Version:  2.0.40 2.2.26 2.4.37 3.11 3.12 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

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  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
 41  *
 42  * NOTE: this latency value is not the same as the concept of
 43  * 'timeslice length' - timeslices in CFS are of variable length
 44  * and have no persistent notion like in traditional, time-slice
 45  * based scheduling concepts.
 46  *
 47  * (to see the precise effective timeslice length of your workload,
 48  *  run vmstat and monitor the context-switches (cs) field)
 49  */
 50 unsigned int sysctl_sched_latency = 6000000ULL;
 51 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
 52 
 53 /*
 54  * The initial- and re-scaling of tunables is configurable
 55  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
 56  *
 57  * Options are:
 58  * SCHED_TUNABLESCALING_NONE - unscaled, always *1
 59  * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
 60  * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
 61  */
 62 enum sched_tunable_scaling sysctl_sched_tunable_scaling
 63         = SCHED_TUNABLESCALING_LOG;
 64 
 65 /*
 66  * Minimal preemption granularity for CPU-bound tasks:
 67  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
 68  */
 69 unsigned int sysctl_sched_min_granularity = 750000ULL;
 70 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
 71 
 72 /*
 73  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
 74  */
 75 static unsigned int sched_nr_latency = 8;
 76 
 77 /*
 78  * After fork, child runs first. If set to 0 (default) then
 79  * parent will (try to) run first.
 80  */
 81 unsigned int sysctl_sched_child_runs_first __read_mostly;
 82 
 83 /*
 84  * SCHED_OTHER wake-up granularity.
 85  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
 86  *
 87  * This option delays the preemption effects of decoupled workloads
 88  * and reduces their over-scheduling. Synchronous workloads will still
 89  * have immediate wakeup/sleep latencies.
 90  */
 91 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
 92 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
 93 
 94 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
 95 
 96 /*
 97  * The exponential sliding  window over which load is averaged for shares
 98  * distribution.
 99  * (default: 10msec)
100  */
101 unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
102 
103 #ifdef CONFIG_CFS_BANDWIDTH
104 /*
105  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
106  * each time a cfs_rq requests quota.
107  *
108  * Note: in the case that the slice exceeds the runtime remaining (either due
109  * to consumption or the quota being specified to be smaller than the slice)
110  * we will always only issue the remaining available time.
111  *
112  * default: 5 msec, units: microseconds
113   */
114 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
115 #endif
116 
117 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
118 {
119         lw->weight += inc;
120         lw->inv_weight = 0;
121 }
122 
123 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
124 {
125         lw->weight -= dec;
126         lw->inv_weight = 0;
127 }
128 
129 static inline void update_load_set(struct load_weight *lw, unsigned long w)
130 {
131         lw->weight = w;
132         lw->inv_weight = 0;
133 }
134 
135 /*
136  * Increase the granularity value when there are more CPUs,
137  * because with more CPUs the 'effective latency' as visible
138  * to users decreases. But the relationship is not linear,
139  * so pick a second-best guess by going with the log2 of the
140  * number of CPUs.
141  *
142  * This idea comes from the SD scheduler of Con Kolivas:
143  */
144 static unsigned int get_update_sysctl_factor(void)
145 {
146         unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
147         unsigned int factor;
148 
149         switch (sysctl_sched_tunable_scaling) {
150         case SCHED_TUNABLESCALING_NONE:
151                 factor = 1;
152                 break;
153         case SCHED_TUNABLESCALING_LINEAR:
154                 factor = cpus;
155                 break;
156         case SCHED_TUNABLESCALING_LOG:
157         default:
158                 factor = 1 + ilog2(cpus);
159                 break;
160         }
161 
162         return factor;
163 }
164 
165 static void update_sysctl(void)
166 {
167         unsigned int factor = get_update_sysctl_factor();
168 
169 #define SET_SYSCTL(name) \
170         (sysctl_##name = (factor) * normalized_sysctl_##name)
171         SET_SYSCTL(sched_min_granularity);
172         SET_SYSCTL(sched_latency);
173         SET_SYSCTL(sched_wakeup_granularity);
174 #undef SET_SYSCTL
175 }
176 
177 void sched_init_granularity(void)
178 {
179         update_sysctl();
180 }
181 
182 #define WMULT_CONST     (~0U)
183 #define WMULT_SHIFT     32
184 
185 static void __update_inv_weight(struct load_weight *lw)
186 {
187         unsigned long w;
188 
189         if (likely(lw->inv_weight))
190                 return;
191 
192         w = scale_load_down(lw->weight);
193 
194         if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
195                 lw->inv_weight = 1;
196         else if (unlikely(!w))
197                 lw->inv_weight = WMULT_CONST;
198         else
199                 lw->inv_weight = WMULT_CONST / w;
200 }
201 
202 /*
203  * delta_exec * weight / lw.weight
204  *   OR
205  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
206  *
207  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
208  * we're guaranteed shift stays positive because inv_weight is guaranteed to
209  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
210  *
211  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
212  * weight/lw.weight <= 1, and therefore our shift will also be positive.
213  */
214 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
215 {
216         u64 fact = scale_load_down(weight);
217         int shift = WMULT_SHIFT;
218 
219         __update_inv_weight(lw);
220 
221         if (unlikely(fact >> 32)) {
222                 while (fact >> 32) {
223                         fact >>= 1;
224                         shift--;
225                 }
226         }
227 
228         /* hint to use a 32x32->64 mul */
229         fact = (u64)(u32)fact * lw->inv_weight;
230 
231         while (fact >> 32) {
232                 fact >>= 1;
233                 shift--;
234         }
235 
236         return mul_u64_u32_shr(delta_exec, fact, shift);
237 }
238 
239 
240 const struct sched_class fair_sched_class;
241 
242 /**************************************************************
243  * CFS operations on generic schedulable entities:
244  */
245 
246 #ifdef CONFIG_FAIR_GROUP_SCHED
247 
248 /* cpu runqueue to which this cfs_rq is attached */
249 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
250 {
251         return cfs_rq->rq;
252 }
253 
254 /* An entity is a task if it doesn't "own" a runqueue */
255 #define entity_is_task(se)      (!se->my_q)
256 
257 static inline struct task_struct *task_of(struct sched_entity *se)
258 {
259 #ifdef CONFIG_SCHED_DEBUG
260         WARN_ON_ONCE(!entity_is_task(se));
261 #endif
262         return container_of(se, struct task_struct, se);
263 }
264 
265 /* Walk up scheduling entities hierarchy */
266 #define for_each_sched_entity(se) \
267                 for (; se; se = se->parent)
268 
269 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
270 {
271         return p->se.cfs_rq;
272 }
273 
274 /* runqueue on which this entity is (to be) queued */
275 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
276 {
277         return se->cfs_rq;
278 }
279 
280 /* runqueue "owned" by this group */
281 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
282 {
283         return grp->my_q;
284 }
285 
286 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 {
288         if (!cfs_rq->on_list) {
289                 /*
290                  * Ensure we either appear before our parent (if already
291                  * enqueued) or force our parent to appear after us when it is
292                  * enqueued.  The fact that we always enqueue bottom-up
293                  * reduces this to two cases.
294                  */
295                 if (cfs_rq->tg->parent &&
296                     cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
297                         list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
298                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
299                 } else {
300                         list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
301                                 &rq_of(cfs_rq)->leaf_cfs_rq_list);
302                 }
303 
304                 cfs_rq->on_list = 1;
305         }
306 }
307 
308 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
309 {
310         if (cfs_rq->on_list) {
311                 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
312                 cfs_rq->on_list = 0;
313         }
314 }
315 
316 /* Iterate thr' all leaf cfs_rq's on a runqueue */
317 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
318         list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
319 
320 /* Do the two (enqueued) entities belong to the same group ? */
321 static inline struct cfs_rq *
322 is_same_group(struct sched_entity *se, struct sched_entity *pse)
323 {
324         if (se->cfs_rq == pse->cfs_rq)
325                 return se->cfs_rq;
326 
327         return NULL;
328 }
329 
330 static inline struct sched_entity *parent_entity(struct sched_entity *se)
331 {
332         return se->parent;
333 }
334 
335 static void
336 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
337 {
338         int se_depth, pse_depth;
339 
340         /*
341          * preemption test can be made between sibling entities who are in the
342          * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
343          * both tasks until we find their ancestors who are siblings of common
344          * parent.
345          */
346 
347         /* First walk up until both entities are at same depth */
348         se_depth = (*se)->depth;
349         pse_depth = (*pse)->depth;
350 
351         while (se_depth > pse_depth) {
352                 se_depth--;
353                 *se = parent_entity(*se);
354         }
355 
356         while (pse_depth > se_depth) {
357                 pse_depth--;
358                 *pse = parent_entity(*pse);
359         }
360 
361         while (!is_same_group(*se, *pse)) {
362                 *se = parent_entity(*se);
363                 *pse = parent_entity(*pse);
364         }
365 }
366 
367 #else   /* !CONFIG_FAIR_GROUP_SCHED */
368 
369 static inline struct task_struct *task_of(struct sched_entity *se)
370 {
371         return container_of(se, struct task_struct, se);
372 }
373 
374 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
375 {
376         return container_of(cfs_rq, struct rq, cfs);
377 }
378 
379 #define entity_is_task(se)      1
380 
381 #define for_each_sched_entity(se) \
382                 for (; se; se = NULL)
383 
384 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
385 {
386         return &task_rq(p)->cfs;
387 }
388 
389 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
390 {
391         struct task_struct *p = task_of(se);
392         struct rq *rq = task_rq(p);
393 
394         return &rq->cfs;
395 }
396 
397 /* runqueue "owned" by this group */
398 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
399 {
400         return NULL;
401 }
402 
403 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
404 {
405 }
406 
407 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
408 {
409 }
410 
411 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
412                 for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
413 
414 static inline struct sched_entity *parent_entity(struct sched_entity *se)
415 {
416         return NULL;
417 }
418 
419 static inline void
420 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
421 {
422 }
423 
424 #endif  /* CONFIG_FAIR_GROUP_SCHED */
425 
426 static __always_inline
427 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
428 
429 /**************************************************************
430  * Scheduling class tree data structure manipulation methods:
431  */
432 
433 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
434 {
435         s64 delta = (s64)(vruntime - max_vruntime);
436         if (delta > 0)
437                 max_vruntime = vruntime;
438 
439         return max_vruntime;
440 }
441 
442 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
443 {
444         s64 delta = (s64)(vruntime - min_vruntime);
445         if (delta < 0)
446                 min_vruntime = vruntime;
447 
448         return min_vruntime;
449 }
450 
451 static inline int entity_before(struct sched_entity *a,
452                                 struct sched_entity *b)
453 {
454         return (s64)(a->vruntime - b->vruntime) < 0;
455 }
456 
457 static void update_min_vruntime(struct cfs_rq *cfs_rq)
458 {
459         u64 vruntime = cfs_rq->min_vruntime;
460 
461         if (cfs_rq->curr)
462                 vruntime = cfs_rq->curr->vruntime;
463 
464         if (cfs_rq->rb_leftmost) {
465                 struct sched_entity *se = rb_entry(cfs_rq->rb_leftmost,
466                                                    struct sched_entity,
467                                                    run_node);
468 
469                 if (!cfs_rq->curr)
470                         vruntime = se->vruntime;
471                 else
472                         vruntime = min_vruntime(vruntime, se->vruntime);
473         }
474 
475         /* ensure we never gain time by being placed backwards. */
476         cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
477 #ifndef CONFIG_64BIT
478         smp_wmb();
479         cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
480 #endif
481 }
482 
483 /*
484  * Enqueue an entity into the rb-tree:
485  */
486 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
487 {
488         struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
489         struct rb_node *parent = NULL;
490         struct sched_entity *entry;
491         int leftmost = 1;
492 
493         /*
494          * Find the right place in the rbtree:
495          */
496         while (*link) {
497                 parent = *link;
498                 entry = rb_entry(parent, struct sched_entity, run_node);
499                 /*
500                  * We dont care about collisions. Nodes with
501                  * the same key stay together.
502                  */
503                 if (entity_before(se, entry)) {
504                         link = &parent->rb_left;
505                 } else {
506                         link = &parent->rb_right;
507                         leftmost = 0;
508                 }
509         }
510 
511         /*
512          * Maintain a cache of leftmost tree entries (it is frequently
513          * used):
514          */
515         if (leftmost)
516                 cfs_rq->rb_leftmost = &se->run_node;
517 
518         rb_link_node(&se->run_node, parent, link);
519         rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
520 }
521 
522 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
523 {
524         if (cfs_rq->rb_leftmost == &se->run_node) {
525                 struct rb_node *next_node;
526 
527                 next_node = rb_next(&se->run_node);
528                 cfs_rq->rb_leftmost = next_node;
529         }
530 
531         rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
532 }
533 
534 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
535 {
536         struct rb_node *left = cfs_rq->rb_leftmost;
537 
538         if (!left)
539                 return NULL;
540 
541         return rb_entry(left, struct sched_entity, run_node);
542 }
543 
544 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
545 {
546         struct rb_node *next = rb_next(&se->run_node);
547 
548         if (!next)
549                 return NULL;
550 
551         return rb_entry(next, struct sched_entity, run_node);
552 }
553 
554 #ifdef CONFIG_SCHED_DEBUG
555 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
556 {
557         struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
558 
559         if (!last)
560                 return NULL;
561 
562         return rb_entry(last, struct sched_entity, run_node);
563 }
564 
565 /**************************************************************
566  * Scheduling class statistics methods:
567  */
568 
569 int sched_proc_update_handler(struct ctl_table *table, int write,
570                 void __user *buffer, size_t *lenp,
571                 loff_t *ppos)
572 {
573         int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
574         unsigned int factor = get_update_sysctl_factor();
575 
576         if (ret || !write)
577                 return ret;
578 
579         sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
580                                         sysctl_sched_min_granularity);
581 
582 #define WRT_SYSCTL(name) \
583         (normalized_sysctl_##name = sysctl_##name / (factor))
584         WRT_SYSCTL(sched_min_granularity);
585         WRT_SYSCTL(sched_latency);
586         WRT_SYSCTL(sched_wakeup_granularity);
587 #undef WRT_SYSCTL
588 
589         return 0;
590 }
591 #endif
592 
593 /*
594  * delta /= w
595  */
596 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
597 {
598         if (unlikely(se->load.weight != NICE_0_LOAD))
599                 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
600 
601         return delta;
602 }
603 
604 /*
605  * The idea is to set a period in which each task runs once.
606  *
607  * When there are too many tasks (sched_nr_latency) we have to stretch
608  * this period because otherwise the slices get too small.
609  *
610  * p = (nr <= nl) ? l : l*nr/nl
611  */
612 static u64 __sched_period(unsigned long nr_running)
613 {
614         if (unlikely(nr_running > sched_nr_latency))
615                 return nr_running * sysctl_sched_min_granularity;
616         else
617                 return sysctl_sched_latency;
618 }
619 
620 /*
621  * We calculate the wall-time slice from the period by taking a part
622  * proportional to the weight.
623  *
624  * s = p*P[w/rw]
625  */
626 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
627 {
628         u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
629 
630         for_each_sched_entity(se) {
631                 struct load_weight *load;
632                 struct load_weight lw;
633 
634                 cfs_rq = cfs_rq_of(se);
635                 load = &cfs_rq->load;
636 
637                 if (unlikely(!se->on_rq)) {
638                         lw = cfs_rq->load;
639 
640                         update_load_add(&lw, se->load.weight);
641                         load = &lw;
642                 }
643                 slice = __calc_delta(slice, se->load.weight, load);
644         }
645         return slice;
646 }
647 
648 /*
649  * We calculate the vruntime slice of a to-be-inserted task.
650  *
651  * vs = s/w
652  */
653 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
654 {
655         return calc_delta_fair(sched_slice(cfs_rq, se), se);
656 }
657 
658 #ifdef CONFIG_SMP
659 static int select_idle_sibling(struct task_struct *p, int cpu);
660 static unsigned long task_h_load(struct task_struct *p);
661 
662 /*
663  * We choose a half-life close to 1 scheduling period.
664  * Note: The tables runnable_avg_yN_inv and runnable_avg_yN_sum are
665  * dependent on this value.
666  */
667 #define LOAD_AVG_PERIOD 32
668 #define LOAD_AVG_MAX 47742 /* maximum possible load avg */
669 #define LOAD_AVG_MAX_N 345 /* number of full periods to produce LOAD_AVG_MAX */
670 
671 /* Give new sched_entity start runnable values to heavy its load in infant time */
672 void init_entity_runnable_average(struct sched_entity *se)
673 {
674         struct sched_avg *sa = &se->avg;
675 
676         sa->last_update_time = 0;
677         /*
678          * sched_avg's period_contrib should be strictly less then 1024, so
679          * we give it 1023 to make sure it is almost a period (1024us), and
680          * will definitely be update (after enqueue).
681          */
682         sa->period_contrib = 1023;
683         sa->load_avg = scale_load_down(se->load.weight);
684         sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
685         /*
686          * At this point, util_avg won't be used in select_task_rq_fair anyway
687          */
688         sa->util_avg = 0;
689         sa->util_sum = 0;
690         /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
691 }
692 
693 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
694 static int update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq);
695 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force);
696 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se);
697 
698 /*
699  * With new tasks being created, their initial util_avgs are extrapolated
700  * based on the cfs_rq's current util_avg:
701  *
702  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
703  *
704  * However, in many cases, the above util_avg does not give a desired
705  * value. Moreover, the sum of the util_avgs may be divergent, such
706  * as when the series is a harmonic series.
707  *
708  * To solve this problem, we also cap the util_avg of successive tasks to
709  * only 1/2 of the left utilization budget:
710  *
711  *   util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
712  *
713  * where n denotes the nth task.
714  *
715  * For example, a simplest series from the beginning would be like:
716  *
717  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
718  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
719  *
720  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
721  * if util_avg > util_avg_cap.
722  */
723 void post_init_entity_util_avg(struct sched_entity *se)
724 {
725         struct cfs_rq *cfs_rq = cfs_rq_of(se);
726         struct sched_avg *sa = &se->avg;
727         long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
728         u64 now = cfs_rq_clock_task(cfs_rq);
729         int tg_update;
730 
731         if (cap > 0) {
732                 if (cfs_rq->avg.util_avg != 0) {
733                         sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
734                         sa->util_avg /= (cfs_rq->avg.load_avg + 1);
735 
736                         if (sa->util_avg > cap)
737                                 sa->util_avg = cap;
738                 } else {
739                         sa->util_avg = cap;
740                 }
741                 sa->util_sum = sa->util_avg * LOAD_AVG_MAX;
742         }
743 
744         if (entity_is_task(se)) {
745                 struct task_struct *p = task_of(se);
746                 if (p->sched_class != &fair_sched_class) {
747                         /*
748                          * For !fair tasks do:
749                          *
750                         update_cfs_rq_load_avg(now, cfs_rq, false);
751                         attach_entity_load_avg(cfs_rq, se);
752                         switched_from_fair(rq, p);
753                          *
754                          * such that the next switched_to_fair() has the
755                          * expected state.
756                          */
757                         se->avg.last_update_time = now;
758                         return;
759                 }
760         }
761 
762         tg_update = update_cfs_rq_load_avg(now, cfs_rq, false);
763         attach_entity_load_avg(cfs_rq, se);
764         if (tg_update)
765                 update_tg_load_avg(cfs_rq, false);
766 }
767 
768 #else /* !CONFIG_SMP */
769 void init_entity_runnable_average(struct sched_entity *se)
770 {
771 }
772 void post_init_entity_util_avg(struct sched_entity *se)
773 {
774 }
775 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
776 {
777 }
778 #endif /* CONFIG_SMP */
779 
780 /*
781  * Update the current task's runtime statistics.
782  */
783 static void update_curr(struct cfs_rq *cfs_rq)
784 {
785         struct sched_entity *curr = cfs_rq->curr;
786         u64 now = rq_clock_task(rq_of(cfs_rq));
787         u64 delta_exec;
788 
789         if (unlikely(!curr))
790                 return;
791 
792         delta_exec = now - curr->exec_start;
793         if (unlikely((s64)delta_exec <= 0))
794                 return;
795 
796         curr->exec_start = now;
797 
798         schedstat_set(curr->statistics.exec_max,
799                       max(delta_exec, curr->statistics.exec_max));
800 
801         curr->sum_exec_runtime += delta_exec;
802         schedstat_add(cfs_rq, exec_clock, delta_exec);
803 
804         curr->vruntime += calc_delta_fair(delta_exec, curr);
805         update_min_vruntime(cfs_rq);
806 
807         if (entity_is_task(curr)) {
808                 struct task_struct *curtask = task_of(curr);
809 
810                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
811                 cpuacct_charge(curtask, delta_exec);
812                 account_group_exec_runtime(curtask, delta_exec);
813         }
814 
815         account_cfs_rq_runtime(cfs_rq, delta_exec);
816 }
817 
818 static void update_curr_fair(struct rq *rq)
819 {
820         update_curr(cfs_rq_of(&rq->curr->se));
821 }
822 
823 #ifdef CONFIG_SCHEDSTATS
824 static inline void
825 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
826 {
827         u64 wait_start = rq_clock(rq_of(cfs_rq));
828 
829         if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
830             likely(wait_start > se->statistics.wait_start))
831                 wait_start -= se->statistics.wait_start;
832 
833         se->statistics.wait_start = wait_start;
834 }
835 
836 static void
837 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
838 {
839         struct task_struct *p;
840         u64 delta;
841 
842         delta = rq_clock(rq_of(cfs_rq)) - se->statistics.wait_start;
843 
844         if (entity_is_task(se)) {
845                 p = task_of(se);
846                 if (task_on_rq_migrating(p)) {
847                         /*
848                          * Preserve migrating task's wait time so wait_start
849                          * time stamp can be adjusted to accumulate wait time
850                          * prior to migration.
851                          */
852                         se->statistics.wait_start = delta;
853                         return;
854                 }
855                 trace_sched_stat_wait(p, delta);
856         }
857 
858         se->statistics.wait_max = max(se->statistics.wait_max, delta);
859         se->statistics.wait_count++;
860         se->statistics.wait_sum += delta;
861         se->statistics.wait_start = 0;
862 }
863 
864 /*
865  * Task is being enqueued - update stats:
866  */
867 static inline void
868 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
869 {
870         /*
871          * Are we enqueueing a waiting task? (for current tasks
872          * a dequeue/enqueue event is a NOP)
873          */
874         if (se != cfs_rq->curr)
875                 update_stats_wait_start(cfs_rq, se);
876 }
877 
878 static inline void
879 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
880 {
881         /*
882          * Mark the end of the wait period if dequeueing a
883          * waiting task:
884          */
885         if (se != cfs_rq->curr)
886                 update_stats_wait_end(cfs_rq, se);
887 
888         if (flags & DEQUEUE_SLEEP) {
889                 if (entity_is_task(se)) {
890                         struct task_struct *tsk = task_of(se);
891 
892                         if (tsk->state & TASK_INTERRUPTIBLE)
893                                 se->statistics.sleep_start = rq_clock(rq_of(cfs_rq));
894                         if (tsk->state & TASK_UNINTERRUPTIBLE)
895                                 se->statistics.block_start = rq_clock(rq_of(cfs_rq));
896                 }
897         }
898 
899 }
900 #else
901 static inline void
902 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
903 {
904 }
905 
906 static inline void
907 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
908 {
909 }
910 
911 static inline void
912 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
913 {
914 }
915 
916 static inline void
917 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
918 {
919 }
920 #endif
921 
922 /*
923  * We are picking a new current task - update its stats:
924  */
925 static inline void
926 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
927 {
928         /*
929          * We are starting a new run period:
930          */
931         se->exec_start = rq_clock_task(rq_of(cfs_rq));
932 }
933 
934 /**************************************************
935  * Scheduling class queueing methods:
936  */
937 
938 #ifdef CONFIG_NUMA_BALANCING
939 /*
940  * Approximate time to scan a full NUMA task in ms. The task scan period is
941  * calculated based on the tasks virtual memory size and
942  * numa_balancing_scan_size.
943  */
944 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
945 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
946 
947 /* Portion of address space to scan in MB */
948 unsigned int sysctl_numa_balancing_scan_size = 256;
949 
950 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
951 unsigned int sysctl_numa_balancing_scan_delay = 1000;
952 
953 static unsigned int task_nr_scan_windows(struct task_struct *p)
954 {
955         unsigned long rss = 0;
956         unsigned long nr_scan_pages;
957 
958         /*
959          * Calculations based on RSS as non-present and empty pages are skipped
960          * by the PTE scanner and NUMA hinting faults should be trapped based
961          * on resident pages
962          */
963         nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
964         rss = get_mm_rss(p->mm);
965         if (!rss)
966                 rss = nr_scan_pages;
967 
968         rss = round_up(rss, nr_scan_pages);
969         return rss / nr_scan_pages;
970 }
971 
972 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
973 #define MAX_SCAN_WINDOW 2560
974 
975 static unsigned int task_scan_min(struct task_struct *p)
976 {
977         unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
978         unsigned int scan, floor;
979         unsigned int windows = 1;
980 
981         if (scan_size < MAX_SCAN_WINDOW)
982                 windows = MAX_SCAN_WINDOW / scan_size;
983         floor = 1000 / windows;
984 
985         scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
986         return max_t(unsigned int, floor, scan);
987 }
988 
989 static unsigned int task_scan_max(struct task_struct *p)
990 {
991         unsigned int smin = task_scan_min(p);
992         unsigned int smax;
993 
994         /* Watch for min being lower than max due to floor calculations */
995         smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
996         return max(smin, smax);
997 }
998 
999 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1000 {
1001         rq->nr_numa_running += (p->numa_preferred_nid != -1);
1002         rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1003 }
1004 
1005 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1006 {
1007         rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1008         rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1009 }
1010 
1011 struct numa_group {
1012         atomic_t refcount;
1013 
1014         spinlock_t lock; /* nr_tasks, tasks */
1015         int nr_tasks;
1016         pid_t gid;
1017         int active_nodes;
1018 
1019         struct rcu_head rcu;
1020         unsigned long total_faults;
1021         unsigned long max_faults_cpu;
1022         /*
1023          * Faults_cpu is used to decide whether memory should move
1024          * towards the CPU. As a consequence, these stats are weighted
1025          * more by CPU use than by memory faults.
1026          */
1027         unsigned long *faults_cpu;
1028         unsigned long faults[0];
1029 };
1030 
1031 /* Shared or private faults. */
1032 #define NR_NUMA_HINT_FAULT_TYPES 2
1033 
1034 /* Memory and CPU locality */
1035 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1036 
1037 /* Averaged statistics, and temporary buffers. */
1038 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1039 
1040 pid_t task_numa_group_id(struct task_struct *p)
1041 {
1042         return p->numa_group ? p->numa_group->gid : 0;
1043 }
1044 
1045 /*
1046  * The averaged statistics, shared & private, memory & cpu,
1047  * occupy the first half of the array. The second half of the
1048  * array is for current counters, which are averaged into the
1049  * first set by task_numa_placement.
1050  */
1051 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1052 {
1053         return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1054 }
1055 
1056 static inline unsigned long task_faults(struct task_struct *p, int nid)
1057 {
1058         if (!p->numa_faults)
1059                 return 0;
1060 
1061         return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1062                 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1063 }
1064 
1065 static inline unsigned long group_faults(struct task_struct *p, int nid)
1066 {
1067         if (!p->numa_group)
1068                 return 0;
1069 
1070         return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1071                 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1072 }
1073 
1074 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1075 {
1076         return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1077                 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1078 }
1079 
1080 /*
1081  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1082  * considered part of a numa group's pseudo-interleaving set. Migrations
1083  * between these nodes are slowed down, to allow things to settle down.
1084  */
1085 #define ACTIVE_NODE_FRACTION 3
1086 
1087 static bool numa_is_active_node(int nid, struct numa_group *ng)
1088 {
1089         return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1090 }
1091 
1092 /* Handle placement on systems where not all nodes are directly connected. */
1093 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1094                                         int maxdist, bool task)
1095 {
1096         unsigned long score = 0;
1097         int node;
1098 
1099         /*
1100          * All nodes are directly connected, and the same distance
1101          * from each other. No need for fancy placement algorithms.
1102          */
1103         if (sched_numa_topology_type == NUMA_DIRECT)
1104                 return 0;
1105 
1106         /*
1107          * This code is called for each node, introducing N^2 complexity,
1108          * which should be ok given the number of nodes rarely exceeds 8.
1109          */
1110         for_each_online_node(node) {
1111                 unsigned long faults;
1112                 int dist = node_distance(nid, node);
1113 
1114                 /*
1115                  * The furthest away nodes in the system are not interesting
1116                  * for placement; nid was already counted.
1117                  */
1118                 if (dist == sched_max_numa_distance || node == nid)
1119                         continue;
1120 
1121                 /*
1122                  * On systems with a backplane NUMA topology, compare groups
1123                  * of nodes, and move tasks towards the group with the most
1124                  * memory accesses. When comparing two nodes at distance
1125                  * "hoplimit", only nodes closer by than "hoplimit" are part
1126                  * of each group. Skip other nodes.
1127                  */
1128                 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1129                                         dist > maxdist)
1130                         continue;
1131 
1132                 /* Add up the faults from nearby nodes. */
1133                 if (task)
1134                         faults = task_faults(p, node);
1135                 else
1136                         faults = group_faults(p, node);
1137 
1138                 /*
1139                  * On systems with a glueless mesh NUMA topology, there are
1140                  * no fixed "groups of nodes". Instead, nodes that are not
1141                  * directly connected bounce traffic through intermediate
1142                  * nodes; a numa_group can occupy any set of nodes.
1143                  * The further away a node is, the less the faults count.
1144                  * This seems to result in good task placement.
1145                  */
1146                 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1147                         faults *= (sched_max_numa_distance - dist);
1148                         faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1149                 }
1150 
1151                 score += faults;
1152         }
1153 
1154         return score;
1155 }
1156 
1157 /*
1158  * These return the fraction of accesses done by a particular task, or
1159  * task group, on a particular numa node.  The group weight is given a
1160  * larger multiplier, in order to group tasks together that are almost
1161  * evenly spread out between numa nodes.
1162  */
1163 static inline unsigned long task_weight(struct task_struct *p, int nid,
1164                                         int dist)
1165 {
1166         unsigned long faults, total_faults;
1167 
1168         if (!p->numa_faults)
1169                 return 0;
1170 
1171         total_faults = p->total_numa_faults;
1172 
1173         if (!total_faults)
1174                 return 0;
1175 
1176         faults = task_faults(p, nid);
1177         faults += score_nearby_nodes(p, nid, dist, true);
1178 
1179         return 1000 * faults / total_faults;
1180 }
1181 
1182 static inline unsigned long group_weight(struct task_struct *p, int nid,
1183                                          int dist)
1184 {
1185         unsigned long faults, total_faults;
1186 
1187         if (!p->numa_group)
1188                 return 0;
1189 
1190         total_faults = p->numa_group->total_faults;
1191 
1192         if (!total_faults)
1193                 return 0;
1194 
1195         faults = group_faults(p, nid);
1196         faults += score_nearby_nodes(p, nid, dist, false);
1197 
1198         return 1000 * faults / total_faults;
1199 }
1200 
1201 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1202                                 int src_nid, int dst_cpu)
1203 {
1204         struct numa_group *ng = p->numa_group;
1205         int dst_nid = cpu_to_node(dst_cpu);
1206         int last_cpupid, this_cpupid;
1207 
1208         this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1209 
1210         /*
1211          * Multi-stage node selection is used in conjunction with a periodic
1212          * migration fault to build a temporal task<->page relation. By using
1213          * a two-stage filter we remove short/unlikely relations.
1214          *
1215          * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1216          * a task's usage of a particular page (n_p) per total usage of this
1217          * page (n_t) (in a given time-span) to a probability.
1218          *
1219          * Our periodic faults will sample this probability and getting the
1220          * same result twice in a row, given these samples are fully
1221          * independent, is then given by P(n)^2, provided our sample period
1222          * is sufficiently short compared to the usage pattern.
1223          *
1224          * This quadric squishes small probabilities, making it less likely we
1225          * act on an unlikely task<->page relation.
1226          */
1227         last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1228         if (!cpupid_pid_unset(last_cpupid) &&
1229                                 cpupid_to_nid(last_cpupid) != dst_nid)
1230                 return false;
1231 
1232         /* Always allow migrate on private faults */
1233         if (cpupid_match_pid(p, last_cpupid))
1234                 return true;
1235 
1236         /* A shared fault, but p->numa_group has not been set up yet. */
1237         if (!ng)
1238                 return true;
1239 
1240         /*
1241          * Destination node is much more heavily used than the source
1242          * node? Allow migration.
