0)
7794                 goto out_balanced;
7795 
7796         sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7797                                                 / sds.total_capacity;
7798 
7799         /*
7800          * If the busiest group is imbalanced the below checks don't
7801          * work because they assume all things are equal, which typically
7802          * isn't true due to cpus_allowed constraints and the like.
7803          */
7804         if (busiest->group_type == group_imbalanced)
7805                 goto force_balance;
7806 
7807         /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7808         if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7809             busiest->group_no_capacity)
7810                 goto force_balance;
7811 
7812         /*
7813          * If the local group is busier than the selected busiest group
7814          * don't try and pull any tasks.
7815          */
7816         if (local->avg_load >= busiest->avg_load)
7817                 goto out_balanced;
7818 
7819         /*
7820          * Don't pull any tasks if this group is already above the domain
7821          * average load.
7822          */
7823         if (local->avg_load >= sds.avg_load)
7824                 goto out_balanced;
7825 
7826         if (env->idle == CPU_IDLE) {
7827                 /*
7828                  * This cpu is idle. If the busiest group is not overloaded
7829                  * and there is no imbalance between this and busiest group
7830                  * wrt idle cpus, it is balanced. The imbalance becomes
7831                  * significant if the diff is greater than 1 otherwise we
7832                  * might end up to just move the imbalance on another group
7833                  */
7834                 if ((busiest->group_type != group_overloaded) &&
7835                                 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7836                         goto out_balanced;
7837         } else {
7838                 /*
7839                  * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7840                  * imbalance_pct to be conservative.
7841                  */
7842                 if (100 * busiest->avg_load <=
7843                                 env->sd->imbalance_pct * local->avg_load)
7844                         goto out_balanced;
7845         }
7846 
7847 force_balance:
7848         /* Looks like there is an imbalance. Compute it */
7849         calculate_imbalance(env, &sds);
7850         return sds.busiest;
7851 
7852 out_balanced:
7853         env->imbalance = 0;
7854         return NULL;
7855 }
7856 
7857 /*
7858  * find_busiest_queue - find the busiest runqueue among the cpus in group.
7859  */
7860 static struct rq *find_busiest_queue(struct lb_env *env,
7861                                      struct sched_group *group)
7862 {
7863         struct rq *busiest = NULL, *rq;
7864         unsigned long busiest_load = 0, busiest_capacity = 1;
7865         int i;
7866 
7867         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7868                 unsigned long capacity, wl;
7869                 enum fbq_type rt;
7870 
7871                 rq = cpu_rq(i);
7872                 rt = fbq_classify_rq(rq);
7873 
7874                 /*
7875                  * We classify groups/runqueues into three groups:
7876                  *  - regular: there are !numa tasks
7877                  *  - remote:  there are numa tasks that run on the 'wrong' node
7878                  *  - all:     there is no distinction
7879                  *
7880                  * In order to avoid migrating ideally placed numa tasks,
7881                  * ignore those when there's better options.
7882                  *
7883                  * If we ignore the actual busiest queue to migrate another
7884                  * task, the next balance pass can still reduce the busiest
7885                  * queue by moving tasks around inside the node.
7886                  *
7887                  * If we cannot move enough load due to this classification
7888                  * the next pass will adjust the group classification and
7889                  * allow migration of more tasks.
7890                  *
7891                  * Both cases only affect the total convergence complexity.
7892                  */
7893                 if (rt > env->fbq_type)
7894                         continue;
7895 
7896                 capacity = capacity_of(i);
7897 
7898                 wl = weighted_cpuload(i);
7899 
7900                 /*
7901                  * When comparing with imbalance, use weighted_cpuload()
7902                  * which is not scaled with the cpu capacity.
7903                  */
7904 
7905                 if (rq->nr_running == 1 && wl > env->imbalance &&
7906                     !check_cpu_capacity(rq, env->sd))
7907                         continue;
7908 
7909                 /*
7910                  * For the load comparisons with the other cpu's, consider
7911                  * the weighted_cpuload() scaled with the cpu capacity, so
7912                  * that the load can be moved away from the cpu that is
7913                  * potentially running at a lower capacity.
7914                  *
7915                  * Thus we're looking for max(wl_i / capacity_i), crosswise
7916                  * multiplication to rid ourselves of the division works out
7917                  * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
7918                  * our previous maximum.
7919                  */
7920                 if (wl * busiest_capacity > busiest_load * capacity) {
7921                         busiest_load = wl;
7922                         busiest_capacity = capacity;
7923                         busiest = rq;
7924                 }
7925         }
7926 
7927         return busiest;
7928 }
7929 
7930 /*
7931  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7932  * so long as it is large enough.