1243          */
1244         if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1245                                         ACTIVE_NODE_FRACTION)
1246                 return true;
1247 
1248         /*
1249          * Distribute memory according to CPU & memory use on each node,
1250          * with 3/4 hysteresis to avoid unnecessary memory migrations:
1251          *
1252          * faults_cpu(dst)   3   faults_cpu(src)
1253          * --------------- * - > ---------------
1254          * faults_mem(dst)   4   faults_mem(src)
1255          */
1256         return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1257                group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1258 }
1259 
1260 static unsigned long weighted_cpuload(const int cpu);
1261 static unsigned long source_load(int cpu, int type);
1262 static unsigned long target_load(int cpu, int type);
1263 static unsigned long capacity_of(int cpu);
1264 static long effective_load(struct task_group *tg, int cpu, long wl, long wg);
1265 
1266 /* Cached statistics for all CPUs within a node */
1267 struct numa_stats {
1268         unsigned long nr_running;
1269         unsigned long load;
1270 
1271         /* Total compute capacity of CPUs on a node */
1272         unsigned long compute_capacity;
1273 
1274         /* Approximate capacity in terms of runnable tasks on a node */
1275         unsigned long task_capacity;
1276         int has_free_capacity;
1277 };
1278 
1279 /*
1280  * XXX borrowed from update_sg_lb_stats
1281  */
1282 static void update_numa_stats(struct numa_stats *ns, int nid)
1283 {
1284         int smt, cpu, cpus = 0;
1285         unsigned long capacity;
1286 
1287         memset(ns, 0, sizeof(*ns));
1288         for_each_cpu(cpu, cpumask_of_node(nid)) {
1289                 struct rq *rq = cpu_rq(cpu);
1290 
1291                 ns->nr_running += rq->nr_running;
1292                 ns->load += weighted_cpuload(cpu);
1293                 ns->compute_capacity += capacity_of(cpu);
1294 
1295                 cpus++;
1296         }
1297 
1298         /*
1299          * If we raced with hotplug and there are no CPUs left in our mask
1300          * the @ns structure is NULL'ed and task_numa_compare() will
1301          * not find this node attractive.
1302          *
1303          * We'll either bail at !has_free_capacity, or we'll detect a huge
1304          * imbalance and bail there.
1305          */
1306         if (!cpus)
1307                 return;
1308 
1309         /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1310         smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1311         capacity = cpus / smt; /* cores */
1312 
1313         ns->task_capacity = min_t(unsigned, capacity,
1314                 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1315         ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1316 }
1317 
1318 struct task_numa_env {
1319         struct task_struct *p;
1320 
1321         int src_cpu, src_nid;
1322         int dst_cpu, dst_nid;
1323 
1324         struct numa_stats src_stats, dst_stats;
1325 
1326         int imbalance_pct;
1327         int dist;
1328 
1329         struct task_struct *best_task;
1330         long best_imp;
1331         int best_cpu;
1332 };
1333 
1334 static void task_numa_assign(struct task_numa_env *env,
1335                              struct task_struct *p, long imp)
1336 {
1337         if (env->best_task)
1338                 put_task_struct(env->best_task);
1339         if (p)
1340                 get_task_struct(p);
1341 
1342         env->best_task = p;
1343         env->best_imp = imp;
1344         env->best_cpu = env->dst_cpu;
1345 }
1346 
1347 static bool load_too_imbalanced(long src_load, long dst_load,
1348                                 struct task_numa_env *env)
1349 {
1350         long imb, old_imb;
1351         long orig_src_load, orig_dst_load;
1352         long src_capacity, dst_capacity;
1353 
1354         /*
1355          * The load is corrected for the CPU capacity available on each node.
1356          *
1357          * src_load        dst_load
1358          * ------------ vs ---------
1359          * src_capacity    dst_capacity
1360          */
1361         src_capacity = env->src_stats.compute_capacity;
1362         dst_capacity = env->dst_stats.compute_capacity;
1363 
1364         /* We care about the slope of the imbalance, not the direction. */
1365         if (dst_load < src_load)
1366                 swap(dst_load, src_load);
1367 
1368         /* Is the difference below the threshold? */
1369         imb = dst_load * src_capacity * 100 -
1370               src_load * dst_capacity * env->imbalance_pct;
1371         if (imb <= 0)
1372                 return false;
1373 
1374         /*
1375          * The imbalance is above the allowed threshold.
1376          * Compare it with the old imbalance.
1377          */
1378         orig_src_load = env->src_stats.load;
1379         orig_dst_load = env->dst_stats.load;
1380 
1381         if (orig_dst_load < orig_src_load)
1382                 swap(orig_dst_load, orig_src_load);
1383 
1384         old_imb = orig_dst_load * src_capacity * 100 -
1385                   orig_src_load * dst_capacity * env->imbalance_pct;
1386 
1387         /* Would this change make things worse? */
1388         return (imb > old_imb);
1389 }
1390 
1391 /*
1392  * This checks if the overall compute and NUMA accesses of the system would
1393  * be improved if the source tasks was migrated to the target dst_cpu taking
1394  * into account that it might be best if task running on the dst_cpu should
1395  * be exchanged with the source task
1396  */
1397 static void task_numa_compare(struct task_numa_env *env,
1398                               long taskimp, long groupimp)
1399 {
1400         struct rq *src_rq = cpu_rq(env->src_cpu);
1401         struct rq *dst_rq = cpu_rq(env->dst_cpu);
1402         struct task_struct *cur;
1403         long src_load, dst_load;
1404         long load;
1405         long imp = env->p->numa_group ? groupimp : taskimp;
1406         long moveimp = imp;
1407         int dist = env->dist;
1408 
1409         rcu_read_lock();
1410         cur = task_rcu_dereference(&dst_rq->curr);
1411         if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1412                 cur = NULL;
1413 
1414         /*
1415          * Because we have preemption enabled we can get migrated around and
1416          * end try selecting ourselves (current == env->p) as a swap candidate.
1417          */
1418         if (cur == env->p)
1419                 goto unlock;
1420 
1421         /*
1422          * "imp" is the fault differential for the source task between the
1423          * source and destination node. Calculate the total differential for
1424          * the source task and potential destination task. The more negative
1425          * the value is, the more rmeote accesses that would be expected to
1426          * be incurred if the tasks were swapped.
1427          */
1428         if (cur) {
1429                 /* Skip this swap candidate if cannot move to the source cpu */
1430                 if (!cpumask_test_cpu(env->src_cpu, tsk_cpus_allowed(cur)))
1431                         goto unlock;
1432 
1433                 /*
1434                  * If dst and source tasks are in the same NUMA group, or not
1435                  * in any group then look only at task weights.
1436                  */
1437                 if (cur->numa_group == env->p->numa_group) {
1438                         imp = taskimp + task_weight(cur, env->src_nid, dist) -
1439                               task_weight(cur, env->dst_nid, dist);
1440                         /*
1441                          * Add some hysteresis to prevent swapping the
1442                          * tasks within a group over tiny differences.
1443                          */
1444                         if (cur->numa_group)
1445                                 imp -= imp/16;
1446                 } else {
1447                         /*
1448                          * Compare the group weights. If a task is all by
1449                          * itself (not part of a group), use the task weight
1450                          * instead.
1451                          */
1452                         if (cur->numa_group)
1453                                 imp += group_weight(cur, env->src_nid, dist) -
1454                                        group_weight(cur, env->dst_nid, dist);
1455                         else
1456                                 imp += task_weight(cur, env->src_nid, dist) -
1457                                        task_weight(cur, env->dst_nid, dist);
1458                 }
1459         }
1460 
1461         if (imp <= env->best_imp && moveimp <= env->best_imp)
1462                 goto unlock;
1463 
1464         if (!cur) {
1465                 /* Is there capacity at our destination? */
1466                 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1467                     !env->dst_stats.has_free_capacity)
1468                         goto unlock;
1469 
1470                 goto balance;
1471         }
1472 
1473         /* Balance doesn't matter much if we're running a task per cpu */
1474         if (imp > env->best_imp && src_rq->nr_running == 1 &&
1475                         dst_rq->nr_running == 1)
1476                 goto assign;
1477 
1478         /*
1479          * In the overloaded case, try and keep the load balanced.
1480          */
1481 balance:
1482         load = task_h_load(env->p);
1483         dst_load = env->dst_stats.load + load;
1484         src_load = env->src_stats.load - load;
1485 
1486         if (moveimp > imp && moveimp > env->best_imp) {
1487                 /*
1488                  * If the improvement from just moving env->p direction is
1489                  * better than swapping tasks around, check if a move is
1490                  * possible. Store a slightly smaller score than moveimp,
1491                  * so an actually idle CPU will win.
1492                  */
1493                 if (!load_too_imbalanced(src_load, dst_load, env)) {
1494                         imp = moveimp - 1;
1495                         cur = NULL;
1496                         goto assign;
1497                 }
1498         }
1499 
1500         if (imp <= env->best_imp)
1501                 goto unlock;
1502 
1503         if (cur) {
1504                 load = task_h_load(cur);
1505                 dst_load -= load;
1506                 src_load += load;
1507         }
1508 
1509         if (load_too_imbalanced(src_load, dst_load, env))
1510                 goto unlock;
1511 
1512         /*
1513          * One idle CPU per node is evaluated for a task numa move.
1514          * Call select_idle_sibling to maybe find a better one.
1515          */
1516         if (!cur)
1517                 env->dst_cpu = select_idle_sibling(env->p, env->dst_cpu);
1518 
1519 assign:
1520         task_numa_assign(env, cur, imp);
1521 unlock:
1522         rcu_read_unlock();
1523 }
1524 
1525 static void task_numa_find_cpu(struct task_numa_env *env,
1526                                 long taskimp, long groupimp)
1527 {
1528         int cpu;
1529 
1530         for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1531                 /* Skip this CPU if the source task cannot migrate */
1532                 if (!cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)))
1533                         continue;
1534 
1535                 env->dst_cpu = cpu;
1536                 task_numa_compare(env, taskimp, groupimp);
1537         }
1538 }
1539 
1540 /* Only move tasks to a NUMA node less busy than the current node. */
1541 static bool numa_has_capacity(struct task_numa_env *env)
1542 {
1543         struct numa_stats *src = &env->src_stats;
1544         struct numa_stats *dst = &env->dst_stats;
1545 
1546         if (src->has_free_capacity && !dst->has_free_capacity)
1547                 return false;
1548 
1549         /*
1550          * Only consider a task move if the source has a higher load
1551          * than the destination, corrected for CPU capacity on each node.
1552          *
1553          *      src->load                dst->load
1554          * --------------------- vs ---------------------
1555          * src->compute_capacity    dst->compute_capacity
1556          */
1557         if (src->load * dst->compute_capacity * env->imbalance_pct >
1558 
1559             dst->load * src->compute_capacity * 100)
1560                 return true;
1561 
1562         return false;
1563 }
1564 
1565 static int task_numa_migrate(struct task_struct *p)
1566 {
1567         struct task_numa_env env = {
1568                 .p = p,
1569 
1570                 .src_cpu = task_cpu(p),
1571                 .src_nid = task_node(p),
1572 
1573                 .imbalance_pct = 112,
1574 
1575                 .best_task = NULL,
1576                 .best_imp = 0,
1577                 .best_cpu = -1,
1578         };
1579         struct sched_domain *sd;
1580         unsigned long taskweight, groupweight;
1581         int nid, ret, dist;
1582         long taskimp, groupimp;
1583 
1584         /*
1585          * Pick the lowest SD_NUMA domain, as that would have the smallest
1586          * imbalance and would be the first to start moving tasks about.
1587          *
1588          * And we want to avoid any moving of tasks about, as that would create
1589          * random movement of tasks -- counter the numa conditions we're trying
1590          * to satisfy here.
1591          */
1592         rcu_read_lock();
1593         sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1594         if (sd)
1595                 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1596         rcu_read_unlock();
1597 
1598         /*
1599          * Cpusets can break the scheduler domain tree into smaller
1600          * balance domains, some of which do not cross NUMA boundaries.
1601          * Tasks that are "trapped" in such domains cannot be migrated
1602          * elsewhere, so there is no point in (re)trying.
1603          */
1604         if (unlikely(!sd)) {
1605                 p->numa_preferred_nid = task_node(p);
1606                 return -EINVAL;
1607         }
1608 
1609         env.dst_nid = p->numa_preferred_nid;
1610         dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1611         taskweight = task_weight(p, env.src_nid, dist);
1612         groupweight = group_weight(p, env.src_nid, dist);
1613         update_numa_stats(&env.src_stats, env.src_nid);
1614         taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1615         groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1616         update_numa_stats(&env.dst_stats, env.dst_nid);
1617 
1618         /* Try to find a spot on the preferred nid. */
1619         if (numa_has_capacity(&env))
1620                 task_numa_find_cpu(&env, taskimp, groupimp);
1621 
1622         /*
1623          * Look at other nodes in these cases:
1624          * - there is no space available on the preferred_nid
1625          * - the task is part of a numa_group that is interleaved across
1626          *   multiple NUMA nodes; in order to better consolidate the group,
1627          *   we need to check other locations.
1628          */
1629         if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1630                 for_each_online_node(nid) {
1631                         if (nid == env.src_nid || nid == p->numa_preferred_nid)
1632                                 continue;
1633 
1634                         dist = node_distance(env.src_nid, env.dst_nid);
1635                         if (sched_numa_topology_type == NUMA_BACKPLANE &&
1636                                                 dist != env.dist) {
1637                                 taskweight = task_weight(p, env.src_nid, dist);
1638                                 groupweight = group_weight(p, env.src_nid, dist);
1639                         }
1640 
1641                         /* Only consider nodes where both task and groups benefit */
1642                         taskimp = task_weight(p, nid, dist) - taskweight;
1643                         groupimp = group_weight(p, nid, dist) - groupweight;
1644                         if (taskimp < 0 && groupimp < 0)
1645                                 continue;
1646 
1647                         env.dist = dist;
1648                         env.dst_nid = nid;
1649                         update_numa_stats(&env.dst_stats, env.dst_nid);
1650                         if (numa_has_capacity(&env))
1651                                 task_numa_find_cpu(&env, taskimp, groupimp);
1652                 }
1653         }
1654 
1655         /*
1656          * If the task is part of a workload that spans multiple NUMA nodes,
1657          * and is migrating into one of the workload's active nodes, remember
1658          * this node as the task's preferred numa node, so the workload can
1659          * settle down.
1660          * A task that migrated to a second choice node will be better off
1661          * trying for a better one later. Do not set the preferred node here.
1662          */
1663         if (p->numa_group) {
1664                 struct numa_group *ng = p->numa_group;
1665 
1666                 if (env.best_cpu == -1)
1667                         nid = env.src_nid;
1668                 else
1669                         nid = env.dst_nid;
1670 
1671                 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1672                         sched_setnuma(p, env.dst_nid);
1673         }
1674 
1675         /* No better CPU than the current one was found. */
1676         if (env.best_cpu == -1)
1677                 return -EAGAIN;
1678 
1679         /*
1680          * Reset the scan period if the task is being rescheduled on an
1681          * alternative node to recheck if the tasks is now properly placed.
1682          */
1683         p->numa_scan_period = task_scan_min(p);
1684 
1685         if (env.best_task == NULL) {
1686                 ret = migrate_task_to(p, env.best_cpu);
1687                 if (ret != 0)
1688                         trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1689                 return ret;
1690         }
1691 
1692         ret = migrate_swap(p, env.best_task);
1693         if (ret != 0)
1694                 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1695         put_task_struct(env.best_task);
1696         return ret;
1697 }
1698 
1699 /* Attempt to migrate a task to a CPU on the preferred node. */
1700 static void numa_migrate_preferred(struct task_struct *p)
1701 {
1702         unsigned long interval = HZ;
1703 
1704         /* This task has no NUMA fault statistics yet */
1705         if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1706                 return;
1707 
1708         /* Periodically retry migrating the task to the preferred node */
1709         interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1710         p->numa_migrate_retry = jiffies + interval;
1711 
1712         /* Success if task is already running on preferred CPU */
1713         if (task_node(p) == p->numa_preferred_nid)
1714                 return;
1715 
1716         /* Otherwise, try migrate to a CPU on the preferred node */
1717         task_numa_migrate(p);
1718 }
1719 
1720 /*
1721  * Find out how many nodes on the workload is actively running on. Do this by
1722  * tracking the nodes from which NUMA hinting faults are triggered. This can
1723  * be different from the set of nodes where the workload's memory is currently
1724  * located.
1725  */
1726 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1727 {
1728         unsigned long faults, max_faults = 0;
1729         int nid, active_nodes = 0;
1730 
1731         for_each_online_node(nid) {
1732                 faults = group_faults_cpu(numa_group, nid);
1733                 if (faults > max_faults)
1734                         max_faults = faults;
1735         }
1736 
1737         for_each_online_node(nid) {
1738                 faults = group_faults_cpu(numa_group, nid);
1739                 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1740                         active_nodes++;
1741         }
1742 
1743         numa_group->max_faults_cpu = max_faults;
1744         numa_group->active_nodes = active_nodes;
1745 }
1746 
1747 /*
1748  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1749  * increments. The more local the fault statistics are, the higher the scan
1750  * period will be for the next scan window. If local/(local+remote) ratio is
1751  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1752  * the scan period will decrease. Aim for 70% local accesses.
1753  */
1754 #define NUMA_PERIOD_SLOTS 10
1755 #define NUMA_PERIOD_THRESHOLD 7
1756 
1757 /*
1758  * Increase the scan period (slow down scanning) if the majority of
1759  * our memory is already on our local node, or if the majority of
1760  * the page accesses are shared with other processes.
1761  * Otherwise, decrease the scan period.
1762  */
1763 static void update_task_scan_period(struct task_struct *p,
1764                         unsigned long shared, unsigned long private)
1765 {
1766         unsigned int period_slot;
1767         int ratio;
1768         int diff;
1769 
1770         unsigned long remote = p->numa_faults_locality[0];
1771         unsigned long local = p->numa_faults_locality[1];
1772 
1773         /*
1774          * If there were no record hinting faults then either the task is
1775          * completely idle or all activity is areas that are not of interest
1776          * to automatic numa balancing. Related to that, if there were failed
1777          * migration then it implies we are migrating too quickly or the local
1778          * node is overloaded. In either case, scan slower
1779          */
1780         if (local + shared == 0 || p->numa_faults_locality[2]) {
1781                 p->numa_scan_period = min(p->numa_scan_period_max,
1782                         p->numa_scan_period << 1);
1783 
1784                 p->mm->numa_next_scan = jiffies +
1785                         msecs_to_jiffies(p->numa_scan_period);
1786 
1787                 return;
1788         }
1789 
1790         /*
1791          * Prepare to scale scan period relative to the current period.
1792          *       == NUMA_PERIOD_THRESHOLD scan period stays the same
1793          *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1794          *       >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1795          */
1796         period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1797         ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1798         if (ratio >= NUMA_PERIOD_THRESHOLD) {
1799                 int slot = ratio - NUMA_PERIOD_THRESHOLD;
1800                 if (!slot)
1801                         slot = 1;
1802                 diff = slot * period_slot;
1803         } else {
1804                 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1805 
1806                 /*
1807                  * Scale scan rate increases based on sharing. There is an
1808                  * inverse relationship between the degree of sharing and
1809                  * the adjustment made to the scanning period. Broadly
1810                  * speaking the intent is that there is little point
1811                  * scanning faster if shared accesses dominate as it may
1812                  * simply bounce migrations uselessly
1813                  */
1814                 ratio = DIV_ROUND_UP(private * NUMA_PERIOD_SLOTS, (private + shared + 1));
1815                 diff = (diff * ratio) / NUMA_PERIOD_SLOTS;
1816         }
1817 
1818         p->numa_scan_period = clamp(p->numa_scan_period + diff,
1819                         task_scan_min(p), task_scan_max(p));
1820         memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1821 }
1822 
1823 /*
1824  * Get the fraction of time the task has been running since the last
1825  * NUMA placement cycle. The scheduler keeps similar statistics, but
1826  * decays those on a 32ms period, which is orders of magnitude off
1827  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1828  * stats only if the task is so new there are no NUMA statistics yet.
1829  */
1830 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1831 {
1832         u64 runtime, delta, now;
1833         /* Use the start of this time slice to avoid calculations. */
1834         now = p->se.exec_start;
1835         runtime = p->se.sum_exec_runtime;
1836 
1837         if (p->last_task_numa_placement) {
1838                 delta = runtime - p->last_sum_exec_runtime;
1839                 *period = now - p->last_task_numa_placement;
1840         } else {
1841                 delta = p->se.avg.load_sum / p->se.load.weight;
1842                 *period = LOAD_AVG_MAX;
1843         }
1844 
1845         p->last_sum_exec_runtime = runtime;
1846         p->last_task_numa_placement = now;
1847 
1848         return delta;
1849 }
1850 
1851 /*
1852  * Determine the preferred nid for a task in a numa_group. This needs to
1853  * be done in a way that produces consistent results with group_weight,
1854  * otherwise workloads might not converge.
1855  */
1856 static int preferred_group_nid(struct task_struct *p, int nid)
1857 {
1858         nodemask_t nodes;
1859         int dist;
1860 
1861         /* Direct connections between all NUMA nodes. */
1862         if (sched_numa_topology_type == NUMA_DIRECT)
1863                 return nid;
1864 
1865         /*
1866          * On a system with glueless mesh NUMA topology, group_weight
1867          * scores nodes according to the number of NUMA hinting faults on
1868          * both the node itself, and on nearby nodes.
1869          */
1870         if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1871                 unsigned long score, max_score = 0;
1872                 int node, max_node = nid;
1873 
1874                 dist = sched_max_numa_distance;
1875 
1876                 for_each_online_node(node) {
1877                         score = group_weight(p, node, dist);
1878                         if (score > max_score) {
1879                                 max_score = score;
1880                                 max_node = node;
1881                         }
1882                 }
1883                 return max_node;
1884         }
1885 
1886         /*
1887          * Finding the preferred nid in a system with NUMA backplane
1888          * interconnect topology is more involved. The goal is to locate
1889          * tasks from numa_groups near each other in the system, and
1890          * untangle workloads from different sides of the system. This requires
1891          * searching down the hierarchy of node groups, recursively searching
1892          * inside the highest scoring group of nodes. The nodemask tricks
1893          * keep the complexity of the search down.
1894          */
1895         nodes = node_online_map;
1896         for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
1897                 unsigned long max_faults = 0;
1898                 nodemask_t max_group = NODE_MASK_NONE;
1899                 int a, b;
1900 
1901                 /* Are there nodes at this distance from each other? */
1902                 if (!find_numa_distance(dist))
1903                         continue;
1904 
1905                 for_each_node_mask(a, nodes) {
1906                         unsigned long faults = 0;
1907                         nodemask_t this_group;
1908                         nodes_clear(this_group);
1909 
1910                         /* Sum group's NUMA faults; includes a==b case. */
1911                         for_each_node_mask(b, nodes) {
1912                                 if (node_distance(a, b) < dist) {
1913                                         faults += group_faults(p, b);
1914                                         node_set(b, this_group);
1915                                         node_clear(b, nodes);
1916                                 }
1917                         }
1918 
1919                         /* Remember the top group. */
1920                         if (faults > max_faults) {
1921                                 max_faults = faults;
1922                                 max_group = this_group;
1923                                 /*
1924                                  * subtle: at the smallest distance there is
1925                                  * just one node left in each "group", the
1926                                  * winner is the preferred nid.
1927                                  */
1928                                 nid = a;
1929                         }
1930                 }
1931                 /* Next round, evaluate the nodes within max_group. */
1932                 if (!max_faults)
1933                         break;
1934                 nodes = max_group;
1935         }
1936         return nid;
1937 }
1938 
1939 static void task_numa_placement(struct task_struct *p)
1940 {
1941         int seq, nid, max_nid = -1, max_group_nid = -1;
1942         unsigned long max_faults = 0, max_group_faults = 0;
1943         unsigned long fault_types[2] = { 0, 0 };
1944         unsigned long total_faults;
1945         u64 runtime, period;
1946         spinlock_t *group_lock = NULL;
1947 
1948         /*
1949          * The p->mm->numa_scan_seq field gets updated without
1950          * exclusive access. Use READ_ONCE() here to ensure
1951          * that the field is read in a single access:
1952          */
1953         seq = READ_ONCE(p->mm->numa_scan_seq);
1954         if (p->numa_scan_seq == seq)
1955                 return;
1956         p->numa_scan_seq = seq;
1957         p->numa_scan_period_max = task_scan_max(p);
1958 
1959         total_faults = p->numa_faults_locality[0] +
1960                        p->numa_faults_locality[1];
1961         runtime = numa_get_avg_runtime(p, &period);
1962 
1963         /* If the task is part of a group prevent parallel updates to group stats */
1964         if (p->numa_group) {
1965                 group_lock = &p->numa_group->lock;
1966                 spin_lock_irq(group_lock);
1967         }
1968 
1969         /* Find the node with the highest number of faults */
1970         for_each_online_node(nid) {
1971                 /* Keep track of the offsets in numa_faults array */
1972                 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
1973                 unsigned long faults = 0, group_faults = 0;
1974                 int priv;
1975 
1976                 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
1977                         long diff, f_diff, f_weight;
1978 
1979                         mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
1980                         membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
1981                         cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
1982                         cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
1983 
1984                         /* Decay existing window, copy faults since last scan */
1985                         diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
1986                         fault_types[priv] += p->numa_faults[membuf_idx];
1987                         p->numa_faults[membuf_idx] = 0;
1988 
1989                         /*
1990                          * Normalize the faults_from, so all tasks in a group
1991                          * count according to CPU use, instead of by the raw
1992                          * number of faults. Tasks with little runtime have
1993                          * little over-all impact on throughput, and thus their
1994                          * faults are less important.
1995                          */
1996                         f_weight = div64_u64(runtime << 16, period + 1);
1997                         f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
1998                                    (total_faults + 1);
1999                         f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2000                         p->numa_faults[cpubuf_idx] = 0;
2001 
2002                         p->numa_faults[mem_idx] += diff;
2003                         p->numa_faults[cpu_idx] += f_diff;
2004                         faults += p->numa_faults[mem_idx];
2005                         p->total_numa_faults += diff;
2006                         if (p->numa_group) {
2007                                 /*
2008                                  * safe because we can only change our own group
2009                                  *
2010                                  * mem_idx represents the offset for a given
2011                                  * nid and priv in a specific region because it
2012                                  * is at the beginning of the numa_faults array.
2013                                  */
2014                                 p->numa_group->faults[mem_idx] += diff;
2015                                 p->numa_group->faults_cpu[mem_idx] += f_diff;
2016                                 p->numa_group->total_faults += diff;
2017                                 group_faults += p->numa_group->faults[mem_idx];
2018                         }
2019                 }
2020 
2021                 if (faults > max_faults) {
2022                         max_faults = faults;
2023                         max_nid = nid;
2024                 }
2025 
2026                 if (group_faults > max_group_faults) {
2027                         max_group_faults = group_faults;
2028                         max_group_nid = nid;
2029                 }
2030         }
2031 
2032         update_task_scan_period(p, fault_types[0], fault_types[1]);
2033 
2034         if (p->numa_group) {
2035                 numa_group_count_active_nodes(p->numa_group);
2036                 spin_unlock_irq(group_lock);
2037                 max_nid = preferred_group_nid(p, max_group_nid);
2038         }
2039 
2040         if (max_faults) {
2041                 /* Set the new preferred node */
2042                 if (max_nid != p->numa_preferred_nid)
2043                         sched_setnuma(p, max_nid);
2044 
2045                 if (task_node(p) != p->numa_preferred_nid)
2046                         numa_migrate_preferred(p);
2047         }
2048 }
2049 
2050 static inline int get_numa_group(struct numa_group *grp)
2051 {
2052         return atomic_inc_not_zero(&grp->refcount);
2053 }
2054 
2055 static inline void put_numa_group(struct numa_group *grp)
2056 {
2057         if (atomic_dec_and_test(&grp->refcount))
2058                 kfree_rcu(grp, rcu);
2059 }
2060 
2061 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2062                         int *priv)
2063 {
2064         struct numa_group *grp, *my_grp;
2065         struct task_struct *tsk;
2066         bool join = false;
2067         int cpu = cpupid_to_cpu(cpupid);
2068         int i;
2069 
2070         if (unlikely(!p->numa_group)) {
2071                 unsigned int size = sizeof(struct numa_group) +
2072                                     4*nr_node_ids*sizeof(unsigned long);
2073 
2074                 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2075                 if (!grp)
2076                         return;
2077 
2078                 atomic_set(&grp->refcount, 1);
2079                 grp->active_nodes = 1;
2080                 grp->max_faults_cpu = 0;
2081                 spin_lock_init(&grp->lock);
2082                 grp->gid = p->pid;
2083                 /* Second half of the array tracks nids where faults happen */
2084                 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2085                                                 nr_node_ids;
2086 
2087                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2088                         grp->faults[i] = p->numa_faults[i];
2089 
2090                 grp->total_faults = p->total_numa_faults;
2091 
2092                 grp->nr_tasks++;
2093                 rcu_assign_pointer(p->numa_group, grp);
2094         }
2095 
2096         rcu_read_lock();
2097         tsk = READ_ONCE(cpu_rq(cpu)->curr);
2098 
2099         if (!cpupid_match_pid(tsk, cpupid))
2100                 goto no_join;
2101 
2102         grp = rcu_dereference(tsk->numa_group);
2103         if (!grp)
2104                 goto no_join;
2105 
2106         my_grp = p->numa_group;
2107         if (grp == my_grp)
2108                 goto no_join;
2109 
2110         /*
2111          * Only join the other group if its bigger; if we're the bigger group,
2112          * the other task will join us.
2113          */
2114         if (my_grp->nr_tasks > grp->nr_tasks)
2115                 goto no_join;
2116 
2117         /*
2118          * Tie-break on the grp address.
2119          */
2120         if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2121                 goto no_join;
2122 
2123         /* Always join threads in the same process. */
2124         if (tsk->mm == current->mm)
2125                 join = true;
2126 
2127         /* Simple filter to avoid false positives due to PID collisions */
2128         if (flags & TNF_SHARED)
2129                 join = true;
2130 
2131         /* Update priv based on whether false sharing was detected */
2132         *priv = !join;
2133 
2134         if (join && !get_numa_group(grp))
2135                 goto no_join;
2136 
2137         rcu_read_unlock();
2138 
2139         if (!join)
2140                 return;
2141 
2142         BUG_ON(irqs_disabled());
2143         double_lock_irq(&my_grp->lock, &grp->lock);
2144 
2145         for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2146                 my_grp->faults[i] -= p->numa_faults[i];
2147                 grp->faults[i] += p->numa_faults[i];
2148         }
2149         my_grp->total_faults -= p->total_numa_faults;
2150         grp->total_faults += p->total_numa_faults;
2151 
2152         my_grp->nr_tasks--;
2153         grp->nr_tasks++;
2154 
2155         spin_unlock(&my_grp->lock);
2156         spin_unlock_irq(&grp->lock);
2157 
2158         rcu_assign_pointer(p->numa_group, grp);
2159 
2160         put_numa_group(my_grp);
2161         return;
2162 
2163 no_join:
2164         rcu_read_unlock();
2165         return;
2166 }
2167 
2168 void task_numa_free(struct task_struct *p)
2169 {
2170         struct numa_group *grp = p->numa_group;
2171         void *numa_faults = p->numa_faults;
2172         unsigned long flags;
2173         int i;
2174 
2175         if (grp) {
2176                 spin_lock_irqsave(&grp->lock, flags);
2177                 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2178                         grp->faults[i] -= p->numa_faults[i];
2179                 grp->total_faults -= p->total_numa_faults;
2180 
2181                 grp->nr_tasks--;
2182                 spin_unlock_irqrestore(&grp->lock, flags);
2183                 RCU_INIT_POINTER(p->numa_group, NULL);
2184                 put_numa_group(grp);
2185         }
2186 
2187         p->numa_faults = NULL;
2188         kfree(numa_faults);
2189 }
2190 
2191 /*
2192  * Got a PROT_NONE fault for a page on @node.
2193  */
2194 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2195 {
2196         struct task_struct *p = current;
2197         bool migrated = flags & TNF_MIGRATED;
2198         int cpu_node = task_node(current);
2199         int local = !!(flags & TNF_FAULT_LOCAL);
2200         struct numa_group *ng;
2201         int priv;
2202 
2203         if (!static_branch_likely(&sched_numa_balancing))
2204                 return;
2205 
2206         /* for example, ksmd faulting in a user's mm */
2207         if (!p->mm)
2208                 return;
2209 
2210         /* Allocate buffer to track faults on a per-node basis */
2211         if (unlikely(!p->numa_faults)) {
2212                 int size = sizeof(*p->numa_faults) *
2213                            NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2214 
2215                 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2216                 if (!p->numa_faults)
2217                         return;
2218 
2219                 p->total_numa_faults = 0;
2220                 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2221         }
2222 
2223         /*
2224          * First accesses are treated as private, otherwise consider accesses
2225          * to be private if the accessing pid has not changed
2226          */
2227         if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2228                 priv = 1;
2229         } else {
2230                 priv = cpupid_match_pid(p, last_cpupid);
2231                 if (!priv && !(flags & TNF_NO_GROUP))
2232                         task_numa_group(p, last_cpupid, flags, &priv);
2233         }
2234 
2235         /*
2236          * If a workload spans multiple NUMA nodes, a shared fault that
2237          * occurs wholly within the set of nodes that the workload is
2238          * actively using should be counted as local. This allows the
2239          * scan rate to slow down when a workload has settled down.