7933  */
7934 #define MAX_PINNED_INTERVAL     512
7935 
7936 static int need_active_balance(struct lb_env *env)
7937 {
7938         struct sched_domain *sd = env->sd;
7939 
7940         if (env->idle == CPU_NEWLY_IDLE) {
7941 
7942                 /*
7943                  * ASYM_PACKING needs to force migrate tasks from busy but
7944                  * lower priority CPUs in order to pack all tasks in the
7945                  * highest priority CPUs.
7946                  */
7947                 if ((sd->flags & SD_ASYM_PACKING) &&
7948                     sched_asym_prefer(env->dst_cpu, env->src_cpu))
7949                         return 1;
7950         }
7951 
7952         /*
7953          * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7954          * It's worth migrating the task if the src_cpu's capacity is reduced
7955          * because of other sched_class or IRQs if more capacity stays
7956          * available on dst_cpu.
7957          */
7958         if ((env->idle != CPU_NOT_IDLE) &&
7959             (env->src_rq->cfs.h_nr_running == 1)) {
7960                 if ((check_cpu_capacity(env->src_rq, sd)) &&
7961                     (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7962                         return 1;
7963         }
7964 
7965         return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7966 }
7967 
7968 static int active_load_balance_cpu_stop(void *data);
7969 
7970 static int should_we_balance(struct lb_env *env)
7971 {
7972         struct sched_group *sg = env->sd->groups;
7973         struct cpumask *sg_cpus, *sg_mask;
7974         int cpu, balance_cpu = -1;
7975 
7976         /*
7977          * In the newly idle case, we will allow all the cpu's
7978          * to do the newly idle load balance.
7979          */
7980         if (env->idle == CPU_NEWLY_IDLE)
7981                 return 1;
7982 
7983         sg_cpus = sched_group_cpus(sg);
7984         sg_mask = sched_group_mask(sg);
7985         /* Try to find first idle cpu */
7986         for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7987                 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7988                         continue;
7989 
7990                 balance_cpu = cpu;
7991                 break;
7992         }
7993 
7994         if (balance_cpu == -1)
7995                 balance_cpu = group_balance_cpu(sg);
7996 
7997         /*
7998          * First idle cpu or the first cpu(busiest) in this sched group
7999          * is eligible for doing load balancing at this and above domains.
8000          */
8001         return balance_cpu == env->dst_cpu;
8002 }
8003 
8004 /*
8005  * Check this_cpu to ensure it is balanced within domain. Attempt to move
8006  * tasks if there is an imbalance.
8007  */
8008 static int load_balance(int this_cpu, struct rq *this_rq,
8009                         struct sched_domain *sd, enum cpu_idle_type idle,
8010                         int *continue_balancing)
8011 {
8012         int ld_moved, cur_ld_moved, active_balance = 0;
8013         struct sched_domain *sd_parent = sd->parent;
8014         struct sched_group *group;
8015         struct rq *busiest;
8016         unsigned long flags;
8017         struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8018 
8019         struct lb_env env = {
8020                 .sd             = sd,
8021                 .dst_cpu        = this_cpu,
8022                 .dst_rq         = this_rq,
8023                 .dst_grpmask    = sched_group_cpus(sd->groups),
8024                 .idle           = idle,
8025                 .loop_break     = sched_nr_migrate_break,
8026                 .cpus           = cpus,
8027                 .fbq_type       = all,
8028                 .tasks          = LIST_HEAD_INIT(env.tasks),
8029         };
8030 
8031         /*
8032          * For NEWLY_IDLE load_balancing, we don't need to consider
8033          * other cpus in our group
8034          */
8035         if (idle == CPU_NEWLY_IDLE)
8036                 env.dst_grpmask = NULL;
8037 
8038         cpumask_copy(cpus, cpu_active_mask);
8039 
8040         schedstat_inc(sd->lb_count[idle]);
8041 
8042 redo:
8043         if (!should_we_balance(&env)) {
8044                 *continue_balancing = 0;
8045                 goto out_balanced;
8046         }
8047 
8048         group = find_busiest_group(&env);
8049         if (!group) {
8050                 schedstat_inc(sd->lb_nobusyg[idle]);
8051                 goto out_balanced;
8052         }
8053 
8054         busiest = find_busiest_queue(&env, group);
8055         if (!busiest) {
8056                 schedstat_inc(sd->lb_nobusyq[idle]);
8057                 goto out_balanced;
8058         }
8059 
8060         BUG_ON(busiest == env.dst_rq);
8061 
8062         schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8063 
8064         env.src_cpu = busiest->cpu;
8065         env.src_rq = busiest;
8066 
8067         ld_moved = 0;
8068         if (busiest->nr_running > 1) {
8069                 /*
8070                  * Attempt to move tasks. If find_busiest_group has found
8071                  * an imbalance but busiest->nr_running <= 1, the group is
8072                  * still unbalanced. ld_moved simply stays zero, so it is
8073                  * correctly treated as an imbalance.