2240          */
2241         ng = p->numa_group;
2242         if (!priv && !local && ng && ng->active_nodes > 1 &&
2243                                 numa_is_active_node(cpu_node, ng) &&
2244                                 numa_is_active_node(mem_node, ng))
2245                 local = 1;
2246 
2247         task_numa_placement(p);
2248 
2249         /*
2250          * Retry task to preferred node migration periodically, in case it
2251          * case it previously failed, or the scheduler moved us.
2252          */
2253         if (time_after(jiffies, p->numa_migrate_retry))
2254                 numa_migrate_preferred(p);
2255 
2256         if (migrated)
2257                 p->numa_pages_migrated += pages;
2258         if (flags & TNF_MIGRATE_FAIL)
2259                 p->numa_faults_locality[2] += pages;
2260 
2261         p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2262         p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2263         p->numa_faults_locality[local] += pages;
2264 }
2265 
2266 static void reset_ptenuma_scan(struct task_struct *p)
2267 {
2268         /*
2269          * We only did a read acquisition of the mmap sem, so
2270          * p->mm->numa_scan_seq is written to without exclusive access
2271          * and the update is not guaranteed to be atomic. That's not
2272          * much of an issue though, since this is just used for
2273          * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2274          * expensive, to avoid any form of compiler optimizations:
2275          */
2276         WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2277         p->mm->numa_scan_offset = 0;
2278 }
2279 
2280 /*
2281  * The expensive part of numa migration is done from task_work context.
2282  * Triggered from task_tick_numa().
2283  */
2284 void task_numa_work(struct callback_head *work)
2285 {
2286         unsigned long migrate, next_scan, now = jiffies;
2287         struct task_struct *p = current;
2288         struct mm_struct *mm = p->mm;
2289         u64 runtime = p->se.sum_exec_runtime;
2290         struct vm_area_struct *vma;
2291         unsigned long start, end;
2292         unsigned long nr_pte_updates = 0;
2293         long pages, virtpages;
2294 
2295         WARN_ON_ONCE(p != container_of(work, struct task_struct, numa_work));
2296 
2297         work->next = work; /* protect against double add */
2298         /*
2299          * Who cares about NUMA placement when they're dying.
2300          *
2301          * NOTE: make sure not to dereference p->mm before this check,
2302          * exit_task_work() happens _after_ exit_mm() so we could be called
2303          * without p->mm even though we still had it when we enqueued this
2304          * work.
2305          */
2306         if (p->flags & PF_EXITING)
2307                 return;
2308 
2309         if (!mm->numa_next_scan) {
2310                 mm->numa_next_scan = now +
2311                         msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2312         }
2313 
2314         /*
2315          * Enforce maximal scan/migration frequency..
2316          */
2317         migrate = mm->numa_next_scan;
2318         if (time_before(now, migrate))
2319                 return;
2320 
2321         if (p->numa_scan_period == 0) {
2322                 p->numa_scan_period_max = task_scan_max(p);
2323                 p->numa_scan_period = task_scan_min(p);
2324         }
2325 
2326         next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2327         if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2328                 return;
2329 
2330         /*
2331          * Delay this task enough that another task of this mm will likely win
2332          * the next time around.
2333          */
2334         p->node_stamp += 2 * TICK_NSEC;
2335 
2336         start = mm->numa_scan_offset;
2337         pages = sysctl_numa_balancing_scan_size;
2338         pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2339         virtpages = pages * 8;     /* Scan up to this much virtual space */
2340         if (!pages)
2341                 return;
2342 
2343 
2344         down_read(&mm->mmap_sem);
2345         vma = find_vma(mm, start);
2346         if (!vma) {
2347                 reset_ptenuma_scan(p);
2348                 start = 0;
2349                 vma = mm->mmap;
2350         }
2351         for (; vma; vma = vma->vm_next) {
2352                 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2353                         is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2354                         continue;
2355                 }
2356 
2357                 /*
2358                  * Shared library pages mapped by multiple processes are not
2359                  * migrated as it is expected they are cache replicated. Avoid
2360                  * hinting faults in read-only file-backed mappings or the vdso
2361                  * as migrating the pages will be of marginal benefit.
2362                  */
2363                 if (!vma->vm_mm ||
2364                     (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2365                         continue;
2366 
2367                 /*
2368                  * Skip inaccessible VMAs to avoid any confusion between
2369                  * PROT_NONE and NUMA hinting ptes
2370                  */
2371                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2372                         continue;
2373 
2374                 do {
2375                         start = max(start, vma->vm_start);
2376                         end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2377                         end = min(end, vma->vm_end);
2378                         nr_pte_updates = change_prot_numa(vma, start, end);
2379 
2380                         /*
2381                          * Try to scan sysctl_numa_balancing_size worth of
2382                          * hpages that have at least one present PTE that
2383                          * is not already pte-numa. If the VMA contains
2384                          * areas that are unused or already full of prot_numa
2385                          * PTEs, scan up to virtpages, to skip through those
2386                          * areas faster.
2387                          */
2388                         if (nr_pte_updates)
2389                                 pages -= (end - start) >> PAGE_SHIFT;
2390                         virtpages -= (end - start) >> PAGE_SHIFT;
2391 
2392                         start = end;
2393                         if (pages <= 0 || virtpages <= 0)
2394                                 goto out;
2395 
2396                         cond_resched();
2397                 } while (end != vma->vm_end);
2398         }
2399 
2400 out:
2401         /*
2402          * It is possible to reach the end of the VMA list but the last few
2403          * VMAs are not guaranteed to the vma_migratable. If they are not, we
2404          * would find the !migratable VMA on the next scan but not reset the
2405          * scanner to the start so check it now.
2406          */
2407         if (vma)
2408                 mm->numa_scan_offset = start;
2409         else
2410                 reset_ptenuma_scan(p);
2411         up_read(&mm->mmap_sem);
2412 
2413         /*
2414          * Make sure tasks use at least 32x as much time to run other code
2415          * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2416          * Usually update_task_scan_period slows down scanning enough; on an
2417          * overloaded system we need to limit overhead on a per task basis.
2418          */
2419         if (unlikely(p->se.sum_exec_runtime != runtime)) {
2420                 u64 diff = p->se.sum_exec_runtime - runtime;
2421                 p->node_stamp += 32 * diff;
2422         }
2423 }
2424 
2425 /*
2426  * Drive the periodic memory faults..
2427  */
2428 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2429 {
2430         struct callback_head *work = &curr->numa_work;
2431         u64 period, now;
2432 
2433         /*
2434          * We don't care about NUMA placement if we don't have memory.
2435          */
2436         if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2437                 return;
2438 
2439         /*
2440          * Using runtime rather than walltime has the dual advantage that
2441          * we (mostly) drive the selection from busy threads and that the
2442          * task needs to have done some actual work before we bother with
2443          * NUMA placement.
2444          */
2445         now = curr->se.sum_exec_runtime;
2446         period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2447 
2448         if (now > curr->node_stamp + period) {
2449                 if (!curr->node_stamp)
2450                         curr->numa_scan_period = task_scan_min(curr);
2451                 curr->node_stamp += period;
2452 
2453                 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2454                         init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2455                         task_work_add(curr, work, true);
2456                 }
2457         }
2458 }
2459 #else
2460 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2461 {
2462 }
2463 
2464 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2465 {
2466 }
2467 
2468 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2469 {
2470 }
2471 #endif /* CONFIG_NUMA_BALANCING */
2472 
2473 static void
2474 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2475 {
2476         update_load_add(&cfs_rq->load, se->load.weight);
2477         if (!parent_entity(se))
2478                 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2479 #ifdef CONFIG_SMP
2480         if (entity_is_task(se)) {
2481                 struct rq *rq = rq_of(cfs_rq);
2482 
2483                 account_numa_enqueue(rq, task_of(se));
2484                 list_add(&se->group_node, &rq->cfs_tasks);
2485         }
2486 #endif
2487         cfs_rq->nr_running++;
2488 }
2489 
2490 static void
2491 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2492 {
2493         update_load_sub(&cfs_rq->load, se->load.weight);
2494         if (!parent_entity(se))
2495                 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2496 #ifdef CONFIG_SMP
2497         if (entity_is_task(se)) {
2498                 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2499                 list_del_init(&se->group_node);
2500         }
2501 #endif
2502         cfs_rq->nr_running--;
2503 }
2504 
2505 #ifdef CONFIG_FAIR_GROUP_SCHED
2506 # ifdef CONFIG_SMP
2507 static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2508 {
2509         long tg_weight, load, shares;
2510 
2511         /*
2512          * This really should be: cfs_rq->avg.load_avg, but instead we use
2513          * cfs_rq->load.weight, which is its upper bound. This helps ramp up
2514          * the shares for small weight interactive tasks.
2515          */
2516         load = scale_load_down(cfs_rq->load.weight);
2517 
2518         tg_weight = atomic_long_read(&tg->load_avg);
2519 
2520         /* Ensure tg_weight >= load */
2521         tg_weight -= cfs_rq->tg_load_avg_contrib;
2522         tg_weight += load;
2523 
2524         shares = (tg->shares * load);
2525         if (tg_weight)
2526                 shares /= tg_weight;
2527 
2528         if (shares < MIN_SHARES)
2529                 shares = MIN_SHARES;
2530         if (shares > tg->shares)
2531                 shares = tg->shares;
2532 
2533         return shares;
2534 }
2535 # else /* CONFIG_SMP */
2536 static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
2537 {
2538         return tg->shares;
2539 }
2540 # endif /* CONFIG_SMP */
2541 
2542 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2543                             unsigned long weight)
2544 {
2545         if (se->on_rq) {
2546                 /* commit outstanding execution time */
2547                 if (cfs_rq->curr == se)
2548                         update_curr(cfs_rq);
2549                 account_entity_dequeue(cfs_rq, se);
2550         }
2551 
2552         update_load_set(&se->load, weight);
2553 
2554         if (se->on_rq)
2555                 account_entity_enqueue(cfs_rq, se);
2556 }
2557 
2558 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2559 
2560 static void update_cfs_shares(struct cfs_rq *cfs_rq)
2561 {
2562         struct task_group *tg;
2563         struct sched_entity *se;
2564         long shares;
2565 
2566         tg = cfs_rq->tg;
2567         se = tg->se[cpu_of(rq_of(cfs_rq))];
2568         if (!se || throttled_hierarchy(cfs_rq))
2569                 return;
2570 #ifndef CONFIG_SMP
2571         if (likely(se->load.weight == tg->shares))
2572                 return;
2573 #endif
2574         shares = calc_cfs_shares(cfs_rq, tg);
2575 
2576         reweight_entity(cfs_rq_of(se), se, shares);
2577 }
2578 #else /* CONFIG_FAIR_GROUP_SCHED */
2579 static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
2580 {
2581 }
2582 #endif /* CONFIG_FAIR_GROUP_SCHED */
2583 
2584 #ifdef CONFIG_SMP
2585 /* Precomputed fixed inverse multiplies for multiplication by y^n */
2586 static const u32 runnable_avg_yN_inv[] = {
2587         0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
2588         0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
2589         0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
2590         0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
2591         0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
2592         0x85aac367, 0x82cd8698,
2593 };
2594 
2595 /*
2596  * Precomputed \Sum y^k { 1<=k<=n }.  These are floor(true_value) to prevent
2597  * over-estimates when re-combining.
2598  */
2599 static const u32 runnable_avg_yN_sum[] = {
2600             0, 1002, 1982, 2941, 3880, 4798, 5697, 6576, 7437, 8279, 9103,
2601          9909,10698,11470,12226,12966,13690,14398,15091,15769,16433,17082,
2602         17718,18340,18949,19545,20128,20698,21256,21802,22336,22859,23371,
2603 };
2604 
2605 /*
2606  * Precomputed \Sum y^k { 1<=k<=n, where n%32=0). Values are rolled down to
2607  * lower integers. See Documentation/scheduler/sched-avg.txt how these
2608  * were generated:
2609  */
2610 static const u32 __accumulated_sum_N32[] = {
2611             0, 23371, 35056, 40899, 43820, 45281,
2612         46011, 46376, 46559, 46650, 46696, 46719,
2613 };
2614 
2615 /*
2616  * Approximate:
2617  *   val * y^n,    where y^32 ~= 0.5 (~1 scheduling period)
2618  */
2619 static __always_inline u64 decay_load(u64 val, u64 n)
2620 {
2621         unsigned int local_n;
2622 
2623         if (!n)
2624                 return val;
2625         else if (unlikely(n > LOAD_AVG_PERIOD * 63))
2626                 return 0;
2627 
2628         /* after bounds checking we can collapse to 32-bit */
2629         local_n = n;
2630 
2631         /*
2632          * As y^PERIOD = 1/2, we can combine
2633          *    y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
2634          * With a look-up table which covers y^n (n<PERIOD)
2635          *
2636          * To achieve constant time decay_load.
2637          */
2638         if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
2639                 val >>= local_n / LOAD_AVG_PERIOD;
2640                 local_n %= LOAD_AVG_PERIOD;
2641         }
2642 
2643         val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
2644         return val;
2645 }
2646 
2647 /*
2648  * For updates fully spanning n periods, the contribution to runnable
2649  * average will be: \Sum 1024*y^n
2650  *
2651  * We can compute this reasonably efficiently by combining:
2652  *   y^PERIOD = 1/2 with precomputed \Sum 1024*y^n {for  n <PERIOD}
2653  */
2654 static u32 __compute_runnable_contrib(u64 n)
2655 {
2656         u32 contrib = 0;
2657 
2658         if (likely(n <= LOAD_AVG_PERIOD))
2659                 return runnable_avg_yN_sum[n];
2660         else if (unlikely(n >= LOAD_AVG_MAX_N))
2661                 return LOAD_AVG_MAX;
2662 
2663         /* Since n < LOAD_AVG_MAX_N, n/LOAD_AVG_PERIOD < 11 */
2664         contrib = __accumulated_sum_N32[n/LOAD_AVG_PERIOD];
2665         n %= LOAD_AVG_PERIOD;
2666         contrib = decay_load(contrib, n);
2667         return contrib + runnable_avg_yN_sum[n];
2668 }
2669 
2670 #define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
2671 
2672 /*
2673  * We can represent the historical contribution to runnable average as the
2674  * coefficients of a geometric series.  To do this we sub-divide our runnable
2675  * history into segments of approximately 1ms (1024us); label the segment that
2676  * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
2677  *
2678  * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
2679  *      p0            p1           p2
2680  *     (now)       (~1ms ago)  (~2ms ago)
2681  *
2682  * Let u_i denote the fraction of p_i that the entity was runnable.
2683  *
2684  * We then designate the fractions u_i as our co-efficients, yielding the
2685  * following representation of historical load:
2686  *   u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
2687  *
2688  * We choose y based on the with of a reasonably scheduling period, fixing:
2689  *   y^32 = 0.5
2690  *
2691  * This means that the contribution to load ~32ms ago (u_32) will be weighted
2692  * approximately half as much as the contribution to load within the last ms
2693  * (u_0).
2694  *
2695  * When a period "rolls over" and we have new u_0`, multiplying the previous
2696  * sum again by y is sufficient to update:
2697  *   load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
2698  *            = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
2699  */
2700 static __always_inline int
2701 __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
2702                   unsigned long weight, int running, struct cfs_rq *cfs_rq)
2703 {
2704         u64 delta, scaled_delta, periods;
2705         u32 contrib;
2706         unsigned int delta_w, scaled_delta_w, decayed = 0;
2707         unsigned long scale_freq, scale_cpu;
2708 
2709         delta = now - sa->last_update_time;
2710         /*
2711          * This should only happen when time goes backwards, which it
2712          * unfortunately does during sched clock init when we swap over to TSC.
2713          */
2714         if ((s64)delta < 0) {
2715                 sa->last_update_time = now;
2716                 return 0;
2717         }
2718 
2719         /*
2720          * Use 1024ns as the unit of measurement since it's a reasonable
2721          * approximation of 1us and fast to compute.
2722          */
2723         delta >>= 10;
2724         if (!delta)
2725                 return 0;
2726         sa->last_update_time = now;
2727 
2728         scale_freq = arch_scale_freq_capacity(NULL, cpu);
2729         scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
2730 
2731         /* delta_w is the amount already accumulated against our next period */
2732         delta_w = sa->period_contrib;
2733         if (delta + delta_w >= 1024) {
2734                 decayed = 1;
2735 
2736                 /* how much left for next period will start over, we don't know yet */
2737                 sa->period_contrib = 0;
2738 
2739                 /*
2740                  * Now that we know we're crossing a period boundary, figure
2741                  * out how much from delta we need to complete the current
2742                  * period and accrue it.
2743                  */
2744                 delta_w = 1024 - delta_w;
2745                 scaled_delta_w = cap_scale(delta_w, scale_freq);
2746                 if (weight) {
2747                         sa->load_sum += weight * scaled_delta_w;
2748                         if (cfs_rq) {
2749                                 cfs_rq->runnable_load_sum +=
2750                                                 weight * scaled_delta_w;
2751                         }
2752                 }
2753                 if (running)
2754                         sa->util_sum += scaled_delta_w * scale_cpu;
2755 
2756                 delta -= delta_w;
2757 
2758                 /* Figure out how many additional periods this update spans */
2759                 periods = delta / 1024;
2760                 delta %= 1024;
2761 
2762                 sa->load_sum = decay_load(sa->load_sum, periods + 1);
2763                 if (cfs_rq) {
2764                         cfs_rq->runnable_load_sum =
2765                                 decay_load(cfs_rq->runnable_load_sum, periods + 1);
2766                 }
2767                 sa->util_sum = decay_load((u64)(sa->util_sum), periods + 1);
2768 
2769                 /* Efficiently calculate \sum (1..n_period) 1024*y^i */
2770                 contrib = __compute_runnable_contrib(periods);
2771                 contrib = cap_scale(contrib, scale_freq);
2772                 if (weight) {
2773                         sa->load_sum += weight * contrib;
2774                         if (cfs_rq)
2775                                 cfs_rq->runnable_load_sum += weight * contrib;
2776                 }
2777                 if (running)
2778                         sa->util_sum += contrib * scale_cpu;
2779         }
2780 
2781         /* Remainder of delta accrued against u_0` */
2782         scaled_delta = cap_scale(delta, scale_freq);
2783         if (weight) {
2784                 sa->load_sum += weight * scaled_delta;
2785                 if (cfs_rq)
2786                         cfs_rq->runnable_load_sum += weight * scaled_delta;
2787         }
2788         if (running)
2789                 sa->util_sum += scaled_delta * scale_cpu;
2790 
2791         sa->period_contrib += delta;
2792 
2793         if (decayed) {
2794                 sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
2795                 if (cfs_rq) {
2796                         cfs_rq->runnable_load_avg =
2797                                 div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
2798                 }
2799                 sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
2800         }
2801 
2802         return decayed;
2803 }
2804 
2805 #ifdef CONFIG_FAIR_GROUP_SCHED
2806 /*
2807  * Updating tg's load_avg is necessary before update_cfs_share (which is done)
2808  * and effective_load (which is not done because it is too costly).
2809  */
2810 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
2811 {
2812         long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
2813 
2814         /*
2815          * No need to update load_avg for root_task_group as it is not used.
2816          */
2817         if (cfs_rq->tg == &root_task_group)
2818                 return;
2819 
2820         if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
2821                 atomic_long_add(delta, &cfs_rq->tg->load_avg);
2822                 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
2823         }
2824 }
2825 
2826 /*
2827  * Called within set_task_rq() right before setting a task's cpu. The
2828  * caller only guarantees p->pi_lock is held; no other assumptions,
2829  * including the state of rq->lock, should be made.
2830  */
2831 void set_task_rq_fair(struct sched_entity *se,
2832                       struct cfs_rq *prev, struct cfs_rq *next)
2833 {
2834         if (!sched_feat(ATTACH_AGE_LOAD))
2835                 return;
2836 
2837         /*
2838          * We are supposed to update the task to "current" time, then its up to
2839          * date and ready to go to new CPU/cfs_rq. But we have difficulty in
2840          * getting what current time is, so simply throw away the out-of-date
2841          * time. This will result in the wakee task is less decayed, but giving
2842          * the wakee more load sounds not bad.
2843          */
2844         if (se->avg.last_update_time && prev) {
2845                 u64 p_last_update_time;
2846                 u64 n_last_update_time;
2847 
2848 #ifndef CONFIG_64BIT
2849                 u64 p_last_update_time_copy;
2850                 u64 n_last_update_time_copy;
2851 
2852                 do {
2853                         p_last_update_time_copy = prev->load_last_update_time_copy;
2854                         n_last_update_time_copy = next->load_last_update_time_copy;
2855 
2856                         smp_rmb();
2857 
2858                         p_last_update_time = prev->avg.last_update_time;
2859                         n_last_update_time = next->avg.last_update_time;
2860 
2861                 } while (p_last_update_time != p_last_update_time_copy ||
2862                          n_last_update_time != n_last_update_time_copy);
2863 #else
2864                 p_last_update_time = prev->avg.last_update_time;
2865                 n_last_update_time = next->avg.last_update_time;
2866 #endif
2867                 __update_load_avg(p_last_update_time, cpu_of(rq_of(prev)),
2868                                   &se->avg, 0, 0, NULL);
2869                 se->avg.last_update_time = n_last_update_time;
2870         }
2871 }
2872 #else /* CONFIG_FAIR_GROUP_SCHED */
2873 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
2874 #endif /* CONFIG_FAIR_GROUP_SCHED */
2875 
2876 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq)
2877 {
2878         struct rq *rq = rq_of(cfs_rq);
2879         int cpu = cpu_of(rq);
2880 
2881         if (cpu == smp_processor_id() && &rq->cfs == cfs_rq) {
2882                 unsigned long max = rq->cpu_capacity_orig;
2883 
2884                 /*
2885                  * There are a few boundary cases this might miss but it should
2886                  * get called often enough that that should (hopefully) not be
2887                  * a real problem -- added to that it only calls on the local
2888                  * CPU, so if we enqueue remotely we'll miss an update, but
2889                  * the next tick/schedule should update.
2890                  *
2891                  * It will not get called when we go idle, because the idle
2892                  * thread is a different class (!fair), nor will the utilization
2893                  * number include things like RT tasks.
2894                  *
2895                  * As is, the util number is not freq-invariant (we'd have to
2896                  * implement arch_scale_freq_capacity() for that).
2897                  *
2898                  * See cpu_util().
2899                  */
2900                 cpufreq_update_util(rq_clock(rq),
2901                                     min(cfs_rq->avg.util_avg, max), max);
2902         }
2903 }
2904 
2905 /*
2906  * Unsigned subtract and clamp on underflow.
2907  *
2908  * Explicitly do a load-store to ensure the intermediate value never hits
2909  * memory. This allows lockless observations without ever seeing the negative
2910  * values.
2911  */
2912 #define sub_positive(_ptr, _val) do {                           \
2913         typeof(_ptr) ptr = (_ptr);                              \
2914         typeof(*ptr) val = (_val);                              \
2915         typeof(*ptr) res, var = READ_ONCE(*ptr);                \
2916         res = var - val;                                        \
2917         if (res > var)                                          \
2918                 res = 0;                                        \
2919         WRITE_ONCE(*ptr, res);                                  \
2920 } while (0)
2921 
2922 /**
2923  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
2924  * @now: current time, as per cfs_rq_clock_task()
2925  * @cfs_rq: cfs_rq to update
2926  * @update_freq: should we call cfs_rq_util_change() or will the call do so
2927  *
2928  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
2929  * avg. The immediate corollary is that all (fair) tasks must be attached, see
2930  * post_init_entity_util_avg().
2931  *
2932  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
2933  *
2934  * Returns true if the load decayed or we removed utilization. It is expected
2935  * that one calls update_tg_load_avg() on this condition, but after you've
2936  * modified the cfs_rq avg (attach/detach), such that we propagate the new
2937  * avg up.
2938  */
2939 static inline int
2940 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
2941 {
2942         struct sched_avg *sa = &cfs_rq->avg;
2943         int decayed, removed_load = 0, removed_util = 0;
2944 
2945         if (atomic_long_read(&cfs_rq->removed_load_avg)) {
2946                 s64 r = atomic_long_xchg(&cfs_rq->removed_load_avg, 0);
2947                 sub_positive(&sa->load_avg, r);
2948                 sub_positive(&sa->load_sum, r * LOAD_AVG_MAX);
2949                 removed_load = 1;
2950         }
2951 
2952         if (atomic_long_read(&cfs_rq->removed_util_avg)) {
2953                 long r = atomic_long_xchg(&cfs_rq->removed_util_avg, 0);
2954                 sub_positive(&sa->util_avg, r);
2955                 sub_positive(&sa->util_sum, r * LOAD_AVG_MAX);
2956                 removed_util = 1;
2957         }
2958 
2959         decayed = __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
2960                 scale_load_down(cfs_rq->load.weight), cfs_rq->curr != NULL, cfs_rq);
2961 
2962 #ifndef CONFIG_64BIT
2963         smp_wmb();
2964         cfs_rq->load_last_update_time_copy = sa->last_update_time;
2965 #endif
2966 
2967         if (update_freq && (decayed || removed_util))
2968                 cfs_rq_util_change(cfs_rq);
2969 
2970         return decayed || removed_load;
2971 }
2972 
2973 /* Update task and its cfs_rq load average */
2974 static inline void update_load_avg(struct sched_entity *se, int update_tg)
2975 {
2976         struct cfs_rq *cfs_rq = cfs_rq_of(se);
2977         u64 now = cfs_rq_clock_task(cfs_rq);
2978         struct rq *rq = rq_of(cfs_rq);
2979         int cpu = cpu_of(rq);
2980 
2981         /*
2982          * Track task load average for carrying it to new CPU after migrated, and
2983          * track group sched_entity load average for task_h_load calc in migration
2984          */
2985         __update_load_avg(now, cpu, &se->avg,
2986                           se->on_rq * scale_load_down(se->load.weight),
2987                           cfs_rq->curr == se, NULL);
2988 
2989         if (update_cfs_rq_load_avg(now, cfs_rq, true) && update_tg)
2990                 update_tg_load_avg(cfs_rq, 0);
2991 }
2992 
2993 /**
2994  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
2995  * @cfs_rq: cfs_rq to attach to
2996  * @se: sched_entity to attach
2997  *
2998  * Must call update_cfs_rq_load_avg() before this, since we rely on
2999  * cfs_rq->avg.last_update_time being current.
3000  */
3001 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3002 {
3003         if (!sched_feat(ATTACH_AGE_LOAD))
3004                 goto skip_aging;
3005 
3006         /*
3007          * If we got migrated (either between CPUs or between cgroups) we'll
3008          * have aged the average right before clearing @last_update_time.
3009          *
3010          * Or we're fresh through post_init_entity_util_avg().
3011          */
3012         if (se->avg.last_update_time) {
3013                 __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3014                                   &se->avg, 0, 0, NULL);
3015 
3016                 /*
3017                  * XXX: we could have just aged the entire load away if we've been
3018                  * absent from the fair class for too long.
3019                  */
3020         }
3021 
3022 skip_aging:
3023         se->avg.last_update_time = cfs_rq->avg.last_update_time;
3024         cfs_rq->avg.load_avg += se->avg.load_avg;
3025         cfs_rq->avg.load_sum += se->avg.load_sum;
3026         cfs_rq->avg.util_avg += se->avg.util_avg;
3027         cfs_rq->avg.util_sum += se->avg.util_sum;
3028 
3029         cfs_rq_util_change(cfs_rq);
3030 }
3031 
3032 /**
3033  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3034  * @cfs_rq: cfs_rq to detach from
3035  * @se: sched_entity to detach
3036  *
3037  * Must call update_cfs_rq_load_avg() before this, since we rely on
3038  * cfs_rq->avg.last_update_time being current.
3039  */
3040 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3041 {
3042         __update_load_avg(cfs_rq->avg.last_update_time, cpu_of(rq_of(cfs_rq)),
3043                           &se->avg, se->on_rq * scale_load_down(se->load.weight),
3044                           cfs_rq->curr == se, NULL);
3045 
3046         sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3047         sub_positive(&cfs_rq->avg.load_sum, se->avg.load_sum);
3048         sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3049         sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3050 
3051         cfs_rq_util_change(cfs_rq);
3052 }
3053 
3054 /* Add the load generated by se into cfs_rq's load average */
3055 static inline void
3056 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3057 {
3058         struct sched_avg *sa = &se->avg;
3059         u64 now = cfs_rq_clock_task(cfs_rq);
3060         int migrated, decayed;
3061 
3062         migrated = !sa->last_update_time;
3063         if (!migrated) {
3064                 __update_load_avg(now, cpu_of(rq_of(cfs_rq)), sa,
3065                         se->on_rq * scale_load_down(se->load.weight),
3066                         cfs_rq->curr == se, NULL);
3067         }
3068 
3069         decayed = update_cfs_rq_load_avg(now, cfs_rq, !migrated);
3070 
3071         cfs_rq->runnable_load_avg += sa->load_avg;
3072         cfs_rq->runnable_load_sum += sa->load_sum;
3073 
3074         if (migrated)
3075                 attach_entity_load_avg(cfs_rq, se);
3076 
3077         if (decayed || migrated)
3078                 update_tg_load_avg(cfs_rq, 0);
3079 }
3080 
3081 /* Remove the runnable load generated by se from cfs_rq's runnable load average */
3082 static inline void
3083 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3084 {
3085         update_load_avg(se, 1);
3086 
3087         cfs_rq->runnable_load_avg =
3088                 max_t(long, cfs_rq->runnable_load_avg - se->avg.load_avg, 0);
3089         cfs_rq->runnable_load_sum =
3090                 max_t(s64,  cfs_rq->runnable_load_sum - se->avg.load_sum, 0);
3091 }
3092 
3093 #ifndef CONFIG_64BIT
3094 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3095 {
3096         u64 last_update_time_copy;
3097         u64 last_update_time;
3098 
3099         do {
3100                 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3101                 smp_rmb();
3102                 last_update_time = cfs_rq->avg.last_update_time;
3103         } while (last_update_time != last_update_time_copy);
3104 
3105         return last_update_time;
3106 }
3107 #else
3108 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3109 {
3110         return cfs_rq->avg.last_update_time;
3111 }
3112 #endif
3113 
3114 /*
3115  * Task first catches up with cfs_rq, and then subtract
3116  * itself from the cfs_rq (task must be off the queue now).
3117  */
3118 void remove_entity_load_avg(struct sched_entity *se)
3119 {
3120         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3121         u64 last_update_time;
3122 
3123         /*
3124          * tasks cannot exit without having gone through wake_up_new_task() ->
3125          * post_init_entity_util_avg() which will have added things to the
3126          * cfs_rq, so we can remove unconditionally.
3127          *
3128          * Similarly for groups, they will have passed through
3129          * post_init_entity_util_avg() before unregister_sched_fair_group()
3130          * calls this.
3131          */
3132 
3133         last_update_time = cfs_rq_last_update_time(cfs_rq);
3134 
3135         __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
3136         atomic_long_add(se->avg.load_avg, &cfs_rq->removed_load_avg);
3137         atomic_long_add(se->avg.util_avg, &cfs_rq->removed_util_avg);
3138 }
3139 
3140 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3141 {
3142         return cfs_rq->runnable_load_avg;
3143 }
3144 
3145 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3146 {
3147         return cfs_rq->avg.load_avg;
3148 }
3149 
3150 static int idle_balance(struct rq *this_rq);
3151 
3152 #else /* CONFIG_SMP */
3153 
3154 static inline int
3155 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq, bool update_freq)
3156 {
3157         return 0;
3158 }
3159 
3160 static inline void update_load_avg(struct sched_entity *se, int not_used)
3161 {
3162         struct cfs_rq *cfs_rq = cfs_rq_of(se);
3163         struct rq *rq = rq_of(cfs_rq);
3164 
3165         cpufreq_trigger_update(rq_clock(rq));
3166 }
3167 
3168 static inline void
3169 enqueue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3170 static inline void
3171 dequeue_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3172 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3173 
3174 static inline void
3175 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3176 static inline void
3177 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3178 
3179 static inline int idle_balance(struct rq *rq)
3180 {
3181         return 0;
3182 }
3183 
3184 #endif /* CONFIG_SMP */
3185 
3186 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
3187 {
3188 #ifdef CONFIG_SCHEDSTATS
3189         struct task_struct *tsk = NULL;
3190 
3191         if (entity_is_task(se))
3192                 tsk = task_of(se);
3193 
3194         if (se->statistics.sleep_start) {
3195                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.sleep_start;
3196 
3197                 if ((s64)delta < 0)
3198                         delta = 0;
3199 
3200                 if (unlikely(delta > se->statistics.sleep_max))
3201                         se->statistics.sleep_max = delta;
3202 
3203                 se->statistics.sleep_start = 0;
3204                 se->statistics.sum_sleep_runtime += delta;
3205 
3206                 if (tsk) {
3207                         account_scheduler_latency(tsk, delta >> 10, 1);
3208                         trace_sched_stat_sleep(tsk, delta);
3209                 }
3210         }
3211         if (se->statistics.block_start) {
3212                 u64 delta = rq_clock(rq_of(cfs_rq)) - se->statistics.block_start;
3213 
3214                 if ((s64)delta < 0)
3215                         delta = 0;
3216 
3217                 if (unlikely(delta > se->statistics.block_max))
3218                         se->statistics.block_max = delta;
3219 
3220                 se->statistics.block_start = 0;
3221                 se->statistics.sum_sleep_runtime += delta;
3222 
3223                 if (tsk) {
3224                         if (tsk->in_iowait) {
3225                                 se->statistics.iowait_sum += delta;
3226                                 se->statistics.iowait_count++;
3227                                 trace_sched_stat_iowait(tsk, delta);
3228                         }
3229 
3230                         trace_sched_stat_blocked(tsk, delta);
3231 
3232                         /*
3233                          * Blocking time is in units of nanosecs, so shift by
3234                          * 20 to get a milliseconds-range estimation of the
3235                          * amount of time that the task spent sleeping:
3236                          */
3237                         if (unlikely(prof_on == SLEEP_PROFILING)) {
3238                                 profile_hits(SLEEP_PROFILING,
3239                                                 (void *)get_wchan(tsk),
3240                                                 delta >> 20);
3241                         }
3242                         account_scheduler_latency(tsk, delta >> 10, 0);
3243                 }
3244         }
3245 #endif
3246 }
3247 
3248 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3249 {
3250 #ifdef CONFIG_SCHED_DEBUG
3251         s64 d = se->vruntime - cfs_rq->min_vruntime;
3252 
3253         if (d < 0)
3254                 d = -d;
3255 
3256         if (d > 3*sysctl_sched_latency)
3257                 schedstat_inc(cfs_rq, nr_spread_over);
3258 #endif
3259 }
3260 
3261 static void
3262 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3263 {
3264         u64 vruntime = cfs_rq->min_vruntime;
3265 
3266         /*
3267          * The 'current' period is already promised to the current tasks,
3268          * however the extra weight of the new task will slow them down a
3269          * little, place the new task so that it fits in the slot that
3270          * stays open at the end.