8074                  */
8075                 env.flags |= LBF_ALL_PINNED;
8076                 env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
8077 
8078 more_balance:
8079                 raw_spin_lock_irqsave(&busiest->lock, flags);
8080 
8081                 /*
8082                  * cur_ld_moved - load moved in current iteration
8083                  * ld_moved     - cumulative load moved across iterations
8084                  */
8085                 cur_ld_moved = detach_tasks(&env);
8086 
8087                 /*
8088                  * We've detached some tasks from busiest_rq. Every
8089                  * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8090                  * unlock busiest->lock, and we are able to be sure
8091                  * that nobody can manipulate the tasks in parallel.
8092                  * See task_rq_lock() family for the details.
8093                  */
8094 
8095                 raw_spin_unlock(&busiest->lock);
8096 
8097                 if (cur_ld_moved) {
8098                         attach_tasks(&env);
8099                         ld_moved += cur_ld_moved;
8100                 }
8101 
8102                 local_irq_restore(flags);
8103 
8104                 if (env.flags & LBF_NEED_BREAK) {
8105                         env.flags &= ~LBF_NEED_BREAK;
8106                         goto more_balance;
8107                 }
8108 
8109                 /*
8110                  * Revisit (affine) tasks on src_cpu that couldn't be moved to
8111                  * us and move them to an alternate dst_cpu in our sched_group
8112                  * where they can run. The upper limit on how many times we
8113                  * iterate on same src_cpu is dependent on number of cpus in our
8114                  * sched_group.
8115                  *
8116                  * This changes load balance semantics a bit on who can move
8117                  * load to a given_cpu. In addition to the given_cpu itself
8118                  * (or a ilb_cpu acting on its behalf where given_cpu is
8119                  * nohz-idle), we now have balance_cpu in a position to move
8120                  * load to given_cpu. In rare situations, this may cause
8121                  * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8122                  * _independently_ and at _same_ time to move some load to
8123                  * given_cpu) causing exceess load to be moved to given_cpu.
8124                  * This however should not happen so much in practice and
8125                  * moreover subsequent load balance cycles should correct the
8126                  * excess load moved.
8127                  */
8128                 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8129 
8130                         /* Prevent to re-select dst_cpu via env's cpus */
8131                         cpumask_clear_cpu(env.dst_cpu, env.cpus);
8132 
8133                         env.dst_rq       = cpu_rq(env.new_dst_cpu);
8134                         env.dst_cpu      = env.new_dst_cpu;
8135                         env.flags       &= ~LBF_DST_PINNED;
8136                         env.loop         = 0;
8137                         env.loop_break   = sched_nr_migrate_break;
8138 
8139                         /*
8140                          * Go back to "more_balance" rather than "redo" since we
8141                          * need to continue with same src_cpu.
8142                          */
8143                         goto more_balance;
8144                 }
8145 
8146                 /*
8147                  * We failed to reach balance because of affinity.
8148                  */
8149                 if (sd_parent) {
8150                         int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8151 
8152                         if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8153                                 *group_imbalance = 1;
8154                 }
8155 
8156                 /* All tasks on this runqueue were pinned by CPU affinity */
8157                 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8158                         cpumask_clear_cpu(cpu_of(busiest), cpus);
8159                         if (!cpumask_empty(cpus)) {
8160                                 env.loop = 0;
8161                                 env.loop_break = sched_nr_migrate_break;
8162                                 goto redo;
8163                         }
8164                         goto out_all_pinned;
8165                 }
8166         }
8167 
8168         if (!ld_moved) {
8169                 schedstat_inc(sd->lb_failed[idle]);
8170                 /*
8171                  * Increment the failure counter only on periodic balance.
8172                  * We do not want newidle balance, which can be very
8173                  * frequent, pollute the failure counter causing
8174                  * excessive cache_hot migrations and active balances.