3271          */
3272         if (initial && sched_feat(START_DEBIT))
3273                 vruntime += sched_vslice(cfs_rq, se);
3274 
3275         /* sleeps up to a single latency don't count. */
3276         if (!initial) {
3277                 unsigned long thresh = sysctl_sched_latency;
3278 
3279                 /*
3280                  * Halve their sleep time's effect, to allow
3281                  * for a gentler effect of sleepers:
3282                  */
3283                 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3284                         thresh >>= 1;
3285 
3286                 vruntime -= thresh;
3287         }
3288 
3289         /* ensure we never gain time by being placed backwards. */
3290         se->vruntime = max_vruntime(se->vruntime, vruntime);
3291 }
3292 
3293 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3294 
3295 static inline void check_schedstat_required(void)
3296 {
3297 #ifdef CONFIG_SCHEDSTATS
3298         if (schedstat_enabled())
3299                 return;
3300 
3301         /* Force schedstat enabled if a dependent tracepoint is active */
3302         if (trace_sched_stat_wait_enabled()    ||
3303                         trace_sched_stat_sleep_enabled()   ||
3304                         trace_sched_stat_iowait_enabled()  ||
3305                         trace_sched_stat_blocked_enabled() ||
3306                         trace_sched_stat_runtime_enabled())  {
3307                 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3308                              "stat_blocked and stat_runtime require the "
3309                              "kernel parameter schedstats=enabled or "
3310                              "kernel.sched_schedstats=1\n");
3311         }
3312 #endif
3313 }
3314 
3315 
3316 /*
3317  * MIGRATION
3318  *
3319  *      dequeue
3320  *        update_curr()
3321  *          update_min_vruntime()
3322  *        vruntime -= min_vruntime
3323  *
3324  *      enqueue
3325  *        update_curr()
3326  *          update_min_vruntime()
3327  *        vruntime += min_vruntime
3328  *
3329  * this way the vruntime transition between RQs is done when both
3330  * min_vruntime are up-to-date.
3331  *
3332  * WAKEUP (remote)
3333  *
3334  *      ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3335  *        vruntime -= min_vruntime
3336  *
3337  *      enqueue
3338  *        update_curr()
3339  *          update_min_vruntime()
3340  *        vruntime += min_vruntime
3341  *
3342  * this way we don't have the most up-to-date min_vruntime on the originating
3343  * CPU and an up-to-date min_vruntime on the destination CPU.
3344  */
3345 
3346 static void
3347 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3348 {
3349         bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3350         bool curr = cfs_rq->curr == se;
3351 
3352         /*
3353          * If we're the current task, we must renormalise before calling
3354          * update_curr().
3355          */
3356         if (renorm && curr)
3357                 se->vruntime += cfs_rq->min_vruntime;
3358 
3359         update_curr(cfs_rq);
3360 
3361         /*
3362          * Otherwise, renormalise after, such that we're placed at the current
3363          * moment in time, instead of some random moment in the past. Being
3364          * placed in the past could significantly boost this task to the
3365          * fairness detriment of existing tasks.
3366          */
3367         if (renorm && !curr)
3368                 se->vruntime += cfs_rq->min_vruntime;
3369 
3370         enqueue_entity_load_avg(cfs_rq, se);
3371         account_entity_enqueue(cfs_rq, se);
3372         update_cfs_shares(cfs_rq);
3373 
3374         if (flags & ENQUEUE_WAKEUP) {
3375                 place_entity(cfs_rq, se, 0);
3376                 if (schedstat_enabled())
3377                         enqueue_sleeper(cfs_rq, se);
3378         }
3379 
3380         check_schedstat_required();
3381         if (schedstat_enabled()) {
3382                 update_stats_enqueue(cfs_rq, se);
3383                 check_spread(cfs_rq, se);
3384         }
3385         if (!curr)
3386                 __enqueue_entity(cfs_rq, se);
3387         se->on_rq = 1;
3388 
3389         if (cfs_rq->nr_running == 1) {
3390                 list_add_leaf_cfs_rq(cfs_rq);
3391                 check_enqueue_throttle(cfs_rq);
3392         }
3393 }
3394 
3395 static void __clear_buddies_last(struct sched_entity *se)
3396 {
3397         for_each_sched_entity(se) {
3398                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3399                 if (cfs_rq->last != se)
3400                         break;
3401 
3402                 cfs_rq->last = NULL;
3403         }
3404 }
3405 
3406 static void __clear_buddies_next(struct sched_entity *se)
3407 {
3408         for_each_sched_entity(se) {
3409                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3410                 if (cfs_rq->next != se)
3411                         break;
3412 
3413                 cfs_rq->next = NULL;
3414         }
3415 }
3416 
3417 static void __clear_buddies_skip(struct sched_entity *se)
3418 {
3419         for_each_sched_entity(se) {
3420                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3421                 if (cfs_rq->skip != se)
3422                         break;
3423 
3424                 cfs_rq->skip = NULL;
3425         }
3426 }
3427 
3428 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3429 {
3430         if (cfs_rq->last == se)
3431                 __clear_buddies_last(se);
3432 
3433         if (cfs_rq->next == se)
3434                 __clear_buddies_next(se);
3435 
3436         if (cfs_rq->skip == se)
3437                 __clear_buddies_skip(se);
3438 }
3439 
3440 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3441 
3442 static void
3443 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3444 {
3445         /*
3446          * Update run-time statistics of the 'current'.
3447          */
3448         update_curr(cfs_rq);
3449         dequeue_entity_load_avg(cfs_rq, se);
3450 
3451         if (schedstat_enabled())
3452                 update_stats_dequeue(cfs_rq, se, flags);
3453 
3454         clear_buddies(cfs_rq, se);
3455 
3456         if (se != cfs_rq->curr)
3457                 __dequeue_entity(cfs_rq, se);
3458         se->on_rq = 0;
3459         account_entity_dequeue(cfs_rq, se);
3460 
3461         /*
3462          * Normalize the entity after updating the min_vruntime because the
3463          * update can refer to the ->curr item and we need to reflect this
3464          * movement in our normalized position.
3465          */
3466         if (!(flags & DEQUEUE_SLEEP))
3467                 se->vruntime -= cfs_rq->min_vruntime;
3468 
3469         /* return excess runtime on last dequeue */
3470         return_cfs_rq_runtime(cfs_rq);
3471 
3472         update_min_vruntime(cfs_rq);
3473         update_cfs_shares(cfs_rq);
3474 }
3475 
3476 /*
3477  * Preempt the current task with a newly woken task if needed:
3478  */
3479 static void
3480 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3481 {
3482         unsigned long ideal_runtime, delta_exec;
3483         struct sched_entity *se;
3484         s64 delta;
3485 
3486         ideal_runtime = sched_slice(cfs_rq, curr);
3487         delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3488         if (delta_exec > ideal_runtime) {
3489                 resched_curr(rq_of(cfs_rq));
3490                 /*
3491                  * The current task ran long enough, ensure it doesn't get
3492                  * re-elected due to buddy favours.
3493                  */
3494                 clear_buddies(cfs_rq, curr);
3495                 return;
3496         }
3497 
3498         /*
3499          * Ensure that a task that missed wakeup preemption by a
3500          * narrow margin doesn't have to wait for a full slice.
3501          * This also mitigates buddy induced latencies under load.
3502          */
3503         if (delta_exec < sysctl_sched_min_granularity)
3504                 return;
3505 
3506         se = __pick_first_entity(cfs_rq);
3507         delta = curr->vruntime - se->vruntime;
3508 
3509         if (delta < 0)
3510                 return;
3511 
3512         if (delta > ideal_runtime)
3513                 resched_curr(rq_of(cfs_rq));
3514 }
3515 
3516 static void
3517 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3518 {
3519         /* 'current' is not kept within the tree. */
3520         if (se->on_rq) {
3521                 /*
3522                  * Any task has to be enqueued before it get to execute on
3523                  * a CPU. So account for the time it spent waiting on the
3524                  * runqueue.
3525                  */
3526                 if (schedstat_enabled())
3527                         update_stats_wait_end(cfs_rq, se);
3528                 __dequeue_entity(cfs_rq, se);
3529                 update_load_avg(se, 1);
3530         }
3531 
3532         update_stats_curr_start(cfs_rq, se);
3533         cfs_rq->curr = se;
3534 #ifdef CONFIG_SCHEDSTATS
3535         /*
3536          * Track our maximum slice length, if the CPU's load is at
3537          * least twice that of our own weight (i.e. dont track it
3538          * when there are only lesser-weight tasks around):
3539          */
3540         if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
3541                 se->statistics.slice_max = max(se->statistics.slice_max,
3542                         se->sum_exec_runtime - se->prev_sum_exec_runtime);
3543         }
3544 #endif
3545         se->prev_sum_exec_runtime = se->sum_exec_runtime;
3546 }
3547 
3548 static int
3549 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
3550 
3551 /*
3552  * Pick the next process, keeping these things in mind, in this order:
3553  * 1) keep things fair between processes/task groups
3554  * 2) pick the "next" process, since someone really wants that to run
3555  * 3) pick the "last" process, for cache locality
3556  * 4) do not run the "skip" process, if something else is available
3557  */
3558 static struct sched_entity *
3559 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3560 {
3561         struct sched_entity *left = __pick_first_entity(cfs_rq);
3562         struct sched_entity *se;
3563 
3564         /*
3565          * If curr is set we have to see if its left of the leftmost entity
3566          * still in the tree, provided there was anything in the tree at all.
3567          */
3568         if (!left || (curr && entity_before(curr, left)))
3569                 left = curr;
3570 
3571         se = left; /* ideally we run the leftmost entity */
3572 
3573         /*
3574          * Avoid running the skip buddy, if running something else can
3575          * be done without getting too unfair.
3576          */
3577         if (cfs_rq->skip == se) {
3578                 struct sched_entity *second;
3579 
3580                 if (se == curr) {
3581                         second = __pick_first_entity(cfs_rq);
3582                 } else {
3583                         second = __pick_next_entity(se);
3584                         if (!second || (curr && entity_before(curr, second)))
3585                                 second = curr;
3586                 }
3587 
3588                 if (second && wakeup_preempt_entity(second, left) < 1)
3589                         se = second;
3590         }
3591 
3592         /*
3593          * Prefer last buddy, try to return the CPU to a preempted task.
3594          */
3595         if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
3596                 se = cfs_rq->last;
3597 
3598         /*
3599          * Someone really wants this to run. If it's not unfair, run it.
3600          */
3601         if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
3602                 se = cfs_rq->next;
3603 
3604         clear_buddies(cfs_rq, se);
3605 
3606         return se;
3607 }
3608 
3609 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3610 
3611 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
3612 {
3613         /*
3614          * If still on the runqueue then deactivate_task()
3615          * was not called and update_curr() has to be done:
3616          */
3617         if (prev->on_rq)
3618                 update_curr(cfs_rq);
3619 
3620         /* throttle cfs_rqs exceeding runtime */
3621         check_cfs_rq_runtime(cfs_rq);
3622 
3623         if (schedstat_enabled()) {
3624                 check_spread(cfs_rq, prev);
3625                 if (prev->on_rq)
3626                         update_stats_wait_start(cfs_rq, prev);
3627         }
3628 
3629         if (prev->on_rq) {
3630                 /* Put 'current' back into the tree. */
3631                 __enqueue_entity(cfs_rq, prev);
3632                 /* in !on_rq case, update occurred at dequeue */
3633                 update_load_avg(prev, 0);
3634         }
3635         cfs_rq->curr = NULL;
3636 }
3637 
3638 static void
3639 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
3640 {
3641         /*
3642          * Update run-time statistics of the 'current'.
3643          */
3644         update_curr(cfs_rq);
3645 
3646         /*
3647          * Ensure that runnable average is periodically updated.
3648          */
3649         update_load_avg(curr, 1);
3650         update_cfs_shares(cfs_rq);
3651 
3652 #ifdef CONFIG_SCHED_HRTICK
3653         /*
3654          * queued ticks are scheduled to match the slice, so don't bother
3655          * validating it and just reschedule.
3656          */
3657         if (queued) {
3658                 resched_curr(rq_of(cfs_rq));
3659                 return;
3660         }
3661         /*
3662          * don't let the period tick interfere with the hrtick preemption
3663          */
3664         if (!sched_feat(DOUBLE_TICK) &&
3665                         hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
3666                 return;
3667 #endif
3668 
3669         if (cfs_rq->nr_running > 1)
3670                 check_preempt_tick(cfs_rq, curr);
3671 }
3672 
3673 
3674 /**************************************************
3675  * CFS bandwidth control machinery
3676  */
3677 
3678 #ifdef CONFIG_CFS_BANDWIDTH
3679 
3680 #ifdef HAVE_JUMP_LABEL
3681 static struct static_key __cfs_bandwidth_used;
3682 
3683 static inline bool cfs_bandwidth_used(void)
3684 {
3685         return static_key_false(&__cfs_bandwidth_used);
3686 }
3687 
3688 void cfs_bandwidth_usage_inc(void)
3689 {
3690         static_key_slow_inc(&__cfs_bandwidth_used);
3691 }
3692 
3693 void cfs_bandwidth_usage_dec(void)
3694 {
3695         static_key_slow_dec(&__cfs_bandwidth_used);
3696 }
3697 #else /* HAVE_JUMP_LABEL */
3698 static bool cfs_bandwidth_used(void)
3699 {
3700         return true;
3701 }
3702 
3703 void cfs_bandwidth_usage_inc(void) {}
3704 void cfs_bandwidth_usage_dec(void) {}
3705 #endif /* HAVE_JUMP_LABEL */
3706 
3707 /*
3708  * default period for cfs group bandwidth.
3709  * default: 0.1s, units: nanoseconds
3710  */
3711 static inline u64 default_cfs_period(void)
3712 {
3713         return 100000000ULL;
3714 }
3715 
3716 static inline u64 sched_cfs_bandwidth_slice(void)
3717 {
3718         return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
3719 }
3720 
3721 /*
3722  * Replenish runtime according to assigned quota and update expiration time.
3723  * We use sched_clock_cpu directly instead of rq->clock to avoid adding
3724  * additional synchronization around rq->lock.
3725  *
3726  * requires cfs_b->lock
3727  */
3728 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
3729 {
3730         u64 now;
3731 
3732         if (cfs_b->quota == RUNTIME_INF)
3733                 return;
3734 
3735         now = sched_clock_cpu(smp_processor_id());
3736         cfs_b->runtime = cfs_b->quota;
3737         cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
3738 }
3739 
3740 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
3741 {
3742         return &tg->cfs_bandwidth;
3743 }
3744 
3745 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
3746 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
3747 {
3748         if (unlikely(cfs_rq->throttle_count))
3749                 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
3750 
3751         return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
3752 }
3753 
3754 /* returns 0 on failure to allocate runtime */
3755 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3756 {
3757         struct task_group *tg = cfs_rq->tg;
3758         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
3759         u64 amount = 0, min_amount, expires;
3760 
3761         /* note: this is a positive sum as runtime_remaining <= 0 */
3762         min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
3763 
3764         raw_spin_lock(&cfs_b->lock);
3765         if (cfs_b->quota == RUNTIME_INF)
3766                 amount = min_amount;
3767         else {
3768                 start_cfs_bandwidth(cfs_b);
3769 
3770                 if (cfs_b->runtime > 0) {
3771                         amount = min(cfs_b->runtime, min_amount);
3772                         cfs_b->runtime -= amount;
3773                         cfs_b->idle = 0;
3774                 }
3775         }
3776         expires = cfs_b->runtime_expires;
3777         raw_spin_unlock(&cfs_b->lock);
3778 
3779         cfs_rq->runtime_remaining += amount;
3780         /*
3781          * we may have advanced our local expiration to account for allowed
3782          * spread between our sched_clock and the one on which runtime was
3783          * issued.
3784          */
3785         if ((s64)(expires - cfs_rq->runtime_expires) > 0)
3786                 cfs_rq->runtime_expires = expires;
3787 
3788         return cfs_rq->runtime_remaining > 0;
3789 }
3790 
3791 /*
3792  * Note: This depends on the synchronization provided by sched_clock and the
3793  * fact that rq->clock snapshots this value.
3794  */
3795 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
3796 {
3797         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3798 
3799         /* if the deadline is ahead of our clock, nothing to do */
3800         if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
3801                 return;
3802 
3803         if (cfs_rq->runtime_remaining < 0)
3804                 return;
3805 
3806         /*
3807          * If the local deadline has passed we have to consider the
3808          * possibility that our sched_clock is 'fast' and the global deadline
3809          * has not truly expired.
3810          *
3811          * Fortunately we can check determine whether this the case by checking
3812          * whether the global deadline has advanced. It is valid to compare
3813          * cfs_b->runtime_expires without any locks since we only care about
3814          * exact equality, so a partial write will still work.
3815          */
3816 
3817         if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
3818                 /* extend local deadline, drift is bounded above by 2 ticks */
3819                 cfs_rq->runtime_expires += TICK_NSEC;
3820         } else {
3821                 /* global deadline is ahead, expiration has passed */
3822                 cfs_rq->runtime_remaining = 0;
3823         }
3824 }
3825 
3826 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3827 {
3828         /* dock delta_exec before expiring quota (as it could span periods) */
3829         cfs_rq->runtime_remaining -= delta_exec;
3830         expire_cfs_rq_runtime(cfs_rq);
3831 
3832         if (likely(cfs_rq->runtime_remaining > 0))
3833                 return;
3834 
3835         /*
3836          * if we're unable to extend our runtime we resched so that the active
3837          * hierarchy can be throttled
3838          */
3839         if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
3840                 resched_curr(rq_of(cfs_rq));
3841 }
3842 
3843 static __always_inline
3844 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
3845 {
3846         if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
3847                 return;
3848 
3849         __account_cfs_rq_runtime(cfs_rq, delta_exec);
3850 }
3851 
3852 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
3853 {
3854         return cfs_bandwidth_used() && cfs_rq->throttled;
3855 }
3856 
3857 /* check whether cfs_rq, or any parent, is throttled */
3858 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
3859 {
3860         return cfs_bandwidth_used() && cfs_rq->throttle_count;
3861 }
3862 
3863 /*
3864  * Ensure that neither of the group entities corresponding to src_cpu or
3865  * dest_cpu are members of a throttled hierarchy when performing group
3866  * load-balance operations.
3867  */
3868 static inline int throttled_lb_pair(struct task_group *tg,
3869                                     int src_cpu, int dest_cpu)
3870 {
3871         struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
3872 
3873         src_cfs_rq = tg->cfs_rq[src_cpu];
3874         dest_cfs_rq = tg->cfs_rq[dest_cpu];
3875 
3876         return throttled_hierarchy(src_cfs_rq) ||
3877                throttled_hierarchy(dest_cfs_rq);
3878 }
3879 
3880 /* updated child weight may affect parent so we have to do this bottom up */
3881 static int tg_unthrottle_up(struct task_group *tg, void *data)
3882 {
3883         struct rq *rq = data;
3884         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3885 
3886         cfs_rq->throttle_count--;
3887         if (!cfs_rq->throttle_count) {
3888                 /* adjust cfs_rq_clock_task() */
3889                 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
3890                                              cfs_rq->throttled_clock_task;
3891         }
3892 
3893         return 0;
3894 }
3895 
3896 static int tg_throttle_down(struct task_group *tg, void *data)
3897 {
3898         struct rq *rq = data;
3899         struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
3900 
3901         /* group is entering throttled state, stop time */
3902         if (!cfs_rq->throttle_count)
3903                 cfs_rq->throttled_clock_task = rq_clock_task(rq);
3904         cfs_rq->throttle_count++;
3905 
3906         return 0;
3907 }
3908 
3909 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
3910 {
3911         struct rq *rq = rq_of(cfs_rq);
3912         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3913         struct sched_entity *se;
3914         long task_delta, dequeue = 1;
3915         bool empty;
3916 
3917         se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
3918 
3919         /* freeze hierarchy runnable averages while throttled */
3920         rcu_read_lock();
3921         walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
3922         rcu_read_unlock();
3923 
3924         task_delta = cfs_rq->h_nr_running;
3925         for_each_sched_entity(se) {
3926                 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
3927                 /* throttled entity or throttle-on-deactivate */
3928                 if (!se->on_rq)
3929                         break;
3930 
3931                 if (dequeue)
3932                         dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
3933                 qcfs_rq->h_nr_running -= task_delta;
3934 
3935                 if (qcfs_rq->load.weight)
3936                         dequeue = 0;
3937         }
3938 
3939         if (!se)
3940                 sub_nr_running(rq, task_delta);
3941 
3942         cfs_rq->throttled = 1;
3943         cfs_rq->throttled_clock = rq_clock(rq);
3944         raw_spin_lock(&cfs_b->lock);
3945         empty = list_empty(&cfs_b->throttled_cfs_rq);
3946 
3947         /*
3948          * Add to the _head_ of the list, so that an already-started
3949          * distribute_cfs_runtime will not see us
3950          */
3951         list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
3952 
3953         /*
3954          * If we're the first throttled task, make sure the bandwidth
3955          * timer is running.
3956          */
3957         if (empty)
3958                 start_cfs_bandwidth(cfs_b);
3959 
3960         raw_spin_unlock(&cfs_b->lock);
3961 }
3962 
3963 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
3964 {
3965         struct rq *rq = rq_of(cfs_rq);
3966         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
3967         struct sched_entity *se;
3968         int enqueue = 1;
3969         long task_delta;
3970 
3971         se = cfs_rq->tg->se[cpu_of(rq)];
3972 
3973         cfs_rq->throttled = 0;
3974 
3975         update_rq_clock(rq);
3976 
3977         raw_spin_lock(&cfs_b->lock);
3978         cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
3979         list_del_rcu(&cfs_rq->throttled_list);
3980         raw_spin_unlock(&cfs_b->lock);
3981 
3982         /* update hierarchical throttle state */
3983         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
3984 
3985         if (!cfs_rq->load.weight)
3986                 return;
3987 
3988         task_delta = cfs_rq->h_nr_running;
3989         for_each_sched_entity(se) {
3990                 if (se->on_rq)
3991                         enqueue = 0;
3992 
3993                 cfs_rq = cfs_rq_of(se);
3994                 if (enqueue)
3995                         enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
3996                 cfs_rq->h_nr_running += task_delta;
3997 
3998                 if (cfs_rq_throttled(cfs_rq))
3999                         break;
4000         }
4001 
4002         if (!se)
4003                 add_nr_running(rq, task_delta);
4004 
4005         /* determine whether we need to wake up potentially idle cpu */
4006         if (rq->curr == rq->idle && rq->cfs.nr_running)
4007                 resched_curr(rq);
4008 }
4009 
4010 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4011                 u64 remaining, u64 expires)
4012 {
4013         struct cfs_rq *cfs_rq;
4014         u64 runtime;
4015         u64 starting_runtime = remaining;
4016 
4017         rcu_read_lock();
4018         list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4019                                 throttled_list) {
4020                 struct rq *rq = rq_of(cfs_rq);
4021 
4022                 raw_spin_lock(&rq->lock);
4023                 if (!cfs_rq_throttled(cfs_rq))
4024                         goto next;
4025 
4026                 runtime = -cfs_rq->runtime_remaining + 1;
4027                 if (runtime > remaining)
4028                         runtime = remaining;
4029                 remaining -= runtime;
4030 
4031                 cfs_rq->runtime_remaining += runtime;
4032                 cfs_rq->runtime_expires = expires;
4033 
4034                 /* we check whether we're throttled above */
4035                 if (cfs_rq->runtime_remaining > 0)
4036                         unthrottle_cfs_rq(cfs_rq);
4037 
4038 next:
4039                 raw_spin_unlock(&rq->lock);
4040 
4041                 if (!remaining)
4042                         break;
4043         }
4044         rcu_read_unlock();
4045 
4046         return starting_runtime - remaining;
4047 }
4048 
4049 /*
4050  * Responsible for refilling a task_group's bandwidth and unthrottling its
4051  * cfs_rqs as appropriate. If there has been no activity within the last
4052  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4053  * used to track this state.
4054  */
4055 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4056 {
4057         u64 runtime, runtime_expires;
4058         int throttled;
4059 
4060         /* no need to continue the timer with no bandwidth constraint */
4061         if (cfs_b->quota == RUNTIME_INF)
4062                 goto out_deactivate;
4063 
4064         throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4065         cfs_b->nr_periods += overrun;
4066 
4067         /*
4068          * idle depends on !throttled (for the case of a large deficit), and if
4069          * we're going inactive then everything else can be deferred
4070          */
4071         if (cfs_b->idle && !throttled)
4072                 goto out_deactivate;
4073 
4074         __refill_cfs_bandwidth_runtime(cfs_b);
4075 
4076         if (!throttled) {
4077                 /* mark as potentially idle for the upcoming period */
4078                 cfs_b->idle = 1;
4079                 return 0;
4080         }
4081 
4082         /* account preceding periods in which throttling occurred */
4083         cfs_b->nr_throttled += overrun;
4084 
4085         runtime_expires = cfs_b->runtime_expires;
4086 
4087         /*
4088          * This check is repeated as we are holding onto the new bandwidth while
4089          * we unthrottle. This can potentially race with an unthrottled group
4090          * trying to acquire new bandwidth from the global pool. This can result
4091          * in us over-using our runtime if it is all used during this loop, but
4092          * only by limited amounts in that extreme case.
4093          */
4094         while (throttled && cfs_b->runtime > 0) {
4095                 runtime = cfs_b->runtime;
4096                 raw_spin_unlock(&cfs_b->lock);
4097                 /* we can't nest cfs_b->lock while distributing bandwidth */
4098                 runtime = distribute_cfs_runtime(cfs_b, runtime,
4099                                                  runtime_expires);
4100                 raw_spin_lock(&cfs_b->lock);
4101 
4102                 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4103 
4104                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4105         }
4106 
4107         /*
4108          * While we are ensured activity in the period following an
4109          * unthrottle, this also covers the case in which the new bandwidth is
4110          * insufficient to cover the existing bandwidth deficit.  (Forcing the
4111          * timer to remain active while there are any throttled entities.)
4112          */
4113         cfs_b->idle = 0;
4114 
4115         return 0;
4116 
4117 out_deactivate:
4118         return 1;
4119 }
4120 
4121 /* a cfs_rq won't donate quota below this amount */
4122 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4123 /* minimum remaining period time to redistribute slack quota */
4124 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4125 /* how long we wait to gather additional slack before distributing */
4126 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4127 
4128 /*
4129  * Are we near the end of the current quota period?
4130  *
4131  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4132  * hrtimer base being cleared by hrtimer_start. In the case of
4133  * migrate_hrtimers, base is never cleared, so we are fine.
4134  */
4135 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4136 {
4137         struct hrtimer *refresh_timer = &cfs_b->period_timer;
4138         u64 remaining;
4139 
4140         /* if the call-back is running a quota refresh is already occurring */
4141         if (hrtimer_callback_running(refresh_timer))
4142                 return 1;
4143 
4144         /* is a quota refresh about to occur? */
4145         remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4146         if (remaining < min_expire)
4147                 return 1;
4148 
4149         return 0;
4150 }
4151 
4152 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4153 {
4154         u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4155 
4156         /* if there's a quota refresh soon don't bother with slack */
4157         if (runtime_refresh_within(cfs_b, min_left))
4158                 return;
4159 
4160         hrtimer_start(&cfs_b->slack_timer,
4161                         ns_to_ktime(cfs_bandwidth_slack_period),
4162                         HRTIMER_MODE_REL);
4163 }
4164 
4165 /* we know any runtime found here is valid as update_curr() precedes return */
4166 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4167 {
4168         struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4169         s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4170 
4171         if (slack_runtime <= 0)
4172                 return;
4173 
4174         raw_spin_lock(&cfs_b->lock);
4175         if (cfs_b->quota != RUNTIME_INF &&
4176             cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4177                 cfs_b->runtime += slack_runtime;
4178 
4179                 /* we are under rq->lock, defer unthrottling using a timer */
4180                 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4181                     !list_empty(&cfs_b->throttled_cfs_rq))
4182                         start_cfs_slack_bandwidth(cfs_b);
4183         }
4184         raw_spin_unlock(&cfs_b->lock);
4185 
4186         /* even if it's not valid for return we don't want to try again */
4187         cfs_rq->runtime_remaining -= slack_runtime;
4188 }
4189 
4190 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4191 {
4192         if (!cfs_bandwidth_used())
4193                 return;
4194 
4195         if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4196                 return;
4197 
4198         __return_cfs_rq_runtime(cfs_rq);
4199 }
4200 
4201 /*
4202  * This is done with a timer (instead of inline with bandwidth return) since
4203  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4204  */
4205 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4206 {
4207         u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4208         u64 expires;
4209 
4210         /* confirm we're still not at a refresh boundary */
4211         raw_spin_lock(&cfs_b->lock);
4212         if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4213                 raw_spin_unlock(&cfs_b->lock);
4214                 return;
4215         }
4216 
4217         if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4218                 runtime = cfs_b->runtime;
4219 
4220         expires = cfs_b->runtime_expires;
4221         raw_spin_unlock(&cfs_b->lock);
4222 
4223         if (!runtime)
4224                 return;
4225 
4226         runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4227 
4228         raw_spin_lock(&cfs_b->lock);
4229         if (expires == cfs_b->runtime_expires)
4230                 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4231         raw_spin_unlock(&cfs_b->lock);
4232 }
4233 
4234 /*
4235  * When a group wakes up we want to make sure that its quota is not already
4236  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4237  * runtime as update_curr() throttling can not not trigger until it's on-rq.
4238  */
4239 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4240 {
4241         if (!cfs_bandwidth_used())
4242                 return;
4243 
4244         /* an active group must be handled by the update_curr()->put() path */
4245         if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4246                 return;
4247 
4248         /* ensure the group is not already throttled */
4249         if (cfs_rq_throttled(cfs_rq))
4250                 return;
4251 
4252         /* update runtime allocation */
4253         account_cfs_rq_runtime(cfs_rq, 0);
4254         if (cfs_rq->runtime_remaining <= 0)
4255                 throttle_cfs_rq(cfs_rq);
4256 }
4257 
4258 static void sync_throttle(struct task_group *tg, int cpu)
4259 {
4260         struct cfs_rq *pcfs_rq, *cfs_rq;
4261 
4262         if (!cfs_bandwidth_used())
4263                 return;
4264 
4265         if (!tg->parent)
4266                 return;
4267 
4268         cfs_rq = tg->cfs_rq[cpu];
4269         pcfs_rq = tg->parent->cfs_rq[cpu];
4270 
4271         cfs_rq->throttle_count = pcfs_rq->throttle_count;
4272         cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4273 }
4274 
4275 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4276 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4277 {
4278         if (!cfs_bandwidth_used())
4279                 return false;
4280 
4281         if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4282                 return false;
4283 
4284         /*
4285          * it's possible for a throttled entity to be forced into a running
4286          * state (e.g. set_curr_task), in this case we're finished.
4287          */
4288         if (cfs_rq_throttled(cfs_rq))
4289                 return true;
4290 
4291         throttle_cfs_rq(cfs_rq);
4292         return true;
4293 }
4294 
4295 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4296 {
4297         struct cfs_bandwidth *cfs_b =
4298                 container_of(timer, struct cfs_bandwidth, slack_timer);
4299 
4300         do_sched_cfs_slack_timer(cfs_b);
4301 
4302         return HRTIMER_NORESTART;
4303 }
4304 
4305 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4306 {
4307         struct cfs_bandwidth *cfs_b =
4308                 container_of(timer, struct cfs_bandwidth, period_timer);
4309         int overrun;
4310         int idle = 0;
4311 
4312         raw_spin_lock(&cfs_b->lock);
4313         for (;;) {
4314                 overrun = hrtimer_forward_now(timer, cfs_b->period);
4315                 if (!overrun)
4316                         break;
4317 
4318                 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4319         }
4320         if (idle)
4321                 cfs_b->period_active = 0;
4322         raw_spin_unlock(&cfs_b->lock);
4323 
4324         return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4325 }
4326 
4327 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4328 {
4329         raw_spin_lock_init(&cfs_b->lock);
4330         cfs_b->runtime = 0;
4331         cfs_b->quota = RUNTIME_INF;
4332         cfs_b->period = ns_to_ktime(default_cfs_period());
4333 
4334         INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4335         hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4336         cfs_b->period_timer.function = sched_cfs_period_timer;
4337         hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4338         cfs_b->slack_timer.function = sched_cfs_slack_timer;
4339 }
4340 
4341 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4342 {
4343         cfs_rq->runtime_enabled = 0;
4344         INIT_LIST_HEAD(&cfs_rq->throttled_list);
4345 }
4346 
4347 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4348 {
4349         lockdep_assert_held(&cfs_b->lock);
4350 
4351         if (!cfs_b->period_active) {
4352                 cfs_b->period_active = 1;
4353                 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4354                 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4355         }
4356 }
4357 
4358 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4359 {
4360         /* init_cfs_bandwidth() was not called */
4361         if (!cfs_b->throttled_cfs_rq.next)
4362                 return;
4363 
4364         hrtimer_cancel(&cfs_b->period_timer);
4365         hrtimer_cancel(&cfs_b->slack_timer);
4366 }
4367 
4368 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4369 {
4370         struct cfs_rq *cfs_rq;
4371 
4372         for_each_leaf_cfs_rq(rq, cfs_rq) {
4373                 struct cfs_bandwidth *cfs_b = &cfs_rq->tg->cfs_bandwidth;
4374 
4375                 raw_spin_lock(&cfs_b->lock);
4376                 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4377                 raw_spin_unlock(&cfs_b->lock);
4378         }
4379 }
4380 
4381 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4382 {
4383         struct cfs_rq *cfs_rq;
4384 
4385         for_each_leaf_cfs_rq(rq, cfs_rq) {
4386                 if (!cfs_rq->runtime_enabled)
4387                         continue;
4388 
4389                 /*
4390                  * clock_task is not advancing so we just need to make sure
4391                  * there's some valid quota amount
4392                  */
4393                 cfs_rq->runtime_remaining = 1;
4394                 /*
4395                  * Offline rq is schedulable till cpu is completely disabled
4396                  * in take_cpu_down(), so we prevent new cfs throttling here.
4397                  */
4398                 cfs_rq->runtime_enabled = 0;
4399 
4400                 if (cfs_rq_throttled(cfs_rq))
4401                         unthrottle_cfs_rq(cfs_rq);
4402         }
4403 }
4404 
4405 #else /* CONFIG_CFS_BANDWIDTH */
4406 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4407 {
4408         return rq_clock_task(rq_of(cfs_rq));
4409 }
4410 
4411 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4412 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4413 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4414 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4415 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4416 
4417 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4418 {
4419         return 0;
4420 }
4421 
4422 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4423 {
4424         return 0;
4425 }
4426 
4427 static inline int throttled_lb_pair(struct task_group *tg,
4428                                     int src_cpu, int dest_cpu)
4429 {
4430         return 0;
4431 }
4432 
4433 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4434 
4435 #ifdef CONFIG_FAIR_GROUP_SCHED
4436 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4437 #endif
4438 
4439 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4440 {
4441         return NULL;
4442 }
4443 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4444 static inline void update_runtime_enabled(struct rq *rq) {}
4445 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4446 
4447 #endif /* CONFIG_CFS_BANDWIDTH */
4448 
4449 /**************************************************
4450  * CFS operations on tasks:
4451  */
4452 
4453 #ifdef CONFIG_SCHED_HRTICK
4454 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4455 {
4456         struct sched_entity *se = &p->se;
4457         struct cfs_rq *cfs_rq = cfs_rq_of(se);
4458 
4459         WARN_ON(task_rq(p) != rq);
4460 
4461         if (cfs_rq->nr_running > 1) {
4462                 u64 slice = sched_slice(cfs_rq, se);
4463                 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4464                 s64 delta = slice - ran;
4465 
4466                 if (delta < 0) {
4467                         if (rq->curr == p)
4468                                 resched_curr(rq);
4469                         return;
4470                 }
4471                 hrtick_start(rq, delta);
4472         }
4473 }
4474 
4475 /*
4476  * called from enqueue/dequeue and updates the hrtick when the
4477  * current task is from our class and nr_running is low enough
4478  * to matter.
4479  */
4480 static void hrtick_update(struct rq *rq)
4481 {
4482         struct task_struct *curr = rq->curr;
4483 
4484         if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4485                 return;
4486 
4487         if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4488                 hrtick_start_fair(rq, curr);
4489 }
4490 #else /* !CONFIG_SCHED_HRTICK */
4491 static inline void
4492 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4493 {
4494 }
4495 
4496 static inline void hrtick_update(struct rq *rq)
4497 {
4498 }
4499 #endif
4500 
4501 /*
4502  * The enqueue_task method is called before nr_running is
4503  * increased. Here we update the fair scheduling stats and
4504  * then put the task into the rbtree:
4505  */
4506 static void
4507 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4508 {
4509         struct cfs_rq *cfs_rq;
4510         struct sched_entity *se = &p->se;
4511 
4512         for_each_sched_entity(se) {
4513                 if (se->on_rq)
4514                         break;
4515                 cfs_rq = cfs_rq_of(se);
4516                 enqueue_entity(cfs_rq, se, flags);
4517 
4518                 /*
4519                  * end evaluation on encountering a throttled cfs_rq
4520                  *
4521                  * note: in the case of encountering a throttled cfs_rq we will
4522                  * post the final h_nr_running increment below.
4523                  */
4524                 if (cfs_rq_throttled(cfs_rq))
4525                         break;
4526                 cfs_rq->h_nr_running++;
4527 
4528                 flags = ENQUEUE_WAKEUP;
4529         }
4530 
4531         for_each_sched_entity(se) {
4532                 cfs_rq = cfs_rq_of(se);
4533                 cfs_rq->h_nr_running++;
4534 
4535                 if (cfs_rq_throttled(cfs_rq))
4536                         break;
4537 
4538                 update_load_avg(se, 1);
4539                 update_cfs_shares(cfs_rq);
4540         }
4541 
4542         if (!se)
4543                 add_nr_running(rq, 1);
4544 
4545         hrtick_update(rq);
4546 }
4547 
4548 static void set_next_buddy(struct sched_entity *se);
4549 
4550 /*
4551  * The dequeue_task method is called before nr_running is
4552  * decreased. We remove the task from the rbtree and
4553  * update the fair scheduling stats:
4554  */
4555 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
4556 {
4557         struct cfs_rq *cfs_rq;
4558         struct sched_entity *se = &p->se;
4559         int task_sleep = flags & DEQUEUE_SLEEP;
4560 
4561         for_each_sched_entity(se) {
4562                 cfs_rq = cfs_rq_of(se);
4563                 dequeue_entity(cfs_rq, se, flags);
4564 
4565                 /*
4566                  * end evaluation on encountering a throttled cfs_rq
4567                  *
4568                  * note: in the case of encountering a throttled cfs_rq we will
4569                  * post the final h_nr_running decrement below.
4570                 */
4571                 if (cfs_rq_throttled(cfs_rq))
4572                         break;
4573                 cfs_rq->h_nr_running--;
4574 
4575                 /* Don't dequeue parent if it has other entities besides us */
4576                 if (cfs_rq->load.weight) {
4577                         /* Avoid re-evaluating load for this entity: */
4578                         se = parent_entity(se);
4579                         /*
4580                          * Bias pick_next to pick a task from this cfs_rq, as
4581                          * p is sleeping when it is within its sched_slice.
4582                          */
4583                         if (task_sleep && se && !throttled_hierarchy(cfs_rq))
4584                                 set_next_buddy(se);
4585                         break;
4586                 }
4587                 flags |= DEQUEUE_SLEEP;
4588         }
4589 
4590         for_each_sched_entity(se) {
4591                 cfs_rq = cfs_rq_of(se);
4592                 cfs_rq->h_nr_running--;
4593 
4594                 if (cfs_rq_throttled(cfs_rq))
4595                         break;
4596 
4597                 update_load_avg(se, 1);
4598                 update_cfs_shares(cfs_rq);
4599         }
4600 
4601         if (!se)
4602                 sub_nr_running(rq, 1);
4603 
4604         hrtick_update(rq);
4605 }
4606 
4607 #ifdef CONFIG_SMP
4608 #ifdef CONFIG_NO_HZ_COMMON
4609 /*
4610  * per rq 'load' arrray crap; XXX kill this.
4611  */
4612 
4613 /*
4614  * The exact cpuload calculated at every tick would be:
4615  *
4616  *   load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
4617  *
4618  * If a cpu misses updates for n ticks (as it was idle) and update gets
4619  * called on the n+1-th tick when cpu may be busy, then we have:
4620  *
4621  *   load_n   = (1 - 1/2^i)^n * load_0
4622  *   load_n+1 = (1 - 1/2^i)   * load_n + (1/2^i) * cur_load
4623  *
4624  * decay_load_missed() below does efficient calculation of
4625  *
4626  *   load' = (1 - 1/2^i)^n * load
4627  *
4628  * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
4629  * This allows us to precompute the above in said factors, thereby allowing the
4630  * reduction of an arbitrary n in O(log_2 n) steps. (See also
4631  * fixed_power_int())
4632  *
4633  * The calculation is approximated on a 128 point scale.
4634  */
4635 #define DEGRADE_SHIFT           7
4636 
4637 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
4638 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
4639         {   0,   0,  0,  0,  0,  0, 0, 0 },
4640         {  64,  32,  8,  0,  0,  0, 0, 0 },
4641         {  96,  72, 40, 12,  1,  0, 0, 0 },
4642         { 112,  98, 75, 43, 15,  1, 0, 0 },
4643         { 120, 112, 98, 76, 45, 16, 2, 0 }
4644 };
4645 
4646 /*
4647  * Update cpu_load for any missed ticks, due to tickless idle. The backlog
4648  * would be when CPU is idle and so we just decay the old load without
4649  * adding any new load.
4650  */
4651 static unsigned long
4652 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
4653 {
4654         int j = 0;
4655 
4656         if (!missed_updates)
4657                 return load;
4658 
4659         if (missed_updates >= degrade_zero_ticks[idx])
4660                 return 0;
4661 
4662         if (idx == 1)
4663                 return load >> missed_updates;
4664 
4665         while (missed_updates) {
4666                 if (missed_updates % 2)
4667                         load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
4668 
4669                 missed_updates >>= 1;
4670                 j++;
4671         }
4672         return load;
4673 }
4674 #endif /* CONFIG_NO_HZ_COMMON */
4675 
4676 /**
4677  * __cpu_load_update - update the rq->cpu_load[] statistics
4678  * @this_rq: The rq to update statistics for
4679  * @this_load: The current load
4680  * @pending_updates: The number of missed updates
4681  *
4682  * Update rq->cpu_load[] statistics. This function is usually called every
4683  * scheduler tick (TICK_NSEC).
4684  *
4685  * This function computes a decaying average:
4686  *
4687  *   load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
4688  *
4689  * Because of NOHZ it might not get called on every tick which gives need for
4690  * the @pending_updates argument.
4691  *
4692  *   load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
4693  *             = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
4694  *             = A * (A * load[i]_n-2 + B) + B
4695  *             = A * (A * (A * load[i]_n-3 + B) + B) + B
4696  *             = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
4697  *             = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
4698  *             = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
4699  *             = (1 - 1/2^i)^n * (load[i]_0 - load) + load
4700  *
4701  * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
4702  * any change in load would have resulted in the tick being turned back on.
4703  *
4704  * For regular NOHZ, this reduces to:
4705  *
4706  *   load[i]_n = (1 - 1/2^i)^n * load[i]_0
4707  *
4708  * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
4709  * term.
4710  */
4711 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
4712                             unsigned long pending_updates)
4713 {
4714         unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
4715         int i, scale;
4716 
4717         this_rq->nr_load_updates++;
4718 
4719         /* Update our load: */
4720         this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
4721         for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
4722                 unsigned long old_load, new_load;
4723 
4724                 /* scale is effectively 1 << i now, and >> i divides by scale */
4725 
4726                 old_load = this_rq->cpu_load[i];
4727 #ifdef CONFIG_NO_HZ_COMMON
4728                 old_load = decay_load_missed(old_load, pending_updates - 1, i);
4729                 if (tickless_load) {
4730                         old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
4731                         /*
4732                          * old_load can never be a negative value because a
4733                          * decayed tickless_load cannot be greater than the
4734                          * original tickless_load.
4735                          */
4736                         old_load += tickless_load;
4737                 }
4738 #endif
4739                 new_load = this_load;
4740                 /*
4741                  * Round up the averaging division if load is increasing. This
4742                  * prevents us from getting stuck on 9 if the load is 10, for
4743                  * example.
4744                  */
4745                 if (new_load > old_load)
4746                         new_load += scale - 1;
4747 
4748                 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
4749         }
4750 
4751         sched_avg_update(this_rq);
4752 }
4753 
4754 /* Used instead of source_load when we know the type == 0 */
4755 static unsigned long weighted_cpuload(const int cpu)
4756 {
4757         return cfs_rq_runnable_load_avg(&cpu_rq(cpu)->cfs);
4758 }
4759 
4760 #ifdef CONFIG_NO_HZ_COMMON
4761 /*
4762  * There is no sane way to deal with nohz on smp when using jiffies because the
4763  * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
4764  * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
4765  *
4766  * Therefore we need to avoid the delta approach from the regular tick when
4767  * possible since that would seriously skew the load calculation. This is why we
4768  * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
4769  * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
4770  * loop exit, nohz_idle_balance, nohz full exit...)
4771  *
4772  * This means we might still be one tick off for nohz periods.
4773  */
4774 
4775 static void cpu_load_update_nohz(struct rq *this_rq,
4776                                  unsigned long curr_jiffies,
4777                                  unsigned long load)
4778 {
4779         unsigned long pending_updates;
4780 
4781         pending_updates = curr_jiffies - this_rq->last_load_update_tick;
4782         if (pending_updates) {
4783                 this_rq->last_load_update_tick = curr_jiffies;
4784                 /*
4785                  * In the regular NOHZ case, we were idle, this means load 0.
4786                  * In the NOHZ_FULL case, we were non-idle, we should consider
4787                  * its weighted load.
4788                  */
4789                 cpu_load_update(this_rq, load, pending_updates);
4790         }
4791 }
4792 
4793 /*
4794  * Called from nohz_idle_balance() to update the load ratings before doing the
4795  * idle balance.
4796  */
4797 static void cpu_load_update_idle(struct rq *this_rq)
4798 {
4799         /*
4800          * bail if there's load or we're actually up-to-date.
4801          */
4802         if (weighted_cpuload(cpu_of(this_rq)))
4803                 return;
4804 
4805         cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
4806 }
4807 
4808 /*
4809  * Record CPU load on nohz entry so we know the tickless load to account
4810  * on nohz exit. cpu_load[0] happens then to be updated more frequently
4811  * than other cpu_load[idx] but it should be fine as cpu_load readers
4812  * shouldn't rely into synchronized cpu_load[*] updates.
4813  */
4814 void cpu_load_update_nohz_start(void)
4815 {
4816         struct rq *this_rq = this_rq();
4817 
4818         /*
4819          * This is all lockless but should be fine. If weighted_cpuload changes
4820          * concurrently we'll exit nohz. And cpu_load write can race with
4821          * cpu_load_update_idle() but both updater would be writing the same.
4822          */
4823         this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
4824 }
4825 
4826 /*
4827  * Account the tickless load in the end of a nohz frame.
4828  */
4829 void cpu_load_update_nohz_stop(void)
4830 {
4831         unsigned long curr_jiffies = READ_ONCE(jiffies);
4832         struct rq *this_rq = this_rq();
4833         unsigned long load;
4834 
4835         if (curr_jiffies == this_rq->last_load_update_tick)
4836                 return;
4837 
4838         load = weighted_cpuload(cpu_of(this_rq));
4839         raw_spin_lock(&this_rq->lock);
4840         update_rq_clock(this_rq);
4841         cpu_load_update_nohz(this_rq, curr_jiffies, load);
4842         raw_spin_unlock(&this_rq->lock);
4843 }
4844 #else /* !CONFIG_NO_HZ_COMMON */
4845 static inline void cpu_load_update_nohz(struct rq *this_rq,
4846                                         unsigned long curr_jiffies,
4847                                         unsigned long load) { }
4848 #endif /* CONFIG_NO_HZ_COMMON */
4849 
4850 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
4851 {
4852 #ifdef CONFIG_NO_HZ_COMMON
4853         /* See the mess around cpu_load_update_nohz(). */
4854         this_rq->last_load_update_tick = READ_ONCE(jiffies);
4855 #endif
4856         cpu_load_update(this_rq, load, 1);
4857 }
4858 
4859 /*
4860  * Called from scheduler_tick()
4861  */
4862 void cpu_load_update_active(struct rq *this_rq)
4863 {
4864         unsigned long load = weighted_cpuload(cpu_of(this_rq));
4865 
4866         if (tick_nohz_tick_stopped())
4867                 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
4868         else
4869                 cpu_load_update_periodic(this_rq, load);
4870 }
4871 
4872 /*
4873  * Return a low guess at the load of a migration-source cpu weighted
4874  * according to the scheduling class and "nice" value.
4875  *
4876  * We want to under-estimate the load of migration sources, to
4877  * balance conservatively.
4878  */
4879 static unsigned long source_load(int cpu, int type)
4880 {
4881         struct rq *rq = cpu_rq(cpu);
4882         unsigned long total = weighted_cpuload(cpu);
4883 
4884         if (type == 0 || !sched_feat(LB_BIAS))
4885                 return total;
4886 
4887         return min(rq->cpu_load[type-1], total);
4888 }
4889 
4890 /*
4891  * Return a high guess at the load of a migration-target cpu weighted
4892  * according to the scheduling class and "nice" value.
4893  */
4894 static unsigned long target_load(int cpu, int type)
4895 {
4896         struct rq *rq = cpu_rq(cpu);
4897         unsigned long total = weighted_cpuload(cpu);
4898 
4899         if (type == 0 || !sched_feat(LB_BIAS))
4900                 return total;
4901 
4902         return max(rq->cpu_load[type-1], total);
4903 }
4904 
4905 static unsigned long capacity_of(int cpu)
4906 {
4907         return cpu_rq(cpu)->cpu_capacity;
4908 }
4909 
4910 static unsigned long capacity_orig_of(int cpu)
4911 {
4912         return cpu_rq(cpu)->cpu_capacity_orig;
4913 }
4914 
4915 static unsigned long cpu_avg_load_per_task(int cpu)
4916 {
4917         struct rq *rq = cpu_rq(cpu);
4918         unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
4919         unsigned long load_avg = weighted_cpuload(cpu);
4920 
4921         if (nr_running)
4922                 return load_avg / nr_running;
4923 
4924         return 0;
4925 }
4926 
4927 #ifdef CONFIG_FAIR_GROUP_SCHED
4928 /*
4929  * effective_load() calculates the load change as seen from the root_task_group
4930  *
4931  * Adding load to a group doesn't make a group heavier, but can cause movement
4932  * of group shares between cpus. Assuming the shares were perfectly aligned one
4933  * can calculate the shift in shares.
4934  *
4935  * Calculate the effective load difference if @wl is added (subtracted) to @tg
4936  * on this @cpu and results in a total addition (subtraction) of @wg to the
4937  * total group weight.
4938  *
4939  * Given a runqueue weight distribution (rw_i) we can compute a shares
4940  * distribution (s_i) using:
4941  *
4942  *   s_i = rw_i / \Sum rw_j                                             (1)
4943  *
4944  * Suppose we have 4 CPUs and our @tg is a direct child of the root group and
4945  * has 7 equal weight tasks, distributed as below (rw_i), with the resulting
4946  * shares distribution (s_i):
4947  *
4948  *   rw_i = {   2,   4,   1,   0 }
4949  *   s_i  = { 2/7, 4/7, 1/7,   0 }
4950  *
4951  * As per wake_affine() we're interested in the load of two CPUs (the CPU the
4952  * task used to run on and the CPU the waker is running on), we need to
4953  * compute the effect of waking a task on either CPU and, in case of a sync
4954  * wakeup, compute the effect of the current task going to sleep.
4955  *
4956  * So for a change of @wl to the local @cpu with an overall group weight change
4957  * of @wl we can compute the new shares distribution (s'_i) using:
4958  *
4959  *   s'_i = (rw_i + @wl) / (@wg + \Sum rw_j)                            (2)
4960  *
4961  * Suppose we're interested in CPUs 0 and 1, and want to compute the load
4962  * differences in waking a task to CPU 0. The additional task changes the
4963  * weight and shares distributions like:
4964  *
4965  *   rw'_i = {   3,   4,   1,   0 }
4966  *   s'_i  = { 3/8, 4/8, 1/8,   0 }
4967  *
4968  * We can then compute the difference in effective weight by using:
4969  *
4970  *   dw_i = S * (s'_i - s_i)                                            (3)
4971  *
4972  * Where 'S' is the group weight as seen by its parent.
4973  *
4974  * Therefore the effective change in loads on CPU 0 would be 5/56 (3/8 - 2/7)
4975  * times the weight of the group. The effect on CPU 1 would be -4/56 (4/8 -
4976  * 4/7) times the weight of the group.
4977  */
4978 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
4979 {
4980         struct sched_entity *se = tg->se[cpu];
4981 
4982         if (!tg->parent)        /* the trivial, non-cgroup case */
4983                 return wl;
4984 
4985         for_each_sched_entity(se) {
4986                 struct cfs_rq *cfs_rq = se->my_q;
4987                 long W, w = cfs_rq_load_avg(cfs_rq);
4988 
4989                 tg = cfs_rq->tg;
4990 
4991                 /*
4992                  * W = @wg + \Sum rw_j
4993                  */
4994                 W = wg + atomic_long_read(&tg->load_avg);
4995 
4996                 /* Ensure \Sum rw_j >= rw_i */
4997                 W -= cfs_rq->tg_load_avg_contrib;
4998                 W += w;
4999 
5000                 /*
5001                  * w = rw_i + @wl
5002                  */
5003                 w += wl;
5004 
5005                 /*
5006                  * wl = S * s'_i; see (2)
5007                  */
5008                 if (W > 0 && w < W)
5009                         wl = (w * (long)tg->shares) / W;
5010                 else
5011                         wl = tg->shares;
5012 
5013                 /*
5014                  * Per the above, wl is the new se->load.weight value; since
5015                  * those are clipped to [MIN_SHARES, ...) do so now. See
5016                  * calc_cfs_shares().
5017                  */
5018                 if (wl < MIN_SHARES)
5019                         wl = MIN_SHARES;
5020 
5021                 /*
5022                  * wl = dw_i = S * (s'_i - s_i); see (3)
5023                  */
5024                 wl -= se->avg.load_avg;
5025 
5026                 /*
5027                  * Recursively apply this logic to all parent groups to compute
5028                  * the final effective load change on the root group. Since
5029                  * only the @tg group gets extra weight, all parent groups can
5030                  * only redistribute existing shares. @wl is the shift in shares
5031                  * resulting from this level per the above.
5032                  */
5033                 wg = 0;
5034         }
5035 
5036         return wl;
5037 }
5038 #else
5039 
5040 static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
5041 {
5042         return wl;
5043 }
5044 
5045 #endif
5046 
5047 static void record_wakee(struct task_struct *p)
5048 {
5049         /*
5050          * Only decay a single time; tasks that have less then 1 wakeup per
5051          * jiffy will not have built up many flips.
5052          */
5053         if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5054                 current->wakee_flips >>= 1;
5055                 current->wakee_flip_decay_ts = jiffies;
5056         }
5057 
5058         if (current->last_wakee != p) {
5059                 current->last_wakee = p;
5060                 current->wakee_flips++;
5061         }
5062 }
5063 
5064 /*
5065  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5066  *
5067  * A waker of many should wake a different task than the one last awakened
5068  * at a frequency roughly N times higher than one of its wakees.
5069  *
5070  * In order to determine whether we should let the load spread vs consolidating
5071  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5072  * partner, and a factor of lls_size higher frequency in the other.
5073  *
5074  * With both conditions met, we can be relatively sure that the relationship is
5075  * non-monogamous, with partner count exceeding socket size.
5076  *
5077  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5078  * whatever is irrelevant, spread criteria is apparent partner count exceeds
5079  * socket size.
5080  */
5081 static int wake_wide(struct task_struct *p)
5082 {
5083         unsigned int master = current->wakee_flips;
5084         unsigned int slave = p->wakee_flips;
5085         int factor = this_cpu_read(sd_llc_size);
5086 
5087         if (master < slave)
5088                 swap(master, slave);
5089         if (slave < factor || master < slave * factor)
5090                 return 0;
5091         return 1;
5092 }
5093 
5094 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
5095 {
5096         s64 this_load, load;
5097         s64 this_eff_load, prev_eff_load;
5098         int idx, this_cpu, prev_cpu;
5099         struct task_group *tg;
5100         unsigned long weight;
5101         int balanced;
5102 
5103         idx       = sd->wake_idx;
5104         this_cpu  = smp_processor_id();
5105         prev_cpu  = task_cpu(p);
5106         load      = source_load(prev_cpu, idx);
5107         this_load = target_load(this_cpu, idx);
5108 
5109         /*
5110          * If sync wakeup then subtract the (maximum possible)
5111          * effect of the currently running task from the load
5112          * of the current CPU:
5113          */
5114         if (sync) {
5115                 tg = task_group(current);
5116                 weight = current->se.avg.load_avg;
5117 
5118                 this_load += effective_load(tg, this_cpu, -weight, -weight);
5119                 load += effective_load(tg, prev_cpu, 0, -weight);
5120         }
5121 
5122         tg = task_group(p);
5123         weight = p->se.avg.load_avg;
5124 
5125         /*
5126          * In low-load situations, where prev_cpu is idle and this_cpu is idle
5127          * due to the sync cause above having dropped this_load to 0, we'll
5128          * always have an imbalance, but there's really nothing you can do
5129          * about that, so that's good too.
5130          *
5131          * Otherwise check if either cpus are near enough in load to allow this
5132          * task to be woken on this_cpu.
5133          */
5134         this_eff_load = 100;
5135         this_eff_load *= capacity_of(prev_cpu);
5136 
5137         prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
5138         prev_eff_load *= capacity_of(this_cpu);
5139 
5140         if (this_load > 0) {
5141                 this_eff_load *= this_load +
5142                         effective_load(tg, this_cpu, weight, weight);
5143 
5144                 prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
5145         }
5146 
5147         balanced = this_eff_load <= prev_eff_load;
5148 
5149         schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
5150 
5151         if (!balanced)
5152                 return 0;
5153 
5154         schedstat_inc(sd, ttwu_move_affine);
5155         schedstat_inc(p, se.statistics.nr_wakeups_affine);
5156 
5157         return 1;
5158 }
5159 
5160 /*
5161  * find_idlest_group finds and returns the least busy CPU group within the
5162  * domain.
5163  */
5164 static struct sched_group *
5165 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5166                   int this_cpu, int sd_flag)
5167 {
5168         struct sched_group *idlest = NULL, *group = sd->groups;
5169         unsigned long min_load = ULONG_MAX, this_load = 0;
5170         int load_idx = sd->forkexec_idx;
5171         int imbalance = 100 + (sd->imbalance_pct-100)/2;
5172 
5173         if (sd_flag & SD_BALANCE_WAKE)
5174                 load_idx = sd->wake_idx;
5175 
5176         do {
5177                 unsigned long load, avg_load;
5178                 int local_group;
5179                 int i;
5180 
5181                 /* Skip over this group if it has no CPUs allowed */
5182                 if (!cpumask_intersects(sched_group_cpus(group),
5183                                         tsk_cpus_allowed(p)))
5184                         continue;
5185 
5186                 local_group = cpumask_test_cpu(this_cpu,
5187                                                sched_group_cpus(group));
5188 
5189                 /* Tally up the load of all CPUs in the group */
5190                 avg_load = 0;
5191 
5192                 for_each_cpu(i, sched_group_cpus(group)) {
5193                         /* Bias balancing toward cpus of our domain */
5194                         if (local_group)
5195                                 load = source_load(i, load_idx);
5196                         else
5197                                 load = target_load(i, load_idx);
5198 
5199                         avg_load += load;
5200                 }
5201 
5202                 /* Adjust by relative CPU capacity of the group */
5203                 avg_load = (avg_load * SCHED_CAPACITY_SCALE) / group->sgc->capacity;
5204 
5205                 if (local_group) {
5206                         this_load = avg_load;
5207                 } else if (avg_load < min_load) {
5208                         min_load = avg_load;
5209                         idlest = group;
5210                 }
5211         } while (group = group->next, group != sd->groups);
5212 
5213         if (!idlest || 100*this_load < imbalance*min_load)
5214                 return NULL;
5215         return idlest;
5216 }
5217 
5218 /*
5219  * find_idlest_cpu - find the idlest cpu among the cpus in group.
5220  */
5221 static int
5222 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5223 {
5224         unsigned long load, min_load = ULONG_MAX;
5225         unsigned int min_exit_latency = UINT_MAX;
5226         u64 latest_idle_timestamp = 0;
5227         int least_loaded_cpu = this_cpu;
5228         int shallowest_idle_cpu = -1;
5229         int i;
5230 
5231         /* Traverse only the allowed CPUs */
5232         for_each_cpu_and(i, sched_group_cpus(group), tsk_cpus_allowed(p)) {
5233                 if (idle_cpu(i)) {
5234                         struct rq *rq = cpu_rq(i);
5235                         struct cpuidle_state *idle = idle_get_state(rq);
5236                         if (idle && idle->exit_latency < min_exit_latency) {
5237                                 /*
5238                                  * We give priority to a CPU whose idle state
5239                                  * has the smallest exit latency irrespective
5240                                  * of any idle timestamp.
5241                                  */
5242                                 min_exit_latency = idle->exit_latency;
5243                                 latest_idle_timestamp = rq->idle_stamp;
5244                                 shallowest_idle_cpu = i;
5245                         } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5246                                    rq->idle_stamp > latest_idle_timestamp) {
5247                                 /*
5248                                  * If equal or no active idle state, then
5249                                  * the most recently idled CPU might have
5250                                  * a warmer cache.
5251                                  */
5252                                 latest_idle_timestamp = rq->idle_stamp;
5253                                 shallowest_idle_cpu = i;
5254                         }
5255                 } else if (shallowest_idle_cpu == -1) {
5256                         load = weighted_cpuload(i);
5257                         if (load < min_load || (load == min_load && i == this_cpu)) {
5258                                 min_load = load;
5259                                 least_loaded_cpu = i;
5260                         }
5261                 }
5262         }
5263 
5264         return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5265 }
5266 
5267 /*
5268  * Try and locate an idle CPU in the sched_domain.
5269  */
5270 static int select_idle_sibling(struct task_struct *p, int target)
5271 {
5272         struct sched_domain *sd;
5273         struct sched_group *sg;
5274         int i = task_cpu(p);
5275 
5276         if (idle_cpu(target))
5277                 return target;
5278 
5279         /*
5280          * If the prevous cpu is cache affine and idle, don't be stupid.
5281          */
5282         if (i != target && cpus_share_cache(i, target) && idle_cpu(i))
5283                 return i;
5284 
5285         /*
5286          * Otherwise, iterate the domains and find an eligible idle cpu.
5287          *
5288          * A completely idle sched group at higher domains is more
5289          * desirable than an idle group at a lower level, because lower
5290          * domains have smaller groups and usually share hardware
5291          * resources which causes tasks to contend on them, e.g. x86
5292          * hyperthread siblings in the lowest domain (SMT) can contend
5293          * on the shared cpu pipeline.
5294          *
5295          * However, while we prefer idle groups at higher domains
5296          * finding an idle cpu at the lowest domain is still better than
5297          * returning 'target', which we've already established, isn't
5298          * idle.
5299          */
5300         sd = rcu_dereference(per_cpu(sd_llc, target));
5301         for_each_lower_domain(sd) {
5302                 sg = sd->groups;
5303                 do {
5304                         if (!cpumask_intersects(sched_group_cpus(sg),
5305                                                 tsk_cpus_allowed(p)))
5306                                 goto next;
5307 
5308                         /* Ensure the entire group is idle */
5309                         for_each_cpu(i, sched_group_cpus(sg)) {
5310                                 if (i == target || !idle_cpu(i))
5311                                         goto next;
5312                         }
5313 
5314                         /*
5315                          * It doesn't matter which cpu we pick, the
5316                          * whole group is idle.