8175                  */
8176                 if (idle != CPU_NEWLY_IDLE)
8177                         sd->nr_balance_failed++;
8178 
8179                 if (need_active_balance(&env)) {
8180                         raw_spin_lock_irqsave(&busiest->lock, flags);
8181 
8182                         /* don't kick the active_load_balance_cpu_stop,
8183                          * if the curr task on busiest cpu can't be
8184                          * moved to this_cpu
8185                          */
8186                         if (!cpumask_test_cpu(this_cpu,
8187                                         tsk_cpus_allowed(busiest->curr))) {
8188                                 raw_spin_unlock_irqrestore(&busiest->lock,
8189                                                             flags);
8190                                 env.flags |= LBF_ALL_PINNED;
8191                                 goto out_one_pinned;
8192                         }
8193 
8194                         /*
8195                          * ->active_balance synchronizes accesses to
8196                          * ->active_balance_work.  Once set, it's cleared
8197                          * only after active load balance is finished.
8198                          */
8199                         if (!busiest->active_balance) {
8200                                 busiest->active_balance = 1;
8201                                 busiest->push_cpu = this_cpu;
8202                                 active_balance = 1;
8203                         }
8204                         raw_spin_unlock_irqrestore(&busiest->lock, flags);
8205 
8206                         if (active_balance) {
8207                                 stop_one_cpu_nowait(cpu_of(busiest),
8208                                         active_load_balance_cpu_stop, busiest,
8209                                         &busiest->active_balance_work);
8210                         }
8211 
8212                         /* We've kicked active balancing, force task migration. */
8213                         sd->nr_balance_failed = sd->cache_nice_tries+1;
8214                 }
8215         } else
8216                 sd->nr_balance_failed = 0;
8217 
8218         if (likely(!active_balance)) {
8219                 /* We were unbalanced, so reset the balancing interval */
8220                 sd->balance_interval = sd->min_interval;
8221         } else {
8222                 /*
8223                  * If we've begun active balancing, start to back off. This
8224                  * case may not be covered by the all_pinned logic if there
8225                  * is only 1 task on the busy runqueue (because we don't call
8226                  * detach_tasks).
8227                  */
8228                 if (sd->balance_interval < sd->max_interval)
8229                         sd->balance_interval *= 2;
8230         }
8231 
8232         goto out;
8233 
8234 out_balanced:
8235         /*
8236          * We reach balance although we may have faced some affinity
8237          * constraints. Clear the imbalance flag if it was set.
8238          */
8239         if (sd_parent) {
8240                 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8241 
8242                 if (*group_imbalance)
8243                         *group_imbalance = 0;
8244         }
8245 
8246 out_all_pinned:
8247         /*
8248          * We reach balance because all tasks are pinned at this level so
8249          * we can't migrate them. Let the imbalance flag set so parent level
8250          * can try to migrate them.
8251          */
8252         schedstat_inc(sd->lb_balanced[idle]);
8253 
8254         sd->nr_balance_failed = 0;
8255 
8256 out_one_pinned:
8257         /* tune up the balancing interval */
8258         if (((env.flags & LBF_ALL_PINNED) &&
8259                         sd->balance_interval < MAX_PINNED_INTERVAL) ||
8260                         (sd->balance_interval < sd->max_interval))
8261                 sd->balance_interval *= 2;
8262 
8263         ld_moved = 0;
8264 out:
8265         return ld_moved;
8266 }
8267 
8268 static inline unsigned long
8269 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8270 {
8271         unsigned long interval = sd->balance_interval;
8272 
8273         if (cpu_busy)
8274                 interval *= sd->busy_factor;
8275 
8276         /* scale ms to jiffies */
8277         interval = msecs_to_jiffies(interval);
8278         interval = clamp(interval, 1UL, max_load_balance_interval);
8279 
8280         return interval;
8281 }
8282 
8283 static inline void
8284 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8285 {
8286         unsigned long interval, next;
8287 
8288         /* used by idle balance, so cpu_busy = 0 */
8289         interval = get_sd_balance_interval(sd, 0);
8290         next = sd->last_balance + interval;
8291 
8292         if (time_after(*next_balance, next))
8293                 *next_balance = next;
8294 }
8295 
8296 /*
8297  * idle_balance is called by schedule() if this_cpu is about to become
8298  * idle. Attempts to pull tasks from other CPUs.
8299  */
8300 static int idle_balance(struct rq *this_rq)
8301 {
8302         unsigned long next_balance = jiffies + HZ;
8303         int this_cpu = this_rq->cpu;
8304         struct sched_domain *sd;
8305         int pulled_task = 0;
8306         u64 curr_cost = 0;
8307 
8308         /*
8309          * We must set idle_stamp _before_ calling idle_balance(), such that we</