5317                          */
5318                         target = cpumask_first_and(sched_group_cpus(sg),
5319                                         tsk_cpus_allowed(p));
5320                         goto done;
5321 next:
5322                         sg = sg->next;
5323                 } while (sg != sd->groups);
5324         }
5325 done:
5326         return target;
5327 }
5328 
5329 /*
5330  * cpu_util returns the amount of capacity of a CPU that is used by CFS
5331  * tasks. The unit of the return value must be the one of capacity so we can
5332  * compare the utilization with the capacity of the CPU that is available for
5333  * CFS task (ie cpu_capacity).
5334  *
5335  * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
5336  * recent utilization of currently non-runnable tasks on a CPU. It represents
5337  * the amount of utilization of a CPU in the range [0..capacity_orig] where
5338  * capacity_orig is the cpu_capacity available at the highest frequency
5339  * (arch_scale_freq_capacity()).
5340  * The utilization of a CPU converges towards a sum equal to or less than the
5341  * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
5342  * the running time on this CPU scaled by capacity_curr.
5343  *
5344  * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
5345  * higher than capacity_orig because of unfortunate rounding in
5346  * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
5347  * the average stabilizes with the new running time. We need to check that the
5348  * utilization stays within the range of [0..capacity_orig] and cap it if
5349  * necessary. Without utilization capping, a group could be seen as overloaded
5350  * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
5351  * available capacity. We allow utilization to overshoot capacity_curr (but not
5352  * capacity_orig) as it useful for predicting the capacity required after task
5353  * migrations (scheduler-driven DVFS).
5354  */
5355 static int cpu_util(int cpu)
5356 {
5357         unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
5358         unsigned long capacity = capacity_orig_of(cpu);
5359 
5360         return (util >= capacity) ? capacity : util;
5361 }
5362 
5363 /*
5364  * select_task_rq_fair: Select target runqueue for the waking task in domains
5365  * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
5366  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
5367  *
5368  * Balances load by selecting the idlest cpu in the idlest group, or under
5369  * certain conditions an idle sibling cpu if the domain has SD_WAKE_AFFINE set.
5370  *
5371  * Returns the target cpu number.
5372  *
5373  * preempt must be disabled.
5374  */
5375 static int
5376 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
5377 {
5378         struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
5379         int cpu = smp_processor_id();
5380         int new_cpu = prev_cpu;
5381         int want_affine = 0;
5382         int sync = wake_flags & WF_SYNC;
5383 
5384         if (sd_flag & SD_BALANCE_WAKE) {
5385                 record_wakee(p);
5386                 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, tsk_cpus_allowed(p));
5387         }
5388 
5389         rcu_read_lock();
5390         for_each_domain(cpu, tmp) {
5391                 if (!(tmp->flags & SD_LOAD_BALANCE))
5392                         break;
5393 
5394                 /*
5395                  * If both cpu and prev_cpu are part of this domain,
5396                  * cpu is a valid SD_WAKE_AFFINE target.
5397                  */
5398                 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
5399                     cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
5400                         affine_sd = tmp;
5401                         break;
5402                 }
5403 
5404                 if (tmp->flags & sd_flag)
5405                         sd = tmp;
5406                 else if (!want_affine)
5407                         break;
5408         }
5409 
5410         if (affine_sd) {
5411                 sd = NULL; /* Prefer wake_affine over balance flags */
5412                 if (cpu != prev_cpu && wake_affine(affine_sd, p, sync))
5413                         new_cpu = cpu;
5414         }
5415 
5416         if (!sd) {
5417                 if (sd_flag & SD_BALANCE_WAKE) /* XXX always ? */
5418                         new_cpu = select_idle_sibling(p, new_cpu);
5419 
5420         } else while (sd) {
5421                 struct sched_group *group;
5422                 int weight;
5423 
5424                 if (!(sd->flags & sd_flag)) {
5425                         sd = sd->child;
5426                         continue;
5427                 }
5428 
5429                 group = find_idlest_group(sd, p, cpu, sd_flag);
5430                 if (!group) {
5431                         sd = sd->child;
5432                         continue;
5433                 }
5434 
5435                 new_cpu = find_idlest_cpu(group, p, cpu);
5436                 if (new_cpu == -1 || new_cpu == cpu) {
5437                         /* Now try balancing at a lower domain level of cpu */
5438                         sd = sd->child;
5439                         continue;
5440                 }
5441 
5442                 /* Now try balancing at a lower domain level of new_cpu */
5443                 cpu = new_cpu;
5444                 weight = sd->span_weight;
5445                 sd = NULL;
5446                 for_each_domain(cpu, tmp) {
5447                         if (weight <= tmp->span_weight)
5448                                 break;
5449                         if (tmp->flags & sd_flag)
5450                                 sd = tmp;
5451                 }
5452                 /* while loop will break here if sd == NULL */
5453         }
5454         rcu_read_unlock();
5455 
5456         return new_cpu;
5457 }
5458 
5459 /*
5460  * Called immediately before a task is migrated to a new cpu; task_cpu(p) and
5461  * cfs_rq_of(p) references at time of call are still valid and identify the
5462  * previous cpu. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
5463  */
5464 static void migrate_task_rq_fair(struct task_struct *p)
5465 {
5466         /*
5467          * As blocked tasks retain absolute vruntime the migration needs to
5468          * deal with this by subtracting the old and adding the new
5469          * min_vruntime -- the latter is done by enqueue_entity() when placing
5470          * the task on the new runqueue.
5471          */
5472         if (p->state == TASK_WAKING) {
5473                 struct sched_entity *se = &p->se;
5474                 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5475                 u64 min_vruntime;
5476 
5477 #ifndef CONFIG_64BIT
5478                 u64 min_vruntime_copy;
5479 
5480                 do {
5481                         min_vruntime_copy = cfs_rq->min_vruntime_copy;
5482                         smp_rmb();
5483                         min_vruntime = cfs_rq->min_vruntime;
5484                 } while (min_vruntime != min_vruntime_copy);
5485 #else
5486                 min_vruntime = cfs_rq->min_vruntime;
5487 #endif
5488 
5489                 se->vruntime -= min_vruntime;
5490         }
5491 
5492         /*
5493          * We are supposed to update the task to "current" time, then its up to date
5494          * and ready to go to new CPU/cfs_rq. But we have difficulty in getting
5495          * what current time is, so simply throw away the out-of-date time. This
5496          * will result in the wakee task is less decayed, but giving the wakee more
5497          * load sounds not bad.
5498          */
5499         remove_entity_load_avg(&p->se);
5500 
5501         /* Tell new CPU we are migrated */
5502         p->se.avg.last_update_time = 0;
5503 
5504         /* We have migrated, no longer consider this task hot */
5505         p->se.exec_start = 0;
5506 }
5507 
5508 static void task_dead_fair(struct task_struct *p)
5509 {
5510         remove_entity_load_avg(&p->se);
5511 }
5512 #endif /* CONFIG_SMP */
5513 
5514 static unsigned long
5515 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
5516 {
5517         unsigned long gran = sysctl_sched_wakeup_granularity;
5518 
5519         /*
5520          * Since its curr running now, convert the gran from real-time
5521          * to virtual-time in his units.
5522          *
5523          * By using 'se' instead of 'curr' we penalize light tasks, so
5524          * they get preempted easier. That is, if 'se' < 'curr' then
5525          * the resulting gran will be larger, therefore penalizing the
5526          * lighter, if otoh 'se' > 'curr' then the resulting gran will
5527          * be smaller, again penalizing the lighter task.
5528          *
5529          * This is especially important for buddies when the leftmost
5530          * task is higher priority than the buddy.
5531          */
5532         return calc_delta_fair(gran, se);
5533 }
5534 
5535 /*
5536  * Should 'se' preempt 'curr'.
5537  *
5538  *             |s1
5539  *        |s2
5540  *   |s3
5541  *         g
5542  *      |<--->|c
5543  *
5544  *  w(c, s1) = -1
5545  *  w(c, s2) =  0
5546  *  w(c, s3) =  1
5547  *
5548  */
5549 static int
5550 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
5551 {
5552         s64 gran, vdiff = curr->vruntime - se->vruntime;
5553 
5554         if (vdiff <= 0)
5555                 return -1;
5556 
5557         gran = wakeup_gran(curr, se);
5558         if (vdiff > gran)
5559                 return 1;
5560 
5561         return 0;
5562 }
5563 
5564 static void set_last_buddy(struct sched_entity *se)
5565 {
5566         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5567                 return;
5568 
5569         for_each_sched_entity(se)
5570                 cfs_rq_of(se)->last = se;
5571 }
5572 
5573 static void set_next_buddy(struct sched_entity *se)
5574 {
5575         if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
5576                 return;
5577 
5578         for_each_sched_entity(se)
5579                 cfs_rq_of(se)->next = se;
5580 }
5581 
5582 static void set_skip_buddy(struct sched_entity *se)
5583 {
5584         for_each_sched_entity(se)
5585                 cfs_rq_of(se)->skip = se;
5586 }
5587 
5588 /*
5589  * Preempt the current task with a newly woken task if needed:
5590  */
5591 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
5592 {
5593         struct task_struct *curr = rq->curr;
5594         struct sched_entity *se = &curr->se, *pse = &p->se;
5595         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5596         int scale = cfs_rq->nr_running >= sched_nr_latency;
5597         int next_buddy_marked = 0;
5598 
5599         if (unlikely(se == pse))
5600                 return;
5601 
5602         /*
5603          * This is possible from callers such as attach_tasks(), in which we
5604          * unconditionally check_prempt_curr() after an enqueue (which may have
5605          * lead to a throttle).  This both saves work and prevents false
5606          * next-buddy nomination below.
5607          */
5608         if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
5609                 return;
5610 
5611         if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
5612                 set_next_buddy(pse);
5613                 next_buddy_marked = 1;
5614         }
5615 
5616         /*
5617          * We can come here with TIF_NEED_RESCHED already set from new task
5618          * wake up path.
5619          *
5620          * Note: this also catches the edge-case of curr being in a throttled
5621          * group (e.g. via set_curr_task), since update_curr() (in the
5622          * enqueue of curr) will have resulted in resched being set.  This
5623          * prevents us from potentially nominating it as a false LAST_BUDDY
5624          * below.
5625          */
5626         if (test_tsk_need_resched(curr))
5627                 return;
5628 
5629         /* Idle tasks are by definition preempted by non-idle tasks. */
5630         if (unlikely(curr->policy == SCHED_IDLE) &&
5631             likely(p->policy != SCHED_IDLE))
5632                 goto preempt;
5633 
5634         /*
5635          * Batch and idle tasks do not preempt non-idle tasks (their preemption
5636          * is driven by the tick):
5637          */
5638         if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
5639                 return;
5640 
5641         find_matching_se(&se, &pse);
5642         update_curr(cfs_rq_of(se));
5643         BUG_ON(!pse);
5644         if (wakeup_preempt_entity(se, pse) == 1) {
5645                 /*
5646                  * Bias pick_next to pick the sched entity that is
5647                  * triggering this preemption.
5648                  */
5649                 if (!next_buddy_marked)
5650                         set_next_buddy(pse);
5651                 goto preempt;
5652         }
5653 
5654         return;
5655 
5656 preempt:
5657         resched_curr(rq);
5658         /*
5659          * Only set the backward buddy when the current task is still
5660          * on the rq. This can happen when a wakeup gets interleaved
5661          * with schedule on the ->pre_schedule() or idle_balance()
5662          * point, either of which can * drop the rq lock.
5663          *
5664          * Also, during early boot the idle thread is in the fair class,
5665          * for obvious reasons its a bad idea to schedule back to it.
5666          */
5667         if (unlikely(!se->on_rq || curr == rq->idle))
5668                 return;
5669 
5670         if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
5671                 set_last_buddy(se);
5672 }
5673 
5674 static struct task_struct *
5675 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct pin_cookie cookie)
5676 {
5677         struct cfs_rq *cfs_rq = &rq->cfs;
5678         struct sched_entity *se;
5679         struct task_struct *p;
5680         int new_tasks;
5681 
5682 again:
5683 #ifdef CONFIG_FAIR_GROUP_SCHED
5684         if (!cfs_rq->nr_running)
5685                 goto idle;
5686 
5687         if (prev->sched_class != &fair_sched_class)
5688                 goto simple;
5689 
5690         /*
5691          * Because of the set_next_buddy() in dequeue_task_fair() it is rather
5692          * likely that a next task is from the same cgroup as the current.
5693          *
5694          * Therefore attempt to avoid putting and setting the entire cgroup
5695          * hierarchy, only change the part that actually changes.
5696          */
5697 
5698         do {
5699                 struct sched_entity *curr = cfs_rq->curr;
5700 
5701                 /*
5702                  * Since we got here without doing put_prev_entity() we also
5703                  * have to consider cfs_rq->curr. If it is still a runnable
5704                  * entity, update_curr() will update its vruntime, otherwise
5705                  * forget we've ever seen it.
5706                  */
5707                 if (curr) {
5708                         if (curr->on_rq)
5709                                 update_curr(cfs_rq);
5710                         else
5711                                 curr = NULL;
5712 
5713                         /*
5714                          * This call to check_cfs_rq_runtime() will do the
5715                          * throttle and dequeue its entity in the parent(s).
5716                          * Therefore the 'simple' nr_running test will indeed
5717                          * be correct.
5718                          */
5719                         if (unlikely(check_cfs_rq_runtime(cfs_rq)))
5720                                 goto simple;
5721                 }
5722 
5723                 se = pick_next_entity(cfs_rq, curr);
5724                 cfs_rq = group_cfs_rq(se);
5725         } while (cfs_rq);
5726 
5727         p = task_of(se);
5728 
5729         /*
5730          * Since we haven't yet done put_prev_entity and if the selected task
5731          * is a different task than we started out with, try and touch the
5732          * least amount of cfs_rqs.
5733          */
5734         if (prev != p) {
5735                 struct sched_entity *pse = &prev->se;
5736 
5737                 while (!(cfs_rq = is_same_group(se, pse))) {
5738                         int se_depth = se->depth;
5739                         int pse_depth = pse->depth;
5740 
5741                         if (se_depth <= pse_depth) {
5742                                 put_prev_entity(cfs_rq_of(pse), pse);
5743                                 pse = parent_entity(pse);
5744                         }
5745                         if (se_depth >= pse_depth) {
5746                                 set_next_entity(cfs_rq_of(se), se);
5747                                 se = parent_entity(se);
5748                         }
5749                 }
5750 
5751                 put_prev_entity(cfs_rq, pse);
5752                 set_next_entity(cfs_rq, se);
5753         }
5754 
5755         if (hrtick_enabled(rq))
5756                 hrtick_start_fair(rq, p);
5757 
5758         return p;
5759 simple:
5760         cfs_rq = &rq->cfs;
5761 #endif
5762 
5763         if (!cfs_rq->nr_running)
5764                 goto idle;
5765 
5766         put_prev_task(rq, prev);
5767 
5768         do {
5769                 se = pick_next_entity(cfs_rq, NULL);
5770                 set_next_entity(cfs_rq, se);
5771                 cfs_rq = group_cfs_rq(se);
5772         } while (cfs_rq);
5773 
5774         p = task_of(se);
5775 
5776         if (hrtick_enabled(rq))
5777                 hrtick_start_fair(rq, p);
5778 
5779         return p;
5780 
5781 idle:
5782         /*
5783          * This is OK, because current is on_cpu, which avoids it being picked
5784          * for load-balance and preemption/IRQs are still disabled avoiding
5785          * further scheduler activity on it and we're being very careful to
5786          * re-start the picking loop.
5787          */
5788         lockdep_unpin_lock(&rq->lock, cookie);
5789         new_tasks = idle_balance(rq);
5790         lockdep_repin_lock(&rq->lock, cookie);
5791         /*
5792          * Because idle_balance() releases (and re-acquires) rq->lock, it is
5793          * possible for any higher priority task to appear. In that case we
5794          * must re-start the pick_next_entity() loop.
5795          */
5796         if (new_tasks < 0)
5797                 return RETRY_TASK;
5798 
5799         if (new_tasks > 0)
5800                 goto again;
5801 
5802         return NULL;
5803 }
5804 
5805 /*
5806  * Account for a descheduled task:
5807  */
5808 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
5809 {
5810         struct sched_entity *se = &prev->se;
5811         struct cfs_rq *cfs_rq;
5812 
5813         for_each_sched_entity(se) {
5814                 cfs_rq = cfs_rq_of(se);
5815                 put_prev_entity(cfs_rq, se);
5816         }
5817 }
5818 
5819 /*
5820  * sched_yield() is very simple
5821  *
5822  * The magic of dealing with the ->skip buddy is in pick_next_entity.
5823  */
5824 static void yield_task_fair(struct rq *rq)
5825 {
5826         struct task_struct *curr = rq->curr;
5827         struct cfs_rq *cfs_rq = task_cfs_rq(curr);
5828         struct sched_entity *se = &curr->se;
5829 
5830         /*
5831          * Are we the only task in the tree?
5832          */
5833         if (unlikely(rq->nr_running == 1))
5834                 return;
5835 
5836         clear_buddies(cfs_rq, se);
5837 
5838         if (curr->policy != SCHED_BATCH) {
5839                 update_rq_clock(rq);
5840                 /*
5841                  * Update run-time statistics of the 'current'.
5842                  */
5843                 update_curr(cfs_rq);
5844                 /*
5845                  * Tell update_rq_clock() that we've just updated,
5846                  * so we don't do microscopic update in schedule()
5847                  * and double the fastpath cost.
5848                  */
5849                 rq_clock_skip_update(rq, true);
5850         }
5851 
5852         set_skip_buddy(se);
5853 }
5854 
5855 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
5856 {
5857         struct sched_entity *se = &p->se;
5858 
5859         /* throttled hierarchies are not runnable */
5860         if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
5861                 return false;
5862 
5863         /* Tell the scheduler that we'd really like pse to run next. */
5864         set_next_buddy(se);
5865 
5866         yield_task_fair(rq);
5867 
5868         return true;
5869 }
5870 
5871 #ifdef CONFIG_SMP
5872 /**************************************************
5873  * Fair scheduling class load-balancing methods.
5874  *
5875  * BASICS
5876  *
5877  * The purpose of load-balancing is to achieve the same basic fairness the
5878  * per-cpu scheduler provides, namely provide a proportional amount of compute
5879  * time to each task. This is expressed in the following equation:
5880  *
5881  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
5882  *
5883  * Where W_i,n is the n-th weight average for cpu i. The instantaneous weight
5884  * W_i,0 is defined as:
5885  *
5886  *   W_i,0 = \Sum_j w_i,j                                             (2)
5887  *
5888  * Where w_i,j is the weight of the j-th runnable task on cpu i. This weight
5889  * is derived from the nice value as per sched_prio_to_weight[].
5890  *
5891  * The weight average is an exponential decay average of the instantaneous
5892  * weight:
5893  *
5894  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
5895  *
5896  * C_i is the compute capacity of cpu i, typically it is the
5897  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
5898  * can also include other factors [XXX].
5899  *
5900  * To achieve this balance we define a measure of imbalance which follows
5901  * directly from (1):
5902  *
5903  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
5904  *
5905  * We them move tasks around to minimize the imbalance. In the continuous
5906  * function space it is obvious this converges, in the discrete case we get
5907  * a few fun cases generally called infeasible weight scenarios.
5908  *
5909  * [XXX expand on:
5910  *     - infeasible weights;
5911  *     - local vs global optima in the discrete case. ]
5912  *
5913  *
5914  * SCHED DOMAINS
5915  *
5916  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
5917  * for all i,j solution, we create a tree of cpus that follows the hardware
5918  * topology where each level pairs two lower groups (or better). This results
5919  * in O(log n) layers. Furthermore we reduce the number of cpus going up the
5920  * tree to only the first of the previous level and we decrease the frequency
5921  * of load-balance at each level inv. proportional to the number of cpus in
5922  * the groups.
5923  *
5924  * This yields:
5925  *
5926  *     log_2 n     1     n
5927  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
5928  *     i = 0      2^i   2^i
5929  *                               `- size of each group
5930  *         |         |     `- number of cpus doing load-balance
5931  *         |         `- freq
5932  *         `- sum over all levels
5933  *
5934  * Coupled with a limit on how many tasks we can migrate every balance pass,
5935  * this makes (5) the runtime complexity of the balancer.
5936  *
5937  * An important property here is that each CPU is still (indirectly) connected
5938  * to every other cpu in at most O(log n) steps:
5939  *
5940  * The adjacency matrix of the resulting graph is given by:
5941  *
5942  *             log_2 n     
5943  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
5944  *             k = 0
5945  *
5946  * And you'll find that:
5947  *
5948  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
5949  *
5950  * Showing there's indeed a path between every cpu in at most O(log n) steps.
5951  * The task movement gives a factor of O(m), giving a convergence complexity
5952  * of:
5953  *
5954  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
5955  *
5956  *
5957  * WORK CONSERVING
5958  *
5959  * In order to avoid CPUs going idle while there's still work to do, new idle
5960  * balancing is more aggressive and has the newly idle cpu iterate up the domain
5961  * tree itself instead of relying on other CPUs to bring it work.
5962  *
5963  * This adds some complexity to both (5) and (8) but it reduces the total idle
5964  * time.
5965  *
5966  * [XXX more?]
5967  *
5968  *
5969  * CGROUPS
5970  *
5971  * Cgroups make a horror show out of (2), instead of a simple sum we get:
5972  *
5973  *                                s_k,i
5974  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
5975  *                                 S_k
5976  *
5977  * Where
5978  *
5979  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
5980  *
5981  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on cpu i.
5982  *
5983  * The big problem is S_k, its a global sum needed to compute a local (W_i)
5984  * property.
5985  *
5986  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
5987  *      rewrite all of this once again.]
5988  */ 
5989 
5990 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
5991 
5992 enum fbq_type { regular, remote, all };
5993 
5994 #define LBF_ALL_PINNED  0x01
5995 #define LBF_NEED_BREAK  0x02
5996 #define LBF_DST_PINNED  0x04
5997 #define LBF_SOME_PINNED 0x08
5998 
5999 struct lb_env {
6000         struct sched_domain     *sd;
6001 
6002         struct rq               *src_rq;
6003         int                     src_cpu;
6004 
6005         int                     dst_cpu;
6006         struct rq               *dst_rq;
6007 
6008         struct cpumask          *dst_grpmask;
6009         int                     new_dst_cpu;
6010         enum cpu_idle_type      idle;
6011         long                    imbalance;
6012         /* The set of CPUs under consideration for load-balancing */
6013         struct cpumask          *cpus;
6014 
6015         unsigned int            flags;
6016 
6017         unsigned int            loop;
6018         unsigned int            loop_break;
6019         unsigned int            loop_max;
6020 
6021         enum fbq_type           fbq_type;
6022         struct list_head        tasks;
6023 };
6024 
6025 /*
6026  * Is this task likely cache-hot:
6027  */
6028 static int task_hot(struct task_struct *p, struct lb_env *env)
6029 {
6030         s64 delta;
6031 
6032         lockdep_assert_held(&env->src_rq->lock);
6033 
6034         if (p->sched_class != &fair_sched_class)
6035                 return 0;
6036 
6037         if (unlikely(p->policy == SCHED_IDLE))
6038                 return 0;
6039 
6040         /*
6041          * Buddy candidates are cache hot:
6042          */
6043         if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6044                         (&p->se == cfs_rq_of(&p->se)->next ||
6045                          &p->se == cfs_rq_of(&p->se)->last))
6046                 return 1;
6047 
6048         if (sysctl_sched_migration_cost == -1)
6049                 return 1;
6050         if (sysctl_sched_migration_cost == 0)
6051                 return 0;
6052 
6053         delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6054 
6055         return delta < (s64)sysctl_sched_migration_cost;
6056 }
6057 
6058 #ifdef CONFIG_NUMA_BALANCING
6059 /*
6060  * Returns 1, if task migration degrades locality
6061  * Returns 0, if task migration improves locality i.e migration preferred.
6062  * Returns -1, if task migration is not affected by locality.
6063  */
6064 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6065 {
6066         struct numa_group *numa_group = rcu_dereference(p->numa_group);
6067         unsigned long src_faults, dst_faults;
6068         int src_nid, dst_nid;
6069 
6070         if (!static_branch_likely(&sched_numa_balancing))
6071                 return -1;
6072 
6073         if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6074                 return -1;
6075 
6076         src_nid = cpu_to_node(env->src_cpu);
6077         dst_nid = cpu_to_node(env->dst_cpu);
6078 
6079         if (src_nid == dst_nid)
6080                 return -1;
6081 
6082         /* Migrating away from the preferred node is always bad. */
6083         if (src_nid == p->numa_preferred_nid) {
6084                 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6085                         return 1;
6086                 else
6087                         return -1;
6088         }
6089 
6090         /* Encourage migration to the preferred node. */
6091         if (dst_nid == p->numa_preferred_nid)
6092                 return 0;
6093 
6094         if (numa_group) {
6095                 src_faults = group_faults(p, src_nid);
6096                 dst_faults = group_faults(p, dst_nid);
6097         } else {
6098                 src_faults = task_faults(p, src_nid);
6099                 dst_faults = task_faults(p, dst_nid);
6100         }
6101 
6102         return dst_faults < src_faults;
6103 }
6104 
6105 #else
6106 static inline int migrate_degrades_locality(struct task_struct *p,
6107                                              struct lb_env *env)
6108 {
6109         return -1;
6110 }
6111 #endif
6112 
6113 /*
6114  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6115  */
6116 static
6117 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6118 {
6119         int tsk_cache_hot;
6120 
6121         lockdep_assert_held(&env->src_rq->lock);
6122 
6123         /*
6124          * We do not migrate tasks that are:
6125          * 1) throttled_lb_pair, or
6126          * 2) cannot be migrated to this CPU due to cpus_allowed, or
6127          * 3) running (obviously), or
6128          * 4) are cache-hot on their current CPU.
6129          */
6130         if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6131                 return 0;
6132 
6133         if (!cpumask_test_cpu(env->dst_cpu, tsk_cpus_allowed(p))) {
6134                 int cpu;
6135 
6136                 schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
6137 
6138                 env->flags |= LBF_SOME_PINNED;
6139 
6140                 /*
6141                  * Remember if this task can be migrated to any other cpu in
6142                  * our sched_group. We may want to revisit it if we couldn't
6143                  * meet load balance goals by pulling other tasks on src_cpu.
6144                  *
6145                  * Also avoid computing new_dst_cpu if we have already computed
6146                  * one in current iteration.
6147                  */
6148                 if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
6149                         return 0;
6150 
6151                 /* Prevent to re-select dst_cpu via env's cpus */
6152                 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
6153                         if (cpumask_test_cpu(cpu, tsk_cpus_allowed(p))) {
6154                                 env->flags |= LBF_DST_PINNED;
6155                                 env->new_dst_cpu = cpu;
6156                                 break;
6157                         }
6158                 }
6159 
6160                 return 0;
6161         }
6162 
6163         /* Record that we found atleast one task that could run on dst_cpu */
6164         env->flags &= ~LBF_ALL_PINNED;
6165 
6166         if (task_running(env->src_rq, p)) {
6167                 schedstat_inc(p, se.statistics.nr_failed_migrations_running);
6168                 return 0;
6169         }
6170 
6171         /*
6172          * Aggressive migration if:
6173          * 1) destination numa is preferred
6174          * 2) task is cache cold, or
6175          * 3) too many balance attempts have failed.
6176          */
6177         tsk_cache_hot = migrate_degrades_locality(p, env);
6178         if (tsk_cache_hot == -1)
6179                 tsk_cache_hot = task_hot(p, env);
6180 
6181         if (tsk_cache_hot <= 0 ||
6182             env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
6183                 if (tsk_cache_hot == 1) {
6184                         schedstat_inc(env->sd, lb_hot_gained[env->idle]);
6185                         schedstat_inc(p, se.statistics.nr_forced_migrations);
6186                 }
6187                 return 1;
6188         }
6189 
6190         schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
6191         return 0;
6192 }
6193 
6194 /*
6195  * detach_task() -- detach the task for the migration specified in env
6196  */
6197 static void detach_task(struct task_struct *p, struct lb_env *env)
6198 {
6199         lockdep_assert_held(&env->src_rq->lock);
6200 
6201         p->on_rq = TASK_ON_RQ_MIGRATING;
6202         deactivate_task(env->src_rq, p, 0);
6203         set_task_cpu(p, env->dst_cpu);
6204 }
6205 
6206 /*
6207  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
6208  * part of active balancing operations within "domain".
6209  *
6210  * Returns a task if successful and NULL otherwise.
6211  */
6212 static struct task_struct *detach_one_task(struct lb_env *env)
6213 {
6214         struct task_struct *p, *n;
6215 
6216         lockdep_assert_held(&env->src_rq->lock);
6217 
6218         list_for_each_entry_safe(p, n, &env->src_rq->cfs_tasks, se.group_node) {
6219                 if (!can_migrate_task(p, env))
6220                         continue;
6221 
6222                 detach_task(p, env);
6223 
6224                 /*
6225                  * Right now, this is only the second place where
6226                  * lb_gained[env->idle] is updated (other is detach_tasks)
6227                  * so we can safely collect stats here rather than
6228                  * inside detach_tasks().
6229                  */
6230                 schedstat_inc(env->sd, lb_gained[env->idle]);
6231                 return p;
6232         }
6233         return NULL;
6234 }
6235 
6236 static const unsigned int sched_nr_migrate_break = 32;
6237 
6238 /*
6239  * detach_tasks() -- tries to detach up to imbalance weighted load from
6240  * busiest_rq, as part of a balancing operation within domain "sd".
6241  *
6242  * Returns number of detached tasks if successful and 0 otherwise.
6243  */
6244 static int detach_tasks(struct lb_env *env)
6245 {
6246         struct list_head *tasks = &env->src_rq->cfs_tasks;
6247         struct task_struct *p;
6248         unsigned long load;
6249         int detached = 0;
6250 
6251         lockdep_assert_held(&env->src_rq->lock);
6252 
6253         if (env->imbalance <= 0)
6254                 return 0;
6255 
6256         while (!list_empty(tasks)) {
6257                 /*
6258                  * We don't want to steal all, otherwise we may be treated likewise,
6259                  * which could at worst lead to a livelock crash.
6260                  */
6261                 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
6262                         break;
6263 
6264                 p = list_first_entry(tasks, struct task_struct, se.group_node);
6265 
6266                 env->loop++;
6267                 /* We've more or less seen every task there is, call it quits */
6268                 if (env->loop > env->loop_max)
6269                         break;
6270 
6271                 /* take a breather every nr_migrate tasks */
6272                 if (env->loop > env->loop_break) {
6273                         env->loop_break += sched_nr_migrate_break;
6274                         env->flags |= LBF_NEED_BREAK;
6275                         break;
6276                 }
6277 
6278                 if (!can_migrate_task(p, env))
6279                         goto next;
6280 
6281                 load = task_h_load(p);
6282 
6283                 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
6284                         goto next;
6285 
6286                 if ((load / 2) > env->imbalance)
6287                         goto next;
6288 
6289                 detach_task(p, env);
6290                 list_add(&p->se.group_node, &env->tasks);
6291 
6292                 detached++;
6293                 env->imbalance -= load;
6294 
6295 #ifdef CONFIG_PREEMPT
6296                 /*
6297                  * NEWIDLE balancing is a source of latency, so preemptible
6298                  * kernels will stop after the first task is detached to minimize
6299                  * the critical section.
6300                  */
6301                 if (env->idle == CPU_NEWLY_IDLE)
6302                         break;
6303 #endif
6304 
6305                 /*
6306                  * We only want to steal up to the prescribed amount of
6307                  * weighted load.
6308                  */
6309                 if (env->imbalance <= 0)
6310                         break;
6311 
6312                 continue;
6313 next:
6314                 list_move_tail(&p->se.group_node, tasks);
6315         }
6316 
6317         /*
6318          * Right now, this is one of only two places we collect this stat
6319          * so we can safely collect detach_one_task() stats here rather
6320          * than inside detach_one_task().
6321          */
6322         schedstat_add(env->sd, lb_gained[env->idle], detached);
6323 
6324         return detached;
6325 }
6326 
6327 /*
6328  * attach_task() -- attach the task detached by detach_task() to its new rq.
6329  */
6330 static void attach_task(struct rq *rq, struct task_struct *p)
6331 {
6332         lockdep_assert_held(&rq->lock);
6333 
6334         BUG_ON(task_rq(p) != rq);
6335         activate_task(rq, p, 0);
6336         p->on_rq = TASK_ON_RQ_QUEUED;
6337         check_preempt_curr(rq, p, 0);
6338 }
6339 
6340 /*
6341  * attach_one_task() -- attaches the task returned from detach_one_task() to
6342  * its new rq.
6343  */
6344 static void attach_one_task(struct rq *rq, struct task_struct *p)
6345 {
6346         raw_spin_lock(&rq->lock);
6347         attach_task(rq, p);
6348         raw_spin_unlock(&rq->lock);
6349 }
6350 
6351 /*
6352  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
6353  * new rq.
6354  */
6355 static void attach_tasks(struct lb_env *env)
6356 {
6357         struct list_head *tasks = &env->tasks;
6358         struct task_struct *p;
6359 
6360         raw_spin_lock(&env->dst_rq->lock);
6361 
6362         while (!list_empty(tasks)) {
6363                 p = list_first_entry(tasks, struct task_struct, se.group_node);
6364                 list_del_init(&p->se.group_node);
6365 
6366                 attach_task(env->dst_rq, p);
6367         }
6368 
6369         raw_spin_unlock(&env->dst_rq->lock);
6370 }
6371 
6372 #ifdef CONFIG_FAIR_GROUP_SCHED
6373 static void update_blocked_averages(int cpu)
6374 {
6375         struct rq *rq = cpu_rq(cpu);
6376         struct cfs_rq *cfs_rq;
6377         unsigned long flags;
6378 
6379         raw_spin_lock_irqsave(&rq->lock, flags);
6380         update_rq_clock(rq);
6381 
6382         /*
6383          * Iterates the task_group tree in a bottom up fashion, see
6384          * list_add_leaf_cfs_rq() for details.
6385          */
6386         for_each_leaf_cfs_rq(rq, cfs_rq) {
6387                 /* throttled entities do not contribute to load */
6388                 if (throttled_hierarchy(cfs_rq))
6389                         continue;
6390 
6391                 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
6392                         update_tg_load_avg(cfs_rq, 0);
6393         }
6394         raw_spin_unlock_irqrestore(&rq->lock, flags);
6395 }
6396 
6397 /*
6398  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
6399  * This needs to be done in a top-down fashion because the load of a child
6400  * group is a fraction of its parents load.
6401  */
6402 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
6403 {
6404         struct rq *rq = rq_of(cfs_rq);
6405         struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
6406         unsigned long now = jiffies;
6407         unsigned long load;
6408 
6409         if (cfs_rq->last_h_load_update == now)
6410                 return;
6411 
6412         cfs_rq->h_load_next = NULL;
6413         for_each_sched_entity(se) {
6414                 cfs_rq = cfs_rq_of(se);
6415                 cfs_rq->h_load_next = se;
6416                 if (cfs_rq->last_h_load_update == now)
6417                         break;
6418         }
6419 
6420         if (!se) {
6421                 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
6422                 cfs_rq->last_h_load_update = now;
6423         }
6424 
6425         while ((se = cfs_rq->h_load_next) != NULL) {
6426                 load = cfs_rq->h_load;
6427                 load = div64_ul(load * se->avg.load_avg,
6428                         cfs_rq_load_avg(cfs_rq) + 1);
6429                 cfs_rq = group_cfs_rq(se);
6430                 cfs_rq->h_load = load;
6431                 cfs_rq->last_h_load_update = now;
6432         }
6433 }
6434 
6435 static unsigned long task_h_load(struct task_struct *p)
6436 {
6437         struct cfs_rq *cfs_rq = task_cfs_rq(p);
6438 
6439         update_cfs_rq_h_load(cfs_rq);
6440         return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
6441                         cfs_rq_load_avg(cfs_rq) + 1);
6442 }
6443 #else
6444 static inline void update_blocked_averages(int cpu)
6445 {
6446         struct rq *rq = cpu_rq(cpu);
6447         struct cfs_rq *cfs_rq = &rq->cfs;
6448         unsigned long flags;
6449 
6450         raw_spin_lock_irqsave(&rq->lock, flags);
6451         update_rq_clock(rq);
6452         update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true);
6453         raw_spin_unlock_irqrestore(&rq->lock, flags);
6454 }
6455 
6456 static unsigned long task_h_load(struct task_struct *p)
6457 {
6458         return p->se.avg.load_avg;
6459 }
6460 #endif
6461 
6462 /********** Helpers for find_busiest_group ************************/
6463 
6464 enum group_type {
6465         group_other = 0,
6466         group_imbalanced,
6467         group_overloaded,
6468 };
6469 
6470 /*
6471  * sg_lb_stats - stats of a sched_group required for load_balancing
6472  */
6473 struct sg_lb_stats {
6474         unsigned long avg_load; /*Avg load across the CPUs of the group */
6475         unsigned long group_load; /* Total load over the CPUs of the group */
6476         unsigned long sum_weighted_load; /* Weighted load of group's tasks */
6477         unsigned long load_per_task;
6478         unsigned long group_capacity;
6479         unsigned long group_util; /* Total utilization of the group */
6480         unsigned int sum_nr_running; /* Nr tasks running in the group */
6481         unsigned int idle_cpus;
6482         unsigned int group_weight;
6483         enum group_type group_type;
6484         int group_no_capacity;
6485 #ifdef CONFIG_NUMA_BALANCING
6486         unsigned int nr_numa_running;
6487         unsigned int nr_preferred_running;
6488 #endif
6489 };
6490 
6491 /*
6492  * sd_lb_stats - Structure to store the statistics of a sched_domain
6493  *               during load balancing.
6494  */
6495 struct sd_lb_stats {
6496         struct sched_group *busiest;    /* Busiest group in this sd */
6497         struct sched_group *local;      /* Local group in this sd */
6498         unsigned long total_load;       /* Total load of all groups in sd */
6499         unsigned long total_capacity;   /* Total capacity of all groups in sd */
6500         unsigned long avg_load; /* Average load across all groups in sd */
6501 
6502         struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
6503         struct sg_lb_stats local_stat;  /* Statistics of the local group */
6504 };
6505 
6506 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
6507 {
6508         /*
6509          * Skimp on the clearing to avoid duplicate work. We can avoid clearing
6510          * local_stat because update_sg_lb_stats() does a full clear/assignment.
6511          * We must however clear busiest_stat::avg_load because
6512          * update_sd_pick_busiest() reads this before assignment.
6513          */
6514         *sds = (struct sd_lb_stats){
6515                 .busiest = NULL,
6516                 .local = NULL,
6517                 .total_load = 0UL,
6518                 .total_capacity = 0UL,
6519                 .busiest_stat = {
6520                         .avg_load = 0UL,
6521                         .sum_nr_running = 0,
6522                         .group_type = group_other,
6523                 },
6524         };
6525 }
6526 
6527 /**
6528  * get_sd_load_idx - Obtain the load index for a given sched domain.
6529  * @sd: The sched_domain whose load_idx is to be obtained.
6530  * @idle: The idle status of the CPU for whose sd load_idx is obtained.
6531  *
6532  * Return: The load index.
6533  */
6534 static inline int get_sd_load_idx(struct sched_domain *sd,
6535                                         enum cpu_idle_type idle)
6536 {
6537         int load_idx;
6538 
6539         switch (idle) {
6540         case CPU_NOT_IDLE:
6541                 load_idx = sd->busy_idx;
6542                 break;
6543 
6544         case CPU_NEWLY_IDLE:
6545                 load_idx = sd->newidle_idx;
6546                 break;
6547         default:
6548                 load_idx = sd->idle_idx;
6549                 break;
6550         }
6551 
6552         return load_idx;
6553 }
6554 
6555 static unsigned long scale_rt_capacity(int cpu)
6556 {
6557         struct rq *rq = cpu_rq(cpu);
6558         u64 total, used, age_stamp, avg;
6559         s64 delta;
6560 
6561         /*
6562          * Since we're reading these variables without serialization make sure
6563          * we read them once before doing sanity checks on them.
6564          */
6565         age_stamp = READ_ONCE(rq->age_stamp);
6566         avg = READ_ONCE(rq->rt_avg);
6567         delta = __rq_clock_broken(rq) - age_stamp;
6568 
6569         if (unlikely(delta < 0))
6570                 delta = 0;
6571 
6572         total = sched_avg_period() + delta;
6573 
6574         used = div_u64(avg, total);
6575 
6576         if (likely(used < SCHED_CAPACITY_SCALE))
6577                 return SCHED_CAPACITY_SCALE - used;
6578 
6579         return 1;
6580 }
6581 
6582 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
6583 {
6584         unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
6585         struct sched_group *sdg = sd->groups;
6586 
6587         cpu_rq(cpu)->cpu_capacity_orig = capacity;
6588 
6589         capacity *= scale_rt_capacity(cpu);
6590         capacity >>= SCHED_CAPACITY_SHIFT;
6591 
6592         if (!capacity)
6593                 capacity = 1;
6594 
6595         cpu_rq(cpu)->cpu_capacity = capacity;
6596         sdg->sgc->capacity = capacity;
6597 }
6598 
6599 void update_group_capacity(struct sched_domain *sd, int cpu)
6600 {
6601         struct sched_domain *child = sd->child;
6602         struct sched_group *group, *sdg = sd->groups;
6603         unsigned long capacity;
6604         unsigned long interval;
6605 
6606         interval = msecs_to_jiffies(sd->balance_interval);
6607         interval = clamp(interval, 1UL, max_load_balance_interval);
6608         sdg->sgc->next_update = jiffies + interval;
6609 
6610         if (!child) {
6611                 update_cpu_capacity(sd, cpu);
6612                 return;
6613         }
6614 
6615         capacity = 0;
6616 
6617         if (child->flags & SD_OVERLAP) {
6618                 /*
6619                  * SD_OVERLAP domains cannot assume that child groups
6620                  * span the current group.
6621                  */
6622 
6623                 for_each_cpu(cpu, sched_group_cpus(sdg)) {
6624                         struct sched_group_capacity *sgc;
6625                         struct rq *rq = cpu_rq(cpu);
6626 
6627                         /*
6628                          * build_sched_domains() -> init_sched_groups_capacity()
6629                          * gets here before we've attached the domains to the
6630                          * runqueues.
6631                          *
6632                          * Use capacity_of(), which is set irrespective of domains
6633                          * in update_cpu_capacity().
6634                          *
6635                          * This avoids capacity from being 0 and
6636                          * causing divide-by-zero issues on boot.
6637                          */
6638                         if (unlikely(!rq->sd)) {
6639                                 capacity += capacity_of(cpu);
6640                                 continue;
6641                         }
6642 
6643                         sgc = rq->sd->groups->sgc;
6644                         capacity += sgc->capacity;
6645                 }
6646         } else  {
6647                 /*
6648                  * !SD_OVERLAP domains can assume that child groups
6649                  * span the current group.
6650                  */ 
6651 
6652                 group = child->groups;
6653                 do {
6654                         capacity += group->sgc->capacity;
6655                         group = group->next;
6656                 } while (group != child->groups);
6657         }
6658 
6659         sdg->sgc->capacity = capacity;
6660 }
6661 
6662 /*
6663  * Check whether the capacity of the rq has been noticeably reduced by side
6664  * activity. The imbalance_pct is used for the threshold.
6665  * Return true is the capacity is reduced
6666  */
6667 static inline int
6668 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
6669 {
6670         return ((rq->cpu_capacity * sd->imbalance_pct) <
6671                                 (rq->cpu_capacity_orig * 100));
6672 }
6673 
6674 /*
6675  * Group imbalance indicates (and tries to solve) the problem where balancing
6676  * groups is inadequate due to tsk_cpus_allowed() constraints.
6677  *
6678  * Imagine a situation of two groups of 4 cpus each and 4 tasks each with a
6679  * cpumask covering 1 cpu of the first group and 3 cpus of the second group.
6680  * Something like:
6681  *
6682  *      { 0 1 2 3 } { 4 5 6 7 }
6683  *              *     * * *
6684  *
6685  * If we were to balance group-wise we'd place two tasks in the first group and
6686  * two tasks in the second group. Clearly this is undesired as it will overload
6687  * cpu 3 and leave one of the cpus in the second group unused.
6688  *
6689  * The current solution to this issue is detecting the skew in the first group
6690  * by noticing the lower domain failed to reach balance and had difficulty
6691  * moving tasks due to affinity constraints.
6692  *
6693  * When this is so detected; this group becomes a candidate for busiest; see
6694  * update_sd_pick_busiest(). And calculate_imbalance() and
6695  * find_busiest_group() avoid some of the usual balance conditions to allow it
6696  * to create an effective group imbalance.
6697  *
6698  * This is a somewhat tricky proposition since the next run might not find the
6699  * group imbalance and decide the groups need to be balanced again. A most
6700  * subtle and fragile situation.
6701  */
6702 
6703 static inline int sg_imbalanced(struct sched_group *group)
6704 {
6705         return group->sgc->imbalance;
6706 }
6707 
6708 /*
6709  * group_has_capacity returns true if the group has spare capacity that could
6710  * be used by some tasks.
6711  * We consider that a group has spare capacity if the  * number of task is
6712  * smaller than the number of CPUs or if the utilization is lower than the
6713  * available capacity for CFS tasks.
6714  * For the latter, we use a threshold to stabilize the state, to take into
6715  * account the variance of the tasks' load and to return true if the available
6716  * capacity in meaningful for the load balancer.
6717  * As an example, an available capacity of 1% can appear but it doesn't make
6718  * any benefit for the load balance.
6719  */
6720 static inline bool
6721 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
6722 {
6723         if (sgs->sum_nr_running < sgs->group_weight)
6724                 return true;
6725 
6726         if ((sgs->group_capacity * 100) >
6727                         (sgs->group_util * env->sd->imbalance_pct))
6728                 return true;
6729 
6730         return false;
6731 }
6732 
6733 /*
6734  *  group_is_overloaded returns true if the group has more tasks than it can
6735  *  handle.
6736  *  group_is_overloaded is not equals to !group_has_capacity because a group
6737  *  with the exact right number of tasks, has no more spare capacity but is not
6738  *  overloaded so both group_has_capacity and group_is_overloaded return
6739  *  false.
6740  */
6741 static inline bool
6742 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
6743 {
6744         if (sgs->sum_nr_running <= sgs->group_weight)
6745                 return false;
6746 
6747         if ((sgs->group_capacity * 100) <
6748                         (sgs->group_util * env->sd->imbalance_pct))
6749                 return true;
6750 
6751         return false;
6752 }
6753 
6754 static inline enum
6755 group_type group_classify(struct sched_group *group,
6756                           struct sg_lb_stats *sgs)
6757 {
6758         if (sgs->group_no_capacity)
6759                 return group_overloaded;
6760 
6761         if (sg_imbalanced(group))
6762                 return group_imbalanced;
6763 
6764         return group_other;
6765 }
6766 
6767 /**
6768  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
6769  * @env: The load balancing environment.
6770  * @group: sched_group whose statistics are to be updated.
6771  * @load_idx: Load index of sched_domain of this_cpu for load calc.
6772  * @local_group: Does group contain this_cpu.
6773  * @sgs: variable to hold the statistics for this group.
6774  * @overload: Indicate more than one runnable task for any CPU.
6775  */
6776 static inline void update_sg_lb_stats(struct lb_env *env,
6777                         struct sched_group *group, int load_idx,
6778                         int local_group, struct sg_lb_stats *sgs,
6779                         bool *overload)
6780 {
6781         unsigned long load;
6782         int i, nr_running;
6783 
6784         memset(sgs, 0, sizeof(*sgs));
6785 
6786         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
6787                 struct rq *rq = cpu_rq(i);
6788 
6789                 /* Bias balancing toward cpus of our domain */
6790                 if (local_group)
6791                         load = target_load(i, load_idx);
6792                 else
6793                         load = source_load(i, load_idx);
6794 
6795                 sgs->group_load += load;
6796                 sgs->group_util += cpu_util(i);
6797                 sgs->sum_nr_running += rq->cfs.h_nr_running;
6798 
6799                 nr_running = rq->nr_running;
6800                 if (nr_running > 1)
6801                         *overload = true;
6802 
6803 #ifdef CONFIG_NUMA_BALANCING
6804                 sgs->nr_numa_running += rq->nr_numa_running;
6805                 sgs->nr_preferred_running += rq->nr_preferred_running;
6806 #endif
6807                 sgs->sum_weighted_load += weighted_cpuload(i);
6808                 /*
6809                  * No need to call idle_cpu() if nr_running is not 0
6810                  */
6811                 if (!nr_running && idle_cpu(i))
6812                         sgs->idle_cpus++;
6813         }
6814 
6815         /* Adjust by relative CPU capacity of the group */
6816         sgs->group_capacity = group->sgc->capacity;
6817         sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
6818 
6819         if (sgs->sum_nr_running)
6820                 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
6821 
6822         sgs->group_weight = group->group_weight;
6823 
6824         sgs->group_no_capacity = group_is_overloaded(env, sgs);
6825         sgs->group_type = group_classify(group, sgs);
6826 }
6827 
6828 /**
6829  * update_sd_pick_busiest - return 1 on busiest group
6830  * @env: The load balancing environment.
6831  * @sds: sched_domain statistics
6832  * @sg: sched_group candidate to be checked for being the busiest
6833  * @sgs: sched_group statistics
6834  *
6835  * Determine if @sg is a busier group than the previously selected
6836  * busiest group.
6837  *
6838  * Return: %true if @sg is a busier group than the previously selected
6839  * busiest group. %false otherwise.
6840  */
6841 static bool update_sd_pick_busiest(struct lb_env *env,
6842                                    struct sd_lb_stats *sds,
6843                                    struct sched_group *sg,
6844                                    struct sg_lb_stats *sgs)
6845 {
6846         struct sg_lb_stats *busiest = &sds->busiest_stat;
6847 
6848         if (sgs->group_type > busiest->group_type)
6849                 return true;
6850 
6851         if (sgs->group_type < busiest->group_type)
6852                 return false;
6853 
6854         if (sgs->avg_load <= busiest->avg_load)
6855                 return false;
6856 
6857         /* This is the busiest node in its class. */
6858         if (!(env->sd->flags & SD_ASYM_PACKING))
6859                 return true;
6860 
6861         /* No ASYM_PACKING if target cpu is already busy */
6862         if (env->idle == CPU_NOT_IDLE)
6863                 return true;
6864         /*
6865          * ASYM_PACKING needs to move all the work to the lowest
6866          * numbered CPUs in the group, therefore mark all groups
6867          * higher than ourself as busy.
6868          */
6869         if (sgs->sum_nr_running && env->dst_cpu < group_first_cpu(sg)) {
6870                 if (!sds->busiest)
6871                         return true;
6872 
6873                 /* Prefer to move from highest possible cpu's work */
6874                 if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
6875                         return true;
6876         }
6877 
6878         return false;
6879 }
6880 
6881 #ifdef CONFIG_NUMA_BALANCING
6882 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6883 {
6884         if (sgs->sum_nr_running > sgs->nr_numa_running)
6885                 return regular;
6886         if (sgs->sum_nr_running > sgs->nr_preferred_running)
6887                 return remote;
6888         return all;
6889 }
6890 
6891 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6892 {
6893         if (rq->nr_running > rq->nr_numa_running)
6894                 return regular;
6895         if (rq->nr_running > rq->nr_preferred_running)
6896                 return remote;
6897         return all;
6898 }
6899 #else
6900 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
6901 {
6902         return all;
6903 }
6904 
6905 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
6906 {
6907         return regular;
6908 }
6909 #endif /* CONFIG_NUMA_BALANCING */
6910 
6911 /**
6912  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
6913  * @env: The load balancing environment.
6914  * @sds: variable to hold the statistics for this sched_domain.
6915  */
6916 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
6917 {
6918         struct sched_domain *child = env->sd->child;
6919         struct sched_group *sg = env->sd->groups;
6920         struct sg_lb_stats tmp_sgs;
6921         int load_idx, prefer_sibling = 0;
6922         bool overload = false;
6923 
6924         if (child && child->flags & SD_PREFER_SIBLING)
6925                 prefer_sibling = 1;
6926 
6927         load_idx = get_sd_load_idx(env->sd, env->idle);
6928 
6929         do {
6930                 struct sg_lb_stats *sgs = &tmp_sgs;
6931                 int local_group;
6932 
6933                 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_cpus(sg));
6934                 if (local_group) {
6935                         sds->local = sg;
6936                         sgs = &sds->local_stat;
6937 
6938                         if (env->idle != CPU_NEWLY_IDLE ||
6939                             time_after_eq(jiffies, sg->sgc->next_update))
6940                                 update_group_capacity(env->sd, env->dst_cpu);
6941                 }
6942 
6943                 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
6944                                                 &overload);
6945 
6946                 if (local_group)
6947                         goto next_group;
6948 
6949                 /*
6950                  * In case the child domain prefers tasks go to siblings
6951                  * first, lower the sg capacity so that we'll try
6952                  * and move all the excess tasks away. We lower the capacity
6953                  * of a group only if the local group has the capacity to fit
6954                  * these excess tasks. The extra check prevents the case where
6955                  * you always pull from the heaviest group when it is already
6956                  * under-utilized (possible with a large weight task outweighs
6957                  * the tasks on the system).
6958                  */
6959                 if (prefer_sibling && sds->local &&
6960                     group_has_capacity(env, &sds->local_stat) &&
6961                     (sgs->sum_nr_running > 1)) {
6962                         sgs->group_no_capacity = 1;
6963                         sgs->group_type = group_classify(sg, sgs);
6964                 }
6965 
6966                 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
6967                         sds->busiest = sg;
6968                         sds->busiest_stat = *sgs;
6969                 }
6970 
6971 next_group:
6972                 /* Now, start updating sd_lb_stats */
6973                 sds->total_load += sgs->group_load;
6974                 sds->total_capacity += sgs->group_capacity;
6975 
6976                 sg = sg->next;
6977         } while (sg != env->sd->groups);
6978 
6979         if (env->sd->flags & SD_NUMA)
6980                 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
6981 
6982         if (!env->sd->parent) {
6983                 /* update overload indicator if we are at root domain */
6984                 if (env->dst_rq->rd->overload != overload)
6985                         env->dst_rq->rd->overload = overload;
6986         }
6987 
6988 }
6989 
6990 /**
6991  * check_asym_packing - Check to see if the group is packed into the
6992  *                      sched doman.
6993  *
6994  * This is primarily intended to used at the sibling level.  Some
6995  * cores like POWER7 prefer to use lower numbered SMT threads.  In the
6996  * case of POWER7, it can move to lower SMT modes only when higher
6997  * threads are idle.  When in lower SMT modes, the threads will
6998  * perform better since they share less core resources.  Hence when we
6999  * have idle threads, we want them to be the higher ones.
7000  *
7001  * This packing function is run on idle threads.  It checks to see if
7002  * the busiest CPU in this domain (core in the P7 case) has a higher
7003  * CPU number than the packing function is being run on.  Here we are
7004  * assuming lower CPU number will be equivalent to lower a SMT thread
7005  * number.
7006  *
7007  * Return: 1 when packing is required and a task should be moved to
7008  * this CPU.  The amount of the imbalance is returned in *imbalance.
7009  *
7010  * @env: The load balancing environment.
7011  * @sds: Statistics of the sched_domain which is to be packed
7012  */
7013 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
7014 {
7015         int busiest_cpu;
7016 
7017         if (!(env->sd->flags & SD_ASYM_PACKING))
7018                 return 0;
7019 
7020         if (env->idle == CPU_NOT_IDLE)
7021                 return 0;
7022 
7023         if (!sds->busiest)
7024                 return 0;
7025 
7026         busiest_cpu = group_first_cpu(sds->busiest);
7027         if (env->dst_cpu > busiest_cpu)
7028                 return 0;
7029 
7030         env->imbalance = DIV_ROUND_CLOSEST(
7031                 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
7032                 SCHED_CAPACITY_SCALE);
7033 
7034         return 1;
7035 }
7036 
7037 /**
7038  * fix_small_imbalance - Calculate the minor imbalance that exists
7039  *                      amongst the groups of a sched_domain, during
7040  *                      load balancing.
7041  * @env: The load balancing environment.
7042  * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
7043  */
7044 static inline
7045 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7046 {
7047         unsigned long tmp, capa_now = 0, capa_move = 0;
7048         unsigned int imbn = 2;
7049         unsigned long scaled_busy_load_per_task;
7050         struct sg_lb_stats *local, *busiest;
7051 
7052         local = &sds->local_stat;
7053         busiest = &sds->busiest_stat;
7054 
7055         if (!local->sum_nr_running)
7056                 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
7057         else if (busiest->load_per_task > local->load_per_task)
7058                 imbn = 1;
7059 
7060         scaled_busy_load_per_task =
7061                 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7062                 busiest->group_capacity;
7063 
7064         if (busiest->avg_load + scaled_busy_load_per_task >=
7065             local->avg_load + (scaled_busy_load_per_task * imbn)) {
7066                 env->imbalance = busiest->load_per_task;
7067                 return;
7068         }
7069 
7070         /*
7071          * OK, we don't have enough imbalance to justify moving tasks,
7072          * however we may be able to increase total CPU capacity used by
7073          * moving them.
7074          */
7075 
7076         capa_now += busiest->group_capacity *
7077                         min(busiest->load_per_task, busiest->avg_load);
7078         capa_now += local->group_capacity *
7079                         min(local->load_per_task, local->avg_load);
7080         capa_now /= SCHED_CAPACITY_SCALE;
7081 
7082         /* Amount of load we'd subtract */
7083         if (busiest->avg_load > scaled_busy_load_per_task) {
7084                 capa_move += busiest->group_capacity *
7085                             min(busiest->load_per_task,
7086                                 busiest->avg_load - scaled_busy_load_per_task);
7087         }
7088 
7089         /* Amount of load we'd add */
7090         if (busiest->avg_load * busiest->group_capacity <
7091             busiest->load_per_task * SCHED_CAPACITY_SCALE) {
7092                 tmp = (busiest->avg_load * busiest->group_capacity) /
7093                       local->group_capacity;
7094         } else {
7095                 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
7096                       local->group_capacity;
7097         }
7098         capa_move += local->group_capacity *
7099                     min(local->load_per_task, local->avg_load + tmp);
7100         capa_move /= SCHED_CAPACITY_SCALE;
7101 
7102         /* Move if we gain throughput */
7103         if (capa_move > capa_now)
7104                 env->imbalance = busiest->load_per_task;
7105 }
7106 
7107 /**
7108  * calculate_imbalance - Calculate the amount of imbalance present within the
7109  *                       groups of a given sched_domain during load balance.
7110  * @env: load balance environment
7111  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
7112  */
7113 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
7114 {
7115         unsigned long max_pull, load_above_capacity = ~0UL;
7116         struct sg_lb_stats *local, *busiest;
7117 
7118         local = &sds->local_stat;
7119         busiest = &sds->busiest_stat;
7120 
7121         if (busiest->group_type == group_imbalanced) {
7122                 /*
7123                  * In the group_imb case we cannot rely on group-wide averages
7124                  * to ensure cpu-load equilibrium, look at wider averages. XXX
7125                  */
7126                 busiest->load_per_task =
7127                         min(busiest->load_per_task, sds->avg_load);
7128         }
7129 
7130         /*
7131          * Avg load of busiest sg can be less and avg load of local sg can
7132          * be greater than avg load across all sgs of sd because avg load
7133          * factors in sg capacity and sgs with smaller group_type are
7134          * skipped when updating the busiest sg:
7135          */
7136         if (busiest->avg_load <= sds->avg_load ||
7137             local->avg_load >= sds->avg_load) {
7138                 env->imbalance = 0;
7139                 return fix_small_imbalance(env, sds);
7140         }
7141 
7142         /*
7143          * If there aren't any idle cpus, avoid creating some.
7144          */
7145         if (busiest->group_type == group_overloaded &&
7146             local->group_type   == group_overloaded) {
7147                 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
7148                 if (load_above_capacity > busiest->group_capacity) {
7149                         load_above_capacity -= busiest->group_capacity;
7150                         load_above_capacity *= NICE_0_LOAD;
7151                         load_above_capacity /= busiest->group_capacity;
7152                 } else
7153                         load_above_capacity = ~0UL;
7154         }
7155 
7156         /*
7157          * We're trying to get all the cpus to the average_load, so we don't
7158          * want to push ourselves above the average load, nor do we wish to
7159          * reduce the max loaded cpu below the average load. At the same time,
7160          * we also don't want to reduce the group load below the group
7161          * capacity. Thus we look for the minimum possible imbalance.
7162          */
7163         max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
7164 
7165         /* How much load to actually move to equalise the imbalance */
7166         env->imbalance = min(
7167                 max_pull * busiest->group_capacity,
7168                 (sds->avg_load - local->avg_load) * local->group_capacity
7169         ) / SCHED_CAPACITY_SCALE;
7170 
7171         /*
7172          * if *imbalance is less than the average load per runnable task
7173          * there is no guarantee that any tasks will be moved so we'll have
7174          * a think about bumping its value to force at least one task to be
7175          * moved
7176          */
7177         if (env->imbalance < busiest->load_per_task)
7178                 return fix_small_imbalance(env, sds);
7179 }
7180 
7181 /******* find_busiest_group() helpers end here *********************/
7182 
7183 /**
7184  * find_busiest_group - Returns the busiest group within the sched_domain
7185  * if there is an imbalance.
7186  *
7187  * Also calculates the amount of weighted load which should be moved
7188  * to restore balance.
7189  *
7190  * @env: The load balancing environment.
7191  *
7192  * Return:      - The busiest group if imbalance exists.
7193  */
7194 static struct sched_group *find_busiest_group(struct lb_env *env)
7195 {
7196         struct sg_lb_stats *local, *busiest;
7197         struct sd_lb_stats sds;
7198 
7199         init_sd_lb_stats(&sds);
7200 
7201         /*
7202          * Compute the various statistics relavent for load balancing at
7203          * this level.
7204          */
7205         update_sd_lb_stats(env, &sds);
7206         local = &sds.local_stat;
7207         busiest = &sds.busiest_stat;
7208 
7209         /* ASYM feature bypasses nice load balance check */
7210         if (check_asym_packing(env, &sds))
7211                 return sds.busiest;
7212 
7213         /* There is no busy sibling group to pull tasks from */
7214         if (!sds.busiest || busiest->sum_nr_running == 0)
7215                 goto out_balanced;
7216 
7217         sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
7218                                                 / sds.total_capacity;
7219 
7220         /*
7221          * If the busiest group is imbalanced the below checks don't
7222          * work because they assume all things are equal, which typically
7223          * isn't true due to cpus_allowed constraints and the like.
7224          */
7225         if (busiest->group_type == group_imbalanced)
7226                 goto force_balance;
7227 
7228         /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
7229         if (env->idle == CPU_NEWLY_IDLE && group_has_capacity(env, local) &&
7230             busiest->group_no_capacity)
7231                 goto force_balance;
7232 
7233         /*
7234          * If the local group is busier than the selected busiest group
7235          * don't try and pull any tasks.
7236          */
7237         if (local->avg_load >= busiest->avg_load)
7238                 goto out_balanced;
7239 
7240         /*
7241          * Don't pull any tasks if this group is already above the domain
7242          * average load.
7243          */
7244         if (local->avg_load >= sds.avg_load)
7245                 goto out_balanced;
7246 
7247         if (env->idle == CPU_IDLE) {
7248                 /*
7249                  * This cpu is idle. If the busiest group is not overloaded
7250                  * and there is no imbalance between this and busiest group
7251                  * wrt idle cpus, it is balanced. The imbalance becomes
7252                  * significant if the diff is greater than 1 otherwise we
7253                  * might end up to just move the imbalance on another group
7254                  */
7255                 if ((busiest->group_type != group_overloaded) &&
7256                                 (local->idle_cpus <= (busiest->idle_cpus + 1)))
7257                         goto out_balanced;
7258         } else {
7259                 /*
7260                  * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
7261                  * imbalance_pct to be conservative.
7262                  */
7263                 if (100 * busiest->avg_load <=
7264                                 env->sd->imbalance_pct * local->avg_load)
7265                         goto out_balanced;
7266         }
7267 
7268 force_balance:
7269         /* Looks like there is an imbalance. Compute it */
7270         calculate_imbalance(env, &sds);
7271         return sds.busiest;
7272 
7273 out_balanced:
7274         env->imbalance = 0;
7275         return NULL;
7276 }
7277 
7278 /*
7279  * find_busiest_queue - find the busiest runqueue among the cpus in group.
7280  */
7281 static struct rq *find_busiest_queue(struct lb_env *env,
7282                                      struct sched_group *group)
7283 {
7284         struct rq *busiest = NULL, *rq;
7285         unsigned long busiest_load = 0, busiest_capacity = 1;
7286         int i;
7287 
7288         for_each_cpu_and(i, sched_group_cpus(group), env->cpus) {
7289                 unsigned long capacity, wl;
7290                 enum fbq_type rt;
7291 
7292                 rq = cpu_rq(i);
7293                 rt = fbq_classify_rq(rq);
7294 
7295                 /*
7296                  * We classify groups/runqueues into three groups:
7297                  *  - regular: there are !numa tasks
7298                  *  - remote:  there are numa tasks that run on the 'wrong' node
7299                  *  - all:     there is no distinction
7300                  *
7301                  * In order to avoid migrating ideally placed numa tasks,
7302                  * ignore those when there's better options.
7303                  *
7304                  * If we ignore the actual busiest queue to migrate another
7305                  * task, the next balance pass can still reduce the busiest
7306                  * queue by moving tasks around inside the node.
7307                  *
7308                  * If we cannot move enough load due to this classification
7309                  * the next pass will adjust the group classification and
7310                  * allow migration of more tasks.
7311                  *
7312                  * Both cases only affect the total convergence complexity.
7313                  */
7314                 if (rt > env->fbq_type)
7315                         continue;
7316 
7317                 capacity = capacity_of(i);
7318 
7319                 wl = weighted_cpuload(i);
7320 
7321                 /*
7322                  * When comparing with imbalance, use weighted_cpuload()
7323                  * which is not scaled with the cpu capacity.
7324                  */
7325 
7326                 if (rq->nr_running == 1 && wl > env->imbalance &&
7327                     !check_cpu_capacity(rq, env->sd))
7328                         continue;
7329 
7330                 /*
7331                  * For the load comparisons with the other cpu's, consider
7332                  * the weighted_cpuload() scaled with the cpu capacity, so
7333                  * that the load can be moved away from the cpu that is
7334                  * potentially running at a lower capacity.
7335                  *
7336                  * Thus we're looking for max(wl_i / capacity_i), crosswise
7337                  * multiplication to rid ourselves of the division works out
7338                  * to: wl_i * capacity_j > wl_j * capacity_i;  where j is
7339                  * our previous maximum.
7340                  */
7341                 if (wl * busiest_capacity > busiest_load * capacity) {
7342                         busiest_load = wl;
7343                         busiest_capacity = capacity;
7344                         busiest = rq;
7345                 }
7346         }
7347 
7348         return busiest;
7349 }
7350 
7351 /*
7352  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
7353  * so long as it is large enough.
7354  */
7355 #define MAX_PINNED_INTERVAL     512
7356 
7357 /* Working cpumask for load_balance and load_balance_newidle. */
7358 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
7359 
7360 static int need_active_balance(struct lb_env *env)
7361 {
7362         struct sched_domain *sd = env->sd;
7363 
7364         if (env->idle == CPU_NEWLY_IDLE) {
7365 
7366                 /*
7367                  * ASYM_PACKING needs to force migrate tasks from busy but
7368                  * higher numbered CPUs in order to pack all tasks in the
7369                  * lowest numbered CPUs.
7370                  */
7371                 if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu)
7372                         return 1;
7373         }
7374 
7375         /*
7376          * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
7377          * It's worth migrating the task if the src_cpu's capacity is reduced
7378          * because of other sched_class or IRQs if more capacity stays
7379          * available on dst_cpu.
7380          */
7381         if ((env->idle != CPU_NOT_IDLE) &&
7382             (env->src_rq->cfs.h_nr_running == 1)) {
7383                 if ((check_cpu_capacity(env->src_rq, sd)) &&
7384                     (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
7385                         return 1;
7386         }
7387 
7388         return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
7389 }
7390 
7391 static int active_load_balance_cpu_stop(void *data);
7392 
7393 static int should_we_balance(struct lb_env *env)
7394 {
7395         struct sched_group *sg = env->sd->groups;
7396         struct cpumask *sg_cpus, *sg_mask;
7397         int cpu, balance_cpu = -1;
7398 
7399         /*
7400          * In the newly idle case, we will allow all the cpu's
7401          * to do the newly idle load balance.
7402          */
7403         if (env->idle == CPU_NEWLY_IDLE)
7404                 return 1;
7405 
7406         sg_cpus = sched_group_cpus(sg);
7407         sg_mask = sched_group_mask(sg);
7408         /* Try to find first idle cpu */
7409         for_each_cpu_and(cpu, sg_cpus, env->cpus) {
7410                 if (!cpumask_test_cpu(cpu, sg_mask) || !idle_cpu(cpu))
7411                         continue;
7412 
7413                 balance_cpu = cpu;
7414                 break;
7415         }
7416 
7417         if (balance_cpu == -1)
7418                 balance_cpu = group_balance_cpu(sg);
7419 
7420         /*
7421          * First idle cpu or the first cpu(busiest) in this sched group
7422          * is eligible for doing load balancing at this and above domains.
7423          */
7424         return balance_cpu == env->dst_cpu;
7425 }
7426 
7427 /*
7428  * Check this_cpu to ensure it is balanced within domain. Attempt to move
7429  * tasks if there is an imbalance.
7430  */
7431 static int load_balance(int this_cpu, struct rq *this_rq,
7432                         struct sched_domain *sd, enum cpu_idle_type idle,
7433                         int *continue_balancing)
7434 {
7435         int ld_moved, cur_ld_moved, active_balance = 0;
7436         struct sched_domain *sd_parent = sd->parent;
7437         struct sched_group *group;
7438         struct rq *busiest;
7439         unsigned long flags;
7440         struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
7441 
7442         struct lb_env env = {
7443                 .sd             = sd,
7444                 .dst_cpu        = this_cpu,
7445                 .dst_rq         = this_rq,
7446                 .dst_grpmask    = sched_group_cpus(sd->groups),
7447                 .idle           = idle,
7448                 .loop_break     = sched_nr_migrate_break,
7449                 .cpus           = cpus,
7450                 .fbq_type       = all,
7451                 .tasks          = LIST_HEAD_INIT(env.tasks),
7452         };
7453 
7454         /*
7455          * For NEWLY_IDLE load_balancing, we don't need to consider
7456          * other cpus in our group
7457          */
7458         if (idle == CPU_NEWLY_IDLE)
7459                 env.dst_grpmask = NULL;
7460 
7461         cpumask_copy(cpus, cpu_active_mask);
7462 
7463         schedstat_inc(sd, lb_count[idle]);
7464 
7465 redo:
7466         if (!should_we_balance(&env)) {
7467                 *continue_balancing = 0;
7468                 goto out_balanced;
7469         }
7470 
7471         group = find_busiest_group(&env);
7472         if (!group) {
7473                 schedstat_inc(sd, lb_nobusyg[idle]);
7474                 goto out_balanced;
7475         }
7476 
7477         busiest = find_busiest_queue(&env, group);
7478         if (!busiest) {
7479                 schedstat_inc(sd, lb_nobusyq[idle]);
7480                 goto out_balanced;
7481         }
7482 
7483         BUG_ON(busiest == env.dst_rq);
7484 
7485         schedstat_add(sd, lb_imbalance[idle], env.imbalance);
7486 
7487         env.src_cpu = busiest->cpu;
7488         env.src_rq = busiest;
7489 
7490         ld_moved = 0;
7491         if (busiest->nr_running > 1) {
7492                 /*
7493                  * Attempt to move tasks. If find_busiest_group has found
7494                  * an imbalance but busiest->nr_running <= 1, the group is
7495                  * still unbalanced. ld_moved simply stays zero, so it is
7496                  * correctly treated as an imbalance.
7497                  */
7498                 env.flags |= LBF_ALL_PINNED;
7499                 env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
7500 
7501 more_balance:
7502                 raw_spin_lock_irqsave(&busiest->lock, flags);
7503 
7504                 /*
7505                  * cur_ld_moved - load moved in current iteration
7506                  * ld_moved     - cumulative load moved across iterations
7507                  */
7508                 cur_ld_moved = detach_tasks(&env);
7509 
7510                 /*
7511                  * We've detached some tasks from busiest_rq. Every
7512                  * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
7513                  * unlock busiest->lock, and we are able to be sure
7514                  * that nobody can manipulate the tasks in parallel.
7515                  * See task_rq_lock() family for the details.
7516                  */
7517 
7518                 raw_spin_unlock(&busiest->lock);
7519 
7520                 if (cur_ld_moved) {
7521                         attach_tasks(&env);
7522                         ld_moved += cur_ld_moved;
7523                 }
7524 
7525                 local_irq_restore(flags);
7526 
7527                 if (env.flags & LBF_NEED_BREAK) {
7528                         env.flags &= ~LBF_NEED_BREAK;
7529                         goto more_balance;
7530                 }
7531 
7532                 /*
7533                  * Revisit (affine) tasks on src_cpu that couldn't be moved to
7534                  * us and move them to an alternate dst_cpu in our sched_group
7535                  * where they can run. The upper limit on how many times we
7536                  * iterate on same src_cpu is dependent on number of cpus in our
7537                  * sched_group.
7538                  *
7539                  * This changes load balance semantics a bit on who can move
7540                  * load to a given_cpu. In addition to the given_cpu itself
7541                  * (or a ilb_cpu acting on its behalf where given_cpu is
7542                  * nohz-idle), we now have balance_cpu in a position to move
7543                  * load to given_cpu. In rare situations, this may cause
7544                  * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
7545                  * _independently_ and at _same_ time to move some load to
7546                  * given_cpu) causing exceess load to be moved to given_cpu.
7547                  * This however should not happen so much in practice and
7548                  * moreover subsequent load balance cycles should correct the
7549                  * excess load moved.
7550                  */
7551                 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
7552 
7553                         /* Prevent to re-select dst_cpu via env's cpus */
7554                         cpumask_clear_cpu(env.dst_cpu, env.cpus);
7555 
7556                         env.dst_rq       = cpu_rq(env.new_dst_cpu);
7557                         env.dst_cpu      = env.new_dst_cpu;
7558                         env.flags       &= ~LBF_DST_PINNED;
7559                         env.loop         = 0;
7560                         env.loop_break   = sched_nr_migrate_break;
7561 
7562                         /*
7563                          * Go back to "more_balance" rather than "redo" since we
7564                          * need to continue with same src_cpu.
7565                          */
7566                         goto more_balance;
7567                 }
7568 
7569                 /*
7570                  * We failed to reach balance because of affinity.
7571                  */
7572                 if (sd_parent) {
7573                         int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7574 
7575                         if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
7576                                 *group_imbalance = 1;
7577                 }
7578 
7579                 /* All tasks on this runqueue were pinned by CPU affinity */
7580                 if (unlikely(env.flags & LBF_ALL_PINNED)) {
7581                         cpumask_clear_cpu(cpu_of(busiest), cpus);
7582                         if (!cpumask_empty(cpus)) {
7583                                 env.loop = 0;
7584                                 env.loop_break = sched_nr_migrate_break;
7585                                 goto redo;
7586                         }
7587                         goto out_all_pinned;
7588                 }
7589         }
7590 
7591         if (!ld_moved) {
7592                 schedstat_inc(sd, lb_failed[idle]);
7593                 /*
7594                  * Increment the failure counter only on periodic balance.
7595                  * We do not want newidle balance, which can be very
7596                  * frequent, pollute the failure counter causing
7597                  * excessive cache_hot migrations and active balances.
7598                  */
7599                 if (idle != CPU_NEWLY_IDLE)
7600                         sd->nr_balance_failed++;
7601 
7602                 if (need_active_balance(&env)) {
7603                         raw_spin_lock_irqsave(&busiest->lock, flags);
7604 
7605                         /* don't kick the active_load_balance_cpu_stop,
7606                          * if the curr task on busiest cpu can't be
7607                          * moved to this_cpu
7608                          */
7609                         if (!cpumask_test_cpu(this_cpu,
7610                                         tsk_cpus_allowed(busiest->curr))) {
7611                                 raw_spin_unlock_irqrestore(&busiest->lock,
7612                                                             flags);
7613                                 env.flags |= LBF_ALL_PINNED;
7614                                 goto out_one_pinned;
7615                         }
7616 
7617                         /*
7618                          * ->active_balance synchronizes accesses to
7619                          * ->active_balance_work.  Once set, it's cleared
7620                          * only after active load balance is finished.
7621                          */
7622                         if (!busiest->active_balance) {
7623                                 busiest->active_balance = 1;
7624                                 busiest->push_cpu = this_cpu;
7625                                 active_balance = 1;
7626                         }
7627                         raw_spin_unlock_irqrestore(&busiest->lock, flags);
7628 
7629                         if (active_balance) {
7630                                 stop_one_cpu_nowait(cpu_of(busiest),
7631                                         active_load_balance_cpu_stop, busiest,
7632                                         &busiest->active_balance_work);
7633                         }
7634 
7635                         /* We've kicked active balancing, force task migration. */
7636                         sd->nr_balance_failed = sd->cache_nice_tries+1;
7637                 }
7638         } else
7639                 sd->nr_balance_failed = 0;
7640 
7641         if (likely(!active_balance)) {
7642                 /* We were unbalanced, so reset the balancing interval */
7643                 sd->balance_interval = sd->min_interval;
7644         } else {
7645                 /*
7646                  * If we've begun active balancing, start to back off. This
7647                  * case may not be covered by the all_pinned logic if there
7648                  * is only 1 task on the busy runqueue (because we don't call
7649                  * detach_tasks).
7650                  */
7651                 if (sd->balance_interval < sd->max_interval)
7652                         sd->balance_interval *= 2;
7653         }
7654 
7655         goto out;
7656 
7657 out_balanced:
7658         /*
7659          * We reach balance although we may have faced some affinity
7660          * constraints. Clear the imbalance flag if it was set.
7661          */
7662         if (sd_parent) {
7663                 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
7664 
7665                 if (*group_imbalance)
7666                         *group_imbalance = 0;
7667         }
7668 
7669 out_all_pinned:
7670         /*
7671          * We reach balance because all tasks are pinned at this level so
7672          * we can't migrate them. Let the imbalance flag set so parent level
7673          * can try to migrate them.
7674          */
7675         schedstat_inc(sd, lb_balanced[idle]);
7676 
7677         sd->nr_balance_failed = 0;
7678 
7679 out_one_pinned:
7680         /* tune up the balancing interval */
7681         if (((env.flags & LBF_ALL_PINNED) &&
7682                         sd->balance_interval < MAX_PINNED_INTERVAL) ||
7683                         (sd->balance_interval < sd->max_interval))
7684                 sd->balance_interval *= 2;
7685 
7686         ld_moved = 0;
7687 out:
7688         return ld_moved;
7689 }
7690 
7691 static inline unsigned long
7692 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
7693 {
7694         unsigned long interval = sd->balance_interval;
7695 
7696         if (cpu_busy)
7697                 interval *= sd->busy_factor;
7698 
7699         /* scale ms to jiffies */
7700         interval = msecs_to_jiffies(interval);
7701         interval = clamp(interval, 1UL, max_load_balance_interval);
7702 
7703         return interval;
7704 }
7705 
7706 static inline void
7707 update_next_balance(struct sched_domain *sd, int cpu_busy, unsigned long *next_balance)
7708 {
7709         unsigned long interval, next;
7710 
7711         interval = get_sd_balance_interval(sd, cpu_busy);
7712         next = sd->last_balance + interval;
7713 
7714         if (time_after(*next_balance, next))
7715                 *next_balance = next;
7716 }
7717 
7718 /*
7719  * idle_balance is called by schedule() if this_cpu is about to become
7720  * idle. Attempts to pull tasks from other CPUs.
7721  */
7722 static int idle_balance(struct rq *this_rq)
7723 {
7724         unsigned long next_balance = jiffies + HZ;
7725         int this_cpu = this_rq->cpu;
7726         struct sched_domain *sd;
7727         int pulled_task = 0;
7728         u64 curr_cost = 0;
7729 
7730         /*
7731          * We must set idle_stamp _before_ calling idle_balance(), such that we
7732          * measure the duration of idle_balance() as idle time.
7733          */
7734         this_rq->idle_stamp = rq_clock(this_rq);
7735 
7736         if (this_rq->avg_idle < sysctl_sched_migration_cost ||
7737             !this_rq->rd->overload) {
7738                 rcu_read_lock();
7739                 sd = rcu_dereference_check_sched_domain(this_rq->sd);
7740                 if (sd)
7741                         update_next_balance(sd, 0, &next_balance);
7742                 rcu_read_unlock();
7743 
7744                 goto out;
7745         }
7746 
7747         raw_spin_unlock(&this_rq->lock);
7748 
7749         update_blocked_averages(this_cpu);
7750         rcu_read_lock();
7751         for_each_domain(this_cpu, sd) {
7752                 int continue_balancing = 1;
7753                 u64 t0, domain_cost;
7754 
7755                 if (!(sd->flags & SD_LOAD_BALANCE))
7756                         continue;
7757 
7758                 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
7759                         update_next_balance(sd, 0, &next_balance);
7760                         break;
7761                 }
7762 
7763                 if (sd->flags & SD_BALANCE_NEWIDLE) {
7764                         t0 = sched_clock_cpu(this_cpu);
7765 
7766                         pulled_task = load_balance(this_cpu, this_rq,
7767                                                    sd, CPU_NEWLY_IDLE,
7768                                                    &continue_balancing);
7769 
7770                         domain_cost = sched_clock_cpu(this_cpu) - t0;
7771                         if (domain_cost > sd->max_newidle_lb_cost)
7772                                 sd->max_newidle_lb_cost = domain_cost;
7773 
7774                         curr_cost += domain_cost;
7775                 }
7776 
7777                 update_next_balance(sd, 0, &next_balance);
7778 
7779                 /*
7780                  * Stop searching for tasks to pull if there are
7781                  * now runnable tasks on this rq.
7782                  */
7783                 if (pulled_task || this_rq->nr_running > 0)
7784                         break;
7785         }
7786         rcu_read_unlock();
7787 
7788         raw_spin_lock(&this_rq->lock);
7789 
7790         if (curr_cost > this_rq->max_idle_balance_cost)
7791                 this_rq->max_idle_balance_cost = curr_cost;
7792 
7793         /*
7794          * While browsing the domains, we released the rq lock, a task could
7795          * have been enqueued in the meantime. Since we're not going idle,
7796          * pretend we pulled a task.
7797          */
7798         if (this_rq->cfs.h_nr_running && !pulled_task)
7799                 pulled_task = 1;
7800 
7801 out:
7802         /* Move the next balance forward */
7803         if (time_after(this_rq->next_balance, next_balance))
7804                 this_rq->next_balance = next_balance;
7805 
7806         /* Is there a task of a high priority class? */
7807         if (this_rq->nr_running != this_rq->cfs.h_nr_running)
7808                 pulled_task = -1;
7809 
7810         if (pulled_task)
7811                 this_rq->idle_stamp = 0;
7812 
7813         return pulled_task;
7814 }
7815 
7816 /*
7817  * active_load_balance_cpu_stop is run by cpu stopper. It pushes
7818  * running tasks off the busiest CPU onto idle CPUs. It requires at
7819  * least 1 task to be running on each physical CPU where possible, and
7820  * avoids physical / logical imbalances.
7821  */
7822 static int active_load_balance_cpu_stop(void *data)
7823 {
7824         struct rq *busiest_rq = data;
7825         int busiest_cpu = cpu_of(busiest_rq);
7826         int target_cpu = busiest_rq->push_cpu;
7827         struct rq *target_rq = cpu_rq(target_cpu);
7828         struct sched_domain *sd;
7829         struct task_struct *p = NULL;
7830 
7831         raw_spin_lock_irq(&busiest_rq->lock);
7832 
7833         /* make sure the requested cpu hasn't gone down in the meantime */
7834         if (unlikely(busiest_cpu != smp_processor_id() ||
7835                      !busiest_rq->active_balance))
7836                 goto out_unlock;
7837 
7838         /* Is there any task to move? */
7839         if (busiest_rq->nr_running <= 1)
7840                 goto out_unlock;
7841 
7842         /*
7843          * This condition is "impossible", if it occurs
7844          * we need to fix it. Originally reported by
7845          * Bjorn Helgaas on a 128-cpu setup.
7846          */
7847         BUG_ON(busiest_rq == target_rq);
7848 
7849         /* Search for an sd spanning us and the target CPU. */
7850         rcu_read_lock();
7851         for_each_domain(target_cpu, sd) {
7852                 if ((sd->flags & SD_LOAD_BALANCE) &&
7853                     cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
7854                                 break;
7855         }
7856 
7857         if (likely(sd)) {
7858                 struct lb_env env = {
7859                         .sd             = sd,
7860                         .dst_cpu        = target_cpu,
7861                         .dst_rq         = target_rq,
7862                         .src_cpu        = busiest_rq->cpu,
7863                         .src_rq         = busiest_rq,
7864                         .idle           = CPU_IDLE,
7865                 };
7866 
7867                 schedstat_inc(sd, alb_count);
7868 
7869                 p = detach_one_task(&env);
7870                 if (p) {
7871                         schedstat_inc(sd, alb_pushed);
7872                         /* Active balancing done, reset the failure counter. */
7873                         sd->nr_balance_failed = 0;
7874                 } else {
7875                         schedstat_inc(sd, alb_failed);
7876                 }
7877         }
7878         rcu_read_unlock();
7879 out_unlock:
7880         busiest_rq->active_balance = 0;
7881         raw_spin_unlock(&busiest_rq->lock);
7882 
7883         if (p)
7884                 attach_one_task(target_rq, p);
7885 
7886         local_irq_enable();
7887 
7888         return 0;
7889 }
7890 
7891 static inline int on_null_domain(struct rq *rq)
7892 {
7893         return unlikely(!rcu_dereference_sched(rq->sd));
7894 }
7895 
7896 #ifdef CONFIG_NO_HZ_COMMON
7897 /*
7898  * idle load balancing details
7899  * - When one of the busy CPUs notice that there may be an idle rebalancing
7900  *   needed, they will kick the idle load balancer, which then does idle
7901  *   load balancing for all the idle CPUs.
7902  */
7903 static struct {
7904         cpumask_var_t idle_cpus_mask;
7905         atomic_t nr_cpus;
7906         unsigned long next_balance;     /* in jiffy units */
7907 } nohz ____cacheline_aligned;
7908 
7909 static inline int find_new_ilb(void)
7910 {
7911         int ilb = cpumask_first(nohz.idle_cpus_mask);
7912 
7913         if (ilb < nr_cpu_ids && idle_cpu(ilb))
7914                 return ilb;
7915 
7916         return nr_cpu_ids;
7917 }
7918 
7919 /*
7920  * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
7921  * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
7922  * CPU (if there is one).
7923  */
7924 static void nohz_balancer_kick(void)
7925 {
7926         int ilb_cpu;
7927 
7928         nohz.next_balance++;
7929 
7930         ilb_cpu = find_new_ilb();
7931 
7932         if (ilb_cpu >= nr_cpu_ids)
7933                 return;
7934 
7935         if (test_and_set_bit(NOHZ_BALANCE_KICK, nohz_flags(ilb_cpu)))
7936                 return;
7937         /*
7938          * Use smp_send_reschedule() instead of resched_cpu().
7939          * This way we generate a sched IPI on the target cpu which
7940          * is idle. And the softirq performing nohz idle load balance
7941          * will be run before returning from the IPI.
7942          */
7943         smp_send_reschedule(ilb_cpu);
7944         return;
7945 }
7946 
7947 void nohz_balance_exit_idle(unsigned int cpu)
7948 {
7949         if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
7950                 /*
7951                  * Completely isolated CPUs don't ever set, so we must test.
7952                  */
7953                 if (likely(cpumask_test_cpu(cpu, nohz.idle_cpus_mask))) {
7954                         cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
7955                         atomic_dec(&nohz.nr_cpus);
7956                 }
7957                 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
7958         }
7959 }
7960 
7961 static inline void set_cpu_sd_state_busy(void)
7962 {
7963         struct sched_domain *sd;
7964         int cpu = smp_processor_id();
7965 
7966         rcu_read_lock();
7967         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7968 
7969         if (!sd || !sd->nohz_idle)
7970                 goto unlock;
7971         sd->nohz_idle = 0;
7972 
7973         atomic_inc(&sd->groups->sgc->nr_busy_cpus);
7974 unlock:
7975         rcu_read_unlock();
7976 }
7977 
7978 void set_cpu_sd_state_idle(void)
7979 {
7980         struct sched_domain *sd;
7981         int cpu = smp_processor_id();
7982 
7983         rcu_read_lock();
7984         sd = rcu_dereference(per_cpu(sd_busy, cpu));
7985 
7986         if (!sd || sd->nohz_idle)
7987                 goto unlock;
7988         sd->nohz_idle = 1;
7989 
7990         atomic_dec(&sd->groups->sgc->nr_busy_cpus);
7991 unlock:
7992         rcu_read_unlock();
7993 }
7994 
7995 /*
7996  * This routine will record that the cpu is going idle with tick stopped.
7997  * This info will be used in performing idle load balancing in the future.
7998  */
7999 void nohz_balance_enter_idle(int cpu)
8000 {
8001         /*
8002          * If this cpu is going down, then nothing needs to be done.
8003          */
8004         if (!cpu_active(cpu))
8005                 return;
8006 
8007         if (test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))
8008                 return;
8009 
8010         /*
8011          * If we're a completely isolated CPU, we don't play.
8012          */
8013         if (on_null_domain(cpu_rq(cpu)))
8014                 return;
8015 
8016         cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
8017         atomic_inc(&nohz.nr_cpus);
8018         set_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
8019 }
8020 #endif
8021 
8022 static DEFINE_SPINLOCK(balancing);
8023 
8024 /*
8025  * Scale the max load_balance interval with the number of CPUs in the system.
8026  * This trades load-balance latency on larger machines for less cross talk.
8027  */
8028 void update_max_interval(void)
8029 {
8030         max_load_balance_interval = HZ*num_online_cpus()/10;
8031 }
8032 
8033 /*
8034  * It checks each scheduling domain to see if it is due to be balanced,
8035  * and initiates a balancing operation if so.
8036  *
8037  * Balancing parameters are set up in init_sched_domains.
8038  */
8039 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8040 {
8041         int continue_balancing = 1;
8042         int cpu = rq->cpu;
8043         unsigned long interval;
8044         struct sched_domain *sd;
8045         /* Earliest time when we have to do rebalance again */
8046         unsigned long next_balance = jiffies + 60*HZ;
8047         int update_next_balance = 0;
8048         int need_serialize, need_decay = 0;
8049         u64 max_cost = 0;
8050 
8051         update_blocked_averages(cpu);
8052 
8053         rcu_read_lock();
8054         for_each_domain(cpu, sd) {
8055                 /*
8056                  * Decay the newidle max times here because this is a regular
8057                  * visit to all the domains. Decay ~1% per second.
8058                  */
8059                 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8060                         sd->max_newidle_lb_cost =
8061                                 (sd->max_newidle_lb_cost * 253) / 256;
8062                         sd->next_decay_max_lb_cost = jiffies + HZ;
8063                         need_decay = 1;
8064                 }
8065                 max_cost += sd->max_newidle_lb_cost;
8066 
8067                 if (!(sd->flags & SD_LOAD_BALANCE))
8068                         continue;
8069 
8070                 /*
8071                  * Stop the load balance at this level. There is another
8072                  * CPU in our sched group which is doing load balancing more
8073                  * actively.
8074                  */
8075                 if (!continue_balancing) {
8076                         if (need_decay)
8077                                 continue;
8078                         break;
8079                 }
8080 
8081                 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8082 
8083                 need_serialize = sd->flags & SD_SERIALIZE;
8084                 if (need_serialize) {
8085                         if (!spin_trylock(&balancing))
8086                                 goto out;
8087                 }
8088 
8089                 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8090                         if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8091                                 /*
8092                                  * The LBF_DST_PINNED logic could have changed
8093                                  * env->dst_cpu, so we can't know our idle
8094                                  * state even if we migrated tasks. Update it.
8095                                  */
8096                                 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8097                         }
8098                         sd->last_balance = jiffies;
8099                         interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8100                 }
8101                 if (need_serialize)
8102                         spin_unlock(&balancing);
8103 out:
8104                 if (time_after(next_balance, sd->last_balance + interval)) {
8105                         next_balance = sd->last_balance + interval;
8106                         update_next_balance = 1;
8107                 }
8108         }
8109         if (need_decay) {
8110                 /*
8111                  * Ensure the rq-wide value also decays but keep it at a
8112                  * reasonable floor to avoid funnies with rq->avg_idle.
8113                  */
8114                 rq->max_idle_balance_cost =
8115                         max((u64)sysctl_sched_migration_cost, max_cost);
8116         }
8117         rcu_read_unlock();
8118 
8119         /*
8120          * next_balance will be updated only when there is a need.
8121          * When the cpu is attached to null domain for ex, it will not be
8122          * updated.
8123          */
8124         if (likely(update_next_balance)) {
8125                 rq->next_balance = next_balance;
8126 
8127 #ifdef CONFIG_NO_HZ_COMMON
8128                 /*
8129                  * If this CPU has been elected to perform the nohz idle
8130                  * balance. Other idle CPUs have already rebalanced with
8131                  * nohz_idle_balance() and nohz.next_balance has been
8132                  * updated accordingly. This CPU is now running the idle load
8133                  * balance for itself and we need to update the
8134                  * nohz.next_balance accordingly.
8135                  */
8136                 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8137                         nohz.next_balance = rq->next_balance;
8138 #endif
8139         }
8140 }
8141 
8142 #ifdef CONFIG_NO_HZ_COMMON
8143 /*
8144  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
8145  * rebalancing for all the cpus for whom scheduler ticks are stopped.
8146  */
8147 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
8148 {
8149         int this_cpu = this_rq->cpu;
8150         struct rq *rq;
8151         int balance_cpu;
8152         /* Earliest time when we have to do rebalance again */
8153         unsigned long next_balance = jiffies + 60*HZ;
8154         int update_next_balance = 0;
8155 
8156         if (idle != CPU_IDLE ||
8157             !test_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu)))
8158                 goto end;
8159 
8160         for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
8161                 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
8162                         continue;
8163 
8164                 /*
8165                  * If this cpu gets work to do, stop the load balancing
8166                  * work being done for other cpus. Next load
8167                  * balancing owner will pick it up.
8168                  */
8169                 if (need_resched())
8170                         break;
8171 
8172                 rq = cpu_rq(balance_cpu);
8173 
8174                 /*
8175                  * If time for next balance is due,
8176                  * do the balance.
8177                  */
8178                 if (time_after_eq(jiffies, rq->next_balance)) {
8179                         raw_spin_lock_irq(&rq->lock);
8180                         update_rq_clock(rq);
8181                         cpu_load_update_idle(rq);
8182                         raw_spin_unlock_irq(&rq->lock);
8183                         rebalance_domains(rq, CPU_IDLE);
8184                 }
8185 
8186                 if (time_after(next_balance, rq->next_balance)) {
8187                         next_balance = rq->next_balance;
8188                         update_next_balance = 1;
8189                 }
8190         }
8191 
8192         /*
8193          * next_balance will be updated only when there is a need.
8194          * When the CPU is attached to null domain for ex, it will not be
8195          * updated.
8196          */
8197         if (likely(update_next_balance))
8198                 nohz.next_balance = next_balance;
8199 end:
8200         clear_bit(NOHZ_BALANCE_KICK, nohz_flags(this_cpu));
8201 }
8202 
8203 /*
8204  * Current heuristic for kicking the idle load balancer in the presence
8205  * of an idle cpu in the system.
8206  *   - This rq has more than one task.
8207  *   - This rq has at least one CFS task and the capacity of the CPU is
8208  *     significantly reduced because of RT tasks or IRQs.
8209  *   - At parent of LLC scheduler domain level, this cpu's scheduler group has
8210  *     multiple busy cpu.
8211  *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
8212  *     domain span are idle.
8213  */
8214 static inline bool nohz_kick_needed(struct rq *rq)
8215 {
8216         unsigned long now = jiffies;
8217         struct sched_domain *sd;
8218         struct sched_group_capacity *sgc;
8219         int nr_busy, cpu = rq->cpu;
8220         bool kick = false;
8221 
8222         if (unlikely(rq->idle_balance))
8223                 return false;
8224 
8225        /*
8226         * We may be recently in ticked or tickless idle mode. At the first
8227         * busy tick after returning from idle, we will update the busy stats.
8228         */
8229         set_cpu_sd_state_busy();
8230         nohz_balance_exit_idle(cpu);
8231 
8232         /*
8233          * None are in tickless mode and hence no need for NOHZ idle load
8234          * balancing.
8235          */
8236         if (likely(!atomic_read(&nohz.nr_cpus)))
8237                 return false;
8238 
8239         if (time_before(now, nohz.next_balance))
8240                 return false;
8241 
8242         if (rq->nr_running >= 2)
8243                 return true;
8244 
8245         rcu_read_lock();
8246         sd = rcu_dereference(per_cpu(sd_busy, cpu));
8247         if (sd) {
8248                 sgc = sd->groups->sgc;
8249                 nr_busy = atomic_read(&sgc->nr_busy_cpus);
8250 
8251                 if (nr_busy > 1) {
8252                         kick = true;
8253                         goto unlock;
8254                 }
8255 
8256         }
8257 
8258         sd = rcu_dereference(rq->sd);
8259         if (sd) {
8260                 if ((rq->cfs.h_nr_running >= 1) &&
8261                                 check_cpu_capacity(rq, sd)) {
8262                         kick = true;
8263                         goto unlock;
8264                 }
8265         }
8266 
8267         sd = rcu_dereference(per_cpu(sd_asym, cpu));
8268         if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
8269                                   sched_domain_span(sd)) < cpu)) {
8270                 kick = true;
8271                 goto unlock;
8272         }
8273 
8274 unlock:
8275         rcu_read_unlock();
8276         return kick;
8277 }
8278 #else
8279 static void nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle) { }
8280 #endif
8281 
8282 /*
8283  * run_rebalance_domains is triggered when needed from the scheduler tick.
8284  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
8285  */
8286 static void run_rebalance_domains(struct softirq_action *h)
8287 {
8288         struct rq *this_rq = this_rq();
8289         enum cpu_idle_type idle = this_rq->idle_balance ?
8290                                                 CPU_IDLE : CPU_NOT_IDLE;
8291 
8292         /*
8293          * If this cpu has a pending nohz_balance_kick, then do the
8294          * balancing on behalf of the other idle cpus whose ticks are
8295          * stopped. Do nohz_idle_balance *before* rebalance_domains to
8296          * give the idle cpus a chance to load balance. Else we may
8297          * load balance only within the local sched_domain hierarchy
8298          * and abort nohz_idle_balance altogether if we pull some load.
8299          */
8300         nohz_idle_balance(this_rq, idle);
8301         rebalance_domains(this_rq, idle);
8302 }
8303 
8304 /*
8305  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
8306  */
8307 void trigger_load_balance(struct rq *rq)
8308 